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 map_generation;
 126         bool lookup_still_valid;
 127         struct vnode *vp;
 128 };
 129
 130 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
 131             int ahead);
 132 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 133             int backward, int forward);
 134
 135 static inline void
 136 release_page(struct faultstate *fs)
 137 {
 138
 139         vm_page_xunbusy(fs->m);
 140         vm_page_lock(fs->m);
 141         vm_page_deactivate(fs->m);
 142         vm_page_unlock(fs->m);
 143         fs->m = NULL;
 144 }
 145
 146 static inline void
 147 unlock_map(struct faultstate *fs)
 148 {
 149
 150         if (fs->lookup_still_valid) {
 151                 vm_map_lookup_done(fs->map, fs->entry);
 152                 fs->lookup_still_valid = false;
 153         }
 154 }
 155
 156 static void
 157 unlock_vp(struct faultstate *fs)
 158 {
 159
 160         if (fs->vp != NULL) {
 161                 vput(fs->vp);
 162                 fs->vp = NULL;
 163         }
 164 }
 165
 166 static void
 167 unlock_and_deallocate(struct faultstate *fs)
 168 {
 169
 170         vm_object_pip_wakeup(fs->object);
 171         VM_OBJECT_WUNLOCK(fs->object);
 172         if (fs->object != fs->first_object) {
 173                 VM_OBJECT_WLOCK(fs->first_object);
 174                 vm_page_lock(fs->first_m);
 175                 vm_page_free(fs->first_m);
 176                 vm_page_unlock(fs->first_m);
 177                 vm_object_pip_wakeup(fs->first_object);
 178                 VM_OBJECT_WUNLOCK(fs->first_object);
 179                 fs->first_m = NULL;
 180         }
 181         vm_object_deallocate(fs->first_object);
 182         unlock_map(fs);
 183         unlock_vp(fs);
 184 }
 185
 186 static void
 187 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 188     vm_prot_t fault_type, int fault_flags, bool set_wd)
 189 {
 190         bool need_dirty;
 191
 192         if (((prot & VM_PROT_WRITE) == 0 &&
 193             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 194             (m->oflags & VPO_UNMANAGED) != 0)
 195                 return;
 196
 197         VM_OBJECT_ASSERT_LOCKED(m->object);
 198
 199         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 200             (fault_flags & VM_FAULT_WIRE) == 0) ||
 201             (fault_flags & VM_FAULT_DIRTY) != 0;
 202
 203         if (set_wd)
 204                 vm_object_set_writeable_dirty(m->object);
 205         else
 206                 /*
 207                  * If two callers of vm_fault_dirty() with set_wd ==
 208                  * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
 209                  * flag set, other with flag clear, race, it is
 210                  * possible for the no-NOSYNC thread to see m->dirty
 211                  * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
 212                  * around manipulation of VPO_NOSYNC and
 213                  * vm_page_dirty() call, to avoid the race and keep
 214                  * m->oflags consistent.
 215                  */
 216                 vm_page_lock(m);
 217
 218         /*
 219          * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 220          * if the page is already dirty to prevent data written with
 221          * the expectation of being synced from not being synced.
 222          * Likewise if this entry does not request NOSYNC then make
 223          * sure the page isn't marked NOSYNC.  Applications sharing
 224          * data should use the same flags to avoid ping ponging.
 225          */
 226         if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
 227                 if (m->dirty == 0) {
 228                         m->oflags |= VPO_NOSYNC;
 229                 }
 230         } else {
 231                 m->oflags &= ~VPO_NOSYNC;
 232         }
 233
 234         /*
 235          * If the fault is a write, we know that this page is being
 236          * written NOW so dirty it explicitly to save on
 237          * pmap_is_modified() calls later.
 238          *
 239          * Also tell the backing pager, if any, that it should remove
 240          * any swap backing since the page is now dirty.
 241          */
 242         if (need_dirty)
 243                 vm_page_dirty(m);
 244         if (!set_wd)
 245                 vm_page_unlock(m);
 246         if (need_dirty)
 247                 vm_pager_page_unswapped(m);
 248 }
 249
 250 static void
 251 vm_fault_fill_hold(vm_page_t *m_hold, vm_page_t m)
 252 {
 253
 254         if (m_hold != NULL) {
 255                 *m_hold = m;
 256                 vm_page_lock(m);
 257                 vm_page_hold(m);
 258                 vm_page_unlock(m);
 259         }
 260 }
 261
 262 /*
 263  * Unlocks fs.first_object and fs.map on success.
 264  */
 265 static int
 266 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
 267     int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
 268 {
 269         vm_page_t m;
 270         int rv;
 271
 272         MPASS(fs->vp == NULL);
 273         m = vm_page_lookup(fs->first_object, fs->first_pindex);
 274         /* A busy page can be mapped for read|execute access. */
 275         if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 276             vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
 277                 return (KERN_FAILURE);
 278         rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
 279             PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), 0);
 280         if (rv != KERN_SUCCESS)
 281                 return (rv);
 282         vm_fault_fill_hold(m_hold, m);
 283         vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false);
 284         VM_OBJECT_RUNLOCK(fs->first_object);
 285         if (!wired)
 286                 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR);
 287         vm_map_lookup_done(fs->map, fs->entry);
 288         curthread->td_ru.ru_minflt++;
 289         return (KERN_SUCCESS);
 290 }
 291
 292 static void
 293 vm_fault_restore_map_lock(struct faultstate *fs)
 294 {
 295
 296         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 297         MPASS(fs->first_object->paging_in_progress > 0);
 298
 299         if (!vm_map_trylock_read(fs->map)) {
 300                 VM_OBJECT_WUNLOCK(fs->first_object);
 301                 vm_map_lock_read(fs->map);
 302                 VM_OBJECT_WLOCK(fs->first_object);
 303         }
 304         fs->lookup_still_valid = true;
 305 }
 306
 307 static void
 308 vm_fault_populate_check_page(vm_page_t m)
 309 {
 310
 311         /*
 312          * Check each page to ensure that the pager is obeying the
 313          * interface: the page must be installed in the object, fully
 314          * valid, and exclusively busied.
 315          */
 316         MPASS(m != NULL);
 317         MPASS(m->valid == VM_PAGE_BITS_ALL);
 318         MPASS(vm_page_xbusied(m));
 319 }
 320
 321 static void
 322 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
 323     vm_pindex_t last)
 324 {
 325         vm_page_t m;
 326         vm_pindex_t pidx;
 327
 328         VM_OBJECT_ASSERT_WLOCKED(object);
 329         MPASS(first <= last);
 330         for (pidx = first, m = vm_page_lookup(object, pidx);
 331             pidx <= last; pidx++, m = vm_page_next(m)) {
 332                 vm_fault_populate_check_page(m);
 333                 vm_page_lock(m);
 334                 vm_page_deactivate(m);
 335                 vm_page_unlock(m);
 336                 vm_page_xunbusy(m);
 337         }
 338 }
 339
 340 static int
 341 vm_fault_populate(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
 342     int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
 343 {
 344         vm_page_t m;
 345         vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
 346         int rv;
 347
 348         MPASS(fs->object == fs->first_object);
 349         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 350         MPASS(fs->first_object->paging_in_progress > 0);
 351         MPASS(fs->first_object->backing_object == NULL);
 352         MPASS(fs->lookup_still_valid);
 353
 354         pager_first = OFF_TO_IDX(fs->entry->offset);
 355         pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
 356         unlock_map(fs);
 357         unlock_vp(fs);
 358
 359         /*
 360          * Call the pager (driver) populate() method.
 361          *
 362          * There is no guarantee that the method will be called again
 363          * if the current fault is for read, and a future fault is
 364          * for write.  Report the entry's maximum allowed protection
 365          * to the driver.
 366          */
 367         rv = vm_pager_populate(fs->first_object, fs->first_pindex,
 368             fault_type, fs->entry->max_protection, &pager_first, &pager_last);
 369
 370         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 371         if (rv == VM_PAGER_BAD) {
 372                 /*
 373                  * VM_PAGER_BAD is the backdoor for a pager to request
 374                  * normal fault handling.
 375                  */
 376                 vm_fault_restore_map_lock(fs);
 377                 if (fs->map->timestamp != fs->map_generation)
 378                         return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 379                 return (KERN_NOT_RECEIVER);
 380         }
 381         if (rv != VM_PAGER_OK)
 382                 return (KERN_FAILURE); /* AKA SIGSEGV */
 383
 384         /* Ensure that the driver is obeying the interface. */
 385         MPASS(pager_first <= pager_last);
 386         MPASS(fs->first_pindex <= pager_last);
 387         MPASS(fs->first_pindex >= pager_first);
 388         MPASS(pager_last < fs->first_object->size);
 389
 390         vm_fault_restore_map_lock(fs);
 391         if (fs->map->timestamp != fs->map_generation) {
 392                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 393                     pager_last);
 394                 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 395         }
 396
 397         /*
 398          * The map is unchanged after our last unlock.  Process the fault.
 399          *
 400          * The range [pager_first, pager_last] that is given to the
 401          * pager is only a hint.  The pager may populate any range
 402          * within the object that includes the requested page index.
 403          * In case the pager expanded the range, clip it to fit into
 404          * the map entry.
 405          */
 406         map_first = OFF_TO_IDX(fs->entry->offset);
 407         if (map_first > pager_first) {
 408                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 409                     map_first - 1);
 410                 pager_first = map_first;
 411         }
 412         map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
 413         if (map_last < pager_last) {
 414                 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
 415                     pager_last);
 416                 pager_last = map_last;
 417         }
 418         for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
 419             pidx <= pager_last; pidx++, m = vm_page_next(m)) {
 420                 vm_fault_populate_check_page(m);
 421                 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags,
 422                     true);
 423                 VM_OBJECT_WUNLOCK(fs->first_object);
 424                 pmap_enter(fs->map->pmap, fs->entry->start + IDX_TO_OFF(pidx) -
 425                     fs->entry->offset, m, prot, fault_type | (wired ?
 426                     PMAP_ENTER_WIRED : 0), 0);
 427                 VM_OBJECT_WLOCK(fs->first_object);
 428                 if (pidx == fs->first_pindex)
 429                         vm_fault_fill_hold(m_hold, m);
 430                 vm_page_lock(m);
 431                 if ((fault_flags & VM_FAULT_WIRE) != 0) {
 432                         KASSERT(wired, ("VM_FAULT_WIRE && !wired"));
 433                         vm_page_wire(m);
 434                 } else {
 435                         vm_page_activate(m);
 436                 }
 437                 vm_page_unlock(m);
 438                 vm_page_xunbusy(m);
 439         }
 440         curthread->td_ru.ru_majflt++;
 441         return (KERN_SUCCESS);
 442 }
 443
 444 /*
 445  *      vm_fault:
 446  *
 447  *      Handle a page fault occurring at the given address,
 448  *      requiring the given permissions, in the map specified.
 449  *      If successful, the page is inserted into the
 450  *      associated physical map.
 451  *
 452  *      NOTE: the given address should be truncated to the
 453  *      proper page address.
 454  *
 455  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 456  *      a standard error specifying why the fault is fatal is returned.
 457  *
 458  *      The map in question must be referenced, and remains so.
 459  *      Caller may hold no locks.
 460  */
 461 int
 462 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 463     int fault_flags)
 464 {
 465         struct thread *td;
 466         int result;
 467
 468         td = curthread;
 469         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 470                 return (KERN_PROTECTION_FAILURE);
 471 #ifdef KTRACE
 472         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 473                 ktrfault(vaddr, fault_type);
 474 #endif
 475         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 476             NULL);
 477 #ifdef KTRACE
 478         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 479                 ktrfaultend(result);
 480 #endif
 481         return (result);
 482 }
 483
 484 int
 485 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 486     int fault_flags, vm_page_t *m_hold)
 487 {
 488         struct faultstate fs;
 489         struct vnode *vp;
 490         vm_object_t next_object, retry_object;
 491         vm_offset_t e_end, e_start;
 492         vm_pindex_t retry_pindex;
 493         vm_prot_t prot, retry_prot;
 494         int ahead, alloc_req, behind, cluster_offset, error, era, faultcount;
 495         int locked, nera, result, rv;
 496         u_char behavior;
 497         boolean_t wired;        /* Passed by reference. */
 498         bool dead, growstack, hardfault, is_first_object_locked;
 499
 500         VM_CNT_INC(v_vm_faults);
 501         fs.vp = NULL;
 502         faultcount = 0;
 503         nera = -1;
 504         growstack = true;
 505         hardfault = false;
 506
 507 RetryFault:;
 508
 509         /*
 510          * Find the backing store object and offset into it to begin the
 511          * search.
 512          */
 513         fs.map = map;
 514         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 515             &fs.first_object, &fs.first_pindex, &prot, &wired);
 516         if (result != KERN_SUCCESS) {
 517                 if (growstack && result == KERN_INVALID_ADDRESS &&
 518                     map != kernel_map) {
 519                         result = vm_map_growstack(curproc, vaddr);
 520                         if (result != KERN_SUCCESS)
 521                                 return (KERN_FAILURE);
 522                         growstack = false;
 523                         goto RetryFault;
 524                 }
 525                 unlock_vp(&fs);
 526                 return (result);
 527         }
 528
 529         fs.map_generation = fs.map->timestamp;
 530
 531         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 532                 panic("%s: fault on nofault entry, addr: %#lx",
 533                     __func__, (u_long)vaddr);
 534         }
 535
 536         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 537             fs.entry->wiring_thread != curthread) {
 538                 vm_map_unlock_read(fs.map);
 539                 vm_map_lock(fs.map);
 540                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 541                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 542                         unlock_vp(&fs);
 543                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 544                         vm_map_unlock_and_wait(fs.map, 0);
 545                 } else
 546                         vm_map_unlock(fs.map);
 547                 goto RetryFault;
 548         }
 549
 550         if (wired)
 551                 fault_type = prot | (fault_type & VM_PROT_COPY);
 552         else
 553                 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
 554                     ("!wired && VM_FAULT_WIRE"));
 555
 556         /*
 557          * Try to avoid lock contention on the top-level object through
 558          * special-case handling of some types of page faults, specifically,
 559          * those that are both (1) mapping an existing page from the top-
 560          * level object and (2) not having to mark that object as containing
 561          * dirty pages.  Under these conditions, a read lock on the top-level
 562          * object suffices, allowing multiple page faults of a similar type to
 563          * run in parallel on the same top-level object.
 564          */
 565         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 566             (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
 567             /* avoid calling vm_object_set_writeable_dirty() */
 568             ((prot & VM_PROT_WRITE) == 0 ||
 569             (fs.first_object->type != OBJT_VNODE &&
 570             (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 571             (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
 572                 VM_OBJECT_RLOCK(fs.first_object);
 573                 if ((prot & VM_PROT_WRITE) == 0 ||
 574                     (fs.first_object->type != OBJT_VNODE &&
 575                     (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 576                     (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) {
 577                         rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
 578                             fault_flags, wired, m_hold);
 579                         if (rv == KERN_SUCCESS)
 580                                 return (rv);
 581                 }
 582                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 583                         VM_OBJECT_RUNLOCK(fs.first_object);
 584                         VM_OBJECT_WLOCK(fs.first_object);
 585                 }
 586         } else {
 587                 VM_OBJECT_WLOCK(fs.first_object);
 588         }
 589
 590         /*
 591          * Make a reference to this object to prevent its disposal while we
 592          * are messing with it.  Once we have the reference, the map is free
 593          * to be diddled.  Since objects reference their shadows (and copies),
 594          * they will stay around as well.
 595          *
 596          * Bump the paging-in-progress count to prevent size changes (e.g.
 597          * truncation operations) during I/O.
 598          */
 599         vm_object_reference_locked(fs.first_object);
 600         vm_object_pip_add(fs.first_object, 1);
 601
 602         fs.lookup_still_valid = true;
 603
 604         fs.first_m = NULL;
 605
 606         /*
 607          * Search for the page at object/offset.
 608          */
 609         fs.object = fs.first_object;
 610         fs.pindex = fs.first_pindex;
 611         while (TRUE) {
 612                 /*
 613                  * If the object is marked for imminent termination,
 614                  * we retry here, since the collapse pass has raced
 615                  * with us.  Otherwise, if we see terminally dead
 616                  * object, return fail.
 617                  */
 618                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 619                         dead = fs.object->type == OBJT_DEAD;
 620                         unlock_and_deallocate(&fs);
 621                         if (dead)
 622                                 return (KERN_PROTECTION_FAILURE);
 623                         pause("vmf_de", 1);
 624                         goto RetryFault;
 625                 }
 626
 627                 /*
 628                  * See if page is resident
 629                  */
 630                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 631                 if (fs.m != NULL) {
 632                         /*
 633                          * Wait/Retry if the page is busy.  We have to do this
 634                          * if the page is either exclusive or shared busy
 635                          * because the vm_pager may be using read busy for
 636                          * pageouts (and even pageins if it is the vnode
 637                          * pager), and we could end up trying to pagein and
 638                          * pageout the same page simultaneously.
 639                          *
 640                          * We can theoretically allow the busy case on a read
 641                          * fault if the page is marked valid, but since such
 642                          * pages are typically already pmap'd, putting that
 643                          * special case in might be more effort then it is
 644                          * worth.  We cannot under any circumstances mess
 645                          * around with a shared busied page except, perhaps,
 646                          * to pmap it.
 647                          */
 648                         if (vm_page_busied(fs.m)) {
 649                                 /*
 650                                  * Reference the page before unlocking and
 651                                  * sleeping so that the page daemon is less
 652                                  * likely to reclaim it.
 653                                  */
 654                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 655                                 if (fs.object != fs.first_object) {
 656                                         if (!VM_OBJECT_TRYWLOCK(
 657                                             fs.first_object)) {
 658                                                 VM_OBJECT_WUNLOCK(fs.object);
 659                                                 VM_OBJECT_WLOCK(fs.first_object);
 660                                                 VM_OBJECT_WLOCK(fs.object);
 661                                         }
 662                                         vm_page_lock(fs.first_m);
 663                                         vm_page_free(fs.first_m);
 664                                         vm_page_unlock(fs.first_m);
 665                                         vm_object_pip_wakeup(fs.first_object);
 666                                         VM_OBJECT_WUNLOCK(fs.first_object);
 667                                         fs.first_m = NULL;
 668                                 }
 669                                 unlock_map(&fs);
 670                                 if (fs.m == vm_page_lookup(fs.object,
 671                                     fs.pindex)) {
 672                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 673                                 }
 674                                 vm_object_pip_wakeup(fs.object);
 675                                 VM_OBJECT_WUNLOCK(fs.object);
 676                                 VM_CNT_INC(v_intrans);
 677                                 vm_object_deallocate(fs.first_object);
 678                                 goto RetryFault;
 679                         }
 680                         vm_page_lock(fs.m);
 681                         vm_page_remque(fs.m);
 682                         vm_page_unlock(fs.m);
 683
 684                         /*
 685                          * Mark page busy for other processes, and the
 686                          * pagedaemon.  If it still isn't completely valid
 687                          * (readable), jump to readrest, else break-out ( we
 688                          * found the page ).
 689                          */
 690                         vm_page_xbusy(fs.m);
 691                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 692                                 goto readrest;
 693                         break;
 694                 }
 695                 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
 696
 697                 /*
 698                  * Page is not resident.  If the pager might contain the page
 699                  * or this is the beginning of the search, allocate a new
 700                  * page.  (Default objects are zero-fill, so there is no real
 701                  * pager for them.)
 702                  */
 703                 if (fs.object->type != OBJT_DEFAULT ||
 704                     fs.object == fs.first_object) {
 705                         if (fs.pindex >= fs.object->size) {
 706                                 unlock_and_deallocate(&fs);
 707                                 return (KERN_PROTECTION_FAILURE);
 708                         }
 709
 710                         if (fs.object == fs.first_object &&
 711                             (fs.first_object->flags & OBJ_POPULATE) != 0 &&
 712                             fs.first_object->shadow_count == 0) {
 713                                 rv = vm_fault_populate(&fs, vaddr, prot,
 714                                     fault_type, fault_flags, wired, m_hold);
 715                                 switch (rv) {
 716                                 case KERN_SUCCESS:
 717                                 case KERN_FAILURE:
 718                                         unlock_and_deallocate(&fs);
 719                                         return (rv);
 720                                 case KERN_RESOURCE_SHORTAGE:
 721                                         unlock_and_deallocate(&fs);
 722                                         goto RetryFault;
 723                                 case KERN_NOT_RECEIVER:
 724                                         /*
 725                                          * Pager's populate() method
 726                                          * returned VM_PAGER_BAD.
 727                                          */
 728                                         break;
 729                                 default:
 730                                         panic("inconsistent return codes");
 731                                 }
 732                         }
 733
 734                         /*
 735                          * Allocate a new page for this object/offset pair.
 736                          *
 737                          * Unlocked read of the p_flag is harmless. At
 738                          * worst, the P_KILLED might be not observed
 739                          * there, and allocation can fail, causing
 740                          * restart and new reading of the p_flag.
 741                          */
 742                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 743 #if VM_NRESERVLEVEL > 0
 744                                 vm_object_color(fs.object, atop(vaddr) -
 745                                     fs.pindex);
 746 #endif
 747                                 alloc_req = P_KILLED(curproc) ?
 748                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 749                                 if (fs.object->type != OBJT_VNODE &&
 750                                     fs.object->backing_object == NULL)
 751                                         alloc_req |= VM_ALLOC_ZERO;
 752                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 753                                     alloc_req);
 754                         }
 755                         if (fs.m == NULL) {
 756                                 unlock_and_deallocate(&fs);
 757                                 VM_WAITPFAULT;
 758                                 goto RetryFault;
 759                         }
 760                 }
 761
 762 readrest:
 763                 /*
 764                  * At this point, we have either allocated a new page or found
 765                  * an existing page that is only partially valid.
 766                  *
 767                  * We hold a reference on the current object and the page is
 768                  * exclusive busied.
 769                  */
 770
 771                 /*
 772                  * If the pager for the current object might have the page,
 773                  * then determine the number of additional pages to read and
 774                  * potentially reprioritize previously read pages for earlier
 775                  * reclamation.  These operations should only be performed
 776                  * once per page fault.  Even if the current pager doesn't
 777                  * have the page, the number of additional pages to read will
 778                  * apply to subsequent objects in the shadow chain.
 779                  */
 780                 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
 781                     !P_KILLED(curproc)) {
 782                         KASSERT(fs.lookup_still_valid, ("map unlocked"));
 783                         era = fs.entry->read_ahead;
 784                         behavior = vm_map_entry_behavior(fs.entry);
 785                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 786                                 nera = 0;
 787                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 788                                 nera = VM_FAULT_READ_AHEAD_MAX;
 789                                 if (vaddr == fs.entry->next_read)
 790                                         vm_fault_dontneed(&fs, vaddr, nera);
 791                         } else if (vaddr == fs.entry->next_read) {
 792                                 /*
 793                                  * This is a sequential fault.  Arithmetically
 794                                  * increase the requested number of pages in
 795                                  * the read-ahead window.  The requested
 796                                  * number of pages is "# of sequential faults
 797                                  * x (read ahead min + 1) + read ahead min"
 798                                  */
 799                                 nera = VM_FAULT_READ_AHEAD_MIN;
 800                                 if (era > 0) {
 801                                         nera += era + 1;
 802                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 803                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 804                                 }
 805                                 if (era == VM_FAULT_READ_AHEAD_MAX)
 806                                         vm_fault_dontneed(&fs, vaddr, nera);
 807                         } else {
 808                                 /*
 809                                  * This is a non-sequential fault.
 810                                  */
 811                                 nera = 0;
 812                         }
 813                         if (era != nera) {
 814                                 /*
 815                                  * A read lock on the map suffices to update
 816                                  * the read ahead count safely.
 817                                  */
 818                                 fs.entry->read_ahead = nera;
 819                         }
 820
 821                         /*
 822                          * Prepare for unlocking the map.  Save the map
 823                          * entry's start and end addresses, which are used to
 824                          * optimize the size of the pager operation below.
 825                          * Even if the map entry's addresses change after
 826                          * unlocking the map, using the saved addresses is
 827                          * safe.
 828                          */
 829                         e_start = fs.entry->start;
 830                         e_end = fs.entry->end;
 831                 }
 832
 833                 /*
 834                  * Call the pager to retrieve the page if there is a chance
 835                  * that the pager has it, and potentially retrieve additional
 836                  * pages at the same time.
 837                  */
 838                 if (fs.object->type != OBJT_DEFAULT) {
 839                         /*
 840                          * Release the map lock before locking the vnode or
 841                          * sleeping in the pager.  (If the current object has
 842                          * a shadow, then an earlier iteration of this loop
 843                          * may have already unlocked the map.)
 844                          */
 845                         unlock_map(&fs);
 846
 847                         if (fs.object->type == OBJT_VNODE &&
 848                             (vp = fs.object->handle) != fs.vp) {
 849                                 /*
 850                                  * Perform an unlock in case the desired vnode
 851                                  * changed while the map was unlocked during a
 852                                  * retry.
 853                                  */
 854                                 unlock_vp(&fs);
 855
 856                                 locked = VOP_ISLOCKED(vp);
 857                                 if (locked != LK_EXCLUSIVE)
 858                                         locked = LK_SHARED;
 859
 860                                 /*
 861                                  * We must not sleep acquiring the vnode lock
 862                                  * while we have the page exclusive busied or
 863                                  * the object's paging-in-progress count
 864                                  * incremented.  Otherwise, we could deadlock.
 865                                  */
 866                                 error = vget(vp, locked | LK_CANRECURSE |
 867                                     LK_NOWAIT, curthread);
 868                                 if (error != 0) {
 869                                         vhold(vp);
 870                                         release_page(&fs);
 871                                         unlock_and_deallocate(&fs);
 872                                         error = vget(vp, locked | LK_RETRY |
 873                                             LK_CANRECURSE, curthread);
 874                                         vdrop(vp);
 875                                         fs.vp = vp;
 876                                         KASSERT(error == 0,
 877                                             ("vm_fault: vget failed"));
 878                                         goto RetryFault;
 879                                 }
 880                                 fs.vp = vp;
 881                         }
 882                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 883                             ("vm_fault: vnode-backed object mapped by system map"));
 884
 885                         /*
 886                          * Page in the requested page and hint the pager,
 887                          * that it may bring up surrounding pages.
 888                          */
 889                         if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
 890                             P_KILLED(curproc)) {
 891                                 behind = 0;
 892                                 ahead = 0;
 893                         } else {
 894                                 /* Is this a sequential fault? */
 895                                 if (nera > 0) {
 896                                         behind = 0;
 897                                         ahead = nera;
 898                                 } else {
 899                                         /*
 900                                          * Request a cluster of pages that is
 901                                          * aligned to a VM_FAULT_READ_DEFAULT
 902                                          * page offset boundary within the
 903                                          * object.  Alignment to a page offset
 904                                          * boundary is more likely to coincide
 905                                          * with the underlying file system
 906                                          * block than alignment to a virtual
 907                                          * address boundary.
 908                                          */
 909                                         cluster_offset = fs.pindex %
 910                                             VM_FAULT_READ_DEFAULT;
 911                                         behind = ulmin(cluster_offset,
 912                                             atop(vaddr - e_start));
 913                                         ahead = VM_FAULT_READ_DEFAULT - 1 -
 914                                             cluster_offset;
 915                                 }
 916                                 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
 917                         }
 918                         rv = vm_pager_get_pages(fs.object, &fs.m, 1,
 919                             &behind, &ahead);
 920                         if (rv == VM_PAGER_OK) {
 921                                 faultcount = behind + 1 + ahead;
 922                                 hardfault = true;
 923                                 break; /* break to PAGE HAS BEEN FOUND */
 924                         }
 925                         if (rv == VM_PAGER_ERROR)
 926                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 927                                     curproc->p_pid, curproc->p_comm);
 928
 929                         /*
 930                          * If an I/O error occurred or the requested page was
 931                          * outside the range of the pager, clean up and return
 932                          * an error.
 933                          */
 934                         if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
 935                                 vm_page_lock(fs.m);
 936                                 if (fs.m->wire_count == 0)
 937                                         vm_page_free(fs.m);
 938                                 else
 939                                         vm_page_xunbusy_maybelocked(fs.m);
 940                                 vm_page_unlock(fs.m);
 941                                 fs.m = NULL;
 942                                 unlock_and_deallocate(&fs);
 943                                 return (rv == VM_PAGER_ERROR ? KERN_FAILURE :
 944                                     KERN_PROTECTION_FAILURE);
 945                         }
 946
 947                         /*
 948                          * The requested page does not exist at this object/
 949                          * offset.  Remove the invalid page from the object,
 950                          * waking up anyone waiting for it, and continue on to
 951                          * the next object.  However, if this is the top-level
 952                          * object, we must leave the busy page in place to
 953                          * prevent another process from rushing past us, and
 954                          * inserting the page in that object at the same time
 955                          * that we are.
 956                          */
 957                         if (fs.object != fs.first_object) {
 958                                 vm_page_lock(fs.m);
 959                                 if (fs.m->wire_count == 0)
 960                                         vm_page_free(fs.m);
 961                                 else
 962                                         vm_page_xunbusy_maybelocked(fs.m);
 963                                 vm_page_unlock(fs.m);
 964                                 fs.m = NULL;
 965                         }
 966                 }
 967
 968                 /*
 969                  * We get here if the object has default pager (or unwiring)
 970                  * or the pager doesn't have the page.
 971                  */
 972                 if (fs.object == fs.first_object)
 973                         fs.first_m = fs.m;
 974
 975                 /*
 976                  * Move on to the next object.  Lock the next object before
 977                  * unlocking the current one.
 978                  */
 979                 next_object = fs.object->backing_object;
 980                 if (next_object == NULL) {
 981                         /*
 982                          * If there's no object left, fill the page in the top
 983                          * object with zeros.
 984                          */
 985                         if (fs.object != fs.first_object) {
 986                                 vm_object_pip_wakeup(fs.object);
 987                                 VM_OBJECT_WUNLOCK(fs.object);
 988
 989                                 fs.object = fs.first_object;
 990                                 fs.pindex = fs.first_pindex;
 991                                 fs.m = fs.first_m;
 992                                 VM_OBJECT_WLOCK(fs.object);
 993                         }
 994                         fs.first_m = NULL;
 995
 996                         /*
 997                          * Zero the page if necessary and mark it valid.
 998                          */
 999                         if ((fs.m->flags & PG_ZERO) == 0) {
1000                                 pmap_zero_page(fs.m);
1001                         } else {
1002                                 VM_CNT_INC(v_ozfod);
1003                         }
1004                         VM_CNT_INC(v_zfod);
1005                         fs.m->valid = VM_PAGE_BITS_ALL;
1006                         /* Don't try to prefault neighboring pages. */
1007                         faultcount = 1;
1008                         break;  /* break to PAGE HAS BEEN FOUND */
1009                 } else {
1010                         KASSERT(fs.object != next_object,
1011                             ("object loop %p", next_object));
1012                         VM_OBJECT_WLOCK(next_object);
1013                         vm_object_pip_add(next_object, 1);
1014                         if (fs.object != fs.first_object)
1015                                 vm_object_pip_wakeup(fs.object);
1016                         fs.pindex +=
1017                             OFF_TO_IDX(fs.object->backing_object_offset);
1018                         VM_OBJECT_WUNLOCK(fs.object);
1019                         fs.object = next_object;
1020                 }
1021         }
1022
1023         vm_page_assert_xbusied(fs.m);
1024
1025         /*
1026          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1027          * is held.]
1028          */
1029
1030         /*
1031          * If the page is being written, but isn't already owned by the
1032          * top-level object, we have to copy it into a new page owned by the
1033          * top-level object.
1034          */
1035         if (fs.object != fs.first_object) {
1036                 /*
1037                  * We only really need to copy if we want to write it.
1038                  */
1039                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1040                         /*
1041                          * This allows pages to be virtually copied from a
1042                          * backing_object into the first_object, where the
1043                          * backing object has no other refs to it, and cannot
1044                          * gain any more refs.  Instead of a bcopy, we just
1045                          * move the page from the backing object to the
1046                          * first object.  Note that we must mark the page
1047                          * dirty in the first object so that it will go out
1048                          * to swap when needed.
1049                          */
1050                         is_first_object_locked = false;
1051                         if (
1052                                 /*
1053                                  * Only one shadow object
1054                                  */
1055                                 (fs.object->shadow_count == 1) &&
1056                                 /*
1057                                  * No COW refs, except us
1058                                  */
1059                                 (fs.object->ref_count == 1) &&
1060                                 /*
1061                                  * No one else can look this object up
1062                                  */
1063                                 (fs.object->handle == NULL) &&
1064                                 /*
1065                                  * No other ways to look the object up
1066                                  */
1067                                 ((fs.object->type == OBJT_DEFAULT) ||
1068                                  (fs.object->type == OBJT_SWAP)) &&
1069                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1070                                 /*
1071                                  * We don't chase down the shadow chain
1072                                  */
1073                             fs.object == fs.first_object->backing_object) {
1074                                 vm_page_lock(fs.m);
1075                                 vm_page_remove(fs.m);
1076                                 vm_page_unlock(fs.m);
1077                                 vm_page_lock(fs.first_m);
1078                                 vm_page_replace_checked(fs.m, fs.first_object,
1079                                     fs.first_pindex, fs.first_m);
1080                                 vm_page_free(fs.first_m);
1081                                 vm_page_unlock(fs.first_m);
1082                                 vm_page_dirty(fs.m);
1083 #if VM_NRESERVLEVEL > 0
1084                                 /*
1085                                  * Rename the reservation.
1086                                  */
1087                                 vm_reserv_rename(fs.m, fs.first_object,
1088                                     fs.object, OFF_TO_IDX(
1089                                     fs.first_object->backing_object_offset));
1090 #endif
1091                                 /*
1092                                  * Removing the page from the backing object
1093                                  * unbusied it.
1094                                  */
1095                                 vm_page_xbusy(fs.m);
1096                                 fs.first_m = fs.m;
1097                                 fs.m = NULL;
1098                                 VM_CNT_INC(v_cow_optim);
1099                         } else {
1100                                 /*
1101                                  * Oh, well, lets copy it.
1102                                  */
1103                                 pmap_copy_page(fs.m, fs.first_m);
1104                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
1105                                 if (wired && (fault_flags &
1106                                     VM_FAULT_WIRE) == 0) {
1107                                         vm_page_lock(fs.first_m);
1108                                         vm_page_wire(fs.first_m);
1109                                         vm_page_unlock(fs.first_m);
1110
1111                                         vm_page_lock(fs.m);
1112                                         vm_page_unwire(fs.m, PQ_INACTIVE);
1113                                         vm_page_unlock(fs.m);
1114                                 }
1115                                 /*
1116                                  * We no longer need the old page or object.
1117                                  */
1118                                 release_page(&fs);
1119                         }
1120                         /*
1121                          * fs.object != fs.first_object due to above
1122                          * conditional
1123                          */
1124                         vm_object_pip_wakeup(fs.object);
1125                         VM_OBJECT_WUNLOCK(fs.object);
1126                         /*
1127                          * Only use the new page below...
1128                          */
1129                         fs.object = fs.first_object;
1130                         fs.pindex = fs.first_pindex;
1131                         fs.m = fs.first_m;
1132                         if (!is_first_object_locked)
1133                                 VM_OBJECT_WLOCK(fs.object);
1134                         VM_CNT_INC(v_cow_faults);
1135                         curthread->td_cow++;
1136                 } else {
1137                         prot &= ~VM_PROT_WRITE;
1138                 }
1139         }
1140
1141         /*
1142          * We must verify that the maps have not changed since our last
1143          * lookup.
1144          */
1145         if (!fs.lookup_still_valid) {
1146                 if (!vm_map_trylock_read(fs.map)) {
1147                         release_page(&fs);
1148                         unlock_and_deallocate(&fs);
1149                         goto RetryFault;
1150                 }
1151                 fs.lookup_still_valid = true;
1152                 if (fs.map->timestamp != fs.map_generation) {
1153                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1154                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1155
1156                         /*
1157                          * If we don't need the page any longer, put it on the inactive
1158                          * list (the easiest thing to do here).  If no one needs it,
1159                          * pageout will grab it eventually.
1160                          */
1161                         if (result != KERN_SUCCESS) {
1162                                 release_page(&fs);
1163                                 unlock_and_deallocate(&fs);
1164
1165                                 /*
1166                                  * If retry of map lookup would have blocked then
1167                                  * retry fault from start.
1168                                  */
1169                                 if (result == KERN_FAILURE)
1170                                         goto RetryFault;
1171                                 return (result);
1172                         }
1173                         if ((retry_object != fs.first_object) ||
1174                             (retry_pindex != fs.first_pindex)) {
1175                                 release_page(&fs);
1176                                 unlock_and_deallocate(&fs);
1177                                 goto RetryFault;
1178                         }
1179
1180                         /*
1181                          * Check whether the protection has changed or the object has
1182                          * been copied while we left the map unlocked. Changing from
1183                          * read to write permission is OK - we leave the page
1184                          * write-protected, and catch the write fault. Changing from
1185                          * write to read permission means that we can't mark the page
1186                          * write-enabled after all.
1187                          */
1188                         prot &= retry_prot;
1189                 }
1190         }
1191
1192         /*
1193          * If the page was filled by a pager, save the virtual address that
1194          * should be faulted on next under a sequential access pattern to the
1195          * map entry.  A read lock on the map suffices to update this address
1196          * safely.
1197          */
1198         if (hardfault)
1199                 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1200
1201         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1202         vm_page_assert_xbusied(fs.m);
1203
1204         /*
1205          * Page must be completely valid or it is not fit to
1206          * map into user space.  vm_pager_get_pages() ensures this.
1207          */
1208         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
1209             ("vm_fault: page %p partially invalid", fs.m));
1210         VM_OBJECT_WUNLOCK(fs.object);
1211
1212         /*
1213          * Put this page into the physical map.  We had to do the unlock above
1214          * because pmap_enter() may sleep.  We don't put the page
1215          * back on the active queue until later so that the pageout daemon
1216          * won't find it (yet).
1217          */
1218         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1219             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1220         if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1221             wired == 0)
1222                 vm_fault_prefault(&fs, vaddr,
1223                     faultcount > 0 ? behind : PFBAK,
1224                     faultcount > 0 ? ahead : PFFOR);
1225         VM_OBJECT_WLOCK(fs.object);
1226         vm_page_lock(fs.m);
1227
1228         /*
1229          * If the page is not wired down, then put it where the pageout daemon
1230          * can find it.
1231          */
1232         if ((fault_flags & VM_FAULT_WIRE) != 0) {
1233                 KASSERT(wired, ("VM_FAULT_WIRE && !wired"));
1234                 vm_page_wire(fs.m);
1235         } else
1236                 vm_page_activate(fs.m);
1237         if (m_hold != NULL) {
1238                 *m_hold = fs.m;
1239                 vm_page_hold(fs.m);
1240         }
1241         vm_page_unlock(fs.m);
1242         vm_page_xunbusy(fs.m);
1243
1244         /*
1245          * Unlock everything, and return
1246          */
1247         unlock_and_deallocate(&fs);
1248         if (hardfault) {
1249                 VM_CNT_INC(v_io_faults);
1250                 curthread->td_ru.ru_majflt++;
1251 #ifdef RACCT
1252                 if (racct_enable && fs.object->type == OBJT_VNODE) {
1253                         PROC_LOCK(curproc);
1254                         if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1255                                 racct_add_force(curproc, RACCT_WRITEBPS,
1256                                     PAGE_SIZE + behind * PAGE_SIZE);
1257                                 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1258                         } else {
1259                                 racct_add_force(curproc, RACCT_READBPS,
1260                                     PAGE_SIZE + ahead * PAGE_SIZE);
1261                                 racct_add_force(curproc, RACCT_READIOPS, 1);
1262                         }
1263                         PROC_UNLOCK(curproc);
1264                 }
1265 #endif
1266         } else
1267                 curthread->td_ru.ru_minflt++;
1268
1269         return (KERN_SUCCESS);
1270 }
1271
1272 /*
1273  * Speed up the reclamation of pages that precede the faulting pindex within
1274  * the first object of the shadow chain.  Essentially, perform the equivalent
1275  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1276  * the faulting pindex by the cluster size when the pages read by vm_fault()
1277  * cross a cluster-size boundary.  The cluster size is the greater of the
1278  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1279  *
1280  * When "fs->first_object" is a shadow object, the pages in the backing object
1281  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1282  * function must only be concerned with pages in the first object.
1283  */
1284 static void
1285 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1286 {
1287         vm_map_entry_t entry;
1288         vm_object_t first_object, object;
1289         vm_offset_t end, start;
1290         vm_page_t m, m_next;
1291         vm_pindex_t pend, pstart;
1292         vm_size_t size;
1293
1294         object = fs->object;
1295         VM_OBJECT_ASSERT_WLOCKED(object);
1296         first_object = fs->first_object;
1297         if (first_object != object) {
1298                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1299                         VM_OBJECT_WUNLOCK(object);
1300                         VM_OBJECT_WLOCK(first_object);
1301                         VM_OBJECT_WLOCK(object);
1302                 }
1303         }
1304         /* Neither fictitious nor unmanaged pages can be reclaimed. */
1305         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1306                 size = VM_FAULT_DONTNEED_MIN;
1307                 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1308                         size = pagesizes[1];
1309                 end = rounddown2(vaddr, size);
1310                 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1311                     (entry = fs->entry)->start < end) {
1312                         if (end - entry->start < size)
1313                                 start = entry->start;
1314                         else
1315                                 start = end - size;
1316                         pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1317                         pstart = OFF_TO_IDX(entry->offset) + atop(start -
1318                             entry->start);
1319                         m_next = vm_page_find_least(first_object, pstart);
1320                         pend = OFF_TO_IDX(entry->offset) + atop(end -
1321                             entry->start);
1322                         while ((m = m_next) != NULL && m->pindex < pend) {
1323                                 m_next = TAILQ_NEXT(m, listq);
1324                                 if (m->valid != VM_PAGE_BITS_ALL ||
1325                                     vm_page_busied(m))
1326                                         continue;
1327
1328                                 /*
1329                                  * Don't clear PGA_REFERENCED, since it would
1330                                  * likely represent a reference by a different
1331                                  * process.
1332                                  *
1333                                  * Typically, at this point, prefetched pages
1334                                  * are still in the inactive queue.  Only
1335                                  * pages that triggered page faults are in the
1336                                  * active queue.
1337                                  */
1338                                 vm_page_lock(m);
1339                                 vm_page_deactivate(m);
1340                                 vm_page_unlock(m);
1341                         }
1342                 }
1343         }
1344         if (first_object != object)
1345                 VM_OBJECT_WUNLOCK(first_object);
1346 }
1347
1348 /*
1349  * vm_fault_prefault provides a quick way of clustering
1350  * pagefaults into a processes address space.  It is a "cousin"
1351  * of vm_map_pmap_enter, except it runs at page fault time instead
1352  * of mmap time.
1353  */
1354 static void
1355 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1356     int backward, int forward)
1357 {
1358         pmap_t pmap;
1359         vm_map_entry_t entry;
1360         vm_object_t backing_object, lobject;
1361         vm_offset_t addr, starta;
1362         vm_pindex_t pindex;
1363         vm_page_t m;
1364         int i;
1365
1366         pmap = fs->map->pmap;
1367         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1368                 return;
1369
1370         entry = fs->entry;
1371
1372         if (addra < backward * PAGE_SIZE) {
1373                 starta = entry->start;
1374         } else {
1375                 starta = addra - backward * PAGE_SIZE;
1376                 if (starta < entry->start)
1377                         starta = entry->start;
1378         }
1379
1380         /*
1381          * Generate the sequence of virtual addresses that are candidates for
1382          * prefaulting in an outward spiral from the faulting virtual address,
1383          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1384          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1385          * If the candidate address doesn't have a backing physical page, then
1386          * the loop immediately terminates.
1387          */
1388         for (i = 0; i < 2 * imax(backward, forward); i++) {
1389                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1390                     PAGE_SIZE);
1391                 if (addr > addra + forward * PAGE_SIZE)
1392                         addr = 0;
1393
1394                 if (addr < starta || addr >= entry->end)
1395                         continue;
1396
1397                 if (!pmap_is_prefaultable(pmap, addr))
1398                         continue;
1399
1400                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1401                 lobject = entry->object.vm_object;
1402                 VM_OBJECT_RLOCK(lobject);
1403                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1404                     lobject->type == OBJT_DEFAULT &&
1405                     (backing_object = lobject->backing_object) != NULL) {
1406                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1407                             0, ("vm_fault_prefault: unaligned object offset"));
1408                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1409                         VM_OBJECT_RLOCK(backing_object);
1410                         VM_OBJECT_RUNLOCK(lobject);
1411                         lobject = backing_object;
1412                 }
1413                 if (m == NULL) {
1414                         VM_OBJECT_RUNLOCK(lobject);
1415                         break;
1416                 }
1417                 if (m->valid == VM_PAGE_BITS_ALL &&
1418                     (m->flags & PG_FICTITIOUS) == 0)
1419                         pmap_enter_quick(pmap, addr, m, entry->protection);
1420                 VM_OBJECT_RUNLOCK(lobject);
1421         }
1422 }
1423
1424 /*
1425  * Hold each of the physical pages that are mapped by the specified range of
1426  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1427  * and allow the specified types of access, "prot".  If all of the implied
1428  * pages are successfully held, then the number of held pages is returned
1429  * together with pointers to those pages in the array "ma".  However, if any
1430  * of the pages cannot be held, -1 is returned.
1431  */
1432 int
1433 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1434     vm_prot_t prot, vm_page_t *ma, int max_count)
1435 {
1436         vm_offset_t end, va;
1437         vm_page_t *mp;
1438         int count;
1439         boolean_t pmap_failed;
1440
1441         if (len == 0)
1442                 return (0);
1443         end = round_page(addr + len);
1444         addr = trunc_page(addr);
1445
1446         /*
1447          * Check for illegal addresses.
1448          */
1449         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1450                 return (-1);
1451
1452         if (atop(end - addr) > max_count)
1453                 panic("vm_fault_quick_hold_pages: count > max_count");
1454         count = atop(end - addr);
1455
1456         /*
1457          * Most likely, the physical pages are resident in the pmap, so it is
1458          * faster to try pmap_extract_and_hold() first.
1459          */
1460         pmap_failed = FALSE;
1461         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1462                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1463                 if (*mp == NULL)
1464                         pmap_failed = TRUE;
1465                 else if ((prot & VM_PROT_WRITE) != 0 &&
1466                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1467                         /*
1468                          * Explicitly dirty the physical page.  Otherwise, the
1469                          * caller's changes may go unnoticed because they are
1470                          * performed through an unmanaged mapping or by a DMA
1471                          * operation.
1472                          *
1473                          * The object lock is not held here.
1474                          * See vm_page_clear_dirty_mask().
1475                          */
1476                         vm_page_dirty(*mp);
1477                 }
1478         }
1479         if (pmap_failed) {
1480                 /*
1481                  * One or more pages could not be held by the pmap.  Either no
1482                  * page was mapped at the specified virtual address or that
1483                  * mapping had insufficient permissions.  Attempt to fault in
1484                  * and hold these pages.
1485                  */
1486                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1487                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1488                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1489                                 goto error;
1490         }
1491         return (count);
1492 error:
1493         for (mp = ma; mp < ma + count; mp++)
1494                 if (*mp != NULL) {
1495                         vm_page_lock(*mp);
1496                         vm_page_unhold(*mp);
1497                         vm_page_unlock(*mp);
1498                 }
1499         return (-1);
1500 }
1501
1502 /*
1503  *      Routine:
1504  *              vm_fault_copy_entry
1505  *      Function:
1506  *              Create new shadow object backing dst_entry with private copy of
1507  *              all underlying pages. When src_entry is equal to dst_entry,
1508  *              function implements COW for wired-down map entry. Otherwise,
1509  *              it forks wired entry into dst_map.
1510  *
1511  *      In/out conditions:
1512  *              The source and destination maps must be locked for write.
1513  *              The source map entry must be wired down (or be a sharing map
1514  *              entry corresponding to a main map entry that is wired down).
1515  */
1516 void
1517 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1518     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1519     vm_ooffset_t *fork_charge)
1520 {
1521         vm_object_t backing_object, dst_object, object, src_object;
1522         vm_pindex_t dst_pindex, pindex, src_pindex;
1523         vm_prot_t access, prot;
1524         vm_offset_t vaddr;
1525         vm_page_t dst_m;
1526         vm_page_t src_m;
1527         boolean_t upgrade;
1528
1529 #ifdef  lint
1530         src_map++;
1531 #endif  /* lint */
1532
1533         upgrade = src_entry == dst_entry;
1534         access = prot = dst_entry->protection;
1535
1536         src_object = src_entry->object.vm_object;
1537         src_pindex = OFF_TO_IDX(src_entry->offset);
1538
1539         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1540                 dst_object = src_object;
1541                 vm_object_reference(dst_object);
1542         } else {
1543                 /*
1544                  * Create the top-level object for the destination entry. (Doesn't
1545                  * actually shadow anything - we copy the pages directly.)
1546                  */
1547                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1548                     atop(dst_entry->end - dst_entry->start));
1549 #if VM_NRESERVLEVEL > 0
1550                 dst_object->flags |= OBJ_COLORED;
1551                 dst_object->pg_color = atop(dst_entry->start);
1552 #endif
1553         }
1554
1555         VM_OBJECT_WLOCK(dst_object);
1556         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1557             ("vm_fault_copy_entry: vm_object not NULL"));
1558         if (src_object != dst_object) {
1559                 dst_entry->object.vm_object = dst_object;
1560                 dst_entry->offset = 0;
1561                 dst_object->charge = dst_entry->end - dst_entry->start;
1562         }
1563         if (fork_charge != NULL) {
1564                 KASSERT(dst_entry->cred == NULL,
1565                     ("vm_fault_copy_entry: leaked swp charge"));
1566                 dst_object->cred = curthread->td_ucred;
1567                 crhold(dst_object->cred);
1568                 *fork_charge += dst_object->charge;
1569         } else if (dst_object->cred == NULL) {
1570                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1571                     dst_entry));
1572                 dst_object->cred = dst_entry->cred;
1573                 dst_entry->cred = NULL;
1574         }
1575
1576         /*
1577          * If not an upgrade, then enter the mappings in the pmap as
1578          * read and/or execute accesses.  Otherwise, enter them as
1579          * write accesses.
1580          *
1581          * A writeable large page mapping is only created if all of
1582          * the constituent small page mappings are modified. Marking
1583          * PTEs as modified on inception allows promotion to happen
1584          * without taking potentially large number of soft faults.
1585          */
1586         if (!upgrade)
1587                 access &= ~VM_PROT_WRITE;
1588
1589         /*
1590          * Loop through all of the virtual pages within the entry's
1591          * range, copying each page from the source object to the
1592          * destination object.  Since the source is wired, those pages
1593          * must exist.  In contrast, the destination is pageable.
1594          * Since the destination object does share any backing storage
1595          * with the source object, all of its pages must be dirtied,
1596          * regardless of whether they can be written.
1597          */
1598         for (vaddr = dst_entry->start, dst_pindex = 0;
1599             vaddr < dst_entry->end;
1600             vaddr += PAGE_SIZE, dst_pindex++) {
1601 again:
1602                 /*
1603                  * Find the page in the source object, and copy it in.
1604                  * Because the source is wired down, the page will be
1605                  * in memory.
1606                  */
1607                 if (src_object != dst_object)
1608                         VM_OBJECT_RLOCK(src_object);
1609                 object = src_object;
1610                 pindex = src_pindex + dst_pindex;
1611                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1612                     (backing_object = object->backing_object) != NULL) {
1613                         /*
1614                          * Unless the source mapping is read-only or
1615                          * it is presently being upgraded from
1616                          * read-only, the first object in the shadow
1617                          * chain should provide all of the pages.  In
1618                          * other words, this loop body should never be
1619                          * executed when the source mapping is already
1620                          * read/write.
1621                          */
1622                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1623                             upgrade,
1624                             ("vm_fault_copy_entry: main object missing page"));
1625
1626                         VM_OBJECT_RLOCK(backing_object);
1627                         pindex += OFF_TO_IDX(object->backing_object_offset);
1628                         if (object != dst_object)
1629                                 VM_OBJECT_RUNLOCK(object);
1630                         object = backing_object;
1631                 }
1632                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1633
1634                 if (object != dst_object) {
1635                         /*
1636                          * Allocate a page in the destination object.
1637                          */
1638                         dst_m = vm_page_alloc(dst_object, (src_object ==
1639                             dst_object ? src_pindex : 0) + dst_pindex,
1640                             VM_ALLOC_NORMAL);
1641                         if (dst_m == NULL) {
1642                                 VM_OBJECT_WUNLOCK(dst_object);
1643                                 VM_OBJECT_RUNLOCK(object);
1644                                 VM_WAIT;
1645                                 VM_OBJECT_WLOCK(dst_object);
1646                                 goto again;
1647                         }
1648                         pmap_copy_page(src_m, dst_m);
1649                         VM_OBJECT_RUNLOCK(object);
1650                         dst_m->valid = VM_PAGE_BITS_ALL;
1651                         dst_m->dirty = VM_PAGE_BITS_ALL;
1652                 } else {
1653                         dst_m = src_m;
1654                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1655                                 goto again;
1656                         vm_page_xbusy(dst_m);
1657                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1658                             ("invalid dst page %p", dst_m));
1659                 }
1660                 VM_OBJECT_WUNLOCK(dst_object);
1661
1662                 /*
1663                  * Enter it in the pmap. If a wired, copy-on-write
1664                  * mapping is being replaced by a write-enabled
1665                  * mapping, then wire that new mapping.
1666                  */
1667                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1668                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1669
1670                 /*
1671                  * Mark it no longer busy, and put it on the active list.
1672                  */
1673                 VM_OBJECT_WLOCK(dst_object);
1674
1675                 if (upgrade) {
1676                         if (src_m != dst_m) {
1677                                 vm_page_lock(src_m);
1678                                 vm_page_unwire(src_m, PQ_INACTIVE);
1679                                 vm_page_unlock(src_m);
1680                                 vm_page_lock(dst_m);
1681                                 vm_page_wire(dst_m);
1682                                 vm_page_unlock(dst_m);
1683                         } else {
1684                                 KASSERT(dst_m->wire_count > 0,
1685                                     ("dst_m %p is not wired", dst_m));
1686                         }
1687                 } else {
1688                         vm_page_lock(dst_m);
1689                         vm_page_activate(dst_m);
1690                         vm_page_unlock(dst_m);
1691                 }
1692                 vm_page_xunbusy(dst_m);
1693         }
1694         VM_OBJECT_WUNLOCK(dst_object);
1695         if (upgrade) {
1696                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1697                 vm_object_deallocate(src_object);
1698         }
1699 }
1700
1701 /*
1702  * Block entry into the machine-independent layer's page fault handler by
1703  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1704  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1705  * spurious page faults.
1706  */
1707 int
1708 vm_fault_disable_pagefaults(void)
1709 {
1710
1711         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1712 }
1713
1714 void
1715 vm_fault_enable_pagefaults(int save)
1716 {
1717
1718         curthread_pflags_restore(save);
1719 }