sys/vm/vm_fault.c

   1 /*
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  * Copyright (c) 1994 John S. Dyson
   5  * All rights reserved.
   6  * Copyright (c) 1994 David Greenman
   7  * All rights reserved.
   8  *
   9  *
  10  * This code is derived from software contributed to Berkeley by
  11  * The Mach Operating System project at Carnegie-Mellon University.
  12  *
  13  * Redistribution and use in source and binary forms, with or without
  14  * modification, are permitted provided that the following conditions
  15  * are met:
  16  * 1. Redistributions of source code must retain the above copyright
  17  *    notice, this list of conditions and the following disclaimer.
  18  * 2. Redistributions in binary form must reproduce the above copyright
  19  *    notice, this list of conditions and the following disclaimer in the
  20  *    documentation and/or other materials provided with the distribution.
  21  * 3. All advertising materials mentioning features or use of this software
  22  *    must display the following acknowledgement:
  23  *      This product includes software developed by the University of
  24  *      California, Berkeley and its contributors.
  25  * 4. Neither the name of the University nor the names of its contributors
  26  *    may be used to endorse or promote products derived from this software
  27  *    without specific prior written permission.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  39  * SUCH DAMAGE.
  40  *
  41  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
  42  *
  43  *
  44  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  45  * All rights reserved.
  46  *
  47  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  48  *
  49  * Permission to use, copy, modify and distribute this software and
  50  * its documentation is hereby granted, provided that both the copyright
  51  * notice and this permission notice appear in all copies of the
  52  * software, derivative works or modified versions, and any portions
  53  * thereof, and that both notices appear in supporting documentation.
  54  *
  55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  58  *
  59  * Carnegie Mellon requests users of this software to return to
  60  *
  61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  62  *  School of Computer Science
  63  *  Carnegie Mellon University
  64  *  Pittsburgh PA 15213-3890
  65  *
  66  * any improvements or extensions that they make and grant Carnegie the
  67  * rights to redistribute these changes.
  68  *
  69  * $FreeBSD$
  70  */
  71
  72 /*
  73  *      Page fault handling module.
  74  */
  75 #include <sys/param.h>
  76 #include <sys/systm.h>
  77 #include <sys/kernel.h>
  78 #include <sys/lock.h>
  79 #include <sys/mutex.h>
  80 #include <sys/proc.h>
  81 #include <sys/resourcevar.h>
  82 #include <sys/sysctl.h>
  83 #include <sys/vmmeter.h>
  84 #include <sys/vnode.h>
  85
  86 #include <vm/vm.h>
  87 #include <vm/vm_param.h>
  88 #include <vm/pmap.h>
  89 #include <vm/vm_map.h>
  90 #include <vm/vm_object.h>
  91 #include <vm/vm_page.h>
  92 #include <vm/vm_pageout.h>
  93 #include <vm/vm_kern.h>
  94 #include <vm/vm_pager.h>
  95 #include <vm/vnode_pager.h>
  96 #include <vm/vm_extern.h>
  97
  98 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
  99
 100 #define VM_FAULT_READ_AHEAD 8
 101 #define VM_FAULT_READ_BEHIND 7
 102 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 103
 104 struct faultstate {
 105         vm_page_t m;
 106         vm_object_t object;
 107         vm_pindex_t pindex;
 108         vm_page_t first_m;
 109         vm_object_t     first_object;
 110         vm_pindex_t first_pindex;
 111         vm_map_t map;
 112         vm_map_entry_t entry;
 113         int lookup_still_valid;
 114         struct vnode *vp;
 115 };
 116
 117 static __inline void
 118 release_page(struct faultstate *fs)
 119 {
 120         vm_page_wakeup(fs->m);
 121         vm_page_deactivate(fs->m);
 122         fs->m = NULL;
 123 }
 124
 125 static __inline void
 126 unlock_map(struct faultstate *fs)
 127 {
 128         if (fs->lookup_still_valid) {
 129                 vm_map_lookup_done(fs->map, fs->entry);
 130                 fs->lookup_still_valid = FALSE;
 131         }
 132 }
 133
 134 static void
 135 _unlock_things(struct faultstate *fs, int dealloc)
 136 {
 137         GIANT_REQUIRED;
 138         vm_object_pip_wakeup(fs->object);
 139         if (fs->object != fs->first_object) {
 140                 vm_page_free(fs->first_m);
 141                 vm_object_pip_wakeup(fs->first_object);
 142                 fs->first_m = NULL;
 143         }
 144         if (dealloc) {
 145                 vm_object_deallocate(fs->first_object);
 146         }
 147         unlock_map(fs);
 148         if (fs->vp != NULL) {
 149                 vput(fs->vp);
 150                 fs->vp = NULL;
 151         }
 152 }
 153
 154 #define unlock_things(fs) _unlock_things(fs, 0)
 155 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
 156
 157 /*
 158  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 159  *            current object *might* contain the page.
 160  *
 161  *            default objects are zero-fill, there is no real pager.
 162  */
 163 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 164                         (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
 165
 166 /*
 167  *      vm_fault:
 168  *
 169  *      Handle a page fault occurring at the given address,
 170  *      requiring the given permissions, in the map specified.
 171  *      If successful, the page is inserted into the
 172  *      associated physical map.
 173  *
 174  *      NOTE: the given address should be truncated to the
 175  *      proper page address.
 176  *
 177  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 178  *      a standard error specifying why the fault is fatal is returned.
 179  *
 180  *
 181  *      The map in question must be referenced, and remains so.
 182  *      Caller may hold no locks.
 183  */
 184 static int vm_fault1(vm_map_t, vm_offset_t, vm_prot_t, int);
 185
 186 int
 187 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 188          int fault_flags)
 189 {
 190         int ret;
 191
 192         mtx_lock(&Giant);
 193         /* GIANT_REQUIRED */
 194
 195         ret = vm_fault1(map, vaddr, fault_type, fault_flags);
 196         mtx_unlock(&Giant);
 197         return (ret);
 198 }
 199
 200 static int
 201 vm_fault1(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 202           int fault_flags)
 203 {
 204         vm_prot_t prot;
 205         int result;
 206         boolean_t wired;
 207         int map_generation;
 208         vm_object_t next_object;
 209         vm_page_t marray[VM_FAULT_READ];
 210         int hardfault;
 211         int faultcount;
 212         struct faultstate fs;
 213
 214         GIANT_REQUIRED;
 215
 216         cnt.v_vm_faults++;
 217         hardfault = 0;
 218
 219 RetryFault:;
 220
 221         /*
 222          * Find the backing store object and offset into it to begin the
 223          * search.
 224          */
 225         fs.map = map;
 226         if ((result = vm_map_lookup(&fs.map, vaddr,
 227                 fault_type, &fs.entry, &fs.first_object,
 228                 &fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) {
 229                 if ((result != KERN_PROTECTION_FAILURE) ||
 230                         ((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) {
 231                         return result;
 232                 }
 233
 234                 /*
 235                  * If we are user-wiring a r/w segment, and it is COW, then
 236                  * we need to do the COW operation.  Note that we don't COW
 237                  * currently RO sections now, because it is NOT desirable
 238                  * to COW .text.  We simply keep .text from ever being COW'ed
 239                  * and take the heat that one cannot debug wired .text sections.
 240                  */
 241                 result = vm_map_lookup(&fs.map, vaddr,
 242                         VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
 243                         &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
 244                 if (result != KERN_SUCCESS) {
 245                         return result;
 246                 }
 247
 248                 /*
 249                  * If we don't COW now, on a user wire, the user will never
 250                  * be able to write to the mapping.  If we don't make this
 251                  * restriction, the bookkeeping would be nearly impossible.
 252                  */
 253                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 254                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 255         }
 256
 257         map_generation = fs.map->timestamp;
 258
 259         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 260                 panic("vm_fault: fault on nofault entry, addr: %lx",
 261                     (u_long)vaddr);
 262         }
 263
 264         /*
 265          * Make a reference to this object to prevent its disposal while we
 266          * are messing with it.  Once we have the reference, the map is free
 267          * to be diddled.  Since objects reference their shadows (and copies),
 268          * they will stay around as well.
 269          *
 270          * Bump the paging-in-progress count to prevent size changes (e.g.
 271          * truncation operations) during I/O.  This must be done after
 272          * obtaining the vnode lock in order to avoid possible deadlocks.
 273          */
 274         vm_object_reference(fs.first_object);
 275         fs.vp = vnode_pager_lock(fs.first_object);
 276         vm_object_pip_add(fs.first_object, 1);
 277
 278         if ((fault_type & VM_PROT_WRITE) &&
 279                 (fs.first_object->type == OBJT_VNODE)) {
 280                 vm_freeze_copyopts(fs.first_object,
 281                         fs.first_pindex, fs.first_pindex + 1);
 282         }
 283
 284         fs.lookup_still_valid = TRUE;
 285
 286         if (wired)
 287                 fault_type = prot;
 288
 289         fs.first_m = NULL;
 290
 291         /*
 292          * Search for the page at object/offset.
 293          */
 294         fs.object = fs.first_object;
 295         fs.pindex = fs.first_pindex;
 296         while (TRUE) {
 297                 /*
 298                  * If the object is dead, we stop here
 299                  */
 300                 if (fs.object->flags & OBJ_DEAD) {
 301                         unlock_and_deallocate(&fs);
 302                         return (KERN_PROTECTION_FAILURE);
 303                 }
 304
 305                 /*
 306                  * See if page is resident
 307                  */
 308                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 309                 if (fs.m != NULL) {
 310                         int queue, s;
 311                         /*
 312                          * Wait/Retry if the page is busy.  We have to do this
 313                          * if the page is busy via either PG_BUSY or
 314                          * vm_page_t->busy because the vm_pager may be using
 315                          * vm_page_t->busy for pageouts ( and even pageins if
 316                          * it is the vnode pager ), and we could end up trying
 317                          * to pagein and pageout the same page simultaneously.
 318                          *
 319                          * We can theoretically allow the busy case on a read
 320                          * fault if the page is marked valid, but since such
 321                          * pages are typically already pmap'd, putting that
 322                          * special case in might be more effort then it is
 323                          * worth.  We cannot under any circumstances mess
 324                          * around with a vm_page_t->busy page except, perhaps,
 325                          * to pmap it.
 326                          */
 327                         if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
 328                                 unlock_things(&fs);
 329                                 (void)vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
 330                                 cnt.v_intrans++;
 331                                 vm_object_deallocate(fs.first_object);
 332                                 goto RetryFault;
 333                         }
 334                         queue = fs.m->queue;
 335
 336                         s = splvm();
 337                         vm_pageq_remove_nowakeup(fs.m);
 338                         splx(s);
 339
 340                         if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
 341                                 vm_page_activate(fs.m);
 342                                 unlock_and_deallocate(&fs);
 343                                 VM_WAITPFAULT;
 344                                 goto RetryFault;
 345                         }
 346
 347                         /*
 348                          * Mark page busy for other processes, and the
 349                          * pagedaemon.  If it still isn't completely valid
 350                          * (readable), jump to readrest, else break-out ( we
 351                          * found the page ).
 352                          */
 353                         vm_page_busy(fs.m);
 354                         if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
 355                                 fs.m->object != kernel_object && fs.m->object != kmem_object) {
 356                                 goto readrest;
 357                         }
 358
 359                         break;
 360                 }
 361
 362                 /*
 363                  * Page is not resident, If this is the search termination
 364                  * or the pager might contain the page, allocate a new page.
 365                  */
 366                 if (TRYPAGER || fs.object == fs.first_object) {
 367                         if (fs.pindex >= fs.object->size) {
 368                                 unlock_and_deallocate(&fs);
 369                                 return (KERN_PROTECTION_FAILURE);
 370                         }
 371
 372                         /*
 373                          * Allocate a new page for this object/offset pair.
 374                          */
 375                         fs.m = NULL;
 376                         if (!vm_page_count_severe()) {
 377                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 378                                     (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
 379                         }
 380                         if (fs.m == NULL) {
 381                                 unlock_and_deallocate(&fs);
 382                                 VM_WAITPFAULT;
 383                                 goto RetryFault;
 384                         }
 385                 }
 386
 387 readrest:
 388                 /*
 389                  * We have found a valid page or we have allocated a new page.
 390                  * The page thus may not be valid or may not be entirely
 391                  * valid.
 392                  *
 393                  * Attempt to fault-in the page if there is a chance that the
 394                  * pager has it, and potentially fault in additional pages
 395                  * at the same time.
 396                  */
 397                 if (TRYPAGER) {
 398                         int rv;
 399                         int reqpage;
 400                         int ahead, behind;
 401                         u_char behavior = vm_map_entry_behavior(fs.entry);
 402
 403                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 404                                 ahead = 0;
 405                                 behind = 0;
 406                         } else {
 407                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 408                                 if (behind > VM_FAULT_READ_BEHIND)
 409                                         behind = VM_FAULT_READ_BEHIND;
 410
 411                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 412                                 if (ahead > VM_FAULT_READ_AHEAD)
 413                                         ahead = VM_FAULT_READ_AHEAD;
 414                         }
 415
 416                         if ((fs.first_object->type != OBJT_DEVICE) &&
 417                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 418                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 419                                 fs.pindex >= fs.entry->lastr &&
 420                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
 421                         ) {
 422                                 vm_pindex_t firstpindex, tmppindex;
 423
 424                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 425                                         firstpindex = 0;
 426                                 else
 427                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 428
 429                                 /*
 430                                  * note: partially valid pages cannot be
 431                                  * included in the lookahead - NFS piecemeal
 432                                  * writes will barf on it badly.
 433                                  */
 434                                 for (tmppindex = fs.first_pindex - 1;
 435                                         tmppindex >= firstpindex;
 436                                         --tmppindex) {
 437                                         vm_page_t mt;
 438
 439                                         mt = vm_page_lookup(fs.first_object, tmppindex);
 440                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
 441                                                 break;
 442                                         if (mt->busy ||
 443                                                 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
 444                                                 mt->hold_count ||
 445                                                 mt->wire_count)
 446                                                 continue;
 447                                         if (mt->dirty == 0)
 448                                                 vm_page_test_dirty(mt);
 449                                         if (mt->dirty) {
 450                                                 vm_page_protect(mt, VM_PROT_NONE);
 451                                                 vm_page_deactivate(mt);
 452                                         } else {
 453                                                 vm_page_cache(mt);
 454                                         }
 455                                 }
 456
 457                                 ahead += behind;
 458                                 behind = 0;
 459                         }
 460
 461                         /*
 462                          * now we find out if any other pages should be paged
 463                          * in at this time this routine checks to see if the
 464                          * pages surrounding this fault reside in the same
 465                          * object as the page for this fault.  If they do,
 466                          * then they are faulted in also into the object.  The
 467                          * array "marray" returned contains an array of
 468                          * vm_page_t structs where one of them is the
 469                          * vm_page_t passed to the routine.  The reqpage
 470                          * return value is the index into the marray for the
 471                          * vm_page_t passed to the routine.
 472                          *
 473                          * fs.m plus the additional pages are PG_BUSY'd.
 474                          */
 475                         faultcount = vm_fault_additional_pages(
 476                             fs.m, behind, ahead, marray, &reqpage);
 477
 478                         /*
 479                          * update lastr imperfectly (we do not know how much
 480                          * getpages will actually read), but good enough.
 481                          */
 482                         fs.entry->lastr = fs.pindex + faultcount - behind;
 483
 484                         /*
 485                          * Call the pager to retrieve the data, if any, after
 486                          * releasing the lock on the map.  We hold a ref on
 487                          * fs.object and the pages are PG_BUSY'd.
 488                          */
 489                         unlock_map(&fs);
 490
 491                         rv = faultcount ?
 492                             vm_pager_get_pages(fs.object, marray, faultcount,
 493                                 reqpage) : VM_PAGER_FAIL;
 494
 495                         if (rv == VM_PAGER_OK) {
 496                                 /*
 497                                  * Found the page. Leave it busy while we play
 498                                  * with it.
 499                                  */
 500
 501                                 /*
 502                                  * Relookup in case pager changed page. Pager
 503                                  * is responsible for disposition of old page
 504                                  * if moved.
 505                                  */
 506                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 507                                 if (!fs.m) {
 508                                         unlock_and_deallocate(&fs);
 509                                         goto RetryFault;
 510                                 }
 511
 512                                 hardfault++;
 513                                 break; /* break to PAGE HAS BEEN FOUND */
 514                         }
 515                         /*
 516                          * Remove the bogus page (which does not exist at this
 517                          * object/offset); before doing so, we must get back
 518                          * our object lock to preserve our invariant.
 519                          *
 520                          * Also wake up any other process that may want to bring
 521                          * in this page.
 522                          *
 523                          * If this is the top-level object, we must leave the
 524                          * busy page to prevent another process from rushing
 525                          * past us, and inserting the page in that object at
 526                          * the same time that we are.
 527                          */
 528                         if (rv == VM_PAGER_ERROR)
 529                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 530                                     curproc->p_pid, curproc->p_comm);
 531                         /*
 532                          * Data outside the range of the pager or an I/O error
 533                          */
 534                         /*
 535                          * XXX - the check for kernel_map is a kludge to work
 536                          * around having the machine panic on a kernel space
 537                          * fault w/ I/O error.
 538                          */
 539                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 540                                 (rv == VM_PAGER_BAD)) {
 541                                 vm_page_free(fs.m);
 542                                 fs.m = NULL;
 543                                 unlock_and_deallocate(&fs);
 544                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 545                         }
 546                         if (fs.object != fs.first_object) {
 547                                 vm_page_free(fs.m);
 548                                 fs.m = NULL;
 549                                 /*
 550                                  * XXX - we cannot just fall out at this
 551                                  * point, m has been freed and is invalid!
 552                                  */
 553                         }
 554                 }
 555
 556                 /*
 557                  * We get here if the object has default pager (or unwiring)
 558                  * or the pager doesn't have the page.
 559                  */
 560                 if (fs.object == fs.first_object)
 561                         fs.first_m = fs.m;
 562
 563                 /*
 564                  * Move on to the next object.  Lock the next object before
 565                  * unlocking the current one.
 566                  */
 567                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 568                 next_object = fs.object->backing_object;
 569                 if (next_object == NULL) {
 570                         /*
 571                          * If there's no object left, fill the page in the top
 572                          * object with zeros.
 573                          */
 574                         if (fs.object != fs.first_object) {
 575                                 vm_object_pip_wakeup(fs.object);
 576
 577                                 fs.object = fs.first_object;
 578                                 fs.pindex = fs.first_pindex;
 579                                 fs.m = fs.first_m;
 580                         }
 581                         fs.first_m = NULL;
 582
 583                         /*
 584                          * Zero the page if necessary and mark it valid.
 585                          */
 586                         if ((fs.m->flags & PG_ZERO) == 0) {
 587                                 vm_page_zero_fill(fs.m);
 588                         } else {
 589                                 cnt.v_ozfod++;
 590                         }
 591                         cnt.v_zfod++;
 592                         fs.m->valid = VM_PAGE_BITS_ALL;
 593                         break;  /* break to PAGE HAS BEEN FOUND */
 594                 } else {
 595                         if (fs.object != fs.first_object) {
 596                                 vm_object_pip_wakeup(fs.object);
 597                         }
 598                         KASSERT(fs.object != next_object, ("object loop %p", next_object));
 599                         fs.object = next_object;
 600                         vm_object_pip_add(fs.object, 1);
 601                 }
 602         }
 603
 604         KASSERT((fs.m->flags & PG_BUSY) != 0,
 605             ("vm_fault: not busy after main loop"));
 606
 607         /*
 608          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 609          * is held.]
 610          */
 611
 612         /*
 613          * If the page is being written, but isn't already owned by the
 614          * top-level object, we have to copy it into a new page owned by the
 615          * top-level object.
 616          */
 617         if (fs.object != fs.first_object) {
 618                 /*
 619                  * We only really need to copy if we want to write it.
 620                  */
 621                 if (fault_type & VM_PROT_WRITE) {
 622                         /*
 623                          * This allows pages to be virtually copied from a
 624                          * backing_object into the first_object, where the
 625                          * backing object has no other refs to it, and cannot
 626                          * gain any more refs.  Instead of a bcopy, we just
 627                          * move the page from the backing object to the
 628                          * first object.  Note that we must mark the page
 629                          * dirty in the first object so that it will go out
 630                          * to swap when needed.
 631                          */
 632                         if (map_generation == fs.map->timestamp &&
 633                                 /*
 634                                  * Only one shadow object
 635                                  */
 636                                 (fs.object->shadow_count == 1) &&
 637                                 /*
 638                                  * No COW refs, except us
 639                                  */
 640                                 (fs.object->ref_count == 1) &&
 641                                 /*
 642                                  * No one else can look this object up
 643                                  */
 644                                 (fs.object->handle == NULL) &&
 645                                 /*
 646                                  * No other ways to look the object up
 647                                  */
 648                                 ((fs.object->type == OBJT_DEFAULT) ||
 649                                  (fs.object->type == OBJT_SWAP)) &&
 650                                 /*
 651                                  * We don't chase down the shadow chain
 652                                  */
 653                                 (fs.object == fs.first_object->backing_object) &&
 654
 655                                 /*
 656                                  * grab the lock if we need to
 657                                  */
 658                                 (fs.lookup_still_valid ||
 659                                  lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT, (void *)0, curthread) == 0)
 660                             ) {
 661
 662                                 fs.lookup_still_valid = 1;
 663                                 /*
 664                                  * get rid of the unnecessary page
 665                                  */
 666                                 vm_page_protect(fs.first_m, VM_PROT_NONE);
 667                                 vm_page_free(fs.first_m);
 668                                 fs.first_m = NULL;
 669
 670                                 /*
 671                                  * grab the page and put it into the
 672                                  * process'es object.  The page is
 673                                  * automatically made dirty.
 674                                  */
 675                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 676                                 fs.first_m = fs.m;
 677                                 vm_page_busy(fs.first_m);
 678                                 fs.m = NULL;
 679                                 cnt.v_cow_optim++;
 680                         } else {
 681                                 /*
 682                                  * Oh, well, lets copy it.
 683                                  */
 684                                 vm_page_copy(fs.m, fs.first_m);
 685                         }
 686
 687                         if (fs.m) {
 688                                 /*
 689                                  * We no longer need the old page or object.
 690                                  */
 691                                 release_page(&fs);
 692                         }
 693
 694                         /*
 695                          * fs.object != fs.first_object due to above
 696                          * conditional
 697                          */
 698                         vm_object_pip_wakeup(fs.object);
 699
 700                         /*
 701                          * Only use the new page below...
 702                          */
 703                         cnt.v_cow_faults++;
 704                         fs.m = fs.first_m;
 705                         fs.object = fs.first_object;
 706                         fs.pindex = fs.first_pindex;
 707
 708                 } else {
 709                         prot &= ~VM_PROT_WRITE;
 710                 }
 711         }
 712
 713         /*
 714          * We must verify that the maps have not changed since our last
 715          * lookup.
 716          */
 717         if (!fs.lookup_still_valid &&
 718                 (fs.map->timestamp != map_generation)) {
 719                 vm_object_t retry_object;
 720                 vm_pindex_t retry_pindex;
 721                 vm_prot_t retry_prot;
 722
 723                 /*
 724                  * Since map entries may be pageable, make sure we can take a
 725                  * page fault on them.
 726                  */
 727
 728                 /*
 729                  * Unlock vnode before the lookup to avoid deadlock.   E.G.
 730                  * avoid a deadlock between the inode and exec_map that can
 731                  * occur due to locks being obtained in different orders.
 732                  */
 733                 if (fs.vp != NULL) {
 734                         vput(fs.vp);
 735                         fs.vp = NULL;
 736                 }
 737
 738                 if (fs.map->infork) {
 739                         release_page(&fs);
 740                         unlock_and_deallocate(&fs);
 741                         goto RetryFault;
 742                 }
 743
 744                 /*
 745                  * To avoid trying to write_lock the map while another process
 746                  * has it read_locked (in vm_map_pageable), we do not try for
 747                  * write permission.  If the page is still writable, we will
 748                  * get write permission.  If it is not, or has been marked
 749                  * needs_copy, we enter the mapping without write permission,
 750                  * and will merely take another fault.
 751                  */
 752                 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
 753                     &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 754                 map_generation = fs.map->timestamp;
 755
 756                 /*
 757                  * If we don't need the page any longer, put it on the active
 758                  * list (the easiest thing to do here).  If no one needs it,
 759                  * pageout will grab it eventually.
 760                  */
 761                 if (result != KERN_SUCCESS) {
 762                         release_page(&fs);
 763                         unlock_and_deallocate(&fs);
 764                         return (result);
 765                 }
 766                 fs.lookup_still_valid = TRUE;
 767
 768                 if ((retry_object != fs.first_object) ||
 769                     (retry_pindex != fs.first_pindex)) {
 770                         release_page(&fs);
 771                         unlock_and_deallocate(&fs);
 772                         goto RetryFault;
 773                 }
 774                 /*
 775                  * Check whether the protection has changed or the object has
 776                  * been copied while we left the map unlocked. Changing from
 777                  * read to write permission is OK - we leave the page
 778                  * write-protected, and catch the write fault. Changing from
 779                  * write to read permission means that we can't mark the page
 780                  * write-enabled after all.
 781                  */
 782                 prot &= retry_prot;
 783         }
 784
 785         /*
 786          * Put this page into the physical map. We had to do the unlock above
 787          * because pmap_enter may cause other faults.   We don't put the page
 788          * back on the active queue until later so that the page-out daemon
 789          * won't find us (yet).
 790          */
 791
 792         if (prot & VM_PROT_WRITE) {
 793                 vm_page_flag_set(fs.m, PG_WRITEABLE);
 794                 vm_object_set_writeable_dirty(fs.m->object);
 795
 796                 /*
 797                  * If the fault is a write, we know that this page is being
 798                  * written NOW so dirty it explicitly to save on
 799                  * pmap_is_modified() calls later.
 800                  *
 801                  * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
 802                  * if the page is already dirty to prevent data written with
 803                  * the expectation of being synced from not being synced.
 804                  * Likewise if this entry does not request NOSYNC then make
 805                  * sure the page isn't marked NOSYNC.  Applications sharing
 806                  * data should use the same flags to avoid ping ponging.
 807                  *
 808                  * Also tell the backing pager, if any, that it should remove
 809                  * any swap backing since the page is now dirty.
 810                  */
 811                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 812                         if (fs.m->dirty == 0)
 813                                 vm_page_flag_set(fs.m, PG_NOSYNC);
 814                 } else {
 815                         vm_page_flag_clear(fs.m, PG_NOSYNC);
 816                 }
 817                 if (fault_flags & VM_FAULT_DIRTY) {
 818                         int s;
 819                         vm_page_dirty(fs.m);
 820                         s = splvm();
 821                         vm_pager_page_unswapped(fs.m);
 822                         splx(s);
 823                 }
 824         }
 825
 826         /*
 827          * Page had better still be busy
 828          */
 829         KASSERT(fs.m->flags & PG_BUSY,
 830                 ("vm_fault: page %p not busy!", fs.m));
 831         unlock_things(&fs);
 832
 833         /*
 834          * Sanity check: page must be completely valid or it is not fit to
 835          * map into user space.  vm_pager_get_pages() ensures this.
 836          */
 837         if (fs.m->valid != VM_PAGE_BITS_ALL) {
 838                 vm_page_zero_invalid(fs.m, TRUE);
 839                 printf("Warning: page %p partially invalid on fault\n", fs.m);
 840         }
 841         pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
 842         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 843                 pmap_prefault(fs.map->pmap, vaddr, fs.entry);
 844         }
 845         vm_page_flag_clear(fs.m, PG_ZERO);
 846         vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
 847         if (fault_flags & VM_FAULT_HOLD)
 848                 vm_page_hold(fs.m);
 849
 850         /*
 851          * If the page is not wired down, then put it where the pageout daemon
 852          * can find it.
 853          */
 854         if (fault_flags & VM_FAULT_WIRE_MASK) {
 855                 if (wired)
 856                         vm_page_wire(fs.m);
 857                 else
 858                         vm_page_unwire(fs.m, 1);
 859         } else {
 860                 vm_page_activate(fs.m);
 861         }
 862
 863         mtx_lock_spin(&sched_lock);
 864         if (curproc && (curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
 865                 if (hardfault) {
 866                         curproc->p_stats->p_ru.ru_majflt++;
 867                 } else {
 868                         curproc->p_stats->p_ru.ru_minflt++;
 869                 }
 870         }
 871         mtx_unlock_spin(&sched_lock);
 872
 873         /*
 874          * Unlock everything, and return
 875          */
 876         vm_page_wakeup(fs.m);
 877         vm_object_deallocate(fs.first_object);
 878         return (KERN_SUCCESS);
 879
 880 }
 881
 882 /*
 883  *      vm_fault_wire:
 884  *
 885  *      Wire down a range of virtual addresses in a map.
 886  */
 887 int
 888 vm_fault_wire(map, start, end)
 889         vm_map_t map;
 890         vm_offset_t start, end;
 891 {
 892
 893         vm_offset_t va;
 894         pmap_t pmap;
 895         int rv;
 896
 897         pmap = vm_map_pmap(map);
 898
 899         /*
 900          * Inform the physical mapping system that the range of addresses may
 901          * not fault, so that page tables and such can be locked down as well.
 902          */
 903         pmap_pageable(pmap, start, end, FALSE);
 904
 905         /*
 906          * We simulate a fault to get the page and enter it in the physical
 907          * map.
 908          */
 909         for (va = start; va < end; va += PAGE_SIZE) {
 910                 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
 911                         VM_FAULT_CHANGE_WIRING);
 912                 if (rv) {
 913                         if (va != start)
 914                                 vm_fault_unwire(map, start, va);
 915                         return (rv);
 916                 }
 917         }
 918         return (KERN_SUCCESS);
 919 }
 920
 921 /*
 922  *      vm_fault_user_wire:
 923  *
 924  *      Wire down a range of virtual addresses in a map.  This
 925  *      is for user mode though, so we only ask for read access
 926  *      on currently read only sections.
 927  */
 928 int
 929 vm_fault_user_wire(map, start, end)
 930         vm_map_t map;
 931         vm_offset_t start, end;
 932 {
 933
 934         vm_offset_t va;
 935         pmap_t pmap;
 936         int rv;
 937
 938         GIANT_REQUIRED;
 939
 940         pmap = vm_map_pmap(map);
 941
 942         /*
 943          * Inform the physical mapping system that the range of addresses may
 944          * not fault, so that page tables and such can be locked down as well.
 945          */
 946         pmap_pageable(pmap, start, end, FALSE);
 947
 948         /*
 949          * We simulate a fault to get the page and enter it in the physical
 950          * map.
 951          */
 952         for (va = start; va < end; va += PAGE_SIZE) {
 953                 rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE);
 954                 if (rv) {
 955                         if (va != start)
 956                                 vm_fault_unwire(map, start, va);
 957                         return (rv);
 958                 }
 959         }
 960         return (KERN_SUCCESS);
 961 }
 962
 963
 964 /*
 965  *      vm_fault_unwire:
 966  *
 967  *      Unwire a range of virtual addresses in a map.
 968  */
 969 void
 970 vm_fault_unwire(map, start, end)
 971         vm_map_t map;
 972         vm_offset_t start, end;
 973 {
 974
 975         vm_offset_t va, pa;
 976         pmap_t pmap;
 977
 978         pmap = vm_map_pmap(map);
 979
 980         /*
 981          * Since the pages are wired down, we must be able to get their
 982          * mappings from the physical map system.
 983          */
 984         for (va = start; va < end; va += PAGE_SIZE) {
 985                 pa = pmap_extract(pmap, va);
 986                 if (pa != (vm_offset_t) 0) {
 987                         pmap_change_wiring(pmap, va, FALSE);
 988                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
 989                 }
 990         }
 991
 992         /*
 993          * Inform the physical mapping system that the range of addresses may
 994          * fault, so that page tables and such may be unwired themselves.
 995          */
 996         pmap_pageable(pmap, start, end, TRUE);
 997
 998 }
 999
1000 /*
1001  *      Routine:
1002  *              vm_fault_copy_entry
1003  *      Function:
1004  *              Copy all of the pages from a wired-down map entry to another.
1005  *
1006  *      In/out conditions:
1007  *              The source and destination maps must be locked for write.
1008  *              The source map entry must be wired down (or be a sharing map
1009  *              entry corresponding to a main map entry that is wired down).
1010  */
1011 void
1012 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1013         vm_map_t dst_map;
1014         vm_map_t src_map;
1015         vm_map_entry_t dst_entry;
1016         vm_map_entry_t src_entry;
1017 {
1018         vm_object_t dst_object;
1019         vm_object_t src_object;
1020         vm_ooffset_t dst_offset;
1021         vm_ooffset_t src_offset;
1022         vm_prot_t prot;
1023         vm_offset_t vaddr;
1024         vm_page_t dst_m;
1025         vm_page_t src_m;
1026
1027 #ifdef  lint
1028         src_map++;
1029 #endif  /* lint */
1030
1031         src_object = src_entry->object.vm_object;
1032         src_offset = src_entry->offset;
1033
1034         /*
1035          * Create the top-level object for the destination entry. (Doesn't
1036          * actually shadow anything - we copy the pages directly.)
1037          */
1038         dst_object = vm_object_allocate(OBJT_DEFAULT,
1039             (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1040
1041         dst_entry->object.vm_object = dst_object;
1042         dst_entry->offset = 0;
1043
1044         prot = dst_entry->max_protection;
1045
1046         /*
1047          * Loop through all of the pages in the entry's range, copying each
1048          * one from the source object (it should be there) to the destination
1049          * object.
1050          */
1051         for (vaddr = dst_entry->start, dst_offset = 0;
1052             vaddr < dst_entry->end;
1053             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1054
1055                 /*
1056                  * Allocate a page in the destination object
1057                  */
1058                 do {
1059                         dst_m = vm_page_alloc(dst_object,
1060                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1061                         if (dst_m == NULL) {
1062                                 VM_WAIT;
1063                         }
1064                 } while (dst_m == NULL);
1065
1066                 /*
1067                  * Find the page in the source object, and copy it in.
1068                  * (Because the source is wired down, the page will be in
1069                  * memory.)
1070                  */
1071                 src_m = vm_page_lookup(src_object,
1072                         OFF_TO_IDX(dst_offset + src_offset));
1073                 if (src_m == NULL)
1074                         panic("vm_fault_copy_wired: page missing");
1075
1076                 vm_page_copy(src_m, dst_m);
1077
1078                 /*
1079                  * Enter it in the pmap...
1080                  */
1081                 vm_page_flag_clear(dst_m, PG_ZERO);
1082                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1083                 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1084
1085                 /*
1086                  * Mark it no longer busy, and put it on the active list.
1087                  */
1088                 vm_page_activate(dst_m);
1089                 vm_page_wakeup(dst_m);
1090         }
1091 }
1092
1093
1094 /*
1095  * This routine checks around the requested page for other pages that
1096  * might be able to be faulted in.  This routine brackets the viable
1097  * pages for the pages to be paged in.
1098  *
1099  * Inputs:
1100  *      m, rbehind, rahead
1101  *
1102  * Outputs:
1103  *  marray (array of vm_page_t), reqpage (index of requested page)
1104  *
1105  * Return value:
1106  *  number of pages in marray
1107  *
1108  * This routine can't block.
1109  */
1110 static int
1111 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1112         vm_page_t m;
1113         int rbehind;
1114         int rahead;
1115         vm_page_t *marray;
1116         int *reqpage;
1117 {
1118         int i,j;
1119         vm_object_t object;
1120         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1121         vm_page_t rtm;
1122         int cbehind, cahead;
1123
1124         GIANT_REQUIRED;
1125
1126         object = m->object;
1127         pindex = m->pindex;
1128
1129         /*
1130          * we don't fault-ahead for device pager
1131          */
1132         if (object->type == OBJT_DEVICE) {
1133                 *reqpage = 0;
1134                 marray[0] = m;
1135                 return 1;
1136         }
1137
1138         /*
1139          * if the requested page is not available, then give up now
1140          */
1141         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1142                 return 0;
1143         }
1144
1145         if ((cbehind == 0) && (cahead == 0)) {
1146                 *reqpage = 0;
1147                 marray[0] = m;
1148                 return 1;
1149         }
1150
1151         if (rahead > cahead) {
1152                 rahead = cahead;
1153         }
1154
1155         if (rbehind > cbehind) {
1156                 rbehind = cbehind;
1157         }
1158
1159         /*
1160          * try to do any readahead that we might have free pages for.
1161          */
1162         if ((rahead + rbehind) >
1163                 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1164                 pagedaemon_wakeup();
1165                 marray[0] = m;
1166                 *reqpage = 0;
1167                 return 1;
1168         }
1169
1170         /*
1171          * scan backward for the read behind pages -- in memory
1172          */
1173         if (pindex > 0) {
1174                 if (rbehind > pindex) {
1175                         rbehind = pindex;
1176                         startpindex = 0;
1177                 } else {
1178                         startpindex = pindex - rbehind;
1179                 }
1180
1181                 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1182                         if (vm_page_lookup(object, tpindex)) {
1183                                 startpindex = tpindex + 1;
1184                                 break;
1185                         }
1186                         if (tpindex == 0)
1187                                 break;
1188                 }
1189
1190                 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1191
1192                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1193                         if (rtm == NULL) {
1194                                 for (j = 0; j < i; j++) {
1195                                         vm_page_free(marray[j]);
1196                                 }
1197                                 marray[0] = m;
1198                                 *reqpage = 0;
1199                                 return 1;
1200                         }
1201
1202                         marray[i] = rtm;
1203                 }
1204         } else {
1205                 startpindex = 0;
1206                 i = 0;
1207         }
1208
1209         marray[i] = m;
1210         /* page offset of the required page */
1211         *reqpage = i;
1212
1213         tpindex = pindex + 1;
1214         i++;
1215
1216         /*
1217          * scan forward for the read ahead pages
1218          */
1219         endpindex = tpindex + rahead;
1220         if (endpindex > object->size)
1221                 endpindex = object->size;
1222
1223         for (; tpindex < endpindex; i++, tpindex++) {
1224
1225                 if (vm_page_lookup(object, tpindex)) {
1226                         break;
1227                 }
1228
1229                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1230                 if (rtm == NULL) {
1231                         break;
1232                 }
1233
1234                 marray[i] = rtm;
1235         }
1236
1237         /* return number of bytes of pages */
1238         return i;
1239 }