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