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