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