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
  70 /*
  71  *      Page fault handling module.
  72  */
  73
  74 #include <sys/cdefs.h>
  75 __FBSDID("$FreeBSD$");
  76
  77 #include "opt_ktrace.h"
  78 #include "opt_vm.h"
  79
  80 #include <sys/param.h>
  81 #include <sys/systm.h>
  82 #include <sys/kernel.h>
  83 #include <sys/lock.h>
  84 #include <sys/mutex.h>
  85 #include <sys/proc.h>
  86 #include <sys/resourcevar.h>
  87 #include <sys/sysctl.h>
  88 #include <sys/vmmeter.h>
  89 #include <sys/vnode.h>
  90 #ifdef KTRACE
  91 #include <sys/ktrace.h>
  92 #endif
  93
  94 #include <vm/vm.h>
  95 #include <vm/vm_param.h>
  96 #include <vm/pmap.h>
  97 #include <vm/vm_map.h>
  98 #include <vm/vm_object.h>
  99 #include <vm/vm_page.h>
 100 #include <vm/vm_pageout.h>
 101 #include <vm/vm_kern.h>
 102 #include <vm/vm_pager.h>
 103 #include <vm/vnode_pager.h>
 104 #include <vm/vm_extern.h>
 105
 106 #include <sys/mount.h>  /* XXX Temporary for VFS_LOCK_GIANT() */
 107
 108 #define PFBAK 4
 109 #define PFFOR 4
 110 #define PAGEORDER_SIZE (PFBAK+PFFOR)
 111
 112 static int prefault_pageorder[] = {
 113         -1 * PAGE_SIZE, 1 * PAGE_SIZE,
 114         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
 115         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
 116         -4 * PAGE_SIZE, 4 * PAGE_SIZE
 117 };
 118
 119 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 120 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
 121 #ifdef KTRACE
 122 static int vm_fault_traced(vm_map_t, vm_offset_t, vm_prot_t, int);
 123 #endif
 124
 125 #define VM_FAULT_READ_AHEAD 8
 126 #define VM_FAULT_READ_BEHIND 7
 127 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 128
 129 struct faultstate {
 130         vm_page_t m;
 131         vm_object_t object;
 132         vm_pindex_t pindex;
 133         vm_page_t first_m;
 134         vm_object_t     first_object;
 135         vm_pindex_t first_pindex;
 136         vm_map_t map;
 137         vm_map_entry_t entry;
 138         int lookup_still_valid;
 139         struct vnode *vp;
 140         int vfslocked;
 141 };
 142
 143 static inline void
 144 release_page(struct faultstate *fs)
 145 {
 146
 147         vm_page_wakeup(fs->m);
 148         vm_page_lock_queues();
 149         vm_page_deactivate(fs->m);
 150         vm_page_unlock_queues();
 151         fs->m = NULL;
 152 }
 153
 154 static inline void
 155 unlock_map(struct faultstate *fs)
 156 {
 157
 158         if (fs->lookup_still_valid) {
 159                 vm_map_lookup_done(fs->map, fs->entry);
 160                 fs->lookup_still_valid = FALSE;
 161         }
 162 }
 163
 164 static void
 165 unlock_and_deallocate(struct faultstate *fs)
 166 {
 167
 168         vm_object_pip_wakeup(fs->object);
 169         VM_OBJECT_UNLOCK(fs->object);
 170         if (fs->object != fs->first_object) {
 171                 VM_OBJECT_LOCK(fs->first_object);
 172                 vm_page_lock_queues();
 173                 vm_page_free(fs->first_m);
 174                 vm_page_unlock_queues();
 175                 vm_object_pip_wakeup(fs->first_object);
 176                 VM_OBJECT_UNLOCK(fs->first_object);
 177                 fs->first_m = NULL;
 178         }
 179         vm_object_deallocate(fs->first_object);
 180         unlock_map(fs);
 181         if (fs->vp != NULL) {
 182                 vput(fs->vp);
 183                 fs->vp = NULL;
 184         }
 185         VFS_UNLOCK_GIANT(fs->vfslocked);
 186         fs->vfslocked = 0;
 187 }
 188
 189 /*
 190  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 191  *            current object *might* contain the page.
 192  *
 193  *            default objects are zero-fill, there is no real pager.
 194  */
 195 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 196                         (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
 197
 198 /*
 199  *      vm_fault:
 200  *
 201  *      Handle a page fault occurring at the given address,
 202  *      requiring the given permissions, in the map specified.
 203  *      If successful, the page is inserted into the
 204  *      associated physical map.
 205  *
 206  *      NOTE: the given address should be truncated to the
 207  *      proper page address.
 208  *
 209  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 210  *      a standard error specifying why the fault is fatal is returned.
 211  *
 212  *
 213  *      The map in question must be referenced, and remains so.
 214  *      Caller may hold no locks.
 215  */
 216 int
 217 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 218          int fault_flags)
 219 #ifdef KTRACE
 220 {
 221         struct thread *td;
 222         int result;
 223
 224         td = curthread;
 225         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 226                 ktrfault(vaddr, fault_type);
 227         result = vm_fault_traced(map, vaddr, fault_type, fault_flags);
 228         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 229                 ktrfaultend(result);
 230         return (result);
 231 }
 232
 233 int
 234 vm_fault_traced(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 235     int fault_flags)
 236 #endif
 237 {
 238         vm_prot_t prot;
 239         int is_first_object_locked, result;
 240         boolean_t are_queues_locked, growstack, wired;
 241         int map_generation;
 242         vm_object_t next_object;
 243         vm_page_t marray[VM_FAULT_READ], mt, mt_prev;
 244         int hardfault;
 245         int faultcount, ahead, behind, alloc_req;
 246         struct faultstate fs;
 247         struct vnode *vp;
 248         int locked, error;
 249
 250         if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
 251                 return (KERN_PROTECTION_FAILURE);
 252
 253         hardfault = 0;
 254         growstack = TRUE;
 255         PCPU_INC(cnt.v_vm_faults);
 256         fs.vp = NULL;
 257         fs.vfslocked = 0;
 258         faultcount = behind = 0;
 259
 260 RetryFault:;
 261
 262         /*
 263          * Find the backing store object and offset into it to begin the
 264          * search.
 265          */
 266         fs.map = map;
 267         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 268             &fs.first_object, &fs.first_pindex, &prot, &wired);
 269         if (result != KERN_SUCCESS) {
 270                 if (result != KERN_PROTECTION_FAILURE ||
 271                     (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
 272                         if (growstack && result == KERN_INVALID_ADDRESS &&
 273                             map != kernel_map && curproc != NULL) {
 274                                 result = vm_map_growstack(curproc, vaddr);
 275                                 if (result != KERN_SUCCESS)
 276                                         return (KERN_FAILURE);
 277                                 growstack = FALSE;
 278                                 goto RetryFault;
 279                         }
 280                         return (result);
 281                 }
 282
 283                 /*
 284                  * If we are user-wiring a r/w segment, and it is COW, then
 285                  * we need to do the COW operation.  Note that we don't COW
 286                  * currently RO sections now, because it is NOT desirable
 287                  * to COW .text.  We simply keep .text from ever being COW'ed
 288                  * and take the heat that one cannot debug wired .text sections.
 289                  */
 290                 result = vm_map_lookup(&fs.map, vaddr,
 291                         VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
 292                         &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
 293                 if (result != KERN_SUCCESS)
 294                         return (result);
 295
 296                 /*
 297                  * If we don't COW now, on a user wire, the user will never
 298                  * be able to write to the mapping.  If we don't make this
 299                  * restriction, the bookkeeping would be nearly impossible.
 300                  *
 301                  * XXX The following assignment modifies the map without
 302                  * holding a write lock on it.
 303                  */
 304                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 305                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 306         }
 307
 308         map_generation = fs.map->timestamp;
 309
 310         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 311                 panic("vm_fault: fault on nofault entry, addr: %lx",
 312                     (u_long)vaddr);
 313         }
 314
 315         /*
 316          * Make a reference to this object to prevent its disposal while we
 317          * are messing with it.  Once we have the reference, the map is free
 318          * to be diddled.  Since objects reference their shadows (and copies),
 319          * they will stay around as well.
 320          *
 321          * Bump the paging-in-progress count to prevent size changes (e.g.
 322          * truncation operations) during I/O.  This must be done after
 323          * obtaining the vnode lock in order to avoid possible deadlocks.
 324          */
 325         VM_OBJECT_LOCK(fs.first_object);
 326         vm_object_reference_locked(fs.first_object);
 327         vm_object_pip_add(fs.first_object, 1);
 328
 329         fs.lookup_still_valid = TRUE;
 330
 331         if (wired)
 332                 fault_type = prot;
 333
 334         fs.first_m = NULL;
 335
 336         /*
 337          * Search for the page at object/offset.
 338          */
 339         fs.object = fs.first_object;
 340         fs.pindex = fs.first_pindex;
 341         while (TRUE) {
 342                 /*
 343                  * If the object is dead, we stop here
 344                  */
 345                 if (fs.object->flags & OBJ_DEAD) {
 346                         unlock_and_deallocate(&fs);
 347                         return (KERN_PROTECTION_FAILURE);
 348                 }
 349
 350                 /*
 351                  * See if page is resident
 352                  */
 353                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 354                 if (fs.m != NULL) {
 355                         /*
 356                          * check for page-based copy on write.
 357                          * We check fs.object == fs.first_object so
 358                          * as to ensure the legacy COW mechanism is
 359                          * used when the page in question is part of
 360                          * a shadow object.  Otherwise, vm_page_cowfault()
 361                          * removes the page from the backing object,
 362                          * which is not what we want.
 363                          */
 364                         vm_page_lock_queues();
 365                         if ((fs.m->cow) &&
 366                             (fault_type & VM_PROT_WRITE) &&
 367                             (fs.object == fs.first_object)) {
 368                                 vm_page_cowfault(fs.m);
 369                                 vm_page_unlock_queues();
 370                                 unlock_and_deallocate(&fs);
 371                                 goto RetryFault;
 372                         }
 373
 374                         /*
 375                          * Wait/Retry if the page is busy.  We have to do this
 376                          * if the page is busy via either VPO_BUSY or
 377                          * vm_page_t->busy because the vm_pager may be using
 378                          * vm_page_t->busy for pageouts ( and even pageins if
 379                          * it is the vnode pager ), and we could end up trying
 380                          * to pagein and pageout the same page simultaneously.
 381                          *
 382                          * We can theoretically allow the busy case on a read
 383                          * fault if the page is marked valid, but since such
 384                          * pages are typically already pmap'd, putting that
 385                          * special case in might be more effort then it is
 386                          * worth.  We cannot under any circumstances mess
 387                          * around with a vm_page_t->busy page except, perhaps,
 388                          * to pmap it.
 389                          */
 390                         if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
 391                                 vm_page_unlock_queues();
 392                                 if (fs.object != fs.first_object) {
 393                                         if (!VM_OBJECT_TRYLOCK(
 394                                             fs.first_object)) {
 395                                                 VM_OBJECT_UNLOCK(fs.object);
 396                                                 VM_OBJECT_LOCK(fs.first_object);
 397                                                 VM_OBJECT_LOCK(fs.object);
 398                                         }
 399                                         vm_page_lock_queues();
 400                                         vm_page_free(fs.first_m);
 401                                         vm_page_unlock_queues();
 402                                         vm_object_pip_wakeup(fs.first_object);
 403                                         VM_OBJECT_UNLOCK(fs.first_object);
 404                                         fs.first_m = NULL;
 405                                 }
 406                                 unlock_map(&fs);
 407                                 if (fs.m == vm_page_lookup(fs.object,
 408                                     fs.pindex)) {
 409                                         vm_page_sleep_if_busy(fs.m, TRUE,
 410                                             "vmpfw");
 411                                 }
 412                                 vm_object_pip_wakeup(fs.object);
 413                                 VM_OBJECT_UNLOCK(fs.object);
 414                                 PCPU_INC(cnt.v_intrans);
 415                                 vm_object_deallocate(fs.first_object);
 416                                 goto RetryFault;
 417                         }
 418                         vm_pageq_remove(fs.m);
 419                         vm_page_unlock_queues();
 420
 421                         /*
 422                          * Mark page busy for other processes, and the
 423                          * pagedaemon.  If it still isn't completely valid
 424                          * (readable), jump to readrest, else break-out ( we
 425                          * found the page ).
 426                          */
 427                         vm_page_busy(fs.m);
 428                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 429                                 goto readrest;
 430                         break;
 431                 }
 432
 433                 /*
 434                  * Page is not resident, If this is the search termination
 435                  * or the pager might contain the page, allocate a new page.
 436                  */
 437                 if (TRYPAGER || fs.object == fs.first_object) {
 438                         if (fs.pindex >= fs.object->size) {
 439                                 unlock_and_deallocate(&fs);
 440                                 return (KERN_PROTECTION_FAILURE);
 441                         }
 442
 443                         /*
 444                          * Allocate a new page for this object/offset pair.
 445                          *
 446                          * Unlocked read of the p_flag is harmless. At
 447                          * worst, the P_KILLED might be not observed
 448                          * there, and allocation can fail, causing
 449                          * restart and new reading of the p_flag.
 450                          */
 451                         fs.m = NULL;
 452                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 453 #if VM_NRESERVLEVEL > 0
 454                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 455                                         fs.object->flags |= OBJ_COLORED;
 456                                         fs.object->pg_color = atop(vaddr) -
 457                                             fs.pindex;
 458                                 }
 459 #endif
 460                                 alloc_req = P_KILLED(curproc) ?
 461                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 462                                 if (fs.object->type != OBJT_VNODE &&
 463                                     fs.object->backing_object == NULL)
 464                                         alloc_req |= VM_ALLOC_ZERO;
 465                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 466                                     alloc_req);
 467                         }
 468                         if (fs.m == NULL) {
 469                                 unlock_and_deallocate(&fs);
 470                                 VM_WAITPFAULT;
 471                                 goto RetryFault;
 472                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 473                                 break;
 474                 }
 475
 476 readrest:
 477                 /*
 478                  * We have found a valid page or we have allocated a new page.
 479                  * The page thus may not be valid or may not be entirely
 480                  * valid.
 481                  *
 482                  * Attempt to fault-in the page if there is a chance that the
 483                  * pager has it, and potentially fault in additional pages
 484                  * at the same time.
 485                  */
 486                 if (TRYPAGER) {
 487                         int rv;
 488                         int reqpage = 0;
 489                         u_char behavior = vm_map_entry_behavior(fs.entry);
 490
 491                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 492                             P_KILLED(curproc)) {
 493                                 ahead = 0;
 494                                 behind = 0;
 495                         } else {
 496                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 497                                 if (behind > VM_FAULT_READ_BEHIND)
 498                                         behind = VM_FAULT_READ_BEHIND;
 499
 500                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 501                                 if (ahead > VM_FAULT_READ_AHEAD)
 502                                         ahead = VM_FAULT_READ_AHEAD;
 503                         }
 504                         is_first_object_locked = FALSE;
 505                         if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 506                              (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 507                               fs.pindex >= fs.entry->lastr &&
 508                               fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
 509                             (fs.first_object == fs.object ||
 510                              (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
 511                             fs.first_object->type != OBJT_DEVICE &&
 512                             fs.first_object->type != OBJT_PHYS &&
 513                             fs.first_object->type != OBJT_SG) {
 514                                 vm_pindex_t firstpindex;
 515
 516                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 517                                         firstpindex = 0;
 518                                 else
 519                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 520                                 mt = fs.first_object != fs.object ?
 521                                     fs.first_m : fs.m;
 522                                 KASSERT(mt != NULL, ("vm_fault: missing mt"));
 523                                 KASSERT((mt->oflags & VPO_BUSY) != 0,
 524                                     ("vm_fault: mt %p not busy", mt));
 525                                 mt_prev = vm_page_prev(mt);
 526
 527                                 are_queues_locked = FALSE;
 528                                 /*
 529                                  * note: partially valid pages cannot be
 530                                  * included in the lookahead - NFS piecemeal
 531                                  * writes will barf on it badly.
 532                                  */
 533                                 while ((mt = mt_prev) != NULL &&
 534                                     mt->pindex >= firstpindex &&
 535                                     mt->valid == VM_PAGE_BITS_ALL) {
 536                                         mt_prev = vm_page_prev(mt);
 537                                         if (mt->busy ||
 538                                             (mt->oflags & VPO_BUSY))
 539                                                 continue;
 540                                         if (!are_queues_locked) {
 541                                                 are_queues_locked = TRUE;
 542                                                 vm_page_lock_queues();
 543                                         }
 544                                         if (mt->hold_count ||
 545                                                 mt->wire_count)
 546                                                 continue;
 547                                         pmap_remove_all(mt);
 548                                         if (mt->dirty) {
 549                                                 vm_page_deactivate(mt);
 550                                         } else {
 551                                                 vm_page_cache(mt);
 552                                         }
 553                                 }
 554                                 if (are_queues_locked)
 555                                         vm_page_unlock_queues();
 556                                 ahead += behind;
 557                                 behind = 0;
 558                         }
 559                         if (is_first_object_locked)
 560                                 VM_OBJECT_UNLOCK(fs.first_object);
 561
 562                         /*
 563                          * Call the pager to retrieve the data, if any, after
 564                          * releasing the lock on the map.  We hold a ref on
 565                          * fs.object and the pages are VPO_BUSY'd.
 566                          */
 567                         unlock_map(&fs);
 568
 569 vnode_lock:
 570                         if (fs.object->type == OBJT_VNODE) {
 571                                 vp = fs.object->handle;
 572                                 if (vp == fs.vp)
 573                                         goto vnode_locked;
 574                                 else if (fs.vp != NULL) {
 575                                         vput(fs.vp);
 576                                         fs.vp = NULL;
 577                                 }
 578                                 locked = VOP_ISLOCKED(vp);
 579
 580                                 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
 581                                         fs.vfslocked = 1;
 582                                         if (!mtx_trylock(&Giant)) {
 583                                                 VM_OBJECT_UNLOCK(fs.object);
 584                                                 mtx_lock(&Giant);
 585                                                 VM_OBJECT_LOCK(fs.object);
 586                                                 goto vnode_lock;
 587                                         }
 588                                 }
 589                                 if (locked != LK_EXCLUSIVE)
 590                                         locked = LK_SHARED;
 591                                 /* Do not sleep for vnode lock while fs.m is busy */
 592                                 error = vget(vp, locked | LK_CANRECURSE |
 593                                     LK_NOWAIT, curthread);
 594                                 if (error != 0) {
 595                                         int vfslocked;
 596
 597                                         vfslocked = fs.vfslocked;
 598                                         fs.vfslocked = 0; /* Keep Giant */
 599                                         vhold(vp);
 600                                         release_page(&fs);
 601                                         unlock_and_deallocate(&fs);
 602                                         error = vget(vp, locked | LK_RETRY |
 603                                             LK_CANRECURSE, curthread);
 604                                         vdrop(vp);
 605                                         fs.vp = vp;
 606                                         fs.vfslocked = vfslocked;
 607                                         KASSERT(error == 0,
 608                                             ("vm_fault: vget failed"));
 609                                         goto RetryFault;
 610                                 }
 611                                 fs.vp = vp;
 612                         }
 613 vnode_locked:
 614                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 615                             ("vm_fault: vnode-backed object mapped by system map"));
 616
 617                         /*
 618                          * now we find out if any other pages should be paged
 619                          * in at this time this routine checks to see if the
 620                          * pages surrounding this fault reside in the same
 621                          * object as the page for this fault.  If they do,
 622                          * then they are faulted in also into the object.  The
 623                          * array "marray" returned contains an array of
 624                          * vm_page_t structs where one of them is the
 625                          * vm_page_t passed to the routine.  The reqpage
 626                          * return value is the index into the marray for the
 627                          * vm_page_t passed to the routine.
 628                          *
 629                          * fs.m plus the additional pages are VPO_BUSY'd.
 630                          */
 631                         faultcount = vm_fault_additional_pages(
 632                             fs.m, behind, ahead, marray, &reqpage);
 633
 634                         rv = faultcount ?
 635                             vm_pager_get_pages(fs.object, marray, faultcount,
 636                                 reqpage) : VM_PAGER_FAIL;
 637
 638                         if (rv == VM_PAGER_OK) {
 639                                 /*
 640                                  * Found the page. Leave it busy while we play
 641                                  * with it.
 642                                  */
 643
 644                                 /*
 645                                  * Relookup in case pager changed page. Pager
 646                                  * is responsible for disposition of old page
 647                                  * if moved.
 648                                  */
 649                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 650                                 if (!fs.m) {
 651                                         unlock_and_deallocate(&fs);
 652                                         goto RetryFault;
 653                                 }
 654
 655                                 hardfault++;
 656                                 break; /* break to PAGE HAS BEEN FOUND */
 657                         }
 658                         /*
 659                          * Remove the bogus page (which does not exist at this
 660                          * object/offset); before doing so, we must get back
 661                          * our object lock to preserve our invariant.
 662                          *
 663                          * Also wake up any other process that may want to bring
 664                          * in this page.
 665                          *
 666                          * If this is the top-level object, we must leave the
 667                          * busy page to prevent another process from rushing
 668                          * past us, and inserting the page in that object at
 669                          * the same time that we are.
 670                          */
 671                         if (rv == VM_PAGER_ERROR)
 672                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 673                                     curproc->p_pid, curproc->p_comm);
 674                         /*
 675                          * Data outside the range of the pager or an I/O error
 676                          */
 677                         /*
 678                          * XXX - the check for kernel_map is a kludge to work
 679                          * around having the machine panic on a kernel space
 680                          * fault w/ I/O error.
 681                          */
 682                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 683                                 (rv == VM_PAGER_BAD)) {
 684                                 vm_page_lock_queues();
 685                                 vm_page_free(fs.m);
 686                                 vm_page_unlock_queues();
 687                                 fs.m = NULL;
 688                                 unlock_and_deallocate(&fs);
 689                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 690                         }
 691                         if (fs.object != fs.first_object) {
 692                                 vm_page_lock_queues();
 693                                 vm_page_free(fs.m);
 694                                 vm_page_unlock_queues();
 695                                 fs.m = NULL;
 696                                 /*
 697                                  * XXX - we cannot just fall out at this
 698                                  * point, m has been freed and is invalid!
 699                                  */
 700                         }
 701                 }
 702
 703                 /*
 704                  * We get here if the object has default pager (or unwiring)
 705                  * or the pager doesn't have the page.
 706                  */
 707                 if (fs.object == fs.first_object)
 708                         fs.first_m = fs.m;
 709
 710                 /*
 711                  * Move on to the next object.  Lock the next object before
 712                  * unlocking the current one.
 713                  */
 714                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 715                 next_object = fs.object->backing_object;
 716                 if (next_object == NULL) {
 717                         /*
 718                          * If there's no object left, fill the page in the top
 719                          * object with zeros.
 720                          */
 721                         if (fs.object != fs.first_object) {
 722                                 vm_object_pip_wakeup(fs.object);
 723                                 VM_OBJECT_UNLOCK(fs.object);
 724
 725                                 fs.object = fs.first_object;
 726                                 fs.pindex = fs.first_pindex;
 727                                 fs.m = fs.first_m;
 728                                 VM_OBJECT_LOCK(fs.object);
 729                         }
 730                         fs.first_m = NULL;
 731
 732                         /*
 733                          * Zero the page if necessary and mark it valid.
 734                          */
 735                         if ((fs.m->flags & PG_ZERO) == 0) {
 736                                 pmap_zero_page(fs.m);
 737                         } else {
 738                                 PCPU_INC(cnt.v_ozfod);
 739                         }
 740                         PCPU_INC(cnt.v_zfod);
 741                         fs.m->valid = VM_PAGE_BITS_ALL;
 742                         break;  /* break to PAGE HAS BEEN FOUND */
 743                 } else {
 744                         KASSERT(fs.object != next_object,
 745                             ("object loop %p", next_object));
 746                         VM_OBJECT_LOCK(next_object);
 747                         vm_object_pip_add(next_object, 1);
 748                         if (fs.object != fs.first_object)
 749                                 vm_object_pip_wakeup(fs.object);
 750                         VM_OBJECT_UNLOCK(fs.object);
 751                         fs.object = next_object;
 752                 }
 753         }
 754
 755         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 756             ("vm_fault: not busy after main loop"));
 757
 758         /*
 759          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 760          * is held.]
 761          */
 762
 763         /*
 764          * If the page is being written, but isn't already owned by the
 765          * top-level object, we have to copy it into a new page owned by the
 766          * top-level object.
 767          */
 768         if (fs.object != fs.first_object) {
 769                 /*
 770                  * We only really need to copy if we want to write it.
 771                  */
 772                 if (fault_type & VM_PROT_WRITE) {
 773                         /*
 774                          * This allows pages to be virtually copied from a
 775                          * backing_object into the first_object, where the
 776                          * backing object has no other refs to it, and cannot
 777                          * gain any more refs.  Instead of a bcopy, we just
 778                          * move the page from the backing object to the
 779                          * first object.  Note that we must mark the page
 780                          * dirty in the first object so that it will go out
 781                          * to swap when needed.
 782                          */
 783                         is_first_object_locked = FALSE;
 784                         if (
 785                                 /*
 786                                  * Only one shadow object
 787                                  */
 788                                 (fs.object->shadow_count == 1) &&
 789                                 /*
 790                                  * No COW refs, except us
 791                                  */
 792                                 (fs.object->ref_count == 1) &&
 793                                 /*
 794                                  * No one else can look this object up
 795                                  */
 796                                 (fs.object->handle == NULL) &&
 797                                 /*
 798                                  * No other ways to look the object up
 799                                  */
 800                                 ((fs.object->type == OBJT_DEFAULT) ||
 801                                  (fs.object->type == OBJT_SWAP)) &&
 802                             (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
 803                                 /*
 804                                  * We don't chase down the shadow chain
 805                                  */
 806                             fs.object == fs.first_object->backing_object) {
 807                                 vm_page_lock_queues();
 808                                 /*
 809                                  * get rid of the unnecessary page
 810                                  */
 811                                 vm_page_free(fs.first_m);
 812                                 /*
 813                                  * grab the page and put it into the
 814                                  * process'es object.  The page is
 815                                  * automatically made dirty.
 816                                  */
 817                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 818                                 vm_page_unlock_queues();
 819                                 vm_page_busy(fs.m);
 820                                 fs.first_m = fs.m;
 821                                 fs.m = NULL;
 822                                 PCPU_INC(cnt.v_cow_optim);
 823                         } else {
 824                                 /*
 825                                  * Oh, well, lets copy it.
 826                                  */
 827                                 pmap_copy_page(fs.m, fs.first_m);
 828                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 829                         }
 830                         if (fs.m) {
 831                                 /*
 832                                  * We no longer need the old page or object.
 833                                  */
 834                                 release_page(&fs);
 835                         }
 836                         /*
 837                          * fs.object != fs.first_object due to above
 838                          * conditional
 839                          */
 840                         vm_object_pip_wakeup(fs.object);
 841                         VM_OBJECT_UNLOCK(fs.object);
 842                         /*
 843                          * Only use the new page below...
 844                          */
 845                         fs.object = fs.first_object;
 846                         fs.pindex = fs.first_pindex;
 847                         fs.m = fs.first_m;
 848                         if (!is_first_object_locked)
 849                                 VM_OBJECT_LOCK(fs.object);
 850                         PCPU_INC(cnt.v_cow_faults);
 851                         curthread->td_cow++;
 852                 } else {
 853                         prot &= ~VM_PROT_WRITE;
 854                 }
 855         }
 856
 857         /*
 858          * We must verify that the maps have not changed since our last
 859          * lookup.
 860          */
 861         if (!fs.lookup_still_valid) {
 862                 vm_object_t retry_object;
 863                 vm_pindex_t retry_pindex;
 864                 vm_prot_t retry_prot;
 865
 866                 if (!vm_map_trylock_read(fs.map)) {
 867                         release_page(&fs);
 868                         unlock_and_deallocate(&fs);
 869                         goto RetryFault;
 870                 }
 871                 fs.lookup_still_valid = TRUE;
 872                 if (fs.map->timestamp != map_generation) {
 873                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 874                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 875
 876                         /*
 877                          * If we don't need the page any longer, put it on the inactive
 878                          * list (the easiest thing to do here).  If no one needs it,
 879                          * pageout will grab it eventually.
 880                          */
 881                         if (result != KERN_SUCCESS) {
 882                                 release_page(&fs);
 883                                 unlock_and_deallocate(&fs);
 884
 885                                 /*
 886                                  * If retry of map lookup would have blocked then
 887                                  * retry fault from start.
 888                                  */
 889                                 if (result == KERN_FAILURE)
 890                                         goto RetryFault;
 891                                 return (result);
 892                         }
 893                         if ((retry_object != fs.first_object) ||
 894                             (retry_pindex != fs.first_pindex)) {
 895                                 release_page(&fs);
 896                                 unlock_and_deallocate(&fs);
 897                                 goto RetryFault;
 898                         }
 899
 900                         /*
 901                          * Check whether the protection has changed or the object has
 902                          * been copied while we left the map unlocked. Changing from
 903                          * read to write permission is OK - we leave the page
 904                          * write-protected, and catch the write fault. Changing from
 905                          * write to read permission means that we can't mark the page
 906                          * write-enabled after all.
 907                          */
 908                         prot &= retry_prot;
 909                 }
 910         }
 911         /*
 912          * If the page was filled by a pager, update the map entry's
 913          * last read offset.  Since the pager does not return the
 914          * actual set of pages that it read, this update is based on
 915          * the requested set.  Typically, the requested and actual
 916          * sets are the same.
 917          *
 918          * XXX The following assignment modifies the map
 919          * without holding a write lock on it.
 920          */
 921         if (hardfault)
 922                 fs.entry->lastr = fs.pindex + faultcount - behind;
 923
 924         if (prot & VM_PROT_WRITE) {
 925                 vm_object_set_writeable_dirty(fs.object);
 926
 927                 /*
 928                  * If the fault is a write, we know that this page is being
 929                  * written NOW so dirty it explicitly to save on
 930                  * pmap_is_modified() calls later.
 931                  *
 932                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 933                  * if the page is already dirty to prevent data written with
 934                  * the expectation of being synced from not being synced.
 935                  * Likewise if this entry does not request NOSYNC then make
 936                  * sure the page isn't marked NOSYNC.  Applications sharing
 937                  * data should use the same flags to avoid ping ponging.
 938                  *
 939                  * Also tell the backing pager, if any, that it should remove
 940                  * any swap backing since the page is now dirty.
 941                  */
 942                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 943                         if (fs.m->dirty == 0)
 944                                 fs.m->oflags |= VPO_NOSYNC;
 945                 } else {
 946                         fs.m->oflags &= ~VPO_NOSYNC;
 947                 }
 948                 if (fault_flags & VM_FAULT_DIRTY) {
 949                         vm_page_dirty(fs.m);
 950                         vm_pager_page_unswapped(fs.m);
 951                 }
 952         }
 953
 954         /*
 955          * Page had better still be busy
 956          */
 957         KASSERT(fs.m->oflags & VPO_BUSY,
 958                 ("vm_fault: page %p not busy!", fs.m));
 959         /*
 960          * Page must be completely valid or it is not fit to
 961          * map into user space.  vm_pager_get_pages() ensures this.
 962          */
 963         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 964             ("vm_fault: page %p partially invalid", fs.m));
 965         VM_OBJECT_UNLOCK(fs.object);
 966
 967         /*
 968          * Put this page into the physical map.  We had to do the unlock above
 969          * because pmap_enter() may sleep.  We don't put the page
 970          * back on the active queue until later so that the pageout daemon
 971          * won't find it (yet).
 972          */
 973         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 974         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 975                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 976         }
 977         VM_OBJECT_LOCK(fs.object);
 978         vm_page_lock_queues();
 979         vm_page_flag_set(fs.m, PG_REFERENCED);
 980
 981         /*
 982          * If the page is not wired down, then put it where the pageout daemon
 983          * can find it.
 984          */
 985         if (fault_flags & VM_FAULT_WIRE_MASK) {
 986                 if (wired)
 987                         vm_page_wire(fs.m);
 988                 else
 989                         vm_page_unwire(fs.m, 1);
 990         } else {
 991                 vm_page_activate(fs.m);
 992         }
 993         vm_page_unlock_queues();
 994         vm_page_wakeup(fs.m);
 995
 996         /*
 997          * Unlock everything, and return
 998          */
 999         unlock_and_deallocate(&fs);
1000         if (hardfault)
1001                 curthread->td_ru.ru_majflt++;
1002         else
1003                 curthread->td_ru.ru_minflt++;
1004
1005         return (KERN_SUCCESS);
1006 }
1007
1008 /*
1009  * vm_fault_prefault provides a quick way of clustering
1010  * pagefaults into a processes address space.  It is a "cousin"
1011  * of vm_map_pmap_enter, except it runs at page fault time instead
1012  * of mmap time.
1013  */
1014 static void
1015 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
1016 {
1017         int i;
1018         vm_offset_t addr, starta;
1019         vm_pindex_t pindex;
1020         vm_page_t m;
1021         vm_object_t object;
1022
1023         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1024                 return;
1025
1026         object = entry->object.vm_object;
1027
1028         starta = addra - PFBAK * PAGE_SIZE;
1029         if (starta < entry->start) {
1030                 starta = entry->start;
1031         } else if (starta > addra) {
1032                 starta = 0;
1033         }
1034
1035         for (i = 0; i < PAGEORDER_SIZE; i++) {
1036                 vm_object_t backing_object, lobject;
1037
1038                 addr = addra + prefault_pageorder[i];
1039                 if (addr > addra + (PFFOR * PAGE_SIZE))
1040                         addr = 0;
1041
1042                 if (addr < starta || addr >= entry->end)
1043                         continue;
1044
1045                 if (!pmap_is_prefaultable(pmap, addr))
1046                         continue;
1047
1048                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1049                 lobject = object;
1050                 VM_OBJECT_LOCK(lobject);
1051                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1052                     lobject->type == OBJT_DEFAULT &&
1053                     (backing_object = lobject->backing_object) != NULL) {
1054                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1055                             0, ("vm_fault_prefault: unaligned object offset"));
1056                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1057                         VM_OBJECT_LOCK(backing_object);
1058                         VM_OBJECT_UNLOCK(lobject);
1059                         lobject = backing_object;
1060                 }
1061                 /*
1062                  * give-up when a page is not in memory
1063                  */
1064                 if (m == NULL) {
1065                         VM_OBJECT_UNLOCK(lobject);
1066                         break;
1067                 }
1068                 if (m->valid == VM_PAGE_BITS_ALL &&
1069                     (m->flags & PG_FICTITIOUS) == 0) {
1070                         vm_page_lock_queues();
1071                         pmap_enter_quick(pmap, addr, m, entry->protection);
1072                         vm_page_unlock_queues();
1073                 }
1074                 VM_OBJECT_UNLOCK(lobject);
1075         }
1076 }
1077
1078 /*
1079  *      vm_fault_quick:
1080  *
1081  *      Ensure that the requested virtual address, which may be in userland,
1082  *      is valid.  Fault-in the page if necessary.  Return -1 on failure.
1083  */
1084 int
1085 vm_fault_quick(caddr_t v, int prot)
1086 {
1087         int r;
1088
1089         if (prot & VM_PROT_WRITE)
1090                 r = subyte(v, fubyte(v));
1091         else
1092                 r = fubyte(v);
1093         return(r);
1094 }
1095
1096 /*
1097  *      vm_fault_wire:
1098  *
1099  *      Wire down a range of virtual addresses in a map.
1100  */
1101 int
1102 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1103     boolean_t user_wire, boolean_t fictitious)
1104 {
1105         vm_offset_t va;
1106         int rv;
1107
1108         /*
1109          * We simulate a fault to get the page and enter it in the physical
1110          * map.  For user wiring, we only ask for read access on currently
1111          * read-only sections.
1112          */
1113         for (va = start; va < end; va += PAGE_SIZE) {
1114                 rv = vm_fault(map, va,
1115                     user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1116                     user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1117                 if (rv) {
1118                         if (va != start)
1119                                 vm_fault_unwire(map, start, va, fictitious);
1120                         return (rv);
1121                 }
1122         }
1123         return (KERN_SUCCESS);
1124 }
1125
1126 /*
1127  *      vm_fault_unwire:
1128  *
1129  *      Unwire a range of virtual addresses in a map.
1130  */
1131 void
1132 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1133     boolean_t fictitious)
1134 {
1135         vm_paddr_t pa;
1136         vm_offset_t va;
1137         pmap_t pmap;
1138
1139         pmap = vm_map_pmap(map);
1140
1141         /*
1142          * Since the pages are wired down, we must be able to get their
1143          * mappings from the physical map system.
1144          */
1145         for (va = start; va < end; va += PAGE_SIZE) {
1146                 pa = pmap_extract(pmap, va);
1147                 if (pa != 0) {
1148                         pmap_change_wiring(pmap, va, FALSE);
1149                         if (!fictitious) {
1150                                 vm_page_lock_queues();
1151                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1152                                 vm_page_unlock_queues();
1153                         }
1154                 }
1155         }
1156 }
1157
1158 /*
1159  *      Routine:
1160  *              vm_fault_copy_entry
1161  *      Function:
1162  *              Create new shadow object backing dst_entry with private copy of
1163  *              all underlying pages. When src_entry is equal to dst_entry,
1164  *              function implements COW for wired-down map entry. Otherwise,
1165  *              it forks wired entry into dst_map.
1166  *
1167  *      In/out conditions:
1168  *              The source and destination maps must be locked for write.
1169  *              The source map entry must be wired down (or be a sharing map
1170  *              entry corresponding to a main map entry that is wired down).
1171  */
1172 void
1173 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1174     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1175     vm_ooffset_t *fork_charge)
1176 {
1177         vm_object_t backing_object, dst_object, object, src_object;
1178         vm_pindex_t dst_pindex, pindex, src_pindex;
1179         vm_prot_t access, prot;
1180         vm_offset_t vaddr;
1181         vm_page_t dst_m;
1182         vm_page_t src_m;
1183         boolean_t src_readonly, upgrade;
1184
1185 #ifdef  lint
1186         src_map++;
1187 #endif  /* lint */
1188
1189         upgrade = src_entry == dst_entry;
1190
1191         src_object = src_entry->object.vm_object;
1192         src_pindex = OFF_TO_IDX(src_entry->offset);
1193         src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1194
1195         /*
1196          * Create the top-level object for the destination entry. (Doesn't
1197          * actually shadow anything - we copy the pages directly.)
1198          */
1199         dst_object = vm_object_allocate(OBJT_DEFAULT,
1200             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1201 #if VM_NRESERVLEVEL > 0
1202         dst_object->flags |= OBJ_COLORED;
1203         dst_object->pg_color = atop(dst_entry->start);
1204 #endif
1205
1206         VM_OBJECT_LOCK(dst_object);
1207         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1208             ("vm_fault_copy_entry: vm_object not NULL"));
1209         dst_entry->object.vm_object = dst_object;
1210         dst_entry->offset = 0;
1211         dst_object->charge = dst_entry->end - dst_entry->start;
1212         if (fork_charge != NULL) {
1213                 KASSERT(dst_entry->uip == NULL,
1214                     ("vm_fault_copy_entry: leaked swp charge"));
1215                 dst_object->uip = curthread->td_ucred->cr_ruidinfo;
1216                 uihold(dst_object->uip);
1217                 *fork_charge += dst_object->charge;
1218         } else {
1219                 dst_object->uip = dst_entry->uip;
1220                 dst_entry->uip = NULL;
1221         }
1222         access = prot = dst_entry->max_protection;
1223         /*
1224          * If not an upgrade, then enter the mappings in the pmap as
1225          * read and/or execute accesses.  Otherwise, enter them as
1226          * write accesses.
1227          *
1228          * A writeable large page mapping is only created if all of
1229          * the constituent small page mappings are modified. Marking
1230          * PTEs as modified on inception allows promotion to happen
1231          * without taking potentially large number of soft faults.
1232          */
1233         if (!upgrade)
1234                 access &= ~VM_PROT_WRITE;
1235
1236         /*
1237          * Loop through all of the pages in the entry's range, copying each
1238          * one from the source object (it should be there) to the destination
1239          * object.
1240          */
1241         for (vaddr = dst_entry->start, dst_pindex = 0;
1242             vaddr < dst_entry->end;
1243             vaddr += PAGE_SIZE, dst_pindex++) {
1244
1245                 /*
1246                  * Allocate a page in the destination object.
1247                  */
1248                 do {
1249                         dst_m = vm_page_alloc(dst_object, dst_pindex,
1250                             VM_ALLOC_NORMAL);
1251                         if (dst_m == NULL) {
1252                                 VM_OBJECT_UNLOCK(dst_object);
1253                                 VM_WAIT;
1254                                 VM_OBJECT_LOCK(dst_object);
1255                         }
1256                 } while (dst_m == NULL);
1257
1258                 /*
1259                  * Find the page in the source object, and copy it in.
1260                  * (Because the source is wired down, the page will be in
1261                  * memory.)
1262                  */
1263                 VM_OBJECT_LOCK(src_object);
1264                 object = src_object;
1265                 pindex = src_pindex + dst_pindex;
1266                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1267                     src_readonly &&
1268                     (backing_object = object->backing_object) != NULL) {
1269                         /*
1270                          * Allow fallback to backing objects if we are reading.
1271                          */
1272                         VM_OBJECT_LOCK(backing_object);
1273                         pindex += OFF_TO_IDX(object->backing_object_offset);
1274                         VM_OBJECT_UNLOCK(object);
1275                         object = backing_object;
1276                 }
1277                 if (src_m == NULL)
1278                         panic("vm_fault_copy_wired: page missing");
1279                 pmap_copy_page(src_m, dst_m);
1280                 VM_OBJECT_UNLOCK(object);
1281                 dst_m->valid = VM_PAGE_BITS_ALL;
1282                 VM_OBJECT_UNLOCK(dst_object);
1283
1284                 /*
1285                  * Enter it in the pmap. If a wired, copy-on-write
1286                  * mapping is being replaced by a write-enabled
1287                  * mapping, then wire that new mapping.
1288                  */
1289                 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1290
1291                 /*
1292                  * Mark it no longer busy, and put it on the active list.
1293                  */
1294                 VM_OBJECT_LOCK(dst_object);
1295                 vm_page_lock_queues();
1296                 if (upgrade) {
1297                         vm_page_unwire(src_m, 0);
1298                         vm_page_wire(dst_m);
1299                 } else
1300                         vm_page_activate(dst_m);
1301                 vm_page_unlock_queues();
1302                 vm_page_wakeup(dst_m);
1303         }
1304         VM_OBJECT_UNLOCK(dst_object);
1305         if (upgrade) {
1306                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1307                 vm_object_deallocate(src_object);
1308         }
1309 }
1310
1311
1312 /*
1313  * This routine checks around the requested page for other pages that
1314  * might be able to be faulted in.  This routine brackets the viable
1315  * pages for the pages to be paged in.
1316  *
1317  * Inputs:
1318  *      m, rbehind, rahead
1319  *
1320  * Outputs:
1321  *  marray (array of vm_page_t), reqpage (index of requested page)
1322  *
1323  * Return value:
1324  *  number of pages in marray
1325  */
1326 static int
1327 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1328         vm_page_t m;
1329         int rbehind;
1330         int rahead;
1331         vm_page_t *marray;
1332         int *reqpage;
1333 {
1334         int i,j;
1335         vm_object_t object;
1336         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1337         vm_page_t rtm;
1338         int cbehind, cahead;
1339
1340         VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1341
1342         object = m->object;
1343         pindex = m->pindex;
1344         cbehind = cahead = 0;
1345
1346         /*
1347          * if the requested page is not available, then give up now
1348          */
1349         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1350                 return 0;
1351         }
1352
1353         if ((cbehind == 0) && (cahead == 0)) {
1354                 *reqpage = 0;
1355                 marray[0] = m;
1356                 return 1;
1357         }
1358
1359         if (rahead > cahead) {
1360                 rahead = cahead;
1361         }
1362
1363         if (rbehind > cbehind) {
1364                 rbehind = cbehind;
1365         }
1366
1367         /*
1368          * scan backward for the read behind pages -- in memory
1369          */
1370         if (pindex > 0) {
1371                 if (rbehind > pindex) {
1372                         rbehind = pindex;
1373                         startpindex = 0;
1374                 } else {
1375                         startpindex = pindex - rbehind;
1376                 }
1377
1378                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1379                     rtm->pindex >= startpindex)
1380                         startpindex = rtm->pindex + 1;
1381
1382                 /* tpindex is unsigned; beware of numeric underflow. */
1383                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1384                     tpindex < pindex; i++, tpindex--) {
1385
1386                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1387                             VM_ALLOC_IFNOTCACHED);
1388                         if (rtm == NULL) {
1389                                 /*
1390                                  * Shift the allocated pages to the
1391                                  * beginning of the array.
1392                                  */
1393                                 for (j = 0; j < i; j++) {
1394                                         marray[j] = marray[j + tpindex + 1 -
1395                                             startpindex];
1396                                 }
1397                                 break;
1398                         }
1399
1400                         marray[tpindex - startpindex] = rtm;
1401                 }
1402         } else {
1403                 startpindex = 0;
1404                 i = 0;
1405         }
1406
1407         marray[i] = m;
1408         /* page offset of the required page */
1409         *reqpage = i;
1410
1411         tpindex = pindex + 1;
1412         i++;
1413
1414         /*
1415          * scan forward for the read ahead pages
1416          */
1417         endpindex = tpindex + rahead;
1418         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1419                 endpindex = rtm->pindex;
1420         if (endpindex > object->size)
1421                 endpindex = object->size;
1422
1423         for (; tpindex < endpindex; i++, tpindex++) {
1424
1425                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1426                     VM_ALLOC_IFNOTCACHED);
1427                 if (rtm == NULL) {
1428                         break;
1429                 }
1430
1431                 marray[i] = rtm;
1432         }
1433
1434         /* return number of pages */
1435         return i;
1436 }
1437
1438 /*
1439  * Block entry into the machine-independent layer's page fault handler by
1440  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1441  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1442  * spurious page faults.
1443  */
1444 int
1445 vm_fault_disable_pagefaults(void)
1446 {
1447
1448         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1449 }
1450
1451 void
1452 vm_fault_enable_pagefaults(int save)
1453 {
1454
1455         curthread_pflags_restore(save);
1456 }