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, growstack, 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         growstack = TRUE;
 349         PCPU_INC(cnt.v_vm_faults);
 350         fs.vp = NULL;
 351         faultcount = reqpage = 0;
 352
 353 RetryFault:;
 354
 355         /*
 356          * Find the backing store object and offset into it to begin the
 357          * search.
 358          */
 359         fs.map = map;
 360         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 361             &fs.first_object, &fs.first_pindex, &prot, &wired);
 362         if (result != KERN_SUCCESS) {
 363                 if (growstack && result == KERN_INVALID_ADDRESS &&
 364                     map != kernel_map) {
 365                         result = vm_map_growstack(curproc, vaddr);
 366                         if (result != KERN_SUCCESS)
 367                                 return (KERN_FAILURE);
 368                         growstack = FALSE;
 369                         goto RetryFault;
 370                 }
 371                 unlock_vp(&fs);
 372                 return (result);
 373         }
 374
 375         fs.map_generation = fs.map->timestamp;
 376
 377         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 378                 panic("vm_fault: fault on nofault entry, addr: %lx",
 379                     (u_long)vaddr);
 380         }
 381
 382         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 383             fs.entry->wiring_thread != curthread) {
 384                 vm_map_unlock_read(fs.map);
 385                 vm_map_lock(fs.map);
 386                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 387                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 388                         unlock_vp(&fs);
 389                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 390                         vm_map_unlock_and_wait(fs.map, 0);
 391                 } else
 392                         vm_map_unlock(fs.map);
 393                 goto RetryFault;
 394         }
 395
 396         if (wired)
 397                 fault_type = prot | (fault_type & VM_PROT_COPY);
 398         else
 399                 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
 400                     ("!wired && VM_FAULT_WIRE"));
 401
 402         /*
 403          * Try to avoid lock contention on the top-level object through
 404          * special-case handling of some types of page faults, specifically,
 405          * those that are both (1) mapping an existing page from the top-
 406          * level object and (2) not having to mark that object as containing
 407          * dirty pages.  Under these conditions, a read lock on the top-level
 408          * object suffices, allowing multiple page faults of a similar type to
 409          * run in parallel on the same top-level object.
 410          */
 411         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 412             (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
 413             /* avoid calling vm_object_set_writeable_dirty() */
 414             ((prot & VM_PROT_WRITE) == 0 ||
 415             (fs.first_object->type != OBJT_VNODE &&
 416             (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 417             (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
 418                 VM_OBJECT_RLOCK(fs.first_object);
 419                 if ((prot & VM_PROT_WRITE) == 0 ||
 420                     (fs.first_object->type != OBJT_VNODE &&
 421                     (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 422                     (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) {
 423                         result = vm_fault_soft_fast(&fs, vaddr, prot,
 424                             fault_type, fault_flags, wired, m_hold);
 425                         if (result == KERN_SUCCESS)
 426                                 return (result);
 427                 }
 428                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 429                         VM_OBJECT_RUNLOCK(fs.first_object);
 430                         VM_OBJECT_WLOCK(fs.first_object);
 431                 }
 432         } else {
 433                 VM_OBJECT_WLOCK(fs.first_object);
 434         }
 435
 436         /*
 437          * Make a reference to this object to prevent its disposal while we
 438          * are messing with it.  Once we have the reference, the map is free
 439          * to be diddled.  Since objects reference their shadows (and copies),
 440          * they will stay around as well.
 441          *
 442          * Bump the paging-in-progress count to prevent size changes (e.g.
 443          * truncation operations) during I/O.  This must be done after
 444          * obtaining the vnode lock in order to avoid possible deadlocks.
 445          */
 446         vm_object_reference_locked(fs.first_object);
 447         vm_object_pip_add(fs.first_object, 1);
 448
 449         fs.lookup_still_valid = TRUE;
 450
 451         fs.first_m = NULL;
 452
 453         /*
 454          * Search for the page at object/offset.
 455          */
 456         fs.object = fs.first_object;
 457         fs.pindex = fs.first_pindex;
 458         while (TRUE) {
 459                 /*
 460                  * If the object is marked for imminent termination,
 461                  * we retry here, since the collapse pass has raced
 462                  * with us.  Otherwise, if we see terminally dead
 463                  * object, return fail.
 464                  */
 465                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 466                         dead = fs.object->type == OBJT_DEAD;
 467                         unlock_and_deallocate(&fs);
 468                         if (dead)
 469                                 return (KERN_PROTECTION_FAILURE);
 470                         pause("vmf_de", 1);
 471                         goto RetryFault;
 472                 }
 473
 474                 /*
 475                  * See if page is resident
 476                  */
 477                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 478                 if (fs.m != NULL) {
 479                         /*
 480                          * Wait/Retry if the page is busy.  We have to do this
 481                          * if the page is either exclusive or shared busy
 482                          * because the vm_pager may be using read busy for
 483                          * pageouts (and even pageins if it is the vnode
 484                          * pager), and we could end up trying to pagein and
 485                          * pageout the same page simultaneously.
 486                          *
 487                          * We can theoretically allow the busy case on a read
 488                          * fault if the page is marked valid, but since such
 489                          * pages are typically already pmap'd, putting that
 490                          * special case in might be more effort then it is
 491                          * worth.  We cannot under any circumstances mess
 492                          * around with a shared busied page except, perhaps,
 493                          * to pmap it.
 494                          */
 495                         if (vm_page_busied(fs.m)) {
 496                                 /*
 497                                  * Reference the page before unlocking and
 498                                  * sleeping so that the page daemon is less
 499                                  * likely to reclaim it.
 500                                  */
 501                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 502                                 if (fs.object != fs.first_object) {
 503                                         if (!VM_OBJECT_TRYWLOCK(
 504                                             fs.first_object)) {
 505                                                 VM_OBJECT_WUNLOCK(fs.object);
 506                                                 VM_OBJECT_WLOCK(fs.first_object);
 507                                                 VM_OBJECT_WLOCK(fs.object);
 508                                         }
 509                                         vm_page_lock(fs.first_m);
 510                                         vm_page_free(fs.first_m);
 511                                         vm_page_unlock(fs.first_m);
 512                                         vm_object_pip_wakeup(fs.first_object);
 513                                         VM_OBJECT_WUNLOCK(fs.first_object);
 514                                         fs.first_m = NULL;
 515                                 }
 516                                 unlock_map(&fs);
 517                                 if (fs.m == vm_page_lookup(fs.object,
 518                                     fs.pindex)) {
 519                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 520                                 }
 521                                 vm_object_pip_wakeup(fs.object);
 522                                 VM_OBJECT_WUNLOCK(fs.object);
 523                                 PCPU_INC(cnt.v_intrans);
 524                                 vm_object_deallocate(fs.first_object);
 525                                 goto RetryFault;
 526                         }
 527                         vm_page_lock(fs.m);
 528                         vm_page_remque(fs.m);
 529                         vm_page_unlock(fs.m);
 530
 531                         /*
 532                          * Mark page busy for other processes, and the
 533                          * pagedaemon.  If it still isn't completely valid
 534                          * (readable), jump to readrest, else break-out ( we
 535                          * found the page ).
 536                          */
 537                         vm_page_xbusy(fs.m);
 538                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 539                                 goto readrest;
 540                         break;
 541                 }
 542
 543                 /*
 544                  * Page is not resident.  If this is the search termination
 545                  * or the pager might contain the page, allocate a new page.
 546                  * Default objects are zero-fill, there is no real pager.
 547                  */
 548                 if (fs.object->type != OBJT_DEFAULT ||
 549                     fs.object == fs.first_object) {
 550                         if (fs.pindex >= fs.object->size) {
 551                                 unlock_and_deallocate(&fs);
 552                                 return (KERN_PROTECTION_FAILURE);
 553                         }
 554
 555                         /*
 556                          * Allocate a new page for this object/offset pair.
 557                          *
 558                          * Unlocked read of the p_flag is harmless. At
 559                          * worst, the P_KILLED might be not observed
 560                          * there, and allocation can fail, causing
 561                          * restart and new reading of the p_flag.
 562                          */
 563                         fs.m = NULL;
 564                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 565 #if VM_NRESERVLEVEL > 0
 566                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 567                                         fs.object->flags |= OBJ_COLORED;
 568                                         fs.object->pg_color = atop(vaddr) -
 569                                             fs.pindex;
 570                                 }
 571 #endif
 572                                 alloc_req = P_KILLED(curproc) ?
 573                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 574                                 if (fs.object->type != OBJT_VNODE &&
 575                                     fs.object->backing_object == NULL)
 576                                         alloc_req |= VM_ALLOC_ZERO;
 577                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 578                                     alloc_req);
 579                         }
 580                         if (fs.m == NULL) {
 581                                 unlock_and_deallocate(&fs);
 582                                 VM_WAITPFAULT;
 583                                 goto RetryFault;
 584                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 585                                 break;
 586                 }
 587
 588 readrest:
 589                 /*
 590                  * We have found a valid page or we have allocated a new page.
 591                  * The page thus may not be valid or may not be entirely
 592                  * valid.
 593                  *
 594                  * Attempt to fault-in the page if there is a chance that the
 595                  * pager has it, and potentially fault in additional pages
 596                  * at the same time.  For default objects simply provide
 597                  * zero-filled pages.
 598                  */
 599                 if (fs.object->type != OBJT_DEFAULT) {
 600                         int rv;
 601                         u_char behavior = vm_map_entry_behavior(fs.entry);
 602
 603                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 604                             P_KILLED(curproc)) {
 605                                 behind = 0;
 606                                 ahead = 0;
 607                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 608                                 behind = 0;
 609                                 ahead = atop(fs.entry->end - vaddr) - 1;
 610                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 611                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 612                                 if (fs.pindex == fs.entry->next_read)
 613                                         vm_fault_cache_behind(&fs,
 614                                             VM_FAULT_READ_MAX);
 615                         } else {
 616                                 /*
 617                                  * If this is a sequential page fault, then
 618                                  * arithmetically increase the number of pages
 619                                  * in the read-ahead window.  Otherwise, reset
 620                                  * the read-ahead window to its smallest size.
 621                                  */
 622                                 behind = atop(vaddr - fs.entry->start);
 623                                 if (behind > VM_FAULT_READ_BEHIND)
 624                                         behind = VM_FAULT_READ_BEHIND;
 625                                 ahead = atop(fs.entry->end - vaddr) - 1;
 626                                 era = fs.entry->read_ahead;
 627                                 if (fs.pindex == fs.entry->next_read) {
 628                                         nera = era + behind;
 629                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 630                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 631                                         behind = 0;
 632                                         if (ahead > nera)
 633                                                 ahead = nera;
 634                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 635                                                 vm_fault_cache_behind(&fs,
 636                                                     VM_FAULT_CACHE_BEHIND);
 637                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 638                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 639                                 if (era != ahead)
 640                                         fs.entry->read_ahead = ahead;
 641                         }
 642
 643                         /*
 644                          * Call the pager to retrieve the data, if any, after
 645                          * releasing the lock on the map.  We hold a ref on
 646                          * fs.object and the pages are exclusive busied.
 647                          */
 648                         unlock_map(&fs);
 649
 650                         if (fs.object->type == OBJT_VNODE &&
 651                             (vp = fs.object->handle) != fs.vp) {
 652                                 unlock_vp(&fs);
 653                                 locked = VOP_ISLOCKED(vp);
 654
 655                                 if (locked != LK_EXCLUSIVE)
 656                                         locked = LK_SHARED;
 657                                 /* Do not sleep for vnode lock while fs.m is busy */
 658                                 error = vget(vp, locked | LK_CANRECURSE |
 659                                     LK_NOWAIT, curthread);
 660                                 if (error != 0) {
 661                                         vhold(vp);
 662                                         release_page(&fs);
 663                                         unlock_and_deallocate(&fs);
 664                                         error = vget(vp, locked | LK_RETRY |
 665                                             LK_CANRECURSE, curthread);
 666                                         vdrop(vp);
 667                                         fs.vp = vp;
 668                                         KASSERT(error == 0,
 669                                             ("vm_fault: vget failed"));
 670                                         goto RetryFault;
 671                                 }
 672                                 fs.vp = vp;
 673                         }
 674                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 675                             ("vm_fault: vnode-backed object mapped by system map"));
 676
 677                         /*
 678                          * now we find out if any other pages should be paged
 679                          * in at this time this routine checks to see if the
 680                          * pages surrounding this fault reside in the same
 681                          * object as the page for this fault.  If they do,
 682                          * then they are faulted in also into the object.  The
 683                          * array "marray" returned contains an array of
 684                          * vm_page_t structs where one of them is the
 685                          * vm_page_t passed to the routine.  The reqpage
 686                          * return value is the index into the marray for the
 687                          * vm_page_t passed to the routine.
 688                          *
 689                          * fs.m plus the additional pages are exclusive busied.
 690                          */
 691                         faultcount = vm_fault_additional_pages(
 692                             fs.m, behind, ahead, marray, &reqpage);
 693
 694                         rv = faultcount ?
 695                             vm_pager_get_pages(fs.object, marray, faultcount,
 696                                 reqpage) : VM_PAGER_FAIL;
 697
 698                         if (rv == VM_PAGER_OK) {
 699                                 /*
 700                                  * Found the page. Leave it busy while we play
 701                                  * with it.
 702                                  */
 703
 704                                 /*
 705                                  * Relookup in case pager changed page. Pager
 706                                  * is responsible for disposition of old page
 707                                  * if moved.
 708                                  */
 709                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 710                                 if (!fs.m) {
 711                                         unlock_and_deallocate(&fs);
 712                                         goto RetryFault;
 713                                 }
 714
 715                                 hardfault++;
 716                                 break; /* break to PAGE HAS BEEN FOUND */
 717                         }
 718                         /*
 719                          * Remove the bogus page (which does not exist at this
 720                          * object/offset); before doing so, we must get back
 721                          * our object lock to preserve our invariant.
 722                          *
 723                          * Also wake up any other process that may want to bring
 724                          * in this page.
 725                          *
 726                          * If this is the top-level object, we must leave the
 727                          * busy page to prevent another process from rushing
 728                          * past us, and inserting the page in that object at
 729                          * the same time that we are.
 730                          */
 731                         if (rv == VM_PAGER_ERROR)
 732                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 733                                     curproc->p_pid, curproc->p_comm);
 734                         /*
 735                          * Data outside the range of the pager or an I/O error
 736                          */
 737                         /*
 738                          * XXX - the check for kernel_map is a kludge to work
 739                          * around having the machine panic on a kernel space
 740                          * fault w/ I/O error.
 741                          */
 742                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 743                                 (rv == VM_PAGER_BAD)) {
 744                                 vm_page_lock(fs.m);
 745                                 vm_page_free(fs.m);
 746                                 vm_page_unlock(fs.m);
 747                                 fs.m = NULL;
 748                                 unlock_and_deallocate(&fs);
 749                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 750                         }
 751                         if (fs.object != fs.first_object) {
 752                                 vm_page_lock(fs.m);
 753                                 vm_page_free(fs.m);
 754                                 vm_page_unlock(fs.m);
 755                                 fs.m = NULL;
 756                                 /*
 757                                  * XXX - we cannot just fall out at this
 758                                  * point, m has been freed and is invalid!
 759                                  */
 760                         }
 761                 }
 762
 763                 /*
 764                  * We get here if the object has default pager (or unwiring)
 765                  * or the pager doesn't have the page.
 766                  */
 767                 if (fs.object == fs.first_object)
 768                         fs.first_m = fs.m;
 769
 770                 /*
 771                  * Move on to the next object.  Lock the next object before
 772                  * unlocking the current one.
 773                  */
 774                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 775                 next_object = fs.object->backing_object;
 776                 if (next_object == NULL) {
 777                         /*
 778                          * If there's no object left, fill the page in the top
 779                          * object with zeros.
 780                          */
 781                         if (fs.object != fs.first_object) {
 782                                 vm_object_pip_wakeup(fs.object);
 783                                 VM_OBJECT_WUNLOCK(fs.object);
 784
 785                                 fs.object = fs.first_object;
 786                                 fs.pindex = fs.first_pindex;
 787                                 fs.m = fs.first_m;
 788                                 VM_OBJECT_WLOCK(fs.object);
 789                         }
 790                         fs.first_m = NULL;
 791
 792                         /*
 793                          * Zero the page if necessary and mark it valid.
 794                          */
 795                         if ((fs.m->flags & PG_ZERO) == 0) {
 796                                 pmap_zero_page(fs.m);
 797                         } else {
 798                                 PCPU_INC(cnt.v_ozfod);
 799                         }
 800                         PCPU_INC(cnt.v_zfod);
 801                         fs.m->valid = VM_PAGE_BITS_ALL;
 802                         /* Don't try to prefault neighboring pages. */
 803                         faultcount = 1;
 804                         break;  /* break to PAGE HAS BEEN FOUND */
 805                 } else {
 806                         KASSERT(fs.object != next_object,
 807                             ("object loop %p", next_object));
 808                         VM_OBJECT_WLOCK(next_object);
 809                         vm_object_pip_add(next_object, 1);
 810                         if (fs.object != fs.first_object)
 811                                 vm_object_pip_wakeup(fs.object);
 812                         VM_OBJECT_WUNLOCK(fs.object);
 813                         fs.object = next_object;
 814                 }
 815         }
 816
 817         vm_page_assert_xbusied(fs.m);
 818
 819         /*
 820          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 821          * is held.]
 822          */
 823
 824         /*
 825          * If the page is being written, but isn't already owned by the
 826          * top-level object, we have to copy it into a new page owned by the
 827          * top-level object.
 828          */
 829         if (fs.object != fs.first_object) {
 830                 /*
 831                  * We only really need to copy if we want to write it.
 832                  */
 833                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 834                         /*
 835                          * This allows pages to be virtually copied from a
 836                          * backing_object into the first_object, where the
 837                          * backing object has no other refs to it, and cannot
 838                          * gain any more refs.  Instead of a bcopy, we just
 839                          * move the page from the backing object to the
 840                          * first object.  Note that we must mark the page
 841                          * dirty in the first object so that it will go out
 842                          * to swap when needed.
 843                          */
 844                         is_first_object_locked = FALSE;
 845                         if (
 846                                 /*
 847                                  * Only one shadow object
 848                                  */
 849                                 (fs.object->shadow_count == 1) &&
 850                                 /*
 851                                  * No COW refs, except us
 852                                  */
 853                                 (fs.object->ref_count == 1) &&
 854                                 /*
 855                                  * No one else can look this object up
 856                                  */
 857                                 (fs.object->handle == NULL) &&
 858                                 /*
 859                                  * No other ways to look the object up
 860                                  */
 861                                 ((fs.object->type == OBJT_DEFAULT) ||
 862                                  (fs.object->type == OBJT_SWAP)) &&
 863                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
 864                                 /*
 865                                  * We don't chase down the shadow chain
 866                                  */
 867                             fs.object == fs.first_object->backing_object) {
 868                                 /*
 869                                  * get rid of the unnecessary page
 870                                  */
 871                                 vm_page_lock(fs.first_m);
 872                                 vm_page_free(fs.first_m);
 873                                 vm_page_unlock(fs.first_m);
 874                                 /*
 875                                  * grab the page and put it into the
 876                                  * process'es object.  The page is
 877                                  * automatically made dirty.
 878                                  */
 879                                 if (vm_page_rename(fs.m, fs.first_object,
 880                                     fs.first_pindex)) {
 881                                         unlock_and_deallocate(&fs);
 882                                         goto RetryFault;
 883                                 }
 884 #if VM_NRESERVLEVEL > 0
 885                                 /*
 886                                  * Rename the reservation.
 887                                  */
 888                                 vm_reserv_rename(fs.m, fs.first_object,
 889                                     fs.object, OFF_TO_IDX(
 890                                     fs.first_object->backing_object_offset));
 891 #endif
 892                                 vm_page_xbusy(fs.m);
 893                                 fs.first_m = fs.m;
 894                                 fs.m = NULL;
 895                                 PCPU_INC(cnt.v_cow_optim);
 896                         } else {
 897                                 /*
 898                                  * Oh, well, lets copy it.
 899                                  */
 900                                 pmap_copy_page(fs.m, fs.first_m);
 901                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 902                                 if (wired && (fault_flags &
 903                                     VM_FAULT_WIRE) == 0) {
 904                                         vm_page_lock(fs.first_m);
 905                                         vm_page_wire(fs.first_m);
 906                                         vm_page_unlock(fs.first_m);
 907
 908                                         vm_page_lock(fs.m);
 909                                         vm_page_unwire(fs.m, FALSE);
 910                                         vm_page_unlock(fs.m);
 911                                 }
 912                                 /*
 913                                  * We no longer need the old page or object.
 914                                  */
 915                                 release_page(&fs);
 916                         }
 917                         /*
 918                          * fs.object != fs.first_object due to above
 919                          * conditional
 920                          */
 921                         vm_object_pip_wakeup(fs.object);
 922                         VM_OBJECT_WUNLOCK(fs.object);
 923                         /*
 924                          * Only use the new page below...
 925                          */
 926                         fs.object = fs.first_object;
 927                         fs.pindex = fs.first_pindex;
 928                         fs.m = fs.first_m;
 929                         if (!is_first_object_locked)
 930                                 VM_OBJECT_WLOCK(fs.object);
 931                         PCPU_INC(cnt.v_cow_faults);
 932                         curthread->td_cow++;
 933                 } else {
 934                         prot &= ~VM_PROT_WRITE;
 935                 }
 936         }
 937
 938         /*
 939          * We must verify that the maps have not changed since our last
 940          * lookup.
 941          */
 942         if (!fs.lookup_still_valid) {
 943                 vm_object_t retry_object;
 944                 vm_pindex_t retry_pindex;
 945                 vm_prot_t retry_prot;
 946
 947                 if (!vm_map_trylock_read(fs.map)) {
 948                         release_page(&fs);
 949                         unlock_and_deallocate(&fs);
 950                         goto RetryFault;
 951                 }
 952                 fs.lookup_still_valid = TRUE;
 953                 if (fs.map->timestamp != fs.map_generation) {
 954                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 955                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 956
 957                         /*
 958                          * If we don't need the page any longer, put it on the inactive
 959                          * list (the easiest thing to do here).  If no one needs it,
 960                          * pageout will grab it eventually.
 961                          */
 962                         if (result != KERN_SUCCESS) {
 963                                 release_page(&fs);
 964                                 unlock_and_deallocate(&fs);
 965
 966                                 /*
 967                                  * If retry of map lookup would have blocked then
 968                                  * retry fault from start.
 969                                  */
 970                                 if (result == KERN_FAILURE)
 971                                         goto RetryFault;
 972                                 return (result);
 973                         }
 974                         if ((retry_object != fs.first_object) ||
 975                             (retry_pindex != fs.first_pindex)) {
 976                                 release_page(&fs);
 977                                 unlock_and_deallocate(&fs);
 978                                 goto RetryFault;
 979                         }
 980
 981                         /*
 982                          * Check whether the protection has changed or the object has
 983                          * been copied while we left the map unlocked. Changing from
 984                          * read to write permission is OK - we leave the page
 985                          * write-protected, and catch the write fault. Changing from
 986                          * write to read permission means that we can't mark the page
 987                          * write-enabled after all.
 988                          */
 989                         prot &= retry_prot;
 990                 }
 991         }
 992         /*
 993          * If the page was filled by a pager, update the map entry's
 994          * last read offset.  Since the pager does not return the
 995          * actual set of pages that it read, this update is based on
 996          * the requested set.  Typically, the requested and actual
 997          * sets are the same.
 998          *
 999          * XXX The following assignment modifies the map
1000          * without holding a write lock on it.
1001          */
1002         if (hardfault)
1003                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
1004
1005         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1006         vm_page_assert_xbusied(fs.m);
1007
1008         /*
1009          * Page must be completely valid or it is not fit to
1010          * map into user space.  vm_pager_get_pages() ensures this.
1011          */
1012         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
1013             ("vm_fault: page %p partially invalid", fs.m));
1014         VM_OBJECT_WUNLOCK(fs.object);
1015
1016         /*
1017          * Put this page into the physical map.  We had to do the unlock above
1018          * because pmap_enter() may sleep.  We don't put the page
1019          * back on the active queue until later so that the pageout daemon
1020          * won't find it (yet).
1021          */
1022         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1023             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1024         if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1025             wired == 0)
1026                 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
1027         VM_OBJECT_WLOCK(fs.object);
1028         vm_page_lock(fs.m);
1029
1030         /*
1031          * If the page is not wired down, then put it where the pageout daemon
1032          * can find it.
1033          */
1034         if ((fault_flags & VM_FAULT_WIRE) != 0) {
1035                 KASSERT(wired, ("VM_FAULT_WIRE && !wired"));
1036                 vm_page_wire(fs.m);
1037         } else
1038                 vm_page_activate(fs.m);
1039         if (m_hold != NULL) {
1040                 *m_hold = fs.m;
1041                 vm_page_hold(fs.m);
1042         }
1043         vm_page_unlock(fs.m);
1044         vm_page_xunbusy(fs.m);
1045
1046         /*
1047          * Unlock everything, and return
1048          */
1049         unlock_and_deallocate(&fs);
1050         if (hardfault) {
1051                 PCPU_INC(cnt.v_io_faults);
1052                 curthread->td_ru.ru_majflt++;
1053         } else
1054                 curthread->td_ru.ru_minflt++;
1055
1056         return (KERN_SUCCESS);
1057 }
1058
1059 /*
1060  * Speed up the reclamation of up to "distance" pages that precede the
1061  * faulting pindex within the first object of the shadow chain.
1062  */
1063 static void
1064 vm_fault_cache_behind(const struct faultstate *fs, int distance)
1065 {
1066         vm_object_t first_object, object;
1067         vm_page_t m, m_prev;
1068         vm_pindex_t pindex;
1069
1070         object = fs->object;
1071         VM_OBJECT_ASSERT_WLOCKED(object);
1072         first_object = fs->first_object;
1073         if (first_object != object) {
1074                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1075                         VM_OBJECT_WUNLOCK(object);
1076                         VM_OBJECT_WLOCK(first_object);
1077                         VM_OBJECT_WLOCK(object);
1078                 }
1079         }
1080         /* Neither fictitious nor unmanaged pages can be cached. */
1081         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1082                 if (fs->first_pindex < distance)
1083                         pindex = 0;
1084                 else
1085                         pindex = fs->first_pindex - distance;
1086                 if (pindex < OFF_TO_IDX(fs->entry->offset))
1087                         pindex = OFF_TO_IDX(fs->entry->offset);
1088                 m = first_object != object ? fs->first_m : fs->m;
1089                 vm_page_assert_xbusied(m);
1090                 m_prev = vm_page_prev(m);
1091                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1092                     m->valid == VM_PAGE_BITS_ALL) {
1093                         m_prev = vm_page_prev(m);
1094                         if (vm_page_busied(m))
1095                                 continue;
1096                         vm_page_lock(m);
1097                         if (m->hold_count == 0 && m->wire_count == 0) {
1098                                 pmap_remove_all(m);
1099                                 vm_page_aflag_clear(m, PGA_REFERENCED);
1100                                 if (m->dirty != 0)
1101                                         vm_page_deactivate(m);
1102                                 else
1103                                         vm_page_cache(m);
1104                         }
1105                         vm_page_unlock(m);
1106                 }
1107         }
1108         if (first_object != object)
1109                 VM_OBJECT_WUNLOCK(first_object);
1110 }
1111
1112 /*
1113  * vm_fault_prefault provides a quick way of clustering
1114  * pagefaults into a processes address space.  It is a "cousin"
1115  * of vm_map_pmap_enter, except it runs at page fault time instead
1116  * of mmap time.
1117  */
1118 static void
1119 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1120     int faultcount, int reqpage)
1121 {
1122         pmap_t pmap;
1123         vm_map_entry_t entry;
1124         vm_object_t backing_object, lobject;
1125         vm_offset_t addr, starta;
1126         vm_pindex_t pindex;
1127         vm_page_t m;
1128         int backward, forward, i;
1129
1130         pmap = fs->map->pmap;
1131         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1132                 return;
1133
1134         if (faultcount > 0) {
1135                 backward = reqpage;
1136                 forward = faultcount - reqpage - 1;
1137         } else {
1138                 backward = PFBAK;
1139                 forward = PFFOR;
1140         }
1141         entry = fs->entry;
1142
1143         if (addra < backward * PAGE_SIZE) {
1144                 starta = entry->start;
1145         } else {
1146                 starta = addra - backward * PAGE_SIZE;
1147                 if (starta < entry->start)
1148                         starta = entry->start;
1149         }
1150
1151         /*
1152          * Generate the sequence of virtual addresses that are candidates for
1153          * prefaulting in an outward spiral from the faulting virtual address,
1154          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1155          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1156          * If the candidate address doesn't have a backing physical page, then
1157          * the loop immediately terminates.
1158          */
1159         for (i = 0; i < 2 * imax(backward, forward); i++) {
1160                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1161                     PAGE_SIZE);
1162                 if (addr > addra + forward * PAGE_SIZE)
1163                         addr = 0;
1164
1165                 if (addr < starta || addr >= entry->end)
1166                         continue;
1167
1168                 if (!pmap_is_prefaultable(pmap, addr))
1169                         continue;
1170
1171                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1172                 lobject = entry->object.vm_object;
1173                 VM_OBJECT_RLOCK(lobject);
1174                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1175                     lobject->type == OBJT_DEFAULT &&
1176                     (backing_object = lobject->backing_object) != NULL) {
1177                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1178                             0, ("vm_fault_prefault: unaligned object offset"));
1179                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1180                         VM_OBJECT_RLOCK(backing_object);
1181                         VM_OBJECT_RUNLOCK(lobject);
1182                         lobject = backing_object;
1183                 }
1184                 if (m == NULL) {
1185                         VM_OBJECT_RUNLOCK(lobject);
1186                         break;
1187                 }
1188                 if (m->valid == VM_PAGE_BITS_ALL &&
1189                     (m->flags & PG_FICTITIOUS) == 0)
1190                         pmap_enter_quick(pmap, addr, m, entry->protection);
1191                 VM_OBJECT_RUNLOCK(lobject);
1192         }
1193 }
1194
1195 /*
1196  * Hold each of the physical pages that are mapped by the specified range of
1197  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1198  * and allow the specified types of access, "prot".  If all of the implied
1199  * pages are successfully held, then the number of held pages is returned
1200  * together with pointers to those pages in the array "ma".  However, if any
1201  * of the pages cannot be held, -1 is returned.
1202  */
1203 int
1204 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1205     vm_prot_t prot, vm_page_t *ma, int max_count)
1206 {
1207         vm_offset_t end, va;
1208         vm_page_t *mp;
1209         int count;
1210         boolean_t pmap_failed;
1211
1212         if (len == 0)
1213                 return (0);
1214         end = round_page(addr + len);
1215         addr = trunc_page(addr);
1216
1217         /*
1218          * Check for illegal addresses.
1219          */
1220         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1221                 return (-1);
1222
1223         if (atop(end - addr) > max_count)
1224                 panic("vm_fault_quick_hold_pages: count > max_count");
1225         count = atop(end - addr);
1226
1227         /*
1228          * Most likely, the physical pages are resident in the pmap, so it is
1229          * faster to try pmap_extract_and_hold() first.
1230          */
1231         pmap_failed = FALSE;
1232         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1233                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1234                 if (*mp == NULL)
1235                         pmap_failed = TRUE;
1236                 else if ((prot & VM_PROT_WRITE) != 0 &&
1237                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1238                         /*
1239                          * Explicitly dirty the physical page.  Otherwise, the
1240                          * caller's changes may go unnoticed because they are
1241                          * performed through an unmanaged mapping or by a DMA
1242                          * operation.
1243                          *
1244                          * The object lock is not held here.
1245                          * See vm_page_clear_dirty_mask().
1246                          */
1247                         vm_page_dirty(*mp);
1248                 }
1249         }
1250         if (pmap_failed) {
1251                 /*
1252                  * One or more pages could not be held by the pmap.  Either no
1253                  * page was mapped at the specified virtual address or that
1254                  * mapping had insufficient permissions.  Attempt to fault in
1255                  * and hold these pages.
1256                  */
1257                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1258                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1259                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1260                                 goto error;
1261         }
1262         return (count);
1263 error:
1264         for (mp = ma; mp < ma + count; mp++)
1265                 if (*mp != NULL) {
1266                         vm_page_lock(*mp);
1267                         vm_page_unhold(*mp);
1268                         vm_page_unlock(*mp);
1269                 }
1270         return (-1);
1271 }
1272
1273 /*
1274  *      Routine:
1275  *              vm_fault_copy_entry
1276  *      Function:
1277  *              Create new shadow object backing dst_entry with private copy of
1278  *              all underlying pages. When src_entry is equal to dst_entry,
1279  *              function implements COW for wired-down map entry. Otherwise,
1280  *              it forks wired entry into dst_map.
1281  *
1282  *      In/out conditions:
1283  *              The source and destination maps must be locked for write.
1284  *              The source map entry must be wired down (or be a sharing map
1285  *              entry corresponding to a main map entry that is wired down).
1286  */
1287 void
1288 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1289     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1290     vm_ooffset_t *fork_charge)
1291 {
1292         vm_object_t backing_object, dst_object, object, src_object;
1293         vm_pindex_t dst_pindex, pindex, src_pindex;
1294         vm_prot_t access, prot;
1295         vm_offset_t vaddr;
1296         vm_page_t dst_m;
1297         vm_page_t src_m;
1298         boolean_t upgrade;
1299
1300 #ifdef  lint
1301         src_map++;
1302 #endif  /* lint */
1303
1304         upgrade = src_entry == dst_entry;
1305         access = prot = dst_entry->protection;
1306
1307         src_object = src_entry->object.vm_object;
1308         src_pindex = OFF_TO_IDX(src_entry->offset);
1309
1310         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1311                 dst_object = src_object;
1312                 vm_object_reference(dst_object);
1313         } else {
1314                 /*
1315                  * Create the top-level object for the destination entry. (Doesn't
1316                  * actually shadow anything - we copy the pages directly.)
1317                  */
1318                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1319                     OFF_TO_IDX(dst_entry->end - dst_entry->start));
1320 #if VM_NRESERVLEVEL > 0
1321                 dst_object->flags |= OBJ_COLORED;
1322                 dst_object->pg_color = atop(dst_entry->start);
1323 #endif
1324         }
1325
1326         VM_OBJECT_WLOCK(dst_object);
1327         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1328             ("vm_fault_copy_entry: vm_object not NULL"));
1329         if (src_object != dst_object) {
1330                 dst_entry->object.vm_object = dst_object;
1331                 dst_entry->offset = 0;
1332                 dst_object->charge = dst_entry->end - dst_entry->start;
1333         }
1334         if (fork_charge != NULL) {
1335                 KASSERT(dst_entry->cred == NULL,
1336                     ("vm_fault_copy_entry: leaked swp charge"));
1337                 dst_object->cred = curthread->td_ucred;
1338                 crhold(dst_object->cred);
1339                 *fork_charge += dst_object->charge;
1340         } else if (dst_object->cred == NULL) {
1341                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1342                     dst_entry));
1343                 dst_object->cred = dst_entry->cred;
1344                 dst_entry->cred = NULL;
1345         }
1346
1347         /*
1348          * If not an upgrade, then enter the mappings in the pmap as
1349          * read and/or execute accesses.  Otherwise, enter them as
1350          * write accesses.
1351          *
1352          * A writeable large page mapping is only created if all of
1353          * the constituent small page mappings are modified. Marking
1354          * PTEs as modified on inception allows promotion to happen
1355          * without taking potentially large number of soft faults.
1356          */
1357         if (!upgrade)
1358                 access &= ~VM_PROT_WRITE;
1359
1360         /*
1361          * Loop through all of the virtual pages within the entry's
1362          * range, copying each page from the source object to the
1363          * destination object.  Since the source is wired, those pages
1364          * must exist.  In contrast, the destination is pageable.
1365          * Since the destination object does share any backing storage
1366          * with the source object, all of its pages must be dirtied,
1367          * regardless of whether they can be written.
1368          */
1369         for (vaddr = dst_entry->start, dst_pindex = 0;
1370             vaddr < dst_entry->end;
1371             vaddr += PAGE_SIZE, dst_pindex++) {
1372 again:
1373                 /*
1374                  * Find the page in the source object, and copy it in.
1375                  * Because the source is wired down, the page will be
1376                  * in memory.
1377                  */
1378                 if (src_object != dst_object)
1379                         VM_OBJECT_RLOCK(src_object);
1380                 object = src_object;
1381                 pindex = src_pindex + dst_pindex;
1382                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1383                     (backing_object = object->backing_object) != NULL) {
1384                         /*
1385                          * Unless the source mapping is read-only or
1386                          * it is presently being upgraded from
1387                          * read-only, the first object in the shadow
1388                          * chain should provide all of the pages.  In
1389                          * other words, this loop body should never be
1390                          * executed when the source mapping is already
1391                          * read/write.
1392                          */
1393                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1394                             upgrade,
1395                             ("vm_fault_copy_entry: main object missing page"));
1396
1397                         VM_OBJECT_RLOCK(backing_object);
1398                         pindex += OFF_TO_IDX(object->backing_object_offset);
1399                         if (object != dst_object)
1400                                 VM_OBJECT_RUNLOCK(object);
1401                         object = backing_object;
1402                 }
1403                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1404
1405                 if (object != dst_object) {
1406                         /*
1407                          * Allocate a page in the destination object.
1408                          */
1409                         dst_m = vm_page_alloc(dst_object, (src_object ==
1410                             dst_object ? src_pindex : 0) + dst_pindex,
1411                             VM_ALLOC_NORMAL);
1412                         if (dst_m == NULL) {
1413                                 VM_OBJECT_WUNLOCK(dst_object);
1414                                 VM_OBJECT_RUNLOCK(object);
1415                                 VM_WAIT;
1416                                 VM_OBJECT_WLOCK(dst_object);
1417                                 goto again;
1418                         }
1419                         pmap_copy_page(src_m, dst_m);
1420                         VM_OBJECT_RUNLOCK(object);
1421                         dst_m->valid = VM_PAGE_BITS_ALL;
1422                         dst_m->dirty = VM_PAGE_BITS_ALL;
1423                 } else {
1424                         dst_m = src_m;
1425                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1426                                 goto again;
1427                         vm_page_xbusy(dst_m);
1428                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1429                             ("invalid dst page %p", dst_m));
1430                 }
1431                 VM_OBJECT_WUNLOCK(dst_object);
1432
1433                 /*
1434                  * Enter it in the pmap. If a wired, copy-on-write
1435                  * mapping is being replaced by a write-enabled
1436                  * mapping, then wire that new mapping.
1437                  */
1438                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1439                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1440
1441                 /*
1442                  * Mark it no longer busy, and put it on the active list.
1443                  */
1444                 VM_OBJECT_WLOCK(dst_object);
1445
1446                 if (upgrade) {
1447                         if (src_m != dst_m) {
1448                                 vm_page_lock(src_m);
1449                                 vm_page_unwire(src_m, 0);
1450                                 vm_page_unlock(src_m);
1451                                 vm_page_lock(dst_m);
1452                                 vm_page_wire(dst_m);
1453                                 vm_page_unlock(dst_m);
1454                         } else {
1455                                 KASSERT(dst_m->wire_count > 0,
1456                                     ("dst_m %p is not wired", dst_m));
1457                         }
1458                 } else {
1459                         vm_page_lock(dst_m);
1460                         vm_page_activate(dst_m);
1461                         vm_page_unlock(dst_m);
1462                 }
1463                 vm_page_xunbusy(dst_m);
1464         }
1465         VM_OBJECT_WUNLOCK(dst_object);
1466         if (upgrade) {
1467                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1468                 vm_object_deallocate(src_object);
1469         }
1470 }
1471
1472
1473 /*
1474  * This routine checks around the requested page for other pages that
1475  * might be able to be faulted in.  This routine brackets the viable
1476  * pages for the pages to be paged in.
1477  *
1478  * Inputs:
1479  *      m, rbehind, rahead
1480  *
1481  * Outputs:
1482  *  marray (array of vm_page_t), reqpage (index of requested page)
1483  *
1484  * Return value:
1485  *  number of pages in marray
1486  */
1487 static int
1488 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1489         vm_page_t m;
1490         int rbehind;
1491         int rahead;
1492         vm_page_t *marray;
1493         int *reqpage;
1494 {
1495         int i,j;
1496         vm_object_t object;
1497         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1498         vm_page_t rtm;
1499         int cbehind, cahead;
1500
1501         VM_OBJECT_ASSERT_WLOCKED(m->object);
1502
1503         object = m->object;
1504         pindex = m->pindex;
1505         cbehind = cahead = 0;
1506
1507         /*
1508          * if the requested page is not available, then give up now
1509          */
1510         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1511                 return 0;
1512         }
1513
1514         if ((cbehind == 0) && (cahead == 0)) {
1515                 *reqpage = 0;
1516                 marray[0] = m;
1517                 return 1;
1518         }
1519
1520         if (rahead > cahead) {
1521                 rahead = cahead;
1522         }
1523
1524         if (rbehind > cbehind) {
1525                 rbehind = cbehind;
1526         }
1527
1528         /*
1529          * scan backward for the read behind pages -- in memory
1530          */
1531         if (pindex > 0) {
1532                 if (rbehind > pindex) {
1533                         rbehind = pindex;
1534                         startpindex = 0;
1535                 } else {
1536                         startpindex = pindex - rbehind;
1537                 }
1538
1539                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1540                     rtm->pindex >= startpindex)
1541                         startpindex = rtm->pindex + 1;
1542
1543                 /* tpindex is unsigned; beware of numeric underflow. */
1544                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1545                     tpindex < pindex; i++, tpindex--) {
1546
1547                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1548                             VM_ALLOC_IFNOTCACHED);
1549                         if (rtm == NULL) {
1550                                 /*
1551                                  * Shift the allocated pages to the
1552                                  * beginning of the array.
1553                                  */
1554                                 for (j = 0; j < i; j++) {
1555                                         marray[j] = marray[j + tpindex + 1 -
1556                                             startpindex];
1557                                 }
1558                                 break;
1559                         }
1560
1561                         marray[tpindex - startpindex] = rtm;
1562                 }
1563         } else {
1564                 startpindex = 0;
1565                 i = 0;
1566         }
1567
1568         marray[i] = m;
1569         /* page offset of the required page */
1570         *reqpage = i;
1571
1572         tpindex = pindex + 1;
1573         i++;
1574
1575         /*
1576          * scan forward for the read ahead pages
1577          */
1578         endpindex = tpindex + rahead;
1579         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1580                 endpindex = rtm->pindex;
1581         if (endpindex > object->size)
1582                 endpindex = object->size;
1583
1584         for (; tpindex < endpindex; i++, tpindex++) {
1585
1586                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1587                     VM_ALLOC_IFNOTCACHED);
1588                 if (rtm == NULL) {
1589                         break;
1590                 }
1591
1592                 marray[i] = rtm;
1593         }
1594
1595         /* return number of pages */
1596         return i;
1597 }
1598
1599 /*
1600  * Block entry into the machine-independent layer's page fault handler by
1601  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1602  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1603  * spurious page faults.
1604  */
1605 int
1606 vm_fault_disable_pagefaults(void)
1607 {
1608
1609         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1610 }
1611
1612 void
1613 vm_fault_enable_pagefaults(int save)
1614 {
1615
1616         curthread_pflags_restore(save);
1617 }