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