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