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/proc.h>
  85 #include <sys/resourcevar.h>
  86 #include <sys/rwlock.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/vm_extern.h>
 104
 105 #define PFBAK 4
 106 #define PFFOR 4
 107
 108 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 109
 110 #define VM_FAULT_READ_BEHIND    8
 111 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 112 #define VM_FAULT_NINCR          (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
 113 #define VM_FAULT_SUM            (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
 114 #define VM_FAULT_CACHE_BEHIND   (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
 115
 116 struct faultstate {
 117         vm_page_t m;
 118         vm_object_t object;
 119         vm_pindex_t pindex;
 120         vm_page_t first_m;
 121         vm_object_t     first_object;
 122         vm_pindex_t first_pindex;
 123         vm_map_t map;
 124         vm_map_entry_t entry;
 125         int lookup_still_valid;
 126         struct vnode *vp;
 127 };
 128
 129 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
 130 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 131             int faultcount, int reqpage);
 132
 133 static inline void
 134 release_page(struct faultstate *fs)
 135 {
 136
 137         vm_page_xunbusy(fs->m);
 138         vm_page_lock(fs->m);
 139         vm_page_deactivate(fs->m);
 140         vm_page_unlock(fs->m);
 141         fs->m = NULL;
 142 }
 143
 144 static inline void
 145 unlock_map(struct faultstate *fs)
 146 {
 147
 148         if (fs->lookup_still_valid) {
 149                 vm_map_lookup_done(fs->map, fs->entry);
 150                 fs->lookup_still_valid = FALSE;
 151         }
 152 }
 153
 154 static void
 155 unlock_and_deallocate(struct faultstate *fs)
 156 {
 157
 158         vm_object_pip_wakeup(fs->object);
 159         VM_OBJECT_WUNLOCK(fs->object);
 160         if (fs->object != fs->first_object) {
 161                 VM_OBJECT_WLOCK(fs->first_object);
 162                 vm_page_lock(fs->first_m);
 163                 vm_page_free(fs->first_m);
 164                 vm_page_unlock(fs->first_m);
 165                 vm_object_pip_wakeup(fs->first_object);
 166                 VM_OBJECT_WUNLOCK(fs->first_object);
 167                 fs->first_m = NULL;
 168         }
 169         vm_object_deallocate(fs->first_object);
 170         unlock_map(fs);
 171         if (fs->vp != NULL) {
 172                 vput(fs->vp);
 173                 fs->vp = NULL;
 174         }
 175 }
 176
 177 /*
 178  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 179  *            current object *might* contain the page.
 180  *
 181  *            default objects are zero-fill, there is no real pager.
 182  */
 183 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 184                         ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
 185
 186 /*
 187  *      vm_fault:
 188  *
 189  *      Handle a page fault occurring at the given address,
 190  *      requiring the given permissions, in the map specified.
 191  *      If successful, the page is inserted into the
 192  *      associated physical map.
 193  *
 194  *      NOTE: the given address should be truncated to the
 195  *      proper page address.
 196  *
 197  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 198  *      a standard error specifying why the fault is fatal is returned.
 199  *
 200  *      The map in question must be referenced, and remains so.
 201  *      Caller may hold no locks.
 202  */
 203 int
 204 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 205     int fault_flags)
 206 {
 207         struct thread *td;
 208         int result;
 209
 210         td = curthread;
 211         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 212                 return (KERN_PROTECTION_FAILURE);
 213 #ifdef KTRACE
 214         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 215                 ktrfault(vaddr, fault_type);
 216 #endif
 217         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 218             NULL);
 219 #ifdef KTRACE
 220         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 221                 ktrfaultend(result);
 222 #endif
 223         return (result);
 224 }
 225
 226 int
 227 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 228     int fault_flags, vm_page_t *m_hold)
 229 {
 230         vm_prot_t prot;
 231         long ahead, behind;
 232         int alloc_req, era, faultcount, nera, reqpage, result;
 233         boolean_t growstack, is_first_object_locked, wired;
 234         int map_generation;
 235         vm_object_t next_object;
 236         vm_page_t marray[VM_FAULT_READ_MAX];
 237         int hardfault;
 238         struct faultstate fs;
 239         struct vnode *vp;
 240         int locked, error;
 241
 242         hardfault = 0;
 243         growstack = TRUE;
 244         PCPU_INC(cnt.v_vm_faults);
 245         fs.vp = NULL;
 246         faultcount = reqpage = 0;
 247
 248 RetryFault:;
 249
 250         /*
 251          * Find the backing store object and offset into it to begin the
 252          * search.
 253          */
 254         fs.map = map;
 255         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 256             &fs.first_object, &fs.first_pindex, &prot, &wired);
 257         if (result != KERN_SUCCESS) {
 258                 if (growstack && result == KERN_INVALID_ADDRESS &&
 259                     map != kernel_map) {
 260                         result = vm_map_growstack(curproc, vaddr);
 261                         if (result != KERN_SUCCESS)
 262                                 return (KERN_FAILURE);
 263                         growstack = FALSE;
 264                         goto RetryFault;
 265                 }
 266                 return (result);
 267         }
 268
 269         map_generation = fs.map->timestamp;
 270
 271         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 272                 if ((curthread->td_pflags & TDP_DEVMEMIO) != 0) {
 273                         vm_map_unlock_read(fs.map);
 274                         return (KERN_FAILURE);
 275                 }
 276                 panic("vm_fault: fault on nofault entry, addr: %lx",
 277                     (u_long)vaddr);
 278         }
 279
 280         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 281             fs.entry->wiring_thread != curthread) {
 282                 vm_map_unlock_read(fs.map);
 283                 vm_map_lock(fs.map);
 284                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 285                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 286                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 287                         vm_map_unlock_and_wait(fs.map, 0);
 288                 } else
 289                         vm_map_unlock(fs.map);
 290                 goto RetryFault;
 291         }
 292
 293         /*
 294          * Make a reference to this object to prevent its disposal while we
 295          * are messing with it.  Once we have the reference, the map is free
 296          * to be diddled.  Since objects reference their shadows (and copies),
 297          * they will stay around as well.
 298          *
 299          * Bump the paging-in-progress count to prevent size changes (e.g.
 300          * truncation operations) during I/O.  This must be done after
 301          * obtaining the vnode lock in order to avoid possible deadlocks.
 302          */
 303         VM_OBJECT_WLOCK(fs.first_object);
 304         vm_object_reference_locked(fs.first_object);
 305         vm_object_pip_add(fs.first_object, 1);
 306
 307         fs.lookup_still_valid = TRUE;
 308
 309         if (wired)
 310                 fault_type = prot | (fault_type & VM_PROT_COPY);
 311
 312         fs.first_m = NULL;
 313
 314         /*
 315          * Search for the page at object/offset.
 316          */
 317         fs.object = fs.first_object;
 318         fs.pindex = fs.first_pindex;
 319         while (TRUE) {
 320                 /*
 321                  * If the object is dead, we stop here
 322                  */
 323                 if (fs.object->flags & OBJ_DEAD) {
 324                         unlock_and_deallocate(&fs);
 325                         return (KERN_PROTECTION_FAILURE);
 326                 }
 327
 328                 /*
 329                  * See if page is resident
 330                  */
 331                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 332                 if (fs.m != NULL) {
 333                         /*
 334                          * Wait/Retry if the page is busy.  We have to do this
 335                          * if the page is either exclusive or shared busy
 336                          * because the vm_pager may be using read busy for
 337                          * pageouts (and even pageins if it is the vnode
 338                          * pager), and we could end up trying to pagein and
 339                          * pageout the same page simultaneously.
 340                          *
 341                          * We can theoretically allow the busy case on a read
 342                          * fault if the page is marked valid, but since such
 343                          * pages are typically already pmap'd, putting that
 344                          * special case in might be more effort then it is
 345                          * worth.  We cannot under any circumstances mess
 346                          * around with a shared busied page except, perhaps,
 347                          * to pmap it.
 348                          */
 349                         if (vm_page_busied(fs.m)) {
 350                                 /*
 351                                  * Reference the page before unlocking and
 352                                  * sleeping so that the page daemon is less
 353                                  * likely to reclaim it.
 354                                  */
 355                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 356                                 if (fs.object != fs.first_object) {
 357                                         if (!VM_OBJECT_TRYWLOCK(
 358                                             fs.first_object)) {
 359                                                 VM_OBJECT_WUNLOCK(fs.object);
 360                                                 VM_OBJECT_WLOCK(fs.first_object);
 361                                                 VM_OBJECT_WLOCK(fs.object);
 362                                         }
 363                                         vm_page_lock(fs.first_m);
 364                                         vm_page_free(fs.first_m);
 365                                         vm_page_unlock(fs.first_m);
 366                                         vm_object_pip_wakeup(fs.first_object);
 367                                         VM_OBJECT_WUNLOCK(fs.first_object);
 368                                         fs.first_m = NULL;
 369                                 }
 370                                 unlock_map(&fs);
 371                                 if (fs.m == vm_page_lookup(fs.object,
 372                                     fs.pindex)) {
 373                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 374                                 }
 375                                 vm_object_pip_wakeup(fs.object);
 376                                 VM_OBJECT_WUNLOCK(fs.object);
 377                                 PCPU_INC(cnt.v_intrans);
 378                                 vm_object_deallocate(fs.first_object);
 379                                 goto RetryFault;
 380                         }
 381                         vm_page_lock(fs.m);
 382                         vm_page_remque(fs.m);
 383                         vm_page_unlock(fs.m);
 384
 385                         /*
 386                          * Mark page busy for other processes, and the
 387                          * pagedaemon.  If it still isn't completely valid
 388                          * (readable), jump to readrest, else break-out ( we
 389                          * found the page ).
 390                          */
 391                         vm_page_xbusy(fs.m);
 392                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 393                                 goto readrest;
 394                         break;
 395                 }
 396
 397                 /*
 398                  * Page is not resident, If this is the search termination
 399                  * or the pager might contain the page, allocate a new page.
 400                  */
 401                 if (TRYPAGER || fs.object == fs.first_object) {
 402                         if (fs.pindex >= fs.object->size) {
 403                                 unlock_and_deallocate(&fs);
 404                                 return (KERN_PROTECTION_FAILURE);
 405                         }
 406
 407                         /*
 408                          * Allocate a new page for this object/offset pair.
 409                          *
 410                          * Unlocked read of the p_flag is harmless. At
 411                          * worst, the P_KILLED might be not observed
 412                          * there, and allocation can fail, causing
 413                          * restart and new reading of the p_flag.
 414                          */
 415                         fs.m = NULL;
 416                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 417 #if VM_NRESERVLEVEL > 0
 418                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 419                                         fs.object->flags |= OBJ_COLORED;
 420                                         fs.object->pg_color = atop(vaddr) -
 421                                             fs.pindex;
 422                                 }
 423 #endif
 424                                 alloc_req = P_KILLED(curproc) ?
 425                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 426                                 if (fs.object->type != OBJT_VNODE &&
 427                                     fs.object->backing_object == NULL)
 428                                         alloc_req |= VM_ALLOC_ZERO;
 429                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 430                                     alloc_req);
 431                         }
 432                         if (fs.m == NULL) {
 433                                 unlock_and_deallocate(&fs);
 434                                 VM_WAITPFAULT;
 435                                 goto RetryFault;
 436                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 437                                 break;
 438                 }
 439
 440 readrest:
 441                 /*
 442                  * We have found a valid page or we have allocated a new page.
 443                  * The page thus may not be valid or may not be entirely
 444                  * valid.
 445                  *
 446                  * Attempt to fault-in the page if there is a chance that the
 447                  * pager has it, and potentially fault in additional pages
 448                  * at the same time.
 449                  */
 450                 if (TRYPAGER) {
 451                         int rv;
 452                         u_char behavior = vm_map_entry_behavior(fs.entry);
 453
 454                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 455                             P_KILLED(curproc)) {
 456                                 behind = 0;
 457                                 ahead = 0;
 458                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 459                                 behind = 0;
 460                                 ahead = atop(fs.entry->end - vaddr) - 1;
 461                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 462                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 463                                 if (fs.pindex == fs.entry->next_read)
 464                                         vm_fault_cache_behind(&fs,
 465                                             VM_FAULT_READ_MAX);
 466                         } else {
 467                                 /*
 468                                  * If this is a sequential page fault, then
 469                                  * arithmetically increase the number of pages
 470                                  * in the read-ahead window.  Otherwise, reset
 471                                  * the read-ahead window to its smallest size.
 472                                  */
 473                                 behind = atop(vaddr - fs.entry->start);
 474                                 if (behind > VM_FAULT_READ_BEHIND)
 475                                         behind = VM_FAULT_READ_BEHIND;
 476                                 ahead = atop(fs.entry->end - vaddr) - 1;
 477                                 era = fs.entry->read_ahead;
 478                                 if (fs.pindex == fs.entry->next_read) {
 479                                         nera = era + behind;
 480                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 481                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 482                                         behind = 0;
 483                                         if (ahead > nera)
 484                                                 ahead = nera;
 485                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 486                                                 vm_fault_cache_behind(&fs,
 487                                                     VM_FAULT_CACHE_BEHIND);
 488                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 489                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 490                                 if (era != ahead)
 491                                         fs.entry->read_ahead = ahead;
 492                         }
 493
 494                         /*
 495                          * Call the pager to retrieve the data, if any, after
 496                          * releasing the lock on the map.  We hold a ref on
 497                          * fs.object and the pages are exclusive busied.
 498                          */
 499                         unlock_map(&fs);
 500
 501                         if (fs.object->type == OBJT_VNODE) {
 502                                 vp = fs.object->handle;
 503                                 if (vp == fs.vp)
 504                                         goto vnode_locked;
 505                                 else if (fs.vp != NULL) {
 506                                         vput(fs.vp);
 507                                         fs.vp = NULL;
 508                                 }
 509                                 locked = VOP_ISLOCKED(vp);
 510
 511                                 if (locked != LK_EXCLUSIVE)
 512                                         locked = LK_SHARED;
 513                                 /* Do not sleep for vnode lock while fs.m is busy */
 514                                 error = vget(vp, locked | LK_CANRECURSE |
 515                                     LK_NOWAIT, curthread);
 516                                 if (error != 0) {
 517                                         vhold(vp);
 518                                         release_page(&fs);
 519                                         unlock_and_deallocate(&fs);
 520                                         error = vget(vp, locked | LK_RETRY |
 521                                             LK_CANRECURSE, curthread);
 522                                         vdrop(vp);
 523                                         fs.vp = vp;
 524                                         KASSERT(error == 0,
 525                                             ("vm_fault: vget failed"));
 526                                         goto RetryFault;
 527                                 }
 528                                 fs.vp = vp;
 529                         }
 530 vnode_locked:
 531                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 532                             ("vm_fault: vnode-backed object mapped by system map"));
 533
 534                         /*
 535                          * now we find out if any other pages should be paged
 536                          * in at this time this routine checks to see if the
 537                          * pages surrounding this fault reside in the same
 538                          * object as the page for this fault.  If they do,
 539                          * then they are faulted in also into the object.  The
 540                          * array "marray" returned contains an array of
 541                          * vm_page_t structs where one of them is the
 542                          * vm_page_t passed to the routine.  The reqpage
 543                          * return value is the index into the marray for the
 544                          * vm_page_t passed to the routine.
 545                          *
 546                          * fs.m plus the additional pages are exclusive busied.
 547                          */
 548                         faultcount = vm_fault_additional_pages(
 549                             fs.m, behind, ahead, marray, &reqpage);
 550
 551                         rv = faultcount ?
 552                             vm_pager_get_pages(fs.object, marray, faultcount,
 553                                 reqpage) : VM_PAGER_FAIL;
 554
 555                         if (rv == VM_PAGER_OK) {
 556                                 /*
 557                                  * Found the page. Leave it busy while we play
 558                                  * with it.
 559                                  */
 560
 561                                 /*
 562                                  * Relookup in case pager changed page. Pager
 563                                  * is responsible for disposition of old page
 564                                  * if moved.
 565                                  */
 566                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 567                                 if (!fs.m) {
 568                                         unlock_and_deallocate(&fs);
 569                                         goto RetryFault;
 570                                 }
 571
 572                                 hardfault++;
 573                                 break; /* break to PAGE HAS BEEN FOUND */
 574                         }
 575                         /*
 576                          * Remove the bogus page (which does not exist at this
 577                          * object/offset); before doing so, we must get back
 578                          * our object lock to preserve our invariant.
 579                          *
 580                          * Also wake up any other process that may want to bring
 581                          * in this page.
 582                          *
 583                          * If this is the top-level object, we must leave the
 584                          * busy page to prevent another process from rushing
 585                          * past us, and inserting the page in that object at
 586                          * the same time that we are.
 587                          */
 588                         if (rv == VM_PAGER_ERROR)
 589                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 590                                     curproc->p_pid, curproc->p_comm);
 591                         /*
 592                          * Data outside the range of the pager or an I/O error
 593                          */
 594                         /*
 595                          * XXX - the check for kernel_map is a kludge to work
 596                          * around having the machine panic on a kernel space
 597                          * fault w/ I/O error.
 598                          */
 599                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 600                                 (rv == VM_PAGER_BAD)) {
 601                                 vm_page_lock(fs.m);
 602                                 vm_page_free(fs.m);
 603                                 vm_page_unlock(fs.m);
 604                                 fs.m = NULL;
 605                                 unlock_and_deallocate(&fs);
 606                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 607                         }
 608                         if (fs.object != fs.first_object) {
 609                                 vm_page_lock(fs.m);
 610                                 vm_page_free(fs.m);
 611                                 vm_page_unlock(fs.m);
 612                                 fs.m = NULL;
 613                                 /*
 614                                  * XXX - we cannot just fall out at this
 615                                  * point, m has been freed and is invalid!
 616                                  */
 617                         }
 618                 }
 619
 620                 /*
 621                  * We get here if the object has default pager (or unwiring)
 622                  * or the pager doesn't have the page.
 623                  */
 624                 if (fs.object == fs.first_object)
 625                         fs.first_m = fs.m;
 626
 627                 /*
 628                  * Move on to the next object.  Lock the next object before
 629                  * unlocking the current one.
 630                  */
 631                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 632                 next_object = fs.object->backing_object;
 633                 if (next_object == NULL) {
 634                         /*
 635                          * If there's no object left, fill the page in the top
 636                          * object with zeros.
 637                          */
 638                         if (fs.object != fs.first_object) {
 639                                 vm_object_pip_wakeup(fs.object);
 640                                 VM_OBJECT_WUNLOCK(fs.object);
 641
 642                                 fs.object = fs.first_object;
 643                                 fs.pindex = fs.first_pindex;
 644                                 fs.m = fs.first_m;
 645                                 VM_OBJECT_WLOCK(fs.object);
 646                         }
 647                         fs.first_m = NULL;
 648
 649                         /*
 650                          * Zero the page if necessary and mark it valid.
 651                          */
 652                         if ((fs.m->flags & PG_ZERO) == 0) {
 653                                 pmap_zero_page(fs.m);
 654                         } else {
 655                                 PCPU_INC(cnt.v_ozfod);
 656                         }
 657                         PCPU_INC(cnt.v_zfod);
 658                         fs.m->valid = VM_PAGE_BITS_ALL;
 659                         break;  /* break to PAGE HAS BEEN FOUND */
 660                 } else {
 661                         KASSERT(fs.object != next_object,
 662                             ("object loop %p", next_object));
 663                         VM_OBJECT_WLOCK(next_object);
 664                         vm_object_pip_add(next_object, 1);
 665                         if (fs.object != fs.first_object)
 666                                 vm_object_pip_wakeup(fs.object);
 667                         VM_OBJECT_WUNLOCK(fs.object);
 668                         fs.object = next_object;
 669                 }
 670         }
 671
 672         vm_page_assert_xbusied(fs.m);
 673
 674         /*
 675          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 676          * is held.]
 677          */
 678
 679         /*
 680          * If the page is being written, but isn't already owned by the
 681          * top-level object, we have to copy it into a new page owned by the
 682          * top-level object.
 683          */
 684         if (fs.object != fs.first_object) {
 685                 /*
 686                  * We only really need to copy if we want to write it.
 687                  */
 688                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 689                         /*
 690                          * This allows pages to be virtually copied from a
 691                          * backing_object into the first_object, where the
 692                          * backing object has no other refs to it, and cannot
 693                          * gain any more refs.  Instead of a bcopy, we just
 694                          * move the page from the backing object to the
 695                          * first object.  Note that we must mark the page
 696                          * dirty in the first object so that it will go out
 697                          * to swap when needed.
 698                          */
 699                         is_first_object_locked = FALSE;
 700                         if (
 701                                 /*
 702                                  * Only one shadow object
 703                                  */
 704                                 (fs.object->shadow_count == 1) &&
 705                                 /*
 706                                  * No COW refs, except us
 707                                  */
 708                                 (fs.object->ref_count == 1) &&
 709                                 /*
 710                                  * No one else can look this object up
 711                                  */
 712                                 (fs.object->handle == NULL) &&
 713                                 /*
 714                                  * No other ways to look the object up
 715                                  */
 716                                 ((fs.object->type == OBJT_DEFAULT) ||
 717                                  (fs.object->type == OBJT_SWAP)) &&
 718                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
 719                                 /*
 720                                  * We don't chase down the shadow chain
 721                                  */
 722                             fs.object == fs.first_object->backing_object) {
 723                                 /*
 724                                  * get rid of the unnecessary page
 725                                  */
 726                                 vm_page_lock(fs.first_m);
 727                                 vm_page_free(fs.first_m);
 728                                 vm_page_unlock(fs.first_m);
 729                                 /*
 730                                  * grab the page and put it into the
 731                                  * process'es object.  The page is
 732                                  * automatically made dirty.
 733                                  */
 734                                 if (vm_page_rename(fs.m, fs.first_object,
 735                                     fs.first_pindex)) {
 736                                         unlock_and_deallocate(&fs);
 737                                         goto RetryFault;
 738                                 }
 739                                 vm_page_xbusy(fs.m);
 740                                 fs.first_m = fs.m;
 741                                 fs.m = NULL;
 742                                 PCPU_INC(cnt.v_cow_optim);
 743                         } else {
 744                                 /*
 745                                  * Oh, well, lets copy it.
 746                                  */
 747                                 pmap_copy_page(fs.m, fs.first_m);
 748                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 749                                 if (wired && (fault_flags &
 750                                     VM_FAULT_CHANGE_WIRING) == 0) {
 751                                         vm_page_lock(fs.first_m);
 752                                         vm_page_wire(fs.first_m);
 753                                         vm_page_unlock(fs.first_m);
 754
 755                                         vm_page_lock(fs.m);
 756                                         vm_page_unwire(fs.m, FALSE);
 757                                         vm_page_unlock(fs.m);
 758                                 }
 759                                 /*
 760                                  * We no longer need the old page or object.
 761                                  */
 762                                 release_page(&fs);
 763                         }
 764                         /*
 765                          * fs.object != fs.first_object due to above
 766                          * conditional
 767                          */
 768                         vm_object_pip_wakeup(fs.object);
 769                         VM_OBJECT_WUNLOCK(fs.object);
 770                         /*
 771                          * Only use the new page below...
 772                          */
 773                         fs.object = fs.first_object;
 774                         fs.pindex = fs.first_pindex;
 775                         fs.m = fs.first_m;
 776                         if (!is_first_object_locked)
 777                                 VM_OBJECT_WLOCK(fs.object);
 778                         PCPU_INC(cnt.v_cow_faults);
 779                         curthread->td_cow++;
 780                 } else {
 781                         prot &= ~VM_PROT_WRITE;
 782                 }
 783         }
 784
 785         /*
 786          * We must verify that the maps have not changed since our last
 787          * lookup.
 788          */
 789         if (!fs.lookup_still_valid) {
 790                 vm_object_t retry_object;
 791                 vm_pindex_t retry_pindex;
 792                 vm_prot_t retry_prot;
 793
 794                 if (!vm_map_trylock_read(fs.map)) {
 795                         release_page(&fs);
 796                         unlock_and_deallocate(&fs);
 797                         goto RetryFault;
 798                 }
 799                 fs.lookup_still_valid = TRUE;
 800                 if (fs.map->timestamp != map_generation) {
 801                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 802                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 803
 804                         /*
 805                          * If we don't need the page any longer, put it on the inactive
 806                          * list (the easiest thing to do here).  If no one needs it,
 807                          * pageout will grab it eventually.
 808                          */
 809                         if (result != KERN_SUCCESS) {
 810                                 release_page(&fs);
 811                                 unlock_and_deallocate(&fs);
 812
 813                                 /*
 814                                  * If retry of map lookup would have blocked then
 815                                  * retry fault from start.
 816                                  */
 817                                 if (result == KERN_FAILURE)
 818                                         goto RetryFault;
 819                                 return (result);
 820                         }
 821                         if ((retry_object != fs.first_object) ||
 822                             (retry_pindex != fs.first_pindex)) {
 823                                 release_page(&fs);
 824                                 unlock_and_deallocate(&fs);
 825                                 goto RetryFault;
 826                         }
 827
 828                         /*
 829                          * Check whether the protection has changed or the object has
 830                          * been copied while we left the map unlocked. Changing from
 831                          * read to write permission is OK - we leave the page
 832                          * write-protected, and catch the write fault. Changing from
 833                          * write to read permission means that we can't mark the page
 834                          * write-enabled after all.
 835                          */
 836                         prot &= retry_prot;
 837                 }
 838         }
 839         /*
 840          * If the page was filled by a pager, update the map entry's
 841          * last read offset.  Since the pager does not return the
 842          * actual set of pages that it read, this update is based on
 843          * the requested set.  Typically, the requested and actual
 844          * sets are the same.
 845          *
 846          * XXX The following assignment modifies the map
 847          * without holding a write lock on it.
 848          */
 849         if (hardfault)
 850                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 851
 852         if (((prot & VM_PROT_WRITE) != 0 ||
 853             (fault_flags & VM_FAULT_DIRTY) != 0) &&
 854             (fs.m->oflags & VPO_UNMANAGED) == 0) {
 855                 vm_object_set_writeable_dirty(fs.object);
 856
 857                 /*
 858                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 859                  * if the page is already dirty to prevent data written with
 860                  * the expectation of being synced from not being synced.
 861                  * Likewise if this entry does not request NOSYNC then make
 862                  * sure the page isn't marked NOSYNC.  Applications sharing
 863                  * data should use the same flags to avoid ping ponging.
 864                  */
 865                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 866                         if (fs.m->dirty == 0)
 867                                 fs.m->oflags |= VPO_NOSYNC;
 868                 } else {
 869                         fs.m->oflags &= ~VPO_NOSYNC;
 870                 }
 871
 872                 /*
 873                  * If the fault is a write, we know that this page is being
 874                  * written NOW so dirty it explicitly to save on
 875                  * pmap_is_modified() calls later.
 876                  *
 877                  * Also tell the backing pager, if any, that it should remove
 878                  * any swap backing since the page is now dirty.
 879                  */
 880                 if (((fault_type & VM_PROT_WRITE) != 0 &&
 881                     (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 882                     (fault_flags & VM_FAULT_DIRTY) != 0) {
 883                         vm_page_dirty(fs.m);
 884                         vm_pager_page_unswapped(fs.m);
 885                 }
 886         }
 887
 888         vm_page_assert_xbusied(fs.m);
 889
 890         /*
 891          * Page must be completely valid or it is not fit to
 892          * map into user space.  vm_pager_get_pages() ensures this.
 893          */
 894         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 895             ("vm_fault: page %p partially invalid", fs.m));
 896         VM_OBJECT_WUNLOCK(fs.object);
 897
 898         /*
 899          * Put this page into the physical map.  We had to do the unlock above
 900          * because pmap_enter() may sleep.  We don't put the page
 901          * back on the active queue until later so that the pageout daemon
 902          * won't find it (yet).
 903          */
 904         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
 905             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
 906         if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
 907                 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
 908         VM_OBJECT_WLOCK(fs.object);
 909         vm_page_lock(fs.m);
 910
 911         /*
 912          * If the page is not wired down, then put it where the pageout daemon
 913          * can find it.
 914          */
 915         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 916                 if (wired)
 917                         vm_page_wire(fs.m);
 918                 else
 919                         vm_page_unwire(fs.m, 1);
 920         } else
 921                 vm_page_activate(fs.m);
 922         if (m_hold != NULL) {
 923                 *m_hold = fs.m;
 924                 vm_page_hold(fs.m);
 925         }
 926         vm_page_unlock(fs.m);
 927         vm_page_xunbusy(fs.m);
 928
 929         /*
 930          * Unlock everything, and return
 931          */
 932         unlock_and_deallocate(&fs);
 933         if (hardfault) {
 934                 PCPU_INC(cnt.v_io_faults);
 935                 curthread->td_ru.ru_majflt++;
 936         } else
 937                 curthread->td_ru.ru_minflt++;
 938
 939         return (KERN_SUCCESS);
 940 }
 941
 942 /*
 943  * Speed up the reclamation of up to "distance" pages that precede the
 944  * faulting pindex within the first object of the shadow chain.
 945  */
 946 static void
 947 vm_fault_cache_behind(const struct faultstate *fs, int distance)
 948 {
 949         vm_object_t first_object, object;
 950         vm_page_t m, m_prev;
 951         vm_pindex_t pindex;
 952
 953         object = fs->object;
 954         VM_OBJECT_ASSERT_WLOCKED(object);
 955         first_object = fs->first_object;
 956         if (first_object != object) {
 957                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
 958                         VM_OBJECT_WUNLOCK(object);
 959                         VM_OBJECT_WLOCK(first_object);
 960                         VM_OBJECT_WLOCK(object);
 961                 }
 962         }
 963         /* Neither fictitious nor unmanaged pages can be cached. */
 964         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
 965                 if (fs->first_pindex < distance)
 966                         pindex = 0;
 967                 else
 968                         pindex = fs->first_pindex - distance;
 969                 if (pindex < OFF_TO_IDX(fs->entry->offset))
 970                         pindex = OFF_TO_IDX(fs->entry->offset);
 971                 m = first_object != object ? fs->first_m : fs->m;
 972                 vm_page_assert_xbusied(m);
 973                 m_prev = vm_page_prev(m);
 974                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
 975                     m->valid == VM_PAGE_BITS_ALL) {
 976                         m_prev = vm_page_prev(m);
 977                         if (vm_page_busied(m))
 978                                 continue;
 979                         vm_page_lock(m);
 980                         if (m->hold_count == 0 && m->wire_count == 0) {
 981                                 pmap_remove_all(m);
 982                                 vm_page_aflag_clear(m, PGA_REFERENCED);
 983                                 if (m->dirty != 0)
 984                                         vm_page_deactivate(m);
 985                                 else
 986                                         vm_page_cache(m);
 987                         }
 988                         vm_page_unlock(m);
 989                 }
 990         }
 991         if (first_object != object)
 992                 VM_OBJECT_WUNLOCK(first_object);
 993 }
 994
 995 /*
 996  * vm_fault_prefault provides a quick way of clustering
 997  * pagefaults into a processes address space.  It is a "cousin"
 998  * of vm_map_pmap_enter, except it runs at page fault time instead
 999  * of mmap time.
1000  */
1001 static void
1002 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1003     int faultcount, int reqpage)
1004 {
1005         pmap_t pmap;
1006         vm_map_entry_t entry;
1007         vm_object_t backing_object, lobject;
1008         vm_offset_t addr, starta;
1009         vm_pindex_t pindex;
1010         vm_page_t m;
1011         int backward, forward, i;
1012
1013         pmap = fs->map->pmap;
1014         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1015                 return;
1016
1017         if (faultcount > 0) {
1018                 backward = reqpage;
1019                 forward = faultcount - reqpage - 1;
1020         } else {
1021                 backward = PFBAK;
1022                 forward = PFFOR;
1023         }
1024         entry = fs->entry;
1025
1026         starta = addra - backward * PAGE_SIZE;
1027         if (starta < entry->start) {
1028                 starta = entry->start;
1029         } else if (starta > addra) {
1030                 starta = 0;
1031         }
1032
1033         /*
1034          * Generate the sequence of virtual addresses that are candidates for
1035          * prefaulting in an outward spiral from the faulting virtual address,
1036          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1037          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1038          * If the candidate address doesn't have a backing physical page, then
1039          * the loop immediately terminates.
1040          */
1041         for (i = 0; i < 2 * imax(backward, forward); i++) {
1042                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1043                     PAGE_SIZE);
1044                 if (addr > addra + forward * PAGE_SIZE)
1045                         addr = 0;
1046
1047                 if (addr < starta || addr >= entry->end)
1048                         continue;
1049
1050                 if (!pmap_is_prefaultable(pmap, addr))
1051                         continue;
1052
1053                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1054                 lobject = entry->object.vm_object;
1055                 VM_OBJECT_RLOCK(lobject);
1056                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1057                     lobject->type == OBJT_DEFAULT &&
1058                     (backing_object = lobject->backing_object) != NULL) {
1059                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1060                             0, ("vm_fault_prefault: unaligned object offset"));
1061                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1062                         VM_OBJECT_RLOCK(backing_object);
1063                         VM_OBJECT_RUNLOCK(lobject);
1064                         lobject = backing_object;
1065                 }
1066                 if (m == NULL) {
1067                         VM_OBJECT_RUNLOCK(lobject);
1068                         break;
1069                 }
1070                 if (m->valid == VM_PAGE_BITS_ALL &&
1071                     (m->flags & PG_FICTITIOUS) == 0)
1072                         pmap_enter_quick(pmap, addr, m, entry->protection);
1073                 VM_OBJECT_RUNLOCK(lobject);
1074         }
1075 }
1076
1077 /*
1078  * Hold each of the physical pages that are mapped by the specified range of
1079  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1080  * and allow the specified types of access, "prot".  If all of the implied
1081  * pages are successfully held, then the number of held pages is returned
1082  * together with pointers to those pages in the array "ma".  However, if any
1083  * of the pages cannot be held, -1 is returned.
1084  */
1085 int
1086 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1087     vm_prot_t prot, vm_page_t *ma, int max_count)
1088 {
1089         vm_offset_t end, va;
1090         vm_page_t *mp;
1091         int count;
1092         boolean_t pmap_failed;
1093
1094         if (len == 0)
1095                 return (0);
1096         end = round_page(addr + len);
1097         addr = trunc_page(addr);
1098
1099         /*
1100          * Check for illegal addresses.
1101          */
1102         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1103                 return (-1);
1104
1105         if (atop(end - addr) > max_count)
1106                 panic("vm_fault_quick_hold_pages: count > max_count");
1107         count = atop(end - addr);
1108
1109         /*
1110          * Most likely, the physical pages are resident in the pmap, so it is
1111          * faster to try pmap_extract_and_hold() first.
1112          */
1113         pmap_failed = FALSE;
1114         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1115                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1116                 if (*mp == NULL)
1117                         pmap_failed = TRUE;
1118                 else if ((prot & VM_PROT_WRITE) != 0 &&
1119                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1120                         /*
1121                          * Explicitly dirty the physical page.  Otherwise, the
1122                          * caller's changes may go unnoticed because they are
1123                          * performed through an unmanaged mapping or by a DMA
1124                          * operation.
1125                          *
1126                          * The object lock is not held here.
1127                          * See vm_page_clear_dirty_mask().
1128                          */
1129                         vm_page_dirty(*mp);
1130                 }
1131         }
1132         if (pmap_failed) {
1133                 /*
1134                  * One or more pages could not be held by the pmap.  Either no
1135                  * page was mapped at the specified virtual address or that
1136                  * mapping had insufficient permissions.  Attempt to fault in
1137                  * and hold these pages.
1138                  */
1139                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1140                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1141                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1142                                 goto error;
1143         }
1144         return (count);
1145 error:
1146         for (mp = ma; mp < ma + count; mp++)
1147                 if (*mp != NULL) {
1148                         vm_page_lock(*mp);
1149                         vm_page_unhold(*mp);
1150                         vm_page_unlock(*mp);
1151                 }
1152         return (-1);
1153 }
1154
1155 /*
1156  *      vm_fault_wire:
1157  *
1158  *      Wire down a range of virtual addresses in a map.
1159  */
1160 int
1161 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1162     boolean_t fictitious)
1163 {
1164         vm_offset_t va;
1165         int rv;
1166
1167         /*
1168          * We simulate a fault to get the page and enter it in the physical
1169          * map.  For user wiring, we only ask for read access on currently
1170          * read-only sections.
1171          */
1172         for (va = start; va < end; va += PAGE_SIZE) {
1173                 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1174                 if (rv) {
1175                         if (va != start)
1176                                 vm_fault_unwire(map, start, va, fictitious);
1177                         return (rv);
1178                 }
1179         }
1180         return (KERN_SUCCESS);
1181 }
1182
1183 /*
1184  *      vm_fault_unwire:
1185  *
1186  *      Unwire a range of virtual addresses in a map.
1187  */
1188 void
1189 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1190     boolean_t fictitious)
1191 {
1192         vm_paddr_t pa;
1193         vm_offset_t va;
1194         vm_page_t m;
1195         pmap_t pmap;
1196
1197         pmap = vm_map_pmap(map);
1198
1199         /*
1200          * Since the pages are wired down, we must be able to get their
1201          * mappings from the physical map system.
1202          */
1203         for (va = start; va < end; va += PAGE_SIZE) {
1204                 pa = pmap_extract(pmap, va);
1205                 if (pa != 0) {
1206                         pmap_change_wiring(pmap, va, FALSE);
1207                         if (!fictitious) {
1208                                 m = PHYS_TO_VM_PAGE(pa);
1209                                 vm_page_lock(m);
1210                                 vm_page_unwire(m, TRUE);
1211                                 vm_page_unlock(m);
1212                         }
1213                 }
1214         }
1215 }
1216
1217 /*
1218  *      Routine:
1219  *              vm_fault_copy_entry
1220  *      Function:
1221  *              Create new shadow object backing dst_entry with private copy of
1222  *              all underlying pages. When src_entry is equal to dst_entry,
1223  *              function implements COW for wired-down map entry. Otherwise,
1224  *              it forks wired entry into dst_map.
1225  *
1226  *      In/out conditions:
1227  *              The source and destination maps must be locked for write.
1228  *              The source map entry must be wired down (or be a sharing map
1229  *              entry corresponding to a main map entry that is wired down).
1230  */
1231 void
1232 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1233     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1234     vm_ooffset_t *fork_charge)
1235 {
1236         vm_object_t backing_object, dst_object, object, src_object;
1237         vm_pindex_t dst_pindex, pindex, src_pindex;
1238         vm_prot_t access, prot;
1239         vm_offset_t vaddr;
1240         vm_page_t dst_m;
1241         vm_page_t src_m;
1242         boolean_t upgrade;
1243
1244 #ifdef  lint
1245         src_map++;
1246 #endif  /* lint */
1247
1248         upgrade = src_entry == dst_entry;
1249         access = prot = dst_entry->protection;
1250
1251         src_object = src_entry->object.vm_object;
1252         src_pindex = OFF_TO_IDX(src_entry->offset);
1253
1254         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1255                 dst_object = src_object;
1256                 vm_object_reference(dst_object);
1257         } else {
1258                 /*
1259                  * Create the top-level object for the destination entry. (Doesn't
1260                  * actually shadow anything - we copy the pages directly.)
1261                  */
1262                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1263                     OFF_TO_IDX(dst_entry->end - dst_entry->start));
1264 #if VM_NRESERVLEVEL > 0
1265                 dst_object->flags |= OBJ_COLORED;
1266                 dst_object->pg_color = atop(dst_entry->start);
1267 #endif
1268         }
1269
1270         VM_OBJECT_WLOCK(dst_object);
1271         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1272             ("vm_fault_copy_entry: vm_object not NULL"));
1273         if (src_object != dst_object) {
1274                 dst_entry->object.vm_object = dst_object;
1275                 dst_entry->offset = 0;
1276                 dst_object->charge = dst_entry->end - dst_entry->start;
1277         }
1278         if (fork_charge != NULL) {
1279                 KASSERT(dst_entry->cred == NULL,
1280                     ("vm_fault_copy_entry: leaked swp charge"));
1281                 dst_object->cred = curthread->td_ucred;
1282                 crhold(dst_object->cred);
1283                 *fork_charge += dst_object->charge;
1284         } else if (dst_object->cred == NULL) {
1285                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1286                     dst_entry));
1287                 dst_object->cred = dst_entry->cred;
1288                 dst_entry->cred = NULL;
1289         }
1290
1291         /*
1292          * If not an upgrade, then enter the mappings in the pmap as
1293          * read and/or execute accesses.  Otherwise, enter them as
1294          * write accesses.
1295          *
1296          * A writeable large page mapping is only created if all of
1297          * the constituent small page mappings are modified. Marking
1298          * PTEs as modified on inception allows promotion to happen
1299          * without taking potentially large number of soft faults.
1300          */
1301         if (!upgrade)
1302                 access &= ~VM_PROT_WRITE;
1303
1304         /*
1305          * Loop through all of the virtual pages within the entry's
1306          * range, copying each page from the source object to the
1307          * destination object.  Since the source is wired, those pages
1308          * must exist.  In contrast, the destination is pageable.
1309          * Since the destination object does share any backing storage
1310          * with the source object, all of its pages must be dirtied,
1311          * regardless of whether they can be written.
1312          */
1313         for (vaddr = dst_entry->start, dst_pindex = 0;
1314             vaddr < dst_entry->end;
1315             vaddr += PAGE_SIZE, dst_pindex++) {
1316 again:
1317                 /*
1318                  * Find the page in the source object, and copy it in.
1319                  * Because the source is wired down, the page will be
1320                  * in memory.
1321                  */
1322                 if (src_object != dst_object)
1323                         VM_OBJECT_RLOCK(src_object);
1324                 object = src_object;
1325                 pindex = src_pindex + dst_pindex;
1326                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1327                     (backing_object = object->backing_object) != NULL) {
1328                         /*
1329                          * Unless the source mapping is read-only or
1330                          * it is presently being upgraded from
1331                          * read-only, the first object in the shadow
1332                          * chain should provide all of the pages.  In
1333                          * other words, this loop body should never be
1334                          * executed when the source mapping is already
1335                          * read/write.
1336                          */
1337                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1338                             upgrade,
1339                             ("vm_fault_copy_entry: main object missing page"));
1340
1341                         VM_OBJECT_RLOCK(backing_object);
1342                         pindex += OFF_TO_IDX(object->backing_object_offset);
1343                         if (object != dst_object)
1344                                 VM_OBJECT_RUNLOCK(object);
1345                         object = backing_object;
1346                 }
1347                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1348
1349                 if (object != dst_object) {
1350                         /*
1351                          * Allocate a page in the destination object.
1352                          */
1353                         dst_m = vm_page_alloc(dst_object, (src_object ==
1354                             dst_object ? src_pindex : 0) + dst_pindex,
1355                             VM_ALLOC_NORMAL);
1356                         if (dst_m == NULL) {
1357                                 VM_OBJECT_WUNLOCK(dst_object);
1358                                 VM_OBJECT_RUNLOCK(object);
1359                                 VM_WAIT;
1360                                 VM_OBJECT_WLOCK(dst_object);
1361                                 goto again;
1362                         }
1363                         pmap_copy_page(src_m, dst_m);
1364                         VM_OBJECT_RUNLOCK(object);
1365                         dst_m->valid = VM_PAGE_BITS_ALL;
1366                         dst_m->dirty = VM_PAGE_BITS_ALL;
1367                 } else {
1368                         dst_m = src_m;
1369                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1370                                 goto again;
1371                         vm_page_xbusy(dst_m);
1372                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1373                             ("invalid dst page %p", dst_m));
1374                 }
1375                 VM_OBJECT_WUNLOCK(dst_object);
1376
1377                 /*
1378                  * Enter it in the pmap. If a wired, copy-on-write
1379                  * mapping is being replaced by a write-enabled
1380                  * mapping, then wire that new mapping.
1381                  */
1382                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1383                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1384
1385                 /*
1386                  * Mark it no longer busy, and put it on the active list.
1387                  */
1388                 VM_OBJECT_WLOCK(dst_object);
1389
1390                 if (upgrade) {
1391                         if (src_m != dst_m) {
1392                                 vm_page_lock(src_m);
1393                                 vm_page_unwire(src_m, 0);
1394                                 vm_page_unlock(src_m);
1395                                 vm_page_lock(dst_m);
1396                                 vm_page_wire(dst_m);
1397                                 vm_page_unlock(dst_m);
1398                         } else {
1399                                 KASSERT(dst_m->wire_count > 0,
1400                                     ("dst_m %p is not wired", dst_m));
1401                         }
1402                 } else {
1403                         vm_page_lock(dst_m);
1404                         vm_page_activate(dst_m);
1405                         vm_page_unlock(dst_m);
1406                 }
1407                 vm_page_xunbusy(dst_m);
1408         }
1409         VM_OBJECT_WUNLOCK(dst_object);
1410         if (upgrade) {
1411                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1412                 vm_object_deallocate(src_object);
1413         }
1414 }
1415
1416
1417 /*
1418  * This routine checks around the requested page for other pages that
1419  * might be able to be faulted in.  This routine brackets the viable
1420  * pages for the pages to be paged in.
1421  *
1422  * Inputs:
1423  *      m, rbehind, rahead
1424  *
1425  * Outputs:
1426  *  marray (array of vm_page_t), reqpage (index of requested page)
1427  *
1428  * Return value:
1429  *  number of pages in marray
1430  */
1431 static int
1432 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1433         vm_page_t m;
1434         int rbehind;
1435         int rahead;
1436         vm_page_t *marray;
1437         int *reqpage;
1438 {
1439         int i,j;
1440         vm_object_t object;
1441         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1442         vm_page_t rtm;
1443         int cbehind, cahead;
1444
1445         VM_OBJECT_ASSERT_WLOCKED(m->object);
1446
1447         object = m->object;
1448         pindex = m->pindex;
1449         cbehind = cahead = 0;
1450
1451         /*
1452          * if the requested page is not available, then give up now
1453          */
1454         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1455                 return 0;
1456         }
1457
1458         if ((cbehind == 0) && (cahead == 0)) {
1459                 *reqpage = 0;
1460                 marray[0] = m;
1461                 return 1;
1462         }
1463
1464         if (rahead > cahead) {
1465                 rahead = cahead;
1466         }
1467
1468         if (rbehind > cbehind) {
1469                 rbehind = cbehind;
1470         }
1471
1472         /*
1473          * scan backward for the read behind pages -- in memory
1474          */
1475         if (pindex > 0) {
1476                 if (rbehind > pindex) {
1477                         rbehind = pindex;
1478                         startpindex = 0;
1479                 } else {
1480                         startpindex = pindex - rbehind;
1481                 }
1482
1483                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1484                     rtm->pindex >= startpindex)
1485                         startpindex = rtm->pindex + 1;
1486
1487                 /* tpindex is unsigned; beware of numeric underflow. */
1488                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1489                     tpindex < pindex; i++, tpindex--) {
1490
1491                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1492                             VM_ALLOC_IFNOTCACHED);
1493                         if (rtm == NULL) {
1494                                 /*
1495                                  * Shift the allocated pages to the
1496                                  * beginning of the array.
1497                                  */
1498                                 for (j = 0; j < i; j++) {
1499                                         marray[j] = marray[j + tpindex + 1 -
1500                                             startpindex];
1501                                 }
1502                                 break;
1503                         }
1504
1505                         marray[tpindex - startpindex] = rtm;
1506                 }
1507         } else {
1508                 startpindex = 0;
1509                 i = 0;
1510         }
1511
1512         marray[i] = m;
1513         /* page offset of the required page */
1514         *reqpage = i;
1515
1516         tpindex = pindex + 1;
1517         i++;
1518
1519         /*
1520          * scan forward for the read ahead pages
1521          */
1522         endpindex = tpindex + rahead;
1523         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1524                 endpindex = rtm->pindex;
1525         if (endpindex > object->size)
1526                 endpindex = object->size;
1527
1528         for (; tpindex < endpindex; i++, tpindex++) {
1529
1530                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1531                     VM_ALLOC_IFNOTCACHED);
1532                 if (rtm == NULL) {
1533                         break;
1534                 }
1535
1536                 marray[i] = rtm;
1537         }
1538
1539         /* return number of pages */
1540         return i;
1541 }
1542
1543 /*
1544  * Block entry into the machine-independent layer's page fault handler by
1545  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1546  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1547  * spurious page faults.
1548  */
1549 int
1550 vm_fault_disable_pagefaults(void)
1551 {
1552
1553         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1554 }
1555
1556 void
1557 vm_fault_enable_pagefaults(int save)
1558 {
1559
1560         curthread_pflags_restore(save);
1561 }