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