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/proc.h>
  85 #include <sys/resourcevar.h>
  86 #include <sys/rwlock.h>
  87 #include <sys/sysctl.h>
  88 #include <sys/vmmeter.h>
  89 #include <sys/vnode.h>
  90 #ifdef KTRACE
  91 #include <sys/ktrace.h>
  92 #endif
  93
  94 #include <vm/vm.h>
  95 #include <vm/vm_param.h>
  96 #include <vm/pmap.h>
  97 #include <vm/vm_map.h>
  98 #include <vm/vm_object.h>
  99 #include <vm/vm_page.h>
 100 #include <vm/vm_pageout.h>
 101 #include <vm/vm_kern.h>
 102 #include <vm/vm_pager.h>
 103 #include <vm/vm_extern.h>
 104 #include <vm/vm_reserv.h>
 105
 106 #define PFBAK 4
 107 #define PFFOR 4
 108
 109 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 110
 111 #define VM_FAULT_READ_BEHIND    8
 112 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 113 #define VM_FAULT_NINCR          (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
 114 #define VM_FAULT_SUM            (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
 115 #define VM_FAULT_CACHE_BEHIND   (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
 116
 117 struct faultstate {
 118         vm_page_t m;
 119         vm_object_t object;
 120         vm_pindex_t pindex;
 121         vm_page_t first_m;
 122         vm_object_t     first_object;
 123         vm_pindex_t first_pindex;
 124         vm_map_t map;
 125         vm_map_entry_t entry;
 126         int lookup_still_valid;
 127         struct vnode *vp;
 128 };
 129
 130 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
 131 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 132             int faultcount, int reqpage);
 133
 134 static inline void
 135 release_page(struct faultstate *fs)
 136 {
 137
 138         vm_page_xunbusy(fs->m);
 139         vm_page_lock(fs->m);
 140         vm_page_deactivate(fs->m);
 141         vm_page_unlock(fs->m);
 142         fs->m = NULL;
 143 }
 144
 145 static inline void
 146 unlock_map(struct faultstate *fs)
 147 {
 148
 149         if (fs->lookup_still_valid) {
 150                 vm_map_lookup_done(fs->map, fs->entry);
 151                 fs->lookup_still_valid = FALSE;
 152         }
 153 }
 154
 155 static void
 156 unlock_and_deallocate(struct faultstate *fs)
 157 {
 158
 159         vm_object_pip_wakeup(fs->object);
 160         VM_OBJECT_WUNLOCK(fs->object);
 161         if (fs->object != fs->first_object) {
 162                 VM_OBJECT_WLOCK(fs->first_object);
 163                 vm_page_lock(fs->first_m);
 164                 vm_page_free(fs->first_m);
 165                 vm_page_unlock(fs->first_m);
 166                 vm_object_pip_wakeup(fs->first_object);
 167                 VM_OBJECT_WUNLOCK(fs->first_object);
 168                 fs->first_m = NULL;
 169         }
 170         vm_object_deallocate(fs->first_object);
 171         unlock_map(fs);
 172         if (fs->vp != NULL) {
 173                 vput(fs->vp);
 174                 fs->vp = NULL;
 175         }
 176 }
 177
 178 static void
 179 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 180     vm_prot_t fault_type, int fault_flags, boolean_t set_wd)
 181 {
 182         boolean_t need_dirty;
 183
 184         if (((prot & VM_PROT_WRITE) == 0 &&
 185             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 186             (m->oflags & VPO_UNMANAGED) != 0)
 187                 return;
 188
 189         VM_OBJECT_ASSERT_LOCKED(m->object);
 190
 191         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 192             (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 193             (fault_flags & VM_FAULT_DIRTY) != 0;
 194
 195         if (set_wd)
 196                 vm_object_set_writeable_dirty(m->object);
 197         else
 198                 /*
 199                  * If two callers of vm_fault_dirty() with set_wd ==
 200                  * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
 201                  * flag set, other with flag clear, race, it is
 202                  * possible for the no-NOSYNC thread to see m->dirty
 203                  * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
 204                  * around manipulation of VPO_NOSYNC and
 205                  * vm_page_dirty() call, to avoid the race and keep
 206                  * m->oflags consistent.
 207                  */
 208                 vm_page_lock(m);
 209
 210         /*
 211          * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 212          * if the page is already dirty to prevent data written with
 213          * the expectation of being synced from not being synced.
 214          * Likewise if this entry does not request NOSYNC then make
 215          * sure the page isn't marked NOSYNC.  Applications sharing
 216          * data should use the same flags to avoid ping ponging.
 217          */
 218         if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
 219                 if (m->dirty == 0) {
 220                         m->oflags |= VPO_NOSYNC;
 221                 }
 222         } else {
 223                 m->oflags &= ~VPO_NOSYNC;
 224         }
 225
 226         /*
 227          * If the fault is a write, we know that this page is being
 228          * written NOW so dirty it explicitly to save on
 229          * pmap_is_modified() calls later.
 230          *
 231          * Also tell the backing pager, if any, that it should remove
 232          * any swap backing since the page is now dirty.
 233          */
 234         if (need_dirty)
 235                 vm_page_dirty(m);
 236         if (!set_wd)
 237                 vm_page_unlock(m);
 238         if (need_dirty)
 239                 vm_pager_page_unswapped(m);
 240 }
 241
 242 /*
 243  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 244  *            current object *might* contain the page.
 245  *
 246  *            default objects are zero-fill, there is no real pager.
 247  */
 248 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 249                         ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
 250
 251 /*
 252  *      vm_fault:
 253  *
 254  *      Handle a page fault occurring at the given address,
 255  *      requiring the given permissions, in the map specified.
 256  *      If successful, the page is inserted into the
 257  *      associated physical map.
 258  *
 259  *      NOTE: the given address should be truncated to the
 260  *      proper page address.
 261  *
 262  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 263  *      a standard error specifying why the fault is fatal is returned.
 264  *
 265  *      The map in question must be referenced, and remains so.
 266  *      Caller may hold no locks.
 267  */
 268 int
 269 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 270     int fault_flags)
 271 {
 272         struct thread *td;
 273         int result;
 274
 275         td = curthread;
 276         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 277                 return (KERN_PROTECTION_FAILURE);
 278 #ifdef KTRACE
 279         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 280                 ktrfault(vaddr, fault_type);
 281 #endif
 282         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 283             NULL);
 284 #ifdef KTRACE
 285         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 286                 ktrfaultend(result);
 287 #endif
 288         return (result);
 289 }
 290
 291 int
 292 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 293     int fault_flags, vm_page_t *m_hold)
 294 {
 295         vm_prot_t prot;
 296         long ahead, behind;
 297         int alloc_req, era, faultcount, nera, reqpage, result;
 298         boolean_t growstack, is_first_object_locked, wired;
 299         int map_generation;
 300         vm_object_t next_object;
 301         vm_page_t marray[VM_FAULT_READ_MAX];
 302         int hardfault;
 303         struct faultstate fs;
 304         struct vnode *vp;
 305         vm_page_t m;
 306         int locked, error;
 307
 308         hardfault = 0;
 309         growstack = TRUE;
 310         PCPU_INC(cnt.v_vm_faults);
 311         fs.vp = NULL;
 312         faultcount = reqpage = 0;
 313
 314 RetryFault:;
 315
 316         /*
 317          * Find the backing store object and offset into it to begin the
 318          * search.
 319          */
 320         fs.map = map;
 321         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 322             &fs.first_object, &fs.first_pindex, &prot, &wired);
 323         if (result != KERN_SUCCESS) {
 324                 if (growstack && result == KERN_INVALID_ADDRESS &&
 325                     map != kernel_map) {
 326                         result = vm_map_growstack(curproc, vaddr);
 327                         if (result != KERN_SUCCESS)
 328                                 return (KERN_FAILURE);
 329                         growstack = FALSE;
 330                         goto RetryFault;
 331                 }
 332                 return (result);
 333         }
 334
 335         map_generation = fs.map->timestamp;
 336
 337         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 338                 panic("vm_fault: fault on nofault entry, addr: %lx",
 339                     (u_long)vaddr);
 340         }
 341
 342         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 343             fs.entry->wiring_thread != curthread) {
 344                 vm_map_unlock_read(fs.map);
 345                 vm_map_lock(fs.map);
 346                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 347                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 348                         if (fs.vp != NULL) {
 349                                 vput(fs.vp);
 350                                 fs.vp = NULL;
 351                         }
 352                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 353                         vm_map_unlock_and_wait(fs.map, 0);
 354                 } else
 355                         vm_map_unlock(fs.map);
 356                 goto RetryFault;
 357         }
 358
 359         if (wired)
 360                 fault_type = prot | (fault_type & VM_PROT_COPY);
 361
 362         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 363             (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 &&
 364             /* avoid calling vm_object_set_writeable_dirty() */
 365             ((prot & VM_PROT_WRITE) == 0 ||
 366             (fs.first_object->type != OBJT_VNODE &&
 367             (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 368             (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
 369                 VM_OBJECT_RLOCK(fs.first_object);
 370                 if ((prot & VM_PROT_WRITE) != 0 &&
 371                     (fs.first_object->type == OBJT_VNODE ||
 372                     (fs.first_object->flags & OBJ_TMPFS_NODE) != 0) &&
 373                     (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
 374                         goto fast_failed;
 375                 m = vm_page_lookup(fs.first_object, fs.first_pindex);
 376                 /* A busy page can be mapped for read|execute access. */
 377                 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 378                     vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
 379                         goto fast_failed;
 380                 result = pmap_enter(fs.map->pmap, vaddr, m, prot,
 381                    fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
 382                    0), 0);
 383                 if (result != KERN_SUCCESS)
 384                         goto fast_failed;
 385                 if (m_hold != NULL) {
 386                         *m_hold = m;
 387                         vm_page_lock(m);
 388                         vm_page_hold(m);
 389                         vm_page_unlock(m);
 390                 }
 391                 vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags,
 392                     FALSE);
 393                 VM_OBJECT_RUNLOCK(fs.first_object);
 394                 if (!wired)
 395                         vm_fault_prefault(&fs, vaddr, 0, 0);
 396                 vm_map_lookup_done(fs.map, fs.entry);
 397                 curthread->td_ru.ru_minflt++;
 398                 return (KERN_SUCCESS);
 399 fast_failed:
 400                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 401                         VM_OBJECT_RUNLOCK(fs.first_object);
 402                         VM_OBJECT_WLOCK(fs.first_object);
 403                 }
 404         } else {
 405                 VM_OBJECT_WLOCK(fs.first_object);
 406         }
 407
 408         /*
 409          * Make a reference to this object to prevent its disposal while we
 410          * are messing with it.  Once we have the reference, the map is free
 411          * to be diddled.  Since objects reference their shadows (and copies),
 412          * they will stay around as well.
 413          *
 414          * Bump the paging-in-progress count to prevent size changes (e.g.
 415          * truncation operations) during I/O.  This must be done after
 416          * obtaining the vnode lock in order to avoid possible deadlocks.
 417          */
 418         vm_object_reference_locked(fs.first_object);
 419         vm_object_pip_add(fs.first_object, 1);
 420
 421         fs.lookup_still_valid = TRUE;
 422
 423         fs.first_m = NULL;
 424
 425         /*
 426          * Search for the page at object/offset.
 427          */
 428         fs.object = fs.first_object;
 429         fs.pindex = fs.first_pindex;
 430         while (TRUE) {
 431                 /*
 432                  * If the object is dead, we stop here
 433                  */
 434                 if (fs.object->flags & OBJ_DEAD) {
 435                         unlock_and_deallocate(&fs);
 436                         return (KERN_PROTECTION_FAILURE);
 437                 }
 438
 439                 /*
 440                  * See if page is resident
 441                  */
 442                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 443                 if (fs.m != NULL) {
 444                         /*
 445                          * Wait/Retry if the page is busy.  We have to do this
 446                          * if the page is either exclusive or shared busy
 447                          * because the vm_pager may be using read busy for
 448                          * pageouts (and even pageins if it is the vnode
 449                          * pager), and we could end up trying to pagein and
 450                          * pageout the same page simultaneously.
 451                          *
 452                          * We can theoretically allow the busy case on a read
 453                          * fault if the page is marked valid, but since such
 454                          * pages are typically already pmap'd, putting that
 455                          * special case in might be more effort then it is
 456                          * worth.  We cannot under any circumstances mess
 457                          * around with a shared busied page except, perhaps,
 458                          * to pmap it.
 459                          */
 460                         if (vm_page_busied(fs.m)) {
 461                                 /*
 462                                  * Reference the page before unlocking and
 463                                  * sleeping so that the page daemon is less
 464                                  * likely to reclaim it.
 465                                  */
 466                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 467                                 if (fs.object != fs.first_object) {
 468                                         if (!VM_OBJECT_TRYWLOCK(
 469                                             fs.first_object)) {
 470                                                 VM_OBJECT_WUNLOCK(fs.object);
 471                                                 VM_OBJECT_WLOCK(fs.first_object);
 472                                                 VM_OBJECT_WLOCK(fs.object);
 473                                         }
 474                                         vm_page_lock(fs.first_m);
 475                                         vm_page_free(fs.first_m);
 476                                         vm_page_unlock(fs.first_m);
 477                                         vm_object_pip_wakeup(fs.first_object);
 478                                         VM_OBJECT_WUNLOCK(fs.first_object);
 479                                         fs.first_m = NULL;
 480                                 }
 481                                 unlock_map(&fs);
 482                                 if (fs.m == vm_page_lookup(fs.object,
 483                                     fs.pindex)) {
 484                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 485                                 }
 486                                 vm_object_pip_wakeup(fs.object);
 487                                 VM_OBJECT_WUNLOCK(fs.object);
 488                                 PCPU_INC(cnt.v_intrans);
 489                                 vm_object_deallocate(fs.first_object);
 490                                 goto RetryFault;
 491                         }
 492                         vm_page_lock(fs.m);
 493                         vm_page_remque(fs.m);
 494                         vm_page_unlock(fs.m);
 495
 496                         /*
 497                          * Mark page busy for other processes, and the
 498                          * pagedaemon.  If it still isn't completely valid
 499                          * (readable), jump to readrest, else break-out ( we
 500                          * found the page ).
 501                          */
 502                         vm_page_xbusy(fs.m);
 503                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 504                                 goto readrest;
 505                         break;
 506                 }
 507
 508                 /*
 509                  * Page is not resident, If this is the search termination
 510                  * or the pager might contain the page, allocate a new page.
 511                  */
 512                 if (TRYPAGER || fs.object == fs.first_object) {
 513                         if (fs.pindex >= fs.object->size) {
 514                                 unlock_and_deallocate(&fs);
 515                                 return (KERN_PROTECTION_FAILURE);
 516                         }
 517
 518                         /*
 519                          * Allocate a new page for this object/offset pair.
 520                          *
 521                          * Unlocked read of the p_flag is harmless. At
 522                          * worst, the P_KILLED might be not observed
 523                          * there, and allocation can fail, causing
 524                          * restart and new reading of the p_flag.
 525                          */
 526                         fs.m = NULL;
 527                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 528 #if VM_NRESERVLEVEL > 0
 529                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 530                                         fs.object->flags |= OBJ_COLORED;
 531                                         fs.object->pg_color = atop(vaddr) -
 532                                             fs.pindex;
 533                                 }
 534 #endif
 535                                 alloc_req = P_KILLED(curproc) ?
 536                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 537                                 if (fs.object->type != OBJT_VNODE &&
 538                                     fs.object->backing_object == NULL)
 539                                         alloc_req |= VM_ALLOC_ZERO;
 540                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 541                                     alloc_req);
 542                         }
 543                         if (fs.m == NULL) {
 544                                 unlock_and_deallocate(&fs);
 545                                 VM_WAITPFAULT;
 546                                 goto RetryFault;
 547                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 548                                 break;
 549                 }
 550
 551 readrest:
 552                 /*
 553                  * We have found a valid page or we have allocated a new page.
 554                  * The page thus may not be valid or may not be entirely
 555                  * valid.
 556                  *
 557                  * Attempt to fault-in the page if there is a chance that the
 558                  * pager has it, and potentially fault in additional pages
 559                  * at the same time.
 560                  */
 561                 if (TRYPAGER) {
 562                         int rv;
 563                         u_char behavior = vm_map_entry_behavior(fs.entry);
 564
 565                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 566                             P_KILLED(curproc)) {
 567                                 behind = 0;
 568                                 ahead = 0;
 569                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 570                                 behind = 0;
 571                                 ahead = atop(fs.entry->end - vaddr) - 1;
 572                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 573                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 574                                 if (fs.pindex == fs.entry->next_read)
 575                                         vm_fault_cache_behind(&fs,
 576                                             VM_FAULT_READ_MAX);
 577                         } else {
 578                                 /*
 579                                  * If this is a sequential page fault, then
 580                                  * arithmetically increase the number of pages
 581                                  * in the read-ahead window.  Otherwise, reset
 582                                  * the read-ahead window to its smallest size.
 583                                  */
 584                                 behind = atop(vaddr - fs.entry->start);
 585                                 if (behind > VM_FAULT_READ_BEHIND)
 586                                         behind = VM_FAULT_READ_BEHIND;
 587                                 ahead = atop(fs.entry->end - vaddr) - 1;
 588                                 era = fs.entry->read_ahead;
 589                                 if (fs.pindex == fs.entry->next_read) {
 590                                         nera = era + behind;
 591                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 592                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 593                                         behind = 0;
 594                                         if (ahead > nera)
 595                                                 ahead = nera;
 596                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 597                                                 vm_fault_cache_behind(&fs,
 598                                                     VM_FAULT_CACHE_BEHIND);
 599                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 600                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 601                                 if (era != ahead)
 602                                         fs.entry->read_ahead = ahead;
 603                         }
 604
 605                         /*
 606                          * Call the pager to retrieve the data, if any, after
 607                          * releasing the lock on the map.  We hold a ref on
 608                          * fs.object and the pages are exclusive busied.
 609                          */
 610                         unlock_map(&fs);
 611
 612                         if (fs.object->type == OBJT_VNODE) {
 613                                 vp = fs.object->handle;
 614                                 if (vp == fs.vp)
 615                                         goto vnode_locked;
 616                                 else if (fs.vp != NULL) {
 617                                         vput(fs.vp);
 618                                         fs.vp = NULL;
 619                                 }
 620                                 locked = VOP_ISLOCKED(vp);
 621
 622                                 if (locked != LK_EXCLUSIVE)
 623                                         locked = LK_SHARED;
 624                                 /* Do not sleep for vnode lock while fs.m is busy */
 625                                 error = vget(vp, locked | LK_CANRECURSE |
 626                                     LK_NOWAIT, curthread);
 627                                 if (error != 0) {
 628                                         vhold(vp);
 629                                         release_page(&fs);
 630                                         unlock_and_deallocate(&fs);
 631                                         error = vget(vp, locked | LK_RETRY |
 632                                             LK_CANRECURSE, curthread);
 633                                         vdrop(vp);
 634                                         fs.vp = vp;
 635                                         KASSERT(error == 0,
 636                                             ("vm_fault: vget failed"));
 637                                         goto RetryFault;
 638                                 }
 639                                 fs.vp = vp;
 640                         }
 641 vnode_locked:
 642                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 643                             ("vm_fault: vnode-backed object mapped by system map"));
 644
 645                         /*
 646                          * now we find out if any other pages should be paged
 647                          * in at this time this routine checks to see if the
 648                          * pages surrounding this fault reside in the same
 649                          * object as the page for this fault.  If they do,
 650                          * then they are faulted in also into the object.  The
 651                          * array "marray" returned contains an array of
 652                          * vm_page_t structs where one of them is the
 653                          * vm_page_t passed to the routine.  The reqpage
 654                          * return value is the index into the marray for the
 655                          * vm_page_t passed to the routine.
 656                          *
 657                          * fs.m plus the additional pages are exclusive busied.
 658                          */
 659                         faultcount = vm_fault_additional_pages(
 660                             fs.m, behind, ahead, marray, &reqpage);
 661
 662                         rv = faultcount ?
 663                             vm_pager_get_pages(fs.object, marray, faultcount,
 664                                 reqpage) : VM_PAGER_FAIL;
 665
 666                         if (rv == VM_PAGER_OK) {
 667                                 /*
 668                                  * Found the page. Leave it busy while we play
 669                                  * with it.
 670                                  */
 671
 672                                 /*
 673                                  * Relookup in case pager changed page. Pager
 674                                  * is responsible for disposition of old page
 675                                  * if moved.
 676                                  */
 677                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 678                                 if (!fs.m) {
 679                                         unlock_and_deallocate(&fs);
 680                                         goto RetryFault;
 681                                 }
 682
 683                                 hardfault++;
 684                                 break; /* break to PAGE HAS BEEN FOUND */
 685                         }
 686                         /*
 687                          * Remove the bogus page (which does not exist at this
 688                          * object/offset); before doing so, we must get back
 689                          * our object lock to preserve our invariant.
 690                          *
 691                          * Also wake up any other process that may want to bring
 692                          * in this page.
 693                          *
 694                          * If this is the top-level object, we must leave the
 695                          * busy page to prevent another process from rushing
 696                          * past us, and inserting the page in that object at
 697                          * the same time that we are.
 698                          */
 699                         if (rv == VM_PAGER_ERROR)
 700                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 701                                     curproc->p_pid, curproc->p_comm);
 702                         /*
 703                          * Data outside the range of the pager or an I/O error
 704                          */
 705                         /*
 706                          * XXX - the check for kernel_map is a kludge to work
 707                          * around having the machine panic on a kernel space
 708                          * fault w/ I/O error.
 709                          */
 710                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 711                                 (rv == VM_PAGER_BAD)) {
 712                                 vm_page_lock(fs.m);
 713                                 vm_page_free(fs.m);
 714                                 vm_page_unlock(fs.m);
 715                                 fs.m = NULL;
 716                                 unlock_and_deallocate(&fs);
 717                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 718                         }
 719                         if (fs.object != fs.first_object) {
 720                                 vm_page_lock(fs.m);
 721                                 vm_page_free(fs.m);
 722                                 vm_page_unlock(fs.m);
 723                                 fs.m = NULL;
 724                                 /*
 725                                  * XXX - we cannot just fall out at this
 726                                  * point, m has been freed and is invalid!
 727                                  */
 728                         }
 729                 }
 730
 731                 /*
 732                  * We get here if the object has default pager (or unwiring)
 733                  * or the pager doesn't have the page.
 734                  */
 735                 if (fs.object == fs.first_object)
 736                         fs.first_m = fs.m;
 737
 738                 /*
 739                  * Move on to the next object.  Lock the next object before
 740                  * unlocking the current one.
 741                  */
 742                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 743                 next_object = fs.object->backing_object;
 744                 if (next_object == NULL) {
 745                         /*
 746                          * If there's no object left, fill the page in the top
 747                          * object with zeros.
 748                          */
 749                         if (fs.object != fs.first_object) {
 750                                 vm_object_pip_wakeup(fs.object);
 751                                 VM_OBJECT_WUNLOCK(fs.object);
 752
 753                                 fs.object = fs.first_object;
 754                                 fs.pindex = fs.first_pindex;
 755                                 fs.m = fs.first_m;
 756                                 VM_OBJECT_WLOCK(fs.object);
 757                         }
 758                         fs.first_m = NULL;
 759
 760                         /*
 761                          * Zero the page if necessary and mark it valid.
 762                          */
 763                         if ((fs.m->flags & PG_ZERO) == 0) {
 764                                 pmap_zero_page(fs.m);
 765                         } else {
 766                                 PCPU_INC(cnt.v_ozfod);
 767                         }
 768                         PCPU_INC(cnt.v_zfod);
 769                         fs.m->valid = VM_PAGE_BITS_ALL;
 770                         /* Don't try to prefault neighboring pages. */
 771                         faultcount = 1;
 772                         break;  /* break to PAGE HAS BEEN FOUND */
 773                 } else {
 774                         KASSERT(fs.object != next_object,
 775                             ("object loop %p", next_object));
 776                         VM_OBJECT_WLOCK(next_object);
 777                         vm_object_pip_add(next_object, 1);
 778                         if (fs.object != fs.first_object)
 779                                 vm_object_pip_wakeup(fs.object);
 780                         VM_OBJECT_WUNLOCK(fs.object);
 781                         fs.object = next_object;
 782                 }
 783         }
 784
 785         vm_page_assert_xbusied(fs.m);
 786
 787         /*
 788          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 789          * is held.]
 790          */
 791
 792         /*
 793          * If the page is being written, but isn't already owned by the
 794          * top-level object, we have to copy it into a new page owned by the
 795          * top-level object.
 796          */
 797         if (fs.object != fs.first_object) {
 798                 /*
 799                  * We only really need to copy if we want to write it.
 800                  */
 801                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 802                         /*
 803                          * This allows pages to be virtually copied from a
 804                          * backing_object into the first_object, where the
 805                          * backing object has no other refs to it, and cannot
 806                          * gain any more refs.  Instead of a bcopy, we just
 807                          * move the page from the backing object to the
 808                          * first object.  Note that we must mark the page
 809                          * dirty in the first object so that it will go out
 810                          * to swap when needed.
 811                          */
 812                         is_first_object_locked = FALSE;
 813                         if (
 814                                 /*
 815                                  * Only one shadow object
 816                                  */
 817                                 (fs.object->shadow_count == 1) &&
 818                                 /*
 819                                  * No COW refs, except us
 820                                  */
 821                                 (fs.object->ref_count == 1) &&
 822                                 /*
 823                                  * No one else can look this object up
 824                                  */
 825                                 (fs.object->handle == NULL) &&
 826                                 /*
 827                                  * No other ways to look the object up
 828                                  */
 829                                 ((fs.object->type == OBJT_DEFAULT) ||
 830                                  (fs.object->type == OBJT_SWAP)) &&
 831                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
 832                                 /*
 833                                  * We don't chase down the shadow chain
 834                                  */
 835                             fs.object == fs.first_object->backing_object) {
 836                                 /*
 837                                  * get rid of the unnecessary page
 838                                  */
 839                                 vm_page_lock(fs.first_m);
 840                                 vm_page_free(fs.first_m);
 841                                 vm_page_unlock(fs.first_m);
 842                                 /*
 843                                  * grab the page and put it into the
 844                                  * process'es object.  The page is
 845                                  * automatically made dirty.
 846                                  */
 847                                 if (vm_page_rename(fs.m, fs.first_object,
 848                                     fs.first_pindex)) {
 849                                         unlock_and_deallocate(&fs);
 850                                         goto RetryFault;
 851                                 }
 852 #if VM_NRESERVLEVEL > 0
 853                                 /*
 854                                  * Rename the reservation.
 855                                  */
 856                                 vm_reserv_rename(fs.m, fs.first_object,
 857                                     fs.object, OFF_TO_IDX(
 858                                     fs.first_object->backing_object_offset));
 859 #endif
 860                                 vm_page_xbusy(fs.m);
 861                                 fs.first_m = fs.m;
 862                                 fs.m = NULL;
 863                                 PCPU_INC(cnt.v_cow_optim);
 864                         } else {
 865                                 /*
 866                                  * Oh, well, lets copy it.
 867                                  */
 868                                 pmap_copy_page(fs.m, fs.first_m);
 869                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 870                                 if (wired && (fault_flags &
 871                                     VM_FAULT_CHANGE_WIRING) == 0) {
 872                                         vm_page_lock(fs.first_m);
 873                                         vm_page_wire(fs.first_m);
 874                                         vm_page_unlock(fs.first_m);
 875
 876                                         vm_page_lock(fs.m);
 877                                         vm_page_unwire(fs.m, FALSE);
 878                                         vm_page_unlock(fs.m);
 879                                 }
 880                                 /*
 881                                  * We no longer need the old page or object.
 882                                  */
 883                                 release_page(&fs);
 884                         }
 885                         /*
 886                          * fs.object != fs.first_object due to above
 887                          * conditional
 888                          */
 889                         vm_object_pip_wakeup(fs.object);
 890                         VM_OBJECT_WUNLOCK(fs.object);
 891                         /*
 892                          * Only use the new page below...
 893                          */
 894                         fs.object = fs.first_object;
 895                         fs.pindex = fs.first_pindex;
 896                         fs.m = fs.first_m;
 897                         if (!is_first_object_locked)
 898                                 VM_OBJECT_WLOCK(fs.object);
 899                         PCPU_INC(cnt.v_cow_faults);
 900                         curthread->td_cow++;
 901                 } else {
 902                         prot &= ~VM_PROT_WRITE;
 903                 }
 904         }
 905
 906         /*
 907          * We must verify that the maps have not changed since our last
 908          * lookup.
 909          */
 910         if (!fs.lookup_still_valid) {
 911                 vm_object_t retry_object;
 912                 vm_pindex_t retry_pindex;
 913                 vm_prot_t retry_prot;
 914
 915                 if (!vm_map_trylock_read(fs.map)) {
 916                         release_page(&fs);
 917                         unlock_and_deallocate(&fs);
 918                         goto RetryFault;
 919                 }
 920                 fs.lookup_still_valid = TRUE;
 921                 if (fs.map->timestamp != map_generation) {
 922                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 923                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 924
 925                         /*
 926                          * If we don't need the page any longer, put it on the inactive
 927                          * list (the easiest thing to do here).  If no one needs it,
 928                          * pageout will grab it eventually.
 929                          */
 930                         if (result != KERN_SUCCESS) {
 931                                 release_page(&fs);
 932                                 unlock_and_deallocate(&fs);
 933
 934                                 /*
 935                                  * If retry of map lookup would have blocked then
 936                                  * retry fault from start.
 937                                  */
 938                                 if (result == KERN_FAILURE)
 939                                         goto RetryFault;
 940                                 return (result);
 941                         }
 942                         if ((retry_object != fs.first_object) ||
 943                             (retry_pindex != fs.first_pindex)) {
 944                                 release_page(&fs);
 945                                 unlock_and_deallocate(&fs);
 946                                 goto RetryFault;
 947                         }
 948
 949                         /*
 950                          * Check whether the protection has changed or the object has
 951                          * been copied while we left the map unlocked. Changing from
 952                          * read to write permission is OK - we leave the page
 953                          * write-protected, and catch the write fault. Changing from
 954                          * write to read permission means that we can't mark the page
 955                          * write-enabled after all.
 956                          */
 957                         prot &= retry_prot;
 958                 }
 959         }
 960         /*
 961          * If the page was filled by a pager, update the map entry's
 962          * last read offset.  Since the pager does not return the
 963          * actual set of pages that it read, this update is based on
 964          * the requested set.  Typically, the requested and actual
 965          * sets are the same.
 966          *
 967          * XXX The following assignment modifies the map
 968          * without holding a write lock on it.
 969          */
 970         if (hardfault)
 971                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 972
 973         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE);
 974         vm_page_assert_xbusied(fs.m);
 975
 976         /*
 977          * Page must be completely valid or it is not fit to
 978          * map into user space.  vm_pager_get_pages() ensures this.
 979          */
 980         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 981             ("vm_fault: page %p partially invalid", fs.m));
 982         VM_OBJECT_WUNLOCK(fs.object);
 983
 984         /*
 985          * Put this page into the physical map.  We had to do the unlock above
 986          * because pmap_enter() may sleep.  We don't put the page
 987          * back on the active queue until later so that the pageout daemon
 988          * won't find it (yet).
 989          */
 990         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
 991             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
 992         if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
 993             wired == 0)
 994                 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
 995         VM_OBJECT_WLOCK(fs.object);
 996         vm_page_lock(fs.m);
 997
 998         /*
 999          * If the page is not wired down, then put it where the pageout daemon
1000          * can find it.
1001          */
1002         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
1003                 if (wired)
1004                         vm_page_wire(fs.m);
1005                 else
1006                         vm_page_unwire(fs.m, 1);
1007         } else
1008                 vm_page_activate(fs.m);
1009         if (m_hold != NULL) {
1010                 *m_hold = fs.m;
1011                 vm_page_hold(fs.m);
1012         }
1013         vm_page_unlock(fs.m);
1014         vm_page_xunbusy(fs.m);
1015
1016         /*
1017          * Unlock everything, and return
1018          */
1019         unlock_and_deallocate(&fs);
1020         if (hardfault) {
1021                 PCPU_INC(cnt.v_io_faults);
1022                 curthread->td_ru.ru_majflt++;
1023         } else
1024                 curthread->td_ru.ru_minflt++;
1025
1026         return (KERN_SUCCESS);
1027 }
1028
1029 /*
1030  * Speed up the reclamation of up to "distance" pages that precede the
1031  * faulting pindex within the first object of the shadow chain.
1032  */
1033 static void
1034 vm_fault_cache_behind(const struct faultstate *fs, int distance)
1035 {
1036         vm_object_t first_object, object;
1037         vm_page_t m, m_prev;
1038         vm_pindex_t pindex;
1039
1040         object = fs->object;
1041         VM_OBJECT_ASSERT_WLOCKED(object);
1042         first_object = fs->first_object;
1043         if (first_object != object) {
1044                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1045                         VM_OBJECT_WUNLOCK(object);
1046                         VM_OBJECT_WLOCK(first_object);
1047                         VM_OBJECT_WLOCK(object);
1048                 }
1049         }
1050         /* Neither fictitious nor unmanaged pages can be cached. */
1051         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1052                 if (fs->first_pindex < distance)
1053                         pindex = 0;
1054                 else
1055                         pindex = fs->first_pindex - distance;
1056                 if (pindex < OFF_TO_IDX(fs->entry->offset))
1057                         pindex = OFF_TO_IDX(fs->entry->offset);
1058                 m = first_object != object ? fs->first_m : fs->m;
1059                 vm_page_assert_xbusied(m);
1060                 m_prev = vm_page_prev(m);
1061                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1062                     m->valid == VM_PAGE_BITS_ALL) {
1063                         m_prev = vm_page_prev(m);
1064                         if (vm_page_busied(m))
1065                                 continue;
1066                         vm_page_lock(m);
1067                         if (m->hold_count == 0 && m->wire_count == 0) {
1068                                 pmap_remove_all(m);
1069                                 vm_page_aflag_clear(m, PGA_REFERENCED);
1070                                 if (m->dirty != 0)
1071                                         vm_page_deactivate(m);
1072                                 else
1073                                         vm_page_cache(m);
1074                         }
1075                         vm_page_unlock(m);
1076                 }
1077         }
1078         if (first_object != object)
1079                 VM_OBJECT_WUNLOCK(first_object);
1080 }
1081
1082 /*
1083  * vm_fault_prefault provides a quick way of clustering
1084  * pagefaults into a processes address space.  It is a "cousin"
1085  * of vm_map_pmap_enter, except it runs at page fault time instead
1086  * of mmap time.
1087  */
1088 static void
1089 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1090     int faultcount, int reqpage)
1091 {
1092         pmap_t pmap;
1093         vm_map_entry_t entry;
1094         vm_object_t backing_object, lobject;
1095         vm_offset_t addr, starta;
1096         vm_pindex_t pindex;
1097         vm_page_t m;
1098         int backward, forward, i;
1099
1100         pmap = fs->map->pmap;
1101         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1102                 return;
1103
1104         if (faultcount > 0) {
1105                 backward = reqpage;
1106                 forward = faultcount - reqpage - 1;
1107         } else {
1108                 backward = PFBAK;
1109                 forward = PFFOR;
1110         }
1111         entry = fs->entry;
1112
1113         starta = addra - backward * PAGE_SIZE;
1114         if (starta < entry->start) {
1115                 starta = entry->start;
1116         } else if (starta > addra) {
1117                 starta = 0;
1118         }
1119
1120         /*
1121          * Generate the sequence of virtual addresses that are candidates for
1122          * prefaulting in an outward spiral from the faulting virtual address,
1123          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1124          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1125          * If the candidate address doesn't have a backing physical page, then
1126          * the loop immediately terminates.
1127          */
1128         for (i = 0; i < 2 * imax(backward, forward); i++) {
1129                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1130                     PAGE_SIZE);
1131                 if (addr > addra + forward * PAGE_SIZE)
1132                         addr = 0;
1133
1134                 if (addr < starta || addr >= entry->end)
1135                         continue;
1136
1137                 if (!pmap_is_prefaultable(pmap, addr))
1138                         continue;
1139
1140                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1141                 lobject = entry->object.vm_object;
1142                 VM_OBJECT_RLOCK(lobject);
1143                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1144                     lobject->type == OBJT_DEFAULT &&
1145                     (backing_object = lobject->backing_object) != NULL) {
1146                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1147                             0, ("vm_fault_prefault: unaligned object offset"));
1148                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1149                         VM_OBJECT_RLOCK(backing_object);
1150                         VM_OBJECT_RUNLOCK(lobject);
1151                         lobject = backing_object;
1152                 }
1153                 if (m == NULL) {
1154                         VM_OBJECT_RUNLOCK(lobject);
1155                         break;
1156                 }
1157                 if (m->valid == VM_PAGE_BITS_ALL &&
1158                     (m->flags & PG_FICTITIOUS) == 0)
1159                         pmap_enter_quick(pmap, addr, m, entry->protection);
1160                 VM_OBJECT_RUNLOCK(lobject);
1161         }
1162 }
1163
1164 /*
1165  * Hold each of the physical pages that are mapped by the specified range of
1166  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1167  * and allow the specified types of access, "prot".  If all of the implied
1168  * pages are successfully held, then the number of held pages is returned
1169  * together with pointers to those pages in the array "ma".  However, if any
1170  * of the pages cannot be held, -1 is returned.
1171  */
1172 int
1173 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1174     vm_prot_t prot, vm_page_t *ma, int max_count)
1175 {
1176         vm_offset_t end, va;
1177         vm_page_t *mp;
1178         int count;
1179         boolean_t pmap_failed;
1180
1181         if (len == 0)
1182                 return (0);
1183         end = round_page(addr + len);
1184         addr = trunc_page(addr);
1185
1186         /*
1187          * Check for illegal addresses.
1188          */
1189         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1190                 return (-1);
1191
1192         if (atop(end - addr) > max_count)
1193                 panic("vm_fault_quick_hold_pages: count > max_count");
1194         count = atop(end - addr);
1195
1196         /*
1197          * Most likely, the physical pages are resident in the pmap, so it is
1198          * faster to try pmap_extract_and_hold() first.
1199          */
1200         pmap_failed = FALSE;
1201         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1202                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1203                 if (*mp == NULL)
1204                         pmap_failed = TRUE;
1205                 else if ((prot & VM_PROT_WRITE) != 0 &&
1206                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1207                         /*
1208                          * Explicitly dirty the physical page.  Otherwise, the
1209                          * caller's changes may go unnoticed because they are
1210                          * performed through an unmanaged mapping or by a DMA
1211                          * operation.
1212                          *
1213                          * The object lock is not held here.
1214                          * See vm_page_clear_dirty_mask().
1215                          */
1216                         vm_page_dirty(*mp);
1217                 }
1218         }
1219         if (pmap_failed) {
1220                 /*
1221                  * One or more pages could not be held by the pmap.  Either no
1222                  * page was mapped at the specified virtual address or that
1223                  * mapping had insufficient permissions.  Attempt to fault in
1224                  * and hold these pages.
1225                  */
1226                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1227                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1228                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1229                                 goto error;
1230         }
1231         return (count);
1232 error:
1233         for (mp = ma; mp < ma + count; mp++)
1234                 if (*mp != NULL) {
1235                         vm_page_lock(*mp);
1236                         vm_page_unhold(*mp);
1237                         vm_page_unlock(*mp);
1238                 }
1239         return (-1);
1240 }
1241
1242 /*
1243  *      Routine:
1244  *              vm_fault_copy_entry
1245  *      Function:
1246  *              Create new shadow object backing dst_entry with private copy of
1247  *              all underlying pages. When src_entry is equal to dst_entry,
1248  *              function implements COW for wired-down map entry. Otherwise,
1249  *              it forks wired entry into dst_map.
1250  *
1251  *      In/out conditions:
1252  *              The source and destination maps must be locked for write.
1253  *              The source map entry must be wired down (or be a sharing map
1254  *              entry corresponding to a main map entry that is wired down).
1255  */
1256 void
1257 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1258     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1259     vm_ooffset_t *fork_charge)
1260 {
1261         vm_object_t backing_object, dst_object, object, src_object;
1262         vm_pindex_t dst_pindex, pindex, src_pindex;
1263         vm_prot_t access, prot;
1264         vm_offset_t vaddr;
1265         vm_page_t dst_m;
1266         vm_page_t src_m;
1267         boolean_t upgrade;
1268
1269 #ifdef  lint
1270         src_map++;
1271 #endif  /* lint */
1272
1273         upgrade = src_entry == dst_entry;
1274         access = prot = dst_entry->protection;
1275
1276         src_object = src_entry->object.vm_object;
1277         src_pindex = OFF_TO_IDX(src_entry->offset);
1278
1279         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1280                 dst_object = src_object;
1281                 vm_object_reference(dst_object);
1282         } else {
1283                 /*
1284                  * Create the top-level object for the destination entry. (Doesn't
1285                  * actually shadow anything - we copy the pages directly.)
1286                  */
1287                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1288                     OFF_TO_IDX(dst_entry->end - dst_entry->start));
1289 #if VM_NRESERVLEVEL > 0
1290                 dst_object->flags |= OBJ_COLORED;
1291                 dst_object->pg_color = atop(dst_entry->start);
1292 #endif
1293         }
1294
1295         VM_OBJECT_WLOCK(dst_object);
1296         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1297             ("vm_fault_copy_entry: vm_object not NULL"));
1298         if (src_object != dst_object) {
1299                 dst_entry->object.vm_object = dst_object;
1300                 dst_entry->offset = 0;
1301                 dst_object->charge = dst_entry->end - dst_entry->start;
1302         }
1303         if (fork_charge != NULL) {
1304                 KASSERT(dst_entry->cred == NULL,
1305                     ("vm_fault_copy_entry: leaked swp charge"));
1306                 dst_object->cred = curthread->td_ucred;
1307                 crhold(dst_object->cred);
1308                 *fork_charge += dst_object->charge;
1309         } else if (dst_object->cred == NULL) {
1310                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1311                     dst_entry));
1312                 dst_object->cred = dst_entry->cred;
1313                 dst_entry->cred = NULL;
1314         }
1315
1316         /*
1317          * If not an upgrade, then enter the mappings in the pmap as
1318          * read and/or execute accesses.  Otherwise, enter them as
1319          * write accesses.
1320          *
1321          * A writeable large page mapping is only created if all of
1322          * the constituent small page mappings are modified. Marking
1323          * PTEs as modified on inception allows promotion to happen
1324          * without taking potentially large number of soft faults.
1325          */
1326         if (!upgrade)
1327                 access &= ~VM_PROT_WRITE;
1328
1329         /*
1330          * Loop through all of the virtual pages within the entry's
1331          * range, copying each page from the source object to the
1332          * destination object.  Since the source is wired, those pages
1333          * must exist.  In contrast, the destination is pageable.
1334          * Since the destination object does share any backing storage
1335          * with the source object, all of its pages must be dirtied,
1336          * regardless of whether they can be written.
1337          */
1338         for (vaddr = dst_entry->start, dst_pindex = 0;
1339             vaddr < dst_entry->end;
1340             vaddr += PAGE_SIZE, dst_pindex++) {
1341 again:
1342                 /*
1343                  * Find the page in the source object, and copy it in.
1344                  * Because the source is wired down, the page will be
1345                  * in memory.
1346                  */
1347                 if (src_object != dst_object)
1348                         VM_OBJECT_RLOCK(src_object);
1349                 object = src_object;
1350                 pindex = src_pindex + dst_pindex;
1351                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1352                     (backing_object = object->backing_object) != NULL) {
1353                         /*
1354                          * Unless the source mapping is read-only or
1355                          * it is presently being upgraded from
1356                          * read-only, the first object in the shadow
1357                          * chain should provide all of the pages.  In
1358                          * other words, this loop body should never be
1359                          * executed when the source mapping is already
1360                          * read/write.
1361                          */
1362                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1363                             upgrade,
1364                             ("vm_fault_copy_entry: main object missing page"));
1365
1366                         VM_OBJECT_RLOCK(backing_object);
1367                         pindex += OFF_TO_IDX(object->backing_object_offset);
1368                         if (object != dst_object)
1369                                 VM_OBJECT_RUNLOCK(object);
1370                         object = backing_object;
1371                 }
1372                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1373
1374                 if (object != dst_object) {
1375                         /*
1376                          * Allocate a page in the destination object.
1377                          */
1378                         dst_m = vm_page_alloc(dst_object, (src_object ==
1379                             dst_object ? src_pindex : 0) + dst_pindex,
1380                             VM_ALLOC_NORMAL);
1381                         if (dst_m == NULL) {
1382                                 VM_OBJECT_WUNLOCK(dst_object);
1383                                 VM_OBJECT_RUNLOCK(object);
1384                                 VM_WAIT;
1385                                 VM_OBJECT_WLOCK(dst_object);
1386                                 goto again;
1387                         }
1388                         pmap_copy_page(src_m, dst_m);
1389                         VM_OBJECT_RUNLOCK(object);
1390                         dst_m->valid = VM_PAGE_BITS_ALL;
1391                         dst_m->dirty = VM_PAGE_BITS_ALL;
1392                 } else {
1393                         dst_m = src_m;
1394                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1395                                 goto again;
1396                         vm_page_xbusy(dst_m);
1397                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1398                             ("invalid dst page %p", dst_m));
1399                 }
1400                 VM_OBJECT_WUNLOCK(dst_object);
1401
1402                 /*
1403                  * Enter it in the pmap. If a wired, copy-on-write
1404                  * mapping is being replaced by a write-enabled
1405                  * mapping, then wire that new mapping.
1406                  */
1407                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1408                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1409
1410                 /*
1411                  * Mark it no longer busy, and put it on the active list.
1412                  */
1413                 VM_OBJECT_WLOCK(dst_object);
1414
1415                 if (upgrade) {
1416                         if (src_m != dst_m) {
1417                                 vm_page_lock(src_m);
1418                                 vm_page_unwire(src_m, 0);
1419                                 vm_page_unlock(src_m);
1420                                 vm_page_lock(dst_m);
1421                                 vm_page_wire(dst_m);
1422                                 vm_page_unlock(dst_m);
1423                         } else {
1424                                 KASSERT(dst_m->wire_count > 0,
1425                                     ("dst_m %p is not wired", dst_m));
1426                         }
1427                 } else {
1428                         vm_page_lock(dst_m);
1429                         vm_page_activate(dst_m);
1430                         vm_page_unlock(dst_m);
1431                 }
1432                 vm_page_xunbusy(dst_m);
1433         }
1434         VM_OBJECT_WUNLOCK(dst_object);
1435         if (upgrade) {
1436                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1437                 vm_object_deallocate(src_object);
1438         }
1439 }
1440
1441
1442 /*
1443  * This routine checks around the requested page for other pages that
1444  * might be able to be faulted in.  This routine brackets the viable
1445  * pages for the pages to be paged in.
1446  *
1447  * Inputs:
1448  *      m, rbehind, rahead
1449  *
1450  * Outputs:
1451  *  marray (array of vm_page_t), reqpage (index of requested page)
1452  *
1453  * Return value:
1454  *  number of pages in marray
1455  */
1456 static int
1457 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1458         vm_page_t m;
1459         int rbehind;
1460         int rahead;
1461         vm_page_t *marray;
1462         int *reqpage;
1463 {
1464         int i,j;
1465         vm_object_t object;
1466         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1467         vm_page_t rtm;
1468         int cbehind, cahead;
1469
1470         VM_OBJECT_ASSERT_WLOCKED(m->object);
1471
1472         object = m->object;
1473         pindex = m->pindex;
1474         cbehind = cahead = 0;
1475
1476         /*
1477          * if the requested page is not available, then give up now
1478          */
1479         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1480                 return 0;
1481         }
1482
1483         if ((cbehind == 0) && (cahead == 0)) {
1484                 *reqpage = 0;
1485                 marray[0] = m;
1486                 return 1;
1487         }
1488
1489         if (rahead > cahead) {
1490                 rahead = cahead;
1491         }
1492
1493         if (rbehind > cbehind) {
1494                 rbehind = cbehind;
1495         }
1496
1497         /*
1498          * scan backward for the read behind pages -- in memory
1499          */
1500         if (pindex > 0) {
1501                 if (rbehind > pindex) {
1502                         rbehind = pindex;
1503                         startpindex = 0;
1504                 } else {
1505                         startpindex = pindex - rbehind;
1506                 }
1507
1508                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1509                     rtm->pindex >= startpindex)
1510                         startpindex = rtm->pindex + 1;
1511
1512                 /* tpindex is unsigned; beware of numeric underflow. */
1513                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1514                     tpindex < pindex; i++, tpindex--) {
1515
1516                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1517                             VM_ALLOC_IFNOTCACHED);
1518                         if (rtm == NULL) {
1519                                 /*
1520                                  * Shift the allocated pages to the
1521                                  * beginning of the array.
1522                                  */
1523                                 for (j = 0; j < i; j++) {
1524                                         marray[j] = marray[j + tpindex + 1 -
1525                                             startpindex];
1526                                 }
1527                                 break;
1528                         }
1529
1530                         marray[tpindex - startpindex] = rtm;
1531                 }
1532         } else {
1533                 startpindex = 0;
1534                 i = 0;
1535         }
1536
1537         marray[i] = m;
1538         /* page offset of the required page */
1539         *reqpage = i;
1540
1541         tpindex = pindex + 1;
1542         i++;
1543
1544         /*
1545          * scan forward for the read ahead pages
1546          */
1547         endpindex = tpindex + rahead;
1548         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1549                 endpindex = rtm->pindex;
1550         if (endpindex > object->size)
1551                 endpindex = object->size;
1552
1553         for (; tpindex < endpindex; i++, tpindex++) {
1554
1555                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1556                     VM_ALLOC_IFNOTCACHED);
1557                 if (rtm == NULL) {
1558                         break;
1559                 }
1560
1561                 marray[i] = rtm;
1562         }
1563
1564         /* return number of pages */
1565         return i;
1566 }
1567
1568 /*
1569  * Block entry into the machine-independent layer's page fault handler by
1570  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1571  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1572  * spurious page faults.
1573  */
1574 int
1575 vm_fault_disable_pagefaults(void)
1576 {
1577
1578         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1579 }
1580
1581 void
1582 vm_fault_enable_pagefaults(int save)
1583 {
1584
1585         curthread_pflags_restore(save);
1586 }