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/mman.h>
  85 #include <sys/proc.h>
  86 #include <sys/racct.h>
  87 #include <sys/resourcevar.h>
  88 #include <sys/rwlock.h>
  89 #include <sys/sysctl.h>
  90 #include <sys/vmmeter.h>
  91 #include <sys/vnode.h>
  92 #ifdef KTRACE
  93 #include <sys/ktrace.h>
  94 #endif
  95
  96 #include <vm/vm.h>
  97 #include <vm/vm_param.h>
  98 #include <vm/pmap.h>
  99 #include <vm/vm_map.h>
 100 #include <vm/vm_object.h>
 101 #include <vm/vm_page.h>
 102 #include <vm/vm_pageout.h>
 103 #include <vm/vm_kern.h>
 104 #include <vm/vm_pager.h>
 105 #include <vm/vm_extern.h>
 106 #include <vm/vm_reserv.h>
 107
 108 #define PFBAK 4
 109 #define PFFOR 4
 110
 111 #define VM_FAULT_READ_DEFAULT   (1 + VM_FAULT_READ_AHEAD_INIT)
 112 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 113
 114 #define VM_FAULT_DONTNEED_MIN   1048576
 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         bool lookup_still_valid;
 126         struct vnode *vp;
 127 };
 128
 129 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
 130             int ahead);
 131 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 132             int backward, int forward);
 133
 134 static inline void
 135 release_page(struct faultstate *fs)
 136 {
 137
 138         vm_page_xunbusy(fs->m);
 139         vm_page_lock(fs->m);
 140         vm_page_deactivate(fs->m);
 141         vm_page_unlock(fs->m);
 142         fs->m = NULL;
 143 }
 144
 145 static inline void
 146 unlock_map(struct faultstate *fs)
 147 {
 148
 149         if (fs->lookup_still_valid) {
 150                 vm_map_lookup_done(fs->map, fs->entry);
 151                 fs->lookup_still_valid = false;
 152         }
 153 }
 154
 155 static void
 156 unlock_vp(struct faultstate *fs)
 157 {
 158
 159         if (fs->vp != NULL) {
 160                 vput(fs->vp);
 161                 fs->vp = NULL;
 162         }
 163 }
 164
 165 static void
 166 unlock_and_deallocate(struct faultstate *fs)
 167 {
 168
 169         vm_object_pip_wakeup(fs->object);
 170         VM_OBJECT_WUNLOCK(fs->object);
 171         if (fs->object != fs->first_object) {
 172                 VM_OBJECT_WLOCK(fs->first_object);
 173                 vm_page_lock(fs->first_m);
 174                 vm_page_free(fs->first_m);
 175                 vm_page_unlock(fs->first_m);
 176                 vm_object_pip_wakeup(fs->first_object);
 177                 VM_OBJECT_WUNLOCK(fs->first_object);
 178                 fs->first_m = NULL;
 179         }
 180         vm_object_deallocate(fs->first_object);
 181         unlock_map(fs);
 182         unlock_vp(fs);
 183 }
 184
 185 static void
 186 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 187     vm_prot_t fault_type, int fault_flags, bool set_wd)
 188 {
 189         bool need_dirty;
 190
 191         if (((prot & VM_PROT_WRITE) == 0 &&
 192             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 193             (m->oflags & VPO_UNMANAGED) != 0)
 194                 return;
 195
 196         VM_OBJECT_ASSERT_LOCKED(m->object);
 197
 198         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 199             (fault_flags & VM_FAULT_WIRE) == 0) ||
 200             (fault_flags & VM_FAULT_DIRTY) != 0;
 201
 202         if (set_wd)
 203                 vm_object_set_writeable_dirty(m->object);
 204         else
 205                 /*
 206                  * If two callers of vm_fault_dirty() with set_wd ==
 207                  * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
 208                  * flag set, other with flag clear, race, it is
 209                  * possible for the no-NOSYNC thread to see m->dirty
 210                  * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
 211                  * around manipulation of VPO_NOSYNC and
 212                  * vm_page_dirty() call, to avoid the race and keep
 213                  * m->oflags consistent.
 214                  */
 215                 vm_page_lock(m);
 216
 217         /*
 218          * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 219          * if the page is already dirty to prevent data written with
 220          * the expectation of being synced from not being synced.
 221          * Likewise if this entry does not request NOSYNC then make
 222          * sure the page isn't marked NOSYNC.  Applications sharing
 223          * data should use the same flags to avoid ping ponging.
 224          */
 225         if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
 226                 if (m->dirty == 0) {
 227                         m->oflags |= VPO_NOSYNC;
 228                 }
 229         } else {
 230                 m->oflags &= ~VPO_NOSYNC;
 231         }
 232
 233         /*
 234          * If the fault is a write, we know that this page is being
 235          * written NOW so dirty it explicitly to save on
 236          * pmap_is_modified() calls later.
 237          *
 238          * Also tell the backing pager, if any, that it should remove
 239          * any swap backing since the page is now dirty.
 240          */
 241         if (need_dirty)
 242                 vm_page_dirty(m);
 243         if (!set_wd)
 244                 vm_page_unlock(m);
 245         if (need_dirty)
 246                 vm_pager_page_unswapped(m);
 247 }
 248
 249 /*
 250  *      vm_fault:
 251  *
 252  *      Handle a page fault occurring at the given address,
 253  *      requiring the given permissions, in the map specified.
 254  *      If successful, the page is inserted into the
 255  *      associated physical map.
 256  *
 257  *      NOTE: the given address should be truncated to the
 258  *      proper page address.
 259  *
 260  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 261  *      a standard error specifying why the fault is fatal is returned.
 262  *
 263  *      The map in question must be referenced, and remains so.
 264  *      Caller may hold no locks.
 265  */
 266 int
 267 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 268     int fault_flags)
 269 {
 270         struct thread *td;
 271         int result;
 272
 273         td = curthread;
 274         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 275                 return (KERN_PROTECTION_FAILURE);
 276 #ifdef KTRACE
 277         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 278                 ktrfault(vaddr, fault_type);
 279 #endif
 280         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 281             NULL);
 282 #ifdef KTRACE
 283         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 284                 ktrfaultend(result);
 285 #endif
 286         return (result);
 287 }
 288
 289 int
 290 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 291     int fault_flags, vm_page_t *m_hold)
 292 {
 293         struct faultstate fs;
 294         struct vnode *vp;
 295         vm_object_t next_object, retry_object;
 296         vm_offset_t e_end, e_start;
 297         vm_page_t m;
 298         vm_pindex_t retry_pindex;
 299         vm_prot_t prot, retry_prot;
 300         int ahead, alloc_req, behind, cluster_offset, error, era, faultcount;
 301         int locked, map_generation, nera, result, rv;
 302         u_char behavior;
 303         boolean_t wired;        /* Passed by reference. */
 304         bool dead, growstack, hardfault, is_first_object_locked;
 305
 306         PCPU_INC(cnt.v_vm_faults);
 307         fs.vp = NULL;
 308         faultcount = 0;
 309         nera = -1;
 310         growstack = true;
 311         hardfault = false;
 312
 313 RetryFault:;
 314
 315         /*
 316          * Find the backing store object and offset into it to begin the
 317          * search.
 318          */
 319         fs.map = map;
 320         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 321             &fs.first_object, &fs.first_pindex, &prot, &wired);
 322         if (result != KERN_SUCCESS) {
 323                 if (growstack && result == KERN_INVALID_ADDRESS &&
 324                     map != kernel_map) {
 325                         result = vm_map_growstack(curproc, vaddr);
 326                         if (result != KERN_SUCCESS)
 327                                 return (KERN_FAILURE);
 328                         growstack = false;
 329                         goto RetryFault;
 330                 }
 331                 unlock_vp(&fs);
 332                 return (result);
 333         }
 334
 335         map_generation = fs.map->timestamp;
 336
 337         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 338                 panic("vm_fault: fault on nofault entry, addr: %lx",
 339                     (u_long)vaddr);
 340         }
 341
 342         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 343             fs.entry->wiring_thread != curthread) {
 344                 vm_map_unlock_read(fs.map);
 345                 vm_map_lock(fs.map);
 346                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 347                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 348                         unlock_vp(&fs);
 349                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 350                         vm_map_unlock_and_wait(fs.map, 0);
 351                 } else
 352                         vm_map_unlock(fs.map);
 353                 goto RetryFault;
 354         }
 355
 356         if (wired)
 357                 fault_type = prot | (fault_type & VM_PROT_COPY);
 358         else
 359                 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
 360                     ("!wired && VM_FAULT_WIRE"));
 361
 362         /*
 363          * Try to avoid lock contention on the top-level object through
 364          * special-case handling of some types of page faults, specifically,
 365          * those that are both (1) mapping an existing page from the top-
 366          * level object and (2) not having to mark that object as containing
 367          * dirty pages.  Under these conditions, a read lock on the top-level
 368          * object suffices, allowing multiple page faults of a similar type to
 369          * run in parallel on the same top-level object.
 370          */
 371         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 372             (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
 373             /* avoid calling vm_object_set_writeable_dirty() */
 374             ((prot & VM_PROT_WRITE) == 0 ||
 375             (fs.first_object->type != OBJT_VNODE &&
 376             (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 377             (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
 378                 VM_OBJECT_RLOCK(fs.first_object);
 379                 if ((prot & VM_PROT_WRITE) != 0 &&
 380                     (fs.first_object->type == OBJT_VNODE ||
 381                     (fs.first_object->flags & OBJ_TMPFS_NODE) != 0) &&
 382                     (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
 383                         goto fast_failed;
 384                 m = vm_page_lookup(fs.first_object, fs.first_pindex);
 385                 /* A busy page can be mapped for read|execute access. */
 386                 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 387                     vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
 388                         goto fast_failed;
 389                 result = pmap_enter(fs.map->pmap, vaddr, m, prot,
 390                    fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
 391                    0), 0);
 392                 if (result != KERN_SUCCESS)
 393                         goto fast_failed;
 394                 if (m_hold != NULL) {
 395                         *m_hold = m;
 396                         vm_page_lock(m);
 397                         vm_page_hold(m);
 398                         vm_page_unlock(m);
 399                 }
 400                 vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags,
 401                     false);
 402                 VM_OBJECT_RUNLOCK(fs.first_object);
 403                 if (!wired)
 404                         vm_fault_prefault(&fs, vaddr, PFBAK, PFFOR);
 405                 vm_map_lookup_done(fs.map, fs.entry);
 406                 curthread->td_ru.ru_minflt++;
 407                 return (KERN_SUCCESS);
 408 fast_failed:
 409                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 410                         VM_OBJECT_RUNLOCK(fs.first_object);
 411                         VM_OBJECT_WLOCK(fs.first_object);
 412                 }
 413         } else {
 414                 VM_OBJECT_WLOCK(fs.first_object);
 415         }
 416
 417         /*
 418          * Make a reference to this object to prevent its disposal while we
 419          * are messing with it.  Once we have the reference, the map is free
 420          * to be diddled.  Since objects reference their shadows (and copies),
 421          * they will stay around as well.
 422          *
 423          * Bump the paging-in-progress count to prevent size changes (e.g.
 424          * truncation operations) during I/O.
 425          */
 426         vm_object_reference_locked(fs.first_object);
 427         vm_object_pip_add(fs.first_object, 1);
 428
 429         fs.lookup_still_valid = true;
 430
 431         fs.first_m = NULL;
 432
 433         /*
 434          * Search for the page at object/offset.
 435          */
 436         fs.object = fs.first_object;
 437         fs.pindex = fs.first_pindex;
 438         while (TRUE) {
 439                 /*
 440                  * If the object is marked for imminent termination,
 441                  * we retry here, since the collapse pass has raced
 442                  * with us.  Otherwise, if we see terminally dead
 443                  * object, return fail.
 444                  */
 445                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 446                         dead = fs.object->type == OBJT_DEAD;
 447                         unlock_and_deallocate(&fs);
 448                         if (dead)
 449                                 return (KERN_PROTECTION_FAILURE);
 450                         pause("vmf_de", 1);
 451                         goto RetryFault;
 452                 }
 453
 454                 /*
 455                  * See if page is resident
 456                  */
 457                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 458                 if (fs.m != NULL) {
 459                         /*
 460                          * Wait/Retry if the page is busy.  We have to do this
 461                          * if the page is either exclusive or shared busy
 462                          * because the vm_pager may be using read busy for
 463                          * pageouts (and even pageins if it is the vnode
 464                          * pager), and we could end up trying to pagein and
 465                          * pageout the same page simultaneously.
 466                          *
 467                          * We can theoretically allow the busy case on a read
 468                          * fault if the page is marked valid, but since such
 469                          * pages are typically already pmap'd, putting that
 470                          * special case in might be more effort then it is
 471                          * worth.  We cannot under any circumstances mess
 472                          * around with a shared busied page except, perhaps,
 473                          * to pmap it.
 474                          */
 475                         if (vm_page_busied(fs.m)) {
 476                                 /*
 477                                  * Reference the page before unlocking and
 478                                  * sleeping so that the page daemon is less
 479                                  * likely to reclaim it.
 480                                  */
 481                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 482                                 if (fs.object != fs.first_object) {
 483                                         if (!VM_OBJECT_TRYWLOCK(
 484                                             fs.first_object)) {
 485                                                 VM_OBJECT_WUNLOCK(fs.object);
 486                                                 VM_OBJECT_WLOCK(fs.first_object);
 487                                                 VM_OBJECT_WLOCK(fs.object);
 488                                         }
 489                                         vm_page_lock(fs.first_m);
 490                                         vm_page_free(fs.first_m);
 491                                         vm_page_unlock(fs.first_m);
 492                                         vm_object_pip_wakeup(fs.first_object);
 493                                         VM_OBJECT_WUNLOCK(fs.first_object);
 494                                         fs.first_m = NULL;
 495                                 }
 496                                 unlock_map(&fs);
 497                                 if (fs.m == vm_page_lookup(fs.object,
 498                                     fs.pindex)) {
 499                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 500                                 }
 501                                 vm_object_pip_wakeup(fs.object);
 502                                 VM_OBJECT_WUNLOCK(fs.object);
 503                                 PCPU_INC(cnt.v_intrans);
 504                                 vm_object_deallocate(fs.first_object);
 505                                 goto RetryFault;
 506                         }
 507                         vm_page_lock(fs.m);
 508                         vm_page_remque(fs.m);
 509                         vm_page_unlock(fs.m);
 510
 511                         /*
 512                          * Mark page busy for other processes, and the
 513                          * pagedaemon.  If it still isn't completely valid
 514                          * (readable), jump to readrest, else break-out ( we
 515                          * found the page ).
 516                          */
 517                         vm_page_xbusy(fs.m);
 518                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 519                                 goto readrest;
 520                         break;
 521                 }
 522                 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
 523
 524                 /*
 525                  * Page is not resident.  If the pager might contain the page
 526                  * or this is the beginning of the search, allocate a new
 527                  * page.  (Default objects are zero-fill, so there is no real
 528                  * pager for them.)
 529                  */
 530                 if (fs.object->type != OBJT_DEFAULT ||
 531                     fs.object == fs.first_object) {
 532                         if (fs.pindex >= fs.object->size) {
 533                                 unlock_and_deallocate(&fs);
 534                                 return (KERN_PROTECTION_FAILURE);
 535                         }
 536
 537                         /*
 538                          * Allocate a new page for this object/offset pair.
 539                          *
 540                          * Unlocked read of the p_flag is harmless. At
 541                          * worst, the P_KILLED might be not observed
 542                          * there, and allocation can fail, causing
 543                          * restart and new reading of the p_flag.
 544                          */
 545                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 546 #if VM_NRESERVLEVEL > 0
 547                                 vm_object_color(fs.object, atop(vaddr) -
 548                                     fs.pindex);
 549 #endif
 550                                 alloc_req = P_KILLED(curproc) ?
 551                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 552                                 if (fs.object->type != OBJT_VNODE &&
 553                                     fs.object->backing_object == NULL)
 554                                         alloc_req |= VM_ALLOC_ZERO;
 555                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 556                                     alloc_req);
 557                         }
 558                         if (fs.m == NULL) {
 559                                 unlock_and_deallocate(&fs);
 560                                 VM_WAITPFAULT;
 561                                 goto RetryFault;
 562                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 563                                 break;
 564                 }
 565
 566 readrest:
 567                 /*
 568                  * At this point, we have either allocated a new page or found
 569                  * an existing page that is only partially valid.
 570                  *
 571                  * We hold a reference on the current object and the page is
 572                  * exclusive busied.
 573                  */
 574
 575                 /*
 576                  * If the pager for the current object might have the page,
 577                  * then determine the number of additional pages to read and
 578                  * potentially reprioritize previously read pages for earlier
 579                  * reclamation.  These operations should only be performed
 580                  * once per page fault.  Even if the current pager doesn't
 581                  * have the page, the number of additional pages to read will
 582                  * apply to subsequent objects in the shadow chain.
 583                  */
 584                 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
 585                     !P_KILLED(curproc)) {
 586                         KASSERT(fs.lookup_still_valid, ("map unlocked"));
 587                         era = fs.entry->read_ahead;
 588                         behavior = vm_map_entry_behavior(fs.entry);
 589                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 590                                 nera = 0;
 591                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 592                                 nera = VM_FAULT_READ_AHEAD_MAX;
 593                                 if (vaddr == fs.entry->next_read)
 594                                         vm_fault_dontneed(&fs, vaddr, nera);
 595                         } else if (vaddr == fs.entry->next_read) {
 596                                 /*
 597                                  * This is a sequential fault.  Arithmetically
 598                                  * increase the requested number of pages in
 599                                  * the read-ahead window.  The requested
 600                                  * number of pages is "# of sequential faults
 601                                  * x (read ahead min + 1) + read ahead min"
 602                                  */
 603                                 nera = VM_FAULT_READ_AHEAD_MIN;
 604                                 if (era > 0) {
 605                                         nera += era + 1;
 606                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 607                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 608                                 }
 609                                 if (era == VM_FAULT_READ_AHEAD_MAX)
 610                                         vm_fault_dontneed(&fs, vaddr, nera);
 611                         } else {
 612                                 /*
 613                                  * This is a non-sequential fault.
 614                                  */
 615                                 nera = 0;
 616                         }
 617                         if (era != nera) {
 618                                 /*
 619                                  * A read lock on the map suffices to update
 620                                  * the read ahead count safely.
 621                                  */
 622                                 fs.entry->read_ahead = nera;
 623                         }
 624
 625                         /*
 626                          * Prepare for unlocking the map.  Save the map
 627                          * entry's start and end addresses, which are used to
 628                          * optimize the size of the pager operation below.
 629                          * Even if the map entry's addresses change after
 630                          * unlocking the map, using the saved addresses is
 631                          * safe.
 632                          */
 633                         e_start = fs.entry->start;
 634                         e_end = fs.entry->end;
 635                 }
 636
 637                 /*
 638                  * Call the pager to retrieve the page if there is a chance
 639                  * that the pager has it, and potentially retrieve additional
 640                  * pages at the same time.
 641                  */
 642                 if (fs.object->type != OBJT_DEFAULT) {
 643                         /*
 644                          * Release the map lock before locking the vnode or
 645                          * sleeping in the pager.  (If the current object has
 646                          * a shadow, then an earlier iteration of this loop
 647                          * may have already unlocked the map.)
 648                          */
 649                         unlock_map(&fs);
 650
 651                         if (fs.object->type == OBJT_VNODE &&
 652                             (vp = fs.object->handle) != fs.vp) {
 653                                 /*
 654                                  * Perform an unlock in case the desired vnode
 655                                  * changed while the map was unlocked during a
 656                                  * retry.
 657                                  */
 658                                 unlock_vp(&fs);
 659
 660                                 locked = VOP_ISLOCKED(vp);
 661                                 if (locked != LK_EXCLUSIVE)
 662                                         locked = LK_SHARED;
 663
 664                                 /*
 665                                  * We must not sleep acquiring the vnode lock
 666                                  * while we have the page exclusive busied or
 667                                  * the object's paging-in-progress count
 668                                  * incremented.  Otherwise, we could deadlock.
 669                                  */
 670                                 error = vget(vp, locked | LK_CANRECURSE |
 671                                     LK_NOWAIT, curthread);
 672                                 if (error != 0) {
 673                                         vhold(vp);
 674                                         release_page(&fs);
 675                                         unlock_and_deallocate(&fs);
 676                                         error = vget(vp, locked | LK_RETRY |
 677                                             LK_CANRECURSE, curthread);
 678                                         vdrop(vp);
 679                                         fs.vp = vp;
 680                                         KASSERT(error == 0,
 681                                             ("vm_fault: vget failed"));
 682                                         goto RetryFault;
 683                                 }
 684                                 fs.vp = vp;
 685                         }
 686                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 687                             ("vm_fault: vnode-backed object mapped by system map"));
 688
 689                         /*
 690                          * Page in the requested page and hint the pager,
 691                          * that it may bring up surrounding pages.
 692                          */
 693                         if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
 694                             P_KILLED(curproc)) {
 695                                 behind = 0;
 696                                 ahead = 0;
 697                         } else {
 698                                 /* Is this a sequential fault? */
 699                                 if (nera > 0) {
 700                                         behind = 0;
 701                                         ahead = nera;
 702                                 } else {
 703                                         /*
 704                                          * Request a cluster of pages that is
 705                                          * aligned to a VM_FAULT_READ_DEFAULT
 706                                          * page offset boundary within the
 707                                          * object.  Alignment to a page offset
 708                                          * boundary is more likely to coincide
 709                                          * with the underlying file system
 710                                          * block than alignment to a virtual
 711                                          * address boundary.
 712                                          */
 713                                         cluster_offset = fs.pindex %
 714                                             VM_FAULT_READ_DEFAULT;
 715                                         behind = ulmin(cluster_offset,
 716                                             atop(vaddr - e_start));
 717                                         ahead = VM_FAULT_READ_DEFAULT - 1 -
 718                                             cluster_offset;
 719                                 }
 720                                 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
 721                         }
 722                         rv = vm_pager_get_pages(fs.object, &fs.m, 1,
 723                             &behind, &ahead);
 724                         if (rv == VM_PAGER_OK) {
 725                                 faultcount = behind + 1 + ahead;
 726                                 hardfault = true;
 727                                 break; /* break to PAGE HAS BEEN FOUND */
 728                         }
 729                         if (rv == VM_PAGER_ERROR)
 730                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 731                                     curproc->p_pid, curproc->p_comm);
 732
 733                         /*
 734                          * If an I/O error occurred or the requested page was
 735                          * outside the range of the pager, clean up and return
 736                          * an error.
 737                          */
 738                         if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
 739                                 vm_page_lock(fs.m);
 740                                 if (fs.m->wire_count == 0)
 741                                         vm_page_free(fs.m);
 742                                 else
 743                                         vm_page_xunbusy_maybelocked(fs.m);
 744                                 vm_page_unlock(fs.m);
 745                                 fs.m = NULL;
 746                                 unlock_and_deallocate(&fs);
 747                                 return (rv == VM_PAGER_ERROR ? KERN_FAILURE :
 748                                     KERN_PROTECTION_FAILURE);
 749                         }
 750
 751                         /*
 752                          * The requested page does not exist at this object/
 753                          * offset.  Remove the invalid page from the object,
 754                          * waking up anyone waiting for it, and continue on to
 755                          * the next object.  However, if this is the top-level
 756                          * object, we must leave the busy page in place to
 757                          * prevent another process from rushing past us, and
 758                          * inserting the page in that object at the same time
 759                          * that we are.
 760                          */
 761                         if (fs.object != fs.first_object) {
 762                                 vm_page_lock(fs.m);
 763                                 if (fs.m->wire_count == 0)
 764                                         vm_page_free(fs.m);
 765                                 else
 766                                         vm_page_xunbusy_maybelocked(fs.m);
 767                                 vm_page_unlock(fs.m);
 768                                 fs.m = NULL;
 769                         }
 770                 }
 771
 772                 /*
 773                  * We get here if the object has default pager (or unwiring)
 774                  * or the pager doesn't have the page.
 775                  */
 776                 if (fs.object == fs.first_object)
 777                         fs.first_m = fs.m;
 778
 779                 /*
 780                  * Move on to the next object.  Lock the next object before
 781                  * unlocking the current one.
 782                  */
 783                 next_object = fs.object->backing_object;
 784                 if (next_object == NULL) {
 785                         /*
 786                          * If there's no object left, fill the page in the top
 787                          * object with zeros.
 788                          */
 789                         if (fs.object != fs.first_object) {
 790                                 vm_object_pip_wakeup(fs.object);
 791                                 VM_OBJECT_WUNLOCK(fs.object);
 792
 793                                 fs.object = fs.first_object;
 794                                 fs.pindex = fs.first_pindex;
 795                                 fs.m = fs.first_m;
 796                                 VM_OBJECT_WLOCK(fs.object);
 797                         }
 798                         fs.first_m = NULL;
 799
 800                         /*
 801                          * Zero the page if necessary and mark it valid.
 802                          */
 803                         if ((fs.m->flags & PG_ZERO) == 0) {
 804                                 pmap_zero_page(fs.m);
 805                         } else {
 806                                 PCPU_INC(cnt.v_ozfod);
 807                         }
 808                         PCPU_INC(cnt.v_zfod);
 809                         fs.m->valid = VM_PAGE_BITS_ALL;
 810                         /* Don't try to prefault neighboring pages. */
 811                         faultcount = 1;
 812                         break;  /* break to PAGE HAS BEEN FOUND */
 813                 } else {
 814                         KASSERT(fs.object != next_object,
 815                             ("object loop %p", next_object));
 816                         VM_OBJECT_WLOCK(next_object);
 817                         vm_object_pip_add(next_object, 1);
 818                         if (fs.object != fs.first_object)
 819                                 vm_object_pip_wakeup(fs.object);
 820                         fs.pindex +=
 821                             OFF_TO_IDX(fs.object->backing_object_offset);
 822                         VM_OBJECT_WUNLOCK(fs.object);
 823                         fs.object = next_object;
 824                 }
 825         }
 826
 827         vm_page_assert_xbusied(fs.m);
 828
 829         /*
 830          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 831          * is held.]
 832          */
 833
 834         /*
 835          * If the page is being written, but isn't already owned by the
 836          * top-level object, we have to copy it into a new page owned by the
 837          * top-level object.
 838          */
 839         if (fs.object != fs.first_object) {
 840                 /*
 841                  * We only really need to copy if we want to write it.
 842                  */
 843                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 844                         /*
 845                          * This allows pages to be virtually copied from a
 846                          * backing_object into the first_object, where the
 847                          * backing object has no other refs to it, and cannot
 848                          * gain any more refs.  Instead of a bcopy, we just
 849                          * move the page from the backing object to the
 850                          * first object.  Note that we must mark the page
 851                          * dirty in the first object so that it will go out
 852                          * to swap when needed.
 853                          */
 854                         is_first_object_locked = false;
 855                         if (
 856                                 /*
 857                                  * Only one shadow object
 858                                  */
 859                                 (fs.object->shadow_count == 1) &&
 860                                 /*
 861                                  * No COW refs, except us
 862                                  */
 863                                 (fs.object->ref_count == 1) &&
 864                                 /*
 865                                  * No one else can look this object up
 866                                  */
 867                                 (fs.object->handle == NULL) &&
 868                                 /*
 869                                  * No other ways to look the object up
 870                                  */
 871                                 ((fs.object->type == OBJT_DEFAULT) ||
 872                                  (fs.object->type == OBJT_SWAP)) &&
 873                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
 874                                 /*
 875                                  * We don't chase down the shadow chain
 876                                  */
 877                             fs.object == fs.first_object->backing_object) {
 878                                 vm_page_lock(fs.m);
 879                                 vm_page_remove(fs.m);
 880                                 vm_page_unlock(fs.m);
 881                                 vm_page_lock(fs.first_m);
 882                                 vm_page_replace_checked(fs.m, fs.first_object,
 883                                     fs.first_pindex, fs.first_m);
 884                                 vm_page_free(fs.first_m);
 885                                 vm_page_unlock(fs.first_m);
 886                                 vm_page_dirty(fs.m);
 887 #if VM_NRESERVLEVEL > 0
 888                                 /*
 889                                  * Rename the reservation.
 890                                  */
 891                                 vm_reserv_rename(fs.m, fs.first_object,
 892                                     fs.object, OFF_TO_IDX(
 893                                     fs.first_object->backing_object_offset));
 894 #endif
 895                                 /*
 896                                  * Removing the page from the backing object
 897                                  * unbusied it.
 898                                  */
 899                                 vm_page_xbusy(fs.m);
 900                                 fs.first_m = fs.m;
 901                                 fs.m = NULL;
 902                                 PCPU_INC(cnt.v_cow_optim);
 903                         } else {
 904                                 /*
 905                                  * Oh, well, lets copy it.
 906                                  */
 907                                 pmap_copy_page(fs.m, fs.first_m);
 908                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 909                                 if (wired && (fault_flags &
 910                                     VM_FAULT_WIRE) == 0) {
 911                                         vm_page_lock(fs.first_m);
 912                                         vm_page_wire(fs.first_m);
 913                                         vm_page_unlock(fs.first_m);
 914
 915                                         vm_page_lock(fs.m);
 916                                         vm_page_unwire(fs.m, PQ_INACTIVE);
 917                                         vm_page_unlock(fs.m);
 918                                 }
 919                                 /*
 920                                  * We no longer need the old page or object.
 921                                  */
 922                                 release_page(&fs);
 923                         }
 924                         /*
 925                          * fs.object != fs.first_object due to above
 926                          * conditional
 927                          */
 928                         vm_object_pip_wakeup(fs.object);
 929                         VM_OBJECT_WUNLOCK(fs.object);
 930                         /*
 931                          * Only use the new page below...
 932                          */
 933                         fs.object = fs.first_object;
 934                         fs.pindex = fs.first_pindex;
 935                         fs.m = fs.first_m;
 936                         if (!is_first_object_locked)
 937                                 VM_OBJECT_WLOCK(fs.object);
 938                         PCPU_INC(cnt.v_cow_faults);
 939                         curthread->td_cow++;
 940                 } else {
 941                         prot &= ~VM_PROT_WRITE;
 942                 }
 943         }
 944
 945         /*
 946          * We must verify that the maps have not changed since our last
 947          * lookup.
 948          */
 949         if (!fs.lookup_still_valid) {
 950                 if (!vm_map_trylock_read(fs.map)) {
 951                         release_page(&fs);
 952                         unlock_and_deallocate(&fs);
 953                         goto RetryFault;
 954                 }
 955                 fs.lookup_still_valid = true;
 956                 if (fs.map->timestamp != map_generation) {
 957                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 958                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 959
 960                         /*
 961                          * If we don't need the page any longer, put it on the inactive
 962                          * list (the easiest thing to do here).  If no one needs it,
 963                          * pageout will grab it eventually.
 964                          */
 965                         if (result != KERN_SUCCESS) {
 966                                 release_page(&fs);
 967                                 unlock_and_deallocate(&fs);
 968
 969                                 /*
 970                                  * If retry of map lookup would have blocked then
 971                                  * retry fault from start.
 972                                  */
 973                                 if (result == KERN_FAILURE)
 974                                         goto RetryFault;
 975                                 return (result);
 976                         }
 977                         if ((retry_object != fs.first_object) ||
 978                             (retry_pindex != fs.first_pindex)) {
 979                                 release_page(&fs);
 980                                 unlock_and_deallocate(&fs);
 981                                 goto RetryFault;
 982                         }
 983
 984                         /*
 985                          * Check whether the protection has changed or the object has
 986                          * been copied while we left the map unlocked. Changing from
 987                          * read to write permission is OK - we leave the page
 988                          * write-protected, and catch the write fault. Changing from
 989                          * write to read permission means that we can't mark the page
 990                          * write-enabled after all.
 991                          */
 992                         prot &= retry_prot;
 993                 }
 994         }
 995
 996         /*
 997          * If the page was filled by a pager, save the virtual address that
 998          * should be faulted on next under a sequential access pattern to the
 999          * map entry.  A read lock on the map suffices to update this address
1000          * safely.
1001          */
1002         if (hardfault)
1003                 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1004
1005         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1006         vm_page_assert_xbusied(fs.m);
1007
1008         /*
1009          * Page must be completely valid or it is not fit to
1010          * map into user space.  vm_pager_get_pages() ensures this.
1011          */
1012         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
1013             ("vm_fault: page %p partially invalid", fs.m));
1014         VM_OBJECT_WUNLOCK(fs.object);
1015
1016         /*
1017          * Put this page into the physical map.  We had to do the unlock above
1018          * because pmap_enter() may sleep.  We don't put the page
1019          * back on the active queue until later so that the pageout daemon
1020          * won't find it (yet).
1021          */
1022         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1023             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1024         if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1025             wired == 0)
1026                 vm_fault_prefault(&fs, vaddr,
1027                     faultcount > 0 ? behind : PFBAK,
1028                     faultcount > 0 ? ahead : PFFOR);
1029         VM_OBJECT_WLOCK(fs.object);
1030         vm_page_lock(fs.m);
1031
1032         /*
1033          * If the page is not wired down, then put it where the pageout daemon
1034          * can find it.
1035          */
1036         if ((fault_flags & VM_FAULT_WIRE) != 0) {
1037                 KASSERT(wired, ("VM_FAULT_WIRE && !wired"));
1038                 vm_page_wire(fs.m);
1039         } else
1040                 vm_page_activate(fs.m);
1041         if (m_hold != NULL) {
1042                 *m_hold = fs.m;
1043                 vm_page_hold(fs.m);
1044         }
1045         vm_page_unlock(fs.m);
1046         vm_page_xunbusy(fs.m);
1047
1048         /*
1049          * Unlock everything, and return
1050          */
1051         unlock_and_deallocate(&fs);
1052         if (hardfault) {
1053                 PCPU_INC(cnt.v_io_faults);
1054                 curthread->td_ru.ru_majflt++;
1055 #ifdef RACCT
1056                 if (racct_enable && fs.object->type == OBJT_VNODE) {
1057                         PROC_LOCK(curproc);
1058                         if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1059                                 racct_add_force(curproc, RACCT_WRITEBPS,
1060                                     PAGE_SIZE + behind * PAGE_SIZE);
1061                                 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1062                         } else {
1063                                 racct_add_force(curproc, RACCT_READBPS,
1064                                     PAGE_SIZE + ahead * PAGE_SIZE);
1065                                 racct_add_force(curproc, RACCT_READIOPS, 1);
1066                         }
1067                         PROC_UNLOCK(curproc);
1068                 }
1069 #endif
1070         } else
1071                 curthread->td_ru.ru_minflt++;
1072
1073         return (KERN_SUCCESS);
1074 }
1075
1076 /*
1077  * Speed up the reclamation of pages that precede the faulting pindex within
1078  * the first object of the shadow chain.  Essentially, perform the equivalent
1079  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1080  * the faulting pindex by the cluster size when the pages read by vm_fault()
1081  * cross a cluster-size boundary.  The cluster size is the greater of the
1082  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1083  *
1084  * When "fs->first_object" is a shadow object, the pages in the backing object
1085  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1086  * function must only be concerned with pages in the first object.
1087  */
1088 static void
1089 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1090 {
1091         vm_map_entry_t entry;
1092         vm_object_t first_object, object;
1093         vm_offset_t end, start;
1094         vm_page_t m, m_next;
1095         vm_pindex_t pend, pstart;
1096         vm_size_t size;
1097
1098         object = fs->object;
1099         VM_OBJECT_ASSERT_WLOCKED(object);
1100         first_object = fs->first_object;
1101         if (first_object != object) {
1102                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1103                         VM_OBJECT_WUNLOCK(object);
1104                         VM_OBJECT_WLOCK(first_object);
1105                         VM_OBJECT_WLOCK(object);
1106                 }
1107         }
1108         /* Neither fictitious nor unmanaged pages can be reclaimed. */
1109         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1110                 size = VM_FAULT_DONTNEED_MIN;
1111                 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1112                         size = pagesizes[1];
1113                 end = rounddown2(vaddr, size);
1114                 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1115                     (entry = fs->entry)->start < end) {
1116                         if (end - entry->start < size)
1117                                 start = entry->start;
1118                         else
1119                                 start = end - size;
1120                         pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1121                         pstart = OFF_TO_IDX(entry->offset) + atop(start -
1122                             entry->start);
1123                         m_next = vm_page_find_least(first_object, pstart);
1124                         pend = OFF_TO_IDX(entry->offset) + atop(end -
1125                             entry->start);
1126                         while ((m = m_next) != NULL && m->pindex < pend) {
1127                                 m_next = TAILQ_NEXT(m, listq);
1128                                 if (m->valid != VM_PAGE_BITS_ALL ||
1129                                     vm_page_busied(m))
1130                                         continue;
1131
1132                                 /*
1133                                  * Don't clear PGA_REFERENCED, since it would
1134                                  * likely represent a reference by a different
1135                                  * process.
1136                                  *
1137                                  * Typically, at this point, prefetched pages
1138                                  * are still in the inactive queue.  Only
1139                                  * pages that triggered page faults are in the
1140                                  * active queue.
1141                                  */
1142                                 vm_page_lock(m);
1143                                 vm_page_deactivate(m);
1144                                 vm_page_unlock(m);
1145                         }
1146                 }
1147         }
1148         if (first_object != object)
1149                 VM_OBJECT_WUNLOCK(first_object);
1150 }
1151
1152 /*
1153  * vm_fault_prefault provides a quick way of clustering
1154  * pagefaults into a processes address space.  It is a "cousin"
1155  * of vm_map_pmap_enter, except it runs at page fault time instead
1156  * of mmap time.
1157  */
1158 static void
1159 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1160     int backward, int forward)
1161 {
1162         pmap_t pmap;
1163         vm_map_entry_t entry;
1164         vm_object_t backing_object, lobject;
1165         vm_offset_t addr, starta;
1166         vm_pindex_t pindex;
1167         vm_page_t m;
1168         int i;
1169
1170         pmap = fs->map->pmap;
1171         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1172                 return;
1173
1174         entry = fs->entry;
1175
1176         starta = addra - backward * PAGE_SIZE;
1177         if (starta < entry->start) {
1178                 starta = entry->start;
1179         } else if (starta > addra) {
1180                 starta = 0;
1181         }
1182
1183         /*
1184          * Generate the sequence of virtual addresses that are candidates for
1185          * prefaulting in an outward spiral from the faulting virtual address,
1186          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1187          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1188          * If the candidate address doesn't have a backing physical page, then
1189          * the loop immediately terminates.
1190          */
1191         for (i = 0; i < 2 * imax(backward, forward); i++) {
1192                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1193                     PAGE_SIZE);
1194                 if (addr > addra + forward * PAGE_SIZE)
1195                         addr = 0;
1196
1197                 if (addr < starta || addr >= entry->end)
1198                         continue;
1199
1200                 if (!pmap_is_prefaultable(pmap, addr))
1201                         continue;
1202
1203                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1204                 lobject = entry->object.vm_object;
1205                 VM_OBJECT_RLOCK(lobject);
1206                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1207                     lobject->type == OBJT_DEFAULT &&
1208                     (backing_object = lobject->backing_object) != NULL) {
1209                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1210                             0, ("vm_fault_prefault: unaligned object offset"));
1211                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1212                         VM_OBJECT_RLOCK(backing_object);
1213                         VM_OBJECT_RUNLOCK(lobject);
1214                         lobject = backing_object;
1215                 }
1216                 if (m == NULL) {
1217                         VM_OBJECT_RUNLOCK(lobject);
1218                         break;
1219                 }
1220                 if (m->valid == VM_PAGE_BITS_ALL &&
1221                     (m->flags & PG_FICTITIOUS) == 0)
1222                         pmap_enter_quick(pmap, addr, m, entry->protection);
1223                 VM_OBJECT_RUNLOCK(lobject);
1224         }
1225 }
1226
1227 /*
1228  * Hold each of the physical pages that are mapped by the specified range of
1229  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1230  * and allow the specified types of access, "prot".  If all of the implied
1231  * pages are successfully held, then the number of held pages is returned
1232  * together with pointers to those pages in the array "ma".  However, if any
1233  * of the pages cannot be held, -1 is returned.
1234  */
1235 int
1236 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1237     vm_prot_t prot, vm_page_t *ma, int max_count)
1238 {
1239         vm_offset_t end, va;
1240         vm_page_t *mp;
1241         int count;
1242         boolean_t pmap_failed;
1243
1244         if (len == 0)
1245                 return (0);
1246         end = round_page(addr + len);
1247         addr = trunc_page(addr);
1248
1249         /*
1250          * Check for illegal addresses.
1251          */
1252         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1253                 return (-1);
1254
1255         if (atop(end - addr) > max_count)
1256                 panic("vm_fault_quick_hold_pages: count > max_count");
1257         count = atop(end - addr);
1258
1259         /*
1260          * Most likely, the physical pages are resident in the pmap, so it is
1261          * faster to try pmap_extract_and_hold() first.
1262          */
1263         pmap_failed = FALSE;
1264         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1265                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1266                 if (*mp == NULL)
1267                         pmap_failed = TRUE;
1268                 else if ((prot & VM_PROT_WRITE) != 0 &&
1269                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1270                         /*
1271                          * Explicitly dirty the physical page.  Otherwise, the
1272                          * caller's changes may go unnoticed because they are
1273                          * performed through an unmanaged mapping or by a DMA
1274                          * operation.
1275                          *
1276                          * The object lock is not held here.
1277                          * See vm_page_clear_dirty_mask().
1278                          */
1279                         vm_page_dirty(*mp);
1280                 }
1281         }
1282         if (pmap_failed) {
1283                 /*
1284                  * One or more pages could not be held by the pmap.  Either no
1285                  * page was mapped at the specified virtual address or that
1286                  * mapping had insufficient permissions.  Attempt to fault in
1287                  * and hold these pages.
1288                  */
1289                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1290                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1291                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1292                                 goto error;
1293         }
1294         return (count);
1295 error:
1296         for (mp = ma; mp < ma + count; mp++)
1297                 if (*mp != NULL) {
1298                         vm_page_lock(*mp);
1299                         vm_page_unhold(*mp);
1300                         vm_page_unlock(*mp);
1301                 }
1302         return (-1);
1303 }
1304
1305 /*
1306  *      Routine:
1307  *              vm_fault_copy_entry
1308  *      Function:
1309  *              Create new shadow object backing dst_entry with private copy of
1310  *              all underlying pages. When src_entry is equal to dst_entry,
1311  *              function implements COW for wired-down map entry. Otherwise,
1312  *              it forks wired entry into dst_map.
1313  *
1314  *      In/out conditions:
1315  *              The source and destination maps must be locked for write.
1316  *              The source map entry must be wired down (or be a sharing map
1317  *              entry corresponding to a main map entry that is wired down).
1318  */
1319 void
1320 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1321     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1322     vm_ooffset_t *fork_charge)
1323 {
1324         vm_object_t backing_object, dst_object, object, src_object;
1325         vm_pindex_t dst_pindex, pindex, src_pindex;
1326         vm_prot_t access, prot;
1327         vm_offset_t vaddr;
1328         vm_page_t dst_m;
1329         vm_page_t src_m;
1330         boolean_t upgrade;
1331
1332 #ifdef  lint
1333         src_map++;
1334 #endif  /* lint */
1335
1336         upgrade = src_entry == dst_entry;
1337         access = prot = dst_entry->protection;
1338
1339         src_object = src_entry->object.vm_object;
1340         src_pindex = OFF_TO_IDX(src_entry->offset);
1341
1342         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1343                 dst_object = src_object;
1344                 vm_object_reference(dst_object);
1345         } else {
1346                 /*
1347                  * Create the top-level object for the destination entry. (Doesn't
1348                  * actually shadow anything - we copy the pages directly.)
1349                  */
1350                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1351                     OFF_TO_IDX(dst_entry->end - dst_entry->start));
1352 #if VM_NRESERVLEVEL > 0
1353                 dst_object->flags |= OBJ_COLORED;
1354                 dst_object->pg_color = atop(dst_entry->start);
1355 #endif
1356         }
1357
1358         VM_OBJECT_WLOCK(dst_object);
1359         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1360             ("vm_fault_copy_entry: vm_object not NULL"));
1361         if (src_object != dst_object) {
1362                 dst_entry->object.vm_object = dst_object;
1363                 dst_entry->offset = 0;
1364                 dst_object->charge = dst_entry->end - dst_entry->start;
1365         }
1366         if (fork_charge != NULL) {
1367                 KASSERT(dst_entry->cred == NULL,
1368                     ("vm_fault_copy_entry: leaked swp charge"));
1369                 dst_object->cred = curthread->td_ucred;
1370                 crhold(dst_object->cred);
1371                 *fork_charge += dst_object->charge;
1372         } else if (dst_object->cred == NULL) {
1373                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1374                     dst_entry));
1375                 dst_object->cred = dst_entry->cred;
1376                 dst_entry->cred = NULL;
1377         }
1378
1379         /*
1380          * If not an upgrade, then enter the mappings in the pmap as
1381          * read and/or execute accesses.  Otherwise, enter them as
1382          * write accesses.
1383          *
1384          * A writeable large page mapping is only created if all of
1385          * the constituent small page mappings are modified. Marking
1386          * PTEs as modified on inception allows promotion to happen
1387          * without taking potentially large number of soft faults.
1388          */
1389         if (!upgrade)
1390                 access &= ~VM_PROT_WRITE;
1391
1392         /*
1393          * Loop through all of the virtual pages within the entry's
1394          * range, copying each page from the source object to the
1395          * destination object.  Since the source is wired, those pages
1396          * must exist.  In contrast, the destination is pageable.
1397          * Since the destination object does share any backing storage
1398          * with the source object, all of its pages must be dirtied,
1399          * regardless of whether they can be written.
1400          */
1401         for (vaddr = dst_entry->start, dst_pindex = 0;
1402             vaddr < dst_entry->end;
1403             vaddr += PAGE_SIZE, dst_pindex++) {
1404 again:
1405                 /*
1406                  * Find the page in the source object, and copy it in.
1407                  * Because the source is wired down, the page will be
1408                  * in memory.
1409                  */
1410                 if (src_object != dst_object)
1411                         VM_OBJECT_RLOCK(src_object);
1412                 object = src_object;
1413                 pindex = src_pindex + dst_pindex;
1414                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1415                     (backing_object = object->backing_object) != NULL) {
1416                         /*
1417                          * Unless the source mapping is read-only or
1418                          * it is presently being upgraded from
1419                          * read-only, the first object in the shadow
1420                          * chain should provide all of the pages.  In
1421                          * other words, this loop body should never be
1422                          * executed when the source mapping is already
1423                          * read/write.
1424                          */
1425                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1426                             upgrade,
1427                             ("vm_fault_copy_entry: main object missing page"));
1428
1429                         VM_OBJECT_RLOCK(backing_object);
1430                         pindex += OFF_TO_IDX(object->backing_object_offset);
1431                         if (object != dst_object)
1432                                 VM_OBJECT_RUNLOCK(object);
1433                         object = backing_object;
1434                 }
1435                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1436
1437                 if (object != dst_object) {
1438                         /*
1439                          * Allocate a page in the destination object.
1440                          */
1441                         dst_m = vm_page_alloc(dst_object, (src_object ==
1442                             dst_object ? src_pindex : 0) + dst_pindex,
1443                             VM_ALLOC_NORMAL);
1444                         if (dst_m == NULL) {
1445                                 VM_OBJECT_WUNLOCK(dst_object);
1446                                 VM_OBJECT_RUNLOCK(object);
1447                                 VM_WAIT;
1448                                 VM_OBJECT_WLOCK(dst_object);
1449                                 goto again;
1450                         }
1451                         pmap_copy_page(src_m, dst_m);
1452                         VM_OBJECT_RUNLOCK(object);
1453                         dst_m->valid = VM_PAGE_BITS_ALL;
1454                         dst_m->dirty = VM_PAGE_BITS_ALL;
1455                 } else {
1456                         dst_m = src_m;
1457                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1458                                 goto again;
1459                         vm_page_xbusy(dst_m);
1460                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1461                             ("invalid dst page %p", dst_m));
1462                 }
1463                 VM_OBJECT_WUNLOCK(dst_object);
1464
1465                 /*
1466                  * Enter it in the pmap. If a wired, copy-on-write
1467                  * mapping is being replaced by a write-enabled
1468                  * mapping, then wire that new mapping.
1469                  */
1470                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1471                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1472
1473                 /*
1474                  * Mark it no longer busy, and put it on the active list.
1475                  */
1476                 VM_OBJECT_WLOCK(dst_object);
1477
1478                 if (upgrade) {
1479                         if (src_m != dst_m) {
1480                                 vm_page_lock(src_m);
1481                                 vm_page_unwire(src_m, PQ_INACTIVE);
1482                                 vm_page_unlock(src_m);
1483                                 vm_page_lock(dst_m);
1484                                 vm_page_wire(dst_m);
1485                                 vm_page_unlock(dst_m);
1486                         } else {
1487                                 KASSERT(dst_m->wire_count > 0,
1488                                     ("dst_m %p is not wired", dst_m));
1489                         }
1490                 } else {
1491                         vm_page_lock(dst_m);
1492                         vm_page_activate(dst_m);
1493                         vm_page_unlock(dst_m);
1494                 }
1495                 vm_page_xunbusy(dst_m);
1496         }
1497         VM_OBJECT_WUNLOCK(dst_object);
1498         if (upgrade) {
1499                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1500                 vm_object_deallocate(src_object);
1501         }
1502 }
1503
1504 /*
1505  * Block entry into the machine-independent layer's page fault handler by
1506  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1507  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1508  * spurious page faults.
1509  */
1510 int
1511 vm_fault_disable_pagefaults(void)
1512 {
1513
1514         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1515 }
1516
1517 void
1518 vm_fault_enable_pagefaults(int save)
1519 {
1520
1521         curthread_pflags_restore(save);
1522 }