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  * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
  70  * $DragonFly: src/sys/vm/vm_fault.c,v 1.42 2007/06/07 23:00:39 dillon Exp $
  71  */
  72
  73 /*
  74  *      Page fault handling module.
  75  */
  76
  77 #include <sys/param.h>
  78 #include <sys/systm.h>
  79 #include <sys/kernel.h>
  80 #include <sys/proc.h>
  81 #include <sys/vnode.h>
  82 #include <sys/resourcevar.h>
  83 #include <sys/vmmeter.h>
  84 #include <sys/vkernel.h>
  85 #include <sys/sfbuf.h>
  86 #include <sys/lock.h>
  87
  88 #include <vm/vm.h>
  89 #include <vm/vm_param.h>
  90 #include <vm/pmap.h>
  91 #include <vm/vm_map.h>
  92 #include <vm/vm_object.h>
  93 #include <vm/vm_page.h>
  94 #include <vm/vm_pageout.h>
  95 #include <vm/vm_kern.h>
  96 #include <vm/vm_pager.h>
  97 #include <vm/vnode_pager.h>
  98 #include <vm/vm_extern.h>
  99
 100 #include <sys/thread2.h>
 101 #include <vm/vm_page2.h>
 102
 103 #define VM_FAULT_READ_AHEAD 8
 104 #define VM_FAULT_READ_BEHIND 7
 105 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 106
 107 struct faultstate {
 108         vm_page_t m;
 109         vm_object_t object;
 110         vm_pindex_t pindex;
 111         vm_prot_t prot;
 112         vm_page_t first_m;
 113         vm_object_t first_object;
 114         vm_prot_t first_prot;
 115         vm_map_t map;
 116         vm_map_entry_t entry;
 117         int lookup_still_valid;
 118         int didlimit;
 119         int hardfault;
 120         int fault_flags;
 121         int map_generation;
 122         boolean_t wired;
 123         struct vnode *vp;
 124 };
 125
 126 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
 127 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
 128 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
 129 static int vm_fault_ratelimit(struct vmspace *);
 130
 131 static __inline void
 132 release_page(struct faultstate *fs)
 133 {
 134         vm_page_wakeup(fs->m);
 135         vm_page_deactivate(fs->m);
 136         fs->m = NULL;
 137 }
 138
 139 static __inline void
 140 unlock_map(struct faultstate *fs)
 141 {
 142         if (fs->lookup_still_valid && fs->map) {
 143                 vm_map_lookup_done(fs->map, fs->entry, 0);
 144                 fs->lookup_still_valid = FALSE;
 145         }
 146 }
 147
 148 /*
 149  * Clean up after a successful call to vm_fault_object() so another call
 150  * to vm_fault_object() can be made.
 151  */
 152 static void
 153 _cleanup_successful_fault(struct faultstate *fs, int relock)
 154 {
 155         if (fs->object != fs->first_object) {
 156                 vm_page_free(fs->first_m);
 157                 vm_object_pip_wakeup(fs->object);
 158                 fs->first_m = NULL;
 159         }
 160         fs->object = fs->first_object;
 161         if (relock && fs->lookup_still_valid == FALSE) {
 162                 if (fs->map)
 163                         vm_map_lock_read(fs->map);
 164                 fs->lookup_still_valid = TRUE;
 165         }
 166 }
 167
 168 static void
 169 _unlock_things(struct faultstate *fs, int dealloc)
 170 {
 171         vm_object_pip_wakeup(fs->first_object);
 172         _cleanup_successful_fault(fs, 0);
 173         if (dealloc) {
 174                 vm_object_deallocate(fs->first_object);
 175         }
 176         unlock_map(fs);
 177         if (fs->vp != NULL) {
 178                 vput(fs->vp);
 179                 fs->vp = NULL;
 180         }
 181 }
 182
 183 #define unlock_things(fs) _unlock_things(fs, 0)
 184 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
 185 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
 186
 187 /*
 188  * TRYPAGER
 189  *
 190  * Determine if the pager for the current object *might* contain the page.
 191  *
 192  * We only need to try the pager if this is not a default object (default
 193  * objects are zero-fill and have no real pager), and if we are not taking
 194  * a wiring fault or if the FS entry is wired.
 195  */
 196 #define TRYPAGER(fs)    \
 197                 (fs->object->type != OBJT_DEFAULT && \
 198                 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
 199
 200 /*
 201  * vm_fault:
 202  *
 203  * Handle a page fault occuring at the given address, requiring the given
 204  * permissions, in the map specified.  If successful, the page is inserted
 205  * into the associated physical map.
 206  *
 207  * NOTE: The given address should be truncated to the proper page address.
 208  *
 209  * KERN_SUCCESS is returned if the page fault is handled; otherwise,
 210  * a standard error specifying why the fault is fatal is returned.
 211  *
 212  * The map in question must be referenced, and remains so.
 213  * The caller may hold no locks.
 214  */
 215 int
 216 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
 217 {
 218         int result;
 219         vm_pindex_t first_pindex;
 220         struct faultstate fs;
 221
 222         mycpu->gd_cnt.v_vm_faults++;
 223
 224         fs.didlimit = 0;
 225         fs.hardfault = 0;
 226         fs.fault_flags = fault_flags;
 227
 228 RetryFault:
 229         /*
 230          * Find the vm_map_entry representing the backing store and resolve
 231          * the top level object and page index.  This may have the side
 232          * effect of executing a copy-on-write on the map entry and/or
 233          * creating a shadow object, but will not COW any actual VM pages.
 234          *
 235          * On success fs.map is left read-locked and various other fields
 236          * are initialized but not otherwise referenced or locked.
 237          *
 238          * NOTE!  vm_map_lookup will try to upgrade the fault_type to
 239          * VM_FAULT_WRITE if the map entry is a virtual page table and also
 240          * writable, so we can set the 'A'accessed bit in the virtual page
 241          * table entry.
 242          */
 243         fs.map = map;
 244         result = vm_map_lookup(&fs.map, vaddr, fault_type,
 245                                &fs.entry, &fs.first_object,
 246                                &first_pindex, &fs.first_prot, &fs.wired);
 247
 248         /*
 249          * If the lookup failed or the map protections are incompatible,
 250          * the fault generally fails.  However, if the caller is trying
 251          * to do a user wiring we have more work to do.
 252          */
 253         if (result != KERN_SUCCESS) {
 254                 if (result != KERN_PROTECTION_FAILURE)
 255                         return result;
 256                 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
 257                         return result;
 258
 259                 /*
 260                  * If we are user-wiring a r/w segment, and it is COW, then
 261                  * we need to do the COW operation.  Note that we don't
 262                  * currently COW RO sections now, because it is NOT desirable
 263                  * to COW .text.  We simply keep .text from ever being COW'ed
 264                  * and take the heat that one cannot debug wired .text sections.
 265                  */
 266                 result = vm_map_lookup(&fs.map, vaddr,
 267                                        VM_PROT_READ|VM_PROT_WRITE|
 268                                         VM_PROT_OVERRIDE_WRITE,
 269                                        &fs.entry, &fs.first_object,
 270                                        &first_pindex, &fs.first_prot,
 271                                        &fs.wired);
 272                 if (result != KERN_SUCCESS)
 273                         return result;
 274
 275                 /*
 276                  * If we don't COW now, on a user wire, the user will never
 277                  * be able to write to the mapping.  If we don't make this
 278                  * restriction, the bookkeeping would be nearly impossible.
 279                  */
 280                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 281                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 282         }
 283
 284         /*
 285          * fs.map is read-locked
 286          *
 287          * Misc checks.  Save the map generation number to detect races.
 288          */
 289         fs.map_generation = fs.map->timestamp;
 290
 291         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 292                 panic("vm_fault: fault on nofault entry, addr: %lx",
 293                     (u_long)vaddr);
 294         }
 295
 296         /*
 297          * A system map entry may return a NULL object.  No object means
 298          * no pager means an unrecoverable kernel fault.
 299          */
 300         if (fs.first_object == NULL) {
 301                 panic("vm_fault: unrecoverable fault at %p in entry %p",
 302                         (void *)vaddr, fs.entry);
 303         }
 304
 305         /*
 306          * Make a reference to this object to prevent its disposal while we
 307          * are messing with it.  Once we have the reference, the map is free
 308          * to be diddled.  Since objects reference their shadows (and copies),
 309          * they will stay around as well.
 310          *
 311          * Bump the paging-in-progress count to prevent size changes (e.g.
 312          * truncation operations) during I/O.  This must be done after
 313          * obtaining the vnode lock in order to avoid possible deadlocks.
 314          */
 315         vm_object_reference(fs.first_object);
 316         fs.vp = vnode_pager_lock(fs.first_object);
 317         vm_object_pip_add(fs.first_object, 1);
 318
 319         fs.lookup_still_valid = TRUE;
 320         fs.first_m = NULL;
 321         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 322
 323         /*
 324          * If the entry is wired we cannot change the page protection.
 325          */
 326         if (fs.wired)
 327                 fault_type = fs.first_prot;
 328
 329         /*
 330          * The page we want is at (first_object, first_pindex), but if the
 331          * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
 332          * page table to figure out the actual pindex.
 333          *
 334          * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
 335          * ONLY
 336          */
 337         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 338                 result = vm_fault_vpagetable(&fs, &first_pindex,
 339                                              fs.entry->aux.master_pde,
 340                                              fault_type);
 341                 if (result == KERN_TRY_AGAIN)
 342                         goto RetryFault;
 343                 if (result != KERN_SUCCESS)
 344                         return (result);
 345         }
 346
 347         /*
 348          * Now we have the actual (object, pindex), fault in the page.  If
 349          * vm_fault_object() fails it will unlock and deallocate the FS
 350          * data.   If it succeeds everything remains locked and fs->object
 351          * will have an additinal PIP count if it is not equal to
 352          * fs->first_object
 353          *
 354          * vm_fault_object will set fs->prot for the pmap operation.  It is
 355          * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
 356          * page can be safely written.  However, it will force a read-only
 357          * mapping for a read fault if the memory is managed by a virtual
 358          * page table.
 359          */
 360         result = vm_fault_object(&fs, first_pindex, fault_type);
 361
 362         if (result == KERN_TRY_AGAIN)
 363                 goto RetryFault;
 364         if (result != KERN_SUCCESS)
 365                 return (result);
 366
 367         /*
 368          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 369          * will contain a busied page.
 370          *
 371          * Enter the page into the pmap and do pmap-related adjustments.
 372          */
 373         unlock_things(&fs);
 374         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
 375
 376         if (((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0) && (fs.wired == 0)) {
 377                 pmap_prefault(fs.map->pmap, vaddr, fs.entry);
 378         }
 379
 380         vm_page_flag_clear(fs.m, PG_ZERO);
 381         vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
 382
 383         /*
 384          * If the page is not wired down, then put it where the pageout daemon
 385          * can find it.
 386          */
 387         if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
 388                 if (fs.wired)
 389                         vm_page_wire(fs.m);
 390                 else
 391                         vm_page_unwire(fs.m, 1);
 392         } else {
 393                 vm_page_activate(fs.m);
 394         }
 395
 396         if (curthread->td_lwp) {
 397                 if (fs.hardfault) {
 398                         curthread->td_lwp->lwp_ru.ru_majflt++;
 399                 } else {
 400                         curthread->td_lwp->lwp_ru.ru_minflt++;
 401                 }
 402         }
 403
 404         /*
 405          * Unlock everything, and return
 406          */
 407         vm_page_wakeup(fs.m);
 408         vm_object_deallocate(fs.first_object);
 409
 410         return (KERN_SUCCESS);
 411 }
 412
 413 /*
 414  * Fault in the specified virtual address in the current process map,
 415  * returning a held VM page or NULL.  See vm_fault_page() for more
 416  * information.
 417  */
 418 vm_page_t
 419 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
 420 {
 421         vm_page_t m;
 422
 423         m = vm_fault_page(&curproc->p_vmspace->vm_map, va,
 424                           fault_type, VM_FAULT_NORMAL, errorp);
 425         return(m);
 426 }
 427
 428 /*
 429  * Fault in the specified virtual address in the specified map, doing all
 430  * necessary manipulation of the object store and all necessary I/O.  Return
 431  * a held VM page or NULL, and set *errorp.  The related pmap is not
 432  * updated.
 433  *
 434  * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
 435  * and marked PG_REFERENCED as well.
 436  */
 437 vm_page_t
 438 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 439               int fault_flags, int *errorp)
 440 {
 441         int result;
 442         vm_pindex_t first_pindex;
 443         struct faultstate fs;
 444
 445         mycpu->gd_cnt.v_vm_faults++;
 446
 447         fs.didlimit = 0;
 448         fs.hardfault = 0;
 449         fs.fault_flags = fault_flags;
 450         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
 451
 452 RetryFault:
 453         /*
 454          * Find the vm_map_entry representing the backing store and resolve
 455          * the top level object and page index.  This may have the side
 456          * effect of executing a copy-on-write on the map entry and/or
 457          * creating a shadow object, but will not COW any actual VM pages.
 458          *
 459          * On success fs.map is left read-locked and various other fields
 460          * are initialized but not otherwise referenced or locked.
 461          *
 462          * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
 463          * if the map entry is a virtual page table and also writable,
 464          * so we can set the 'A'accessed bit in the virtual page table entry.
 465          */
 466         fs.map = map;
 467         result = vm_map_lookup(&fs.map, vaddr, fault_type,
 468                                &fs.entry, &fs.first_object,
 469                                &first_pindex, &fs.first_prot, &fs.wired);
 470
 471         if (result != KERN_SUCCESS) {
 472                 *errorp = result;
 473                 return (NULL);
 474         }
 475
 476         /*
 477          * fs.map is read-locked
 478          *
 479          * Misc checks.  Save the map generation number to detect races.
 480          */
 481         fs.map_generation = fs.map->timestamp;
 482
 483         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 484                 panic("vm_fault: fault on nofault entry, addr: %lx",
 485                     (u_long)vaddr);
 486         }
 487
 488         /*
 489          * A system map entry may return a NULL object.  No object means
 490          * no pager means an unrecoverable kernel fault.
 491          */
 492         if (fs.first_object == NULL) {
 493                 panic("vm_fault: unrecoverable fault at %p in entry %p",
 494                         (void *)vaddr, fs.entry);
 495         }
 496
 497         /*
 498          * Make a reference to this object to prevent its disposal while we
 499          * are messing with it.  Once we have the reference, the map is free
 500          * to be diddled.  Since objects reference their shadows (and copies),
 501          * they will stay around as well.
 502          *
 503          * Bump the paging-in-progress count to prevent size changes (e.g.
 504          * truncation operations) during I/O.  This must be done after
 505          * obtaining the vnode lock in order to avoid possible deadlocks.
 506          */
 507         vm_object_reference(fs.first_object);
 508         fs.vp = vnode_pager_lock(fs.first_object);
 509         vm_object_pip_add(fs.first_object, 1);
 510
 511         fs.lookup_still_valid = TRUE;
 512         fs.first_m = NULL;
 513         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 514
 515         /*
 516          * If the entry is wired we cannot change the page protection.
 517          */
 518         if (fs.wired)
 519                 fault_type = fs.first_prot;
 520
 521         /*
 522          * The page we want is at (first_object, first_pindex), but if the
 523          * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
 524          * page table to figure out the actual pindex.
 525          *
 526          * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
 527          * ONLY
 528          */
 529         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 530                 result = vm_fault_vpagetable(&fs, &first_pindex,
 531                                              fs.entry->aux.master_pde,
 532                                              fault_type);
 533                 if (result == KERN_TRY_AGAIN)
 534                         goto RetryFault;
 535                 if (result != KERN_SUCCESS) {
 536                         *errorp = result;
 537                         return (NULL);
 538                 }
 539         }
 540
 541         /*
 542          * Now we have the actual (object, pindex), fault in the page.  If
 543          * vm_fault_object() fails it will unlock and deallocate the FS
 544          * data.   If it succeeds everything remains locked and fs->object
 545          * will have an additinal PIP count if it is not equal to
 546          * fs->first_object
 547          */
 548         result = vm_fault_object(&fs, first_pindex, fault_type);
 549
 550         if (result == KERN_TRY_AGAIN)
 551                 goto RetryFault;
 552         if (result != KERN_SUCCESS) {
 553                 *errorp = result;
 554                 return(NULL);
 555         }
 556
 557         /*
 558          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 559          * will contain a busied page.
 560          */
 561         unlock_things(&fs);
 562
 563         /*
 564          * Return a held page.  We are not doing any pmap manipulation so do
 565          * not set PG_MAPPED.  However, adjust the page flags according to
 566          * the fault type because the caller may not use a managed pmapping
 567          * (so we don't want to lose the fact that the page will be dirtied
 568          * if a write fault was specified).
 569          */
 570         vm_page_hold(fs.m);
 571         vm_page_flag_clear(fs.m, PG_ZERO);
 572         if (fault_type & VM_PROT_WRITE)
 573                 vm_page_dirty(fs.m);
 574
 575         /*
 576          * Update the pmap.  We really only have to do this if a COW
 577          * occured to replace the read-only page with the new page.  For
 578          * now just do it unconditionally. XXX
 579          */
 580         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
 581         vm_page_flag_set(fs.m, PG_REFERENCED|PG_MAPPED);
 582
 583         /*
 584          * Unbusy the page by activating it.  It remains held and will not
 585          * be reclaimed.
 586          */
 587         vm_page_activate(fs.m);
 588
 589         if (curthread->td_lwp) {
 590                 if (fs.hardfault) {
 591                         curthread->td_lwp->lwp_ru.ru_majflt++;
 592                 } else {
 593                         curthread->td_lwp->lwp_ru.ru_minflt++;
 594                 }
 595         }
 596
 597         /*
 598          * Unlock everything, and return the held page.
 599          */
 600         vm_page_wakeup(fs.m);
 601         vm_object_deallocate(fs.first_object);
 602
 603         *errorp = 0;
 604         return(fs.m);
 605 }
 606
 607 /*
 608  * Fault in the specified
 609  */
 610 vm_page_t
 611 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
 612                      vm_prot_t fault_type, int fault_flags, int *errorp)
 613 {
 614         int result;
 615         vm_pindex_t first_pindex;
 616         struct faultstate fs;
 617         struct vm_map_entry entry;
 618
 619         bzero(&entry, sizeof(entry));
 620         entry.object.vm_object = object;
 621         entry.maptype = VM_MAPTYPE_NORMAL;
 622         entry.protection = entry.max_protection = fault_type;
 623
 624         fs.didlimit = 0;
 625         fs.hardfault = 0;
 626         fs.fault_flags = fault_flags;
 627         fs.map = NULL;
 628         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
 629
 630 RetryFault:
 631
 632         fs.first_object = object;
 633         first_pindex = OFF_TO_IDX(offset);
 634         fs.entry = &entry;
 635         fs.first_prot = fault_type;
 636         fs.wired = 0;
 637         /*fs.map_generation = 0; unused */
 638
 639         /*
 640          * Make a reference to this object to prevent its disposal while we
 641          * are messing with it.  Once we have the reference, the map is free
 642          * to be diddled.  Since objects reference their shadows (and copies),
 643          * they will stay around as well.
 644          *
 645          * Bump the paging-in-progress count to prevent size changes (e.g.
 646          * truncation operations) during I/O.  This must be done after
 647          * obtaining the vnode lock in order to avoid possible deadlocks.
 648          */
 649         vm_object_reference(fs.first_object);
 650         fs.vp = vnode_pager_lock(fs.first_object);
 651         vm_object_pip_add(fs.first_object, 1);
 652
 653         fs.lookup_still_valid = TRUE;
 654         fs.first_m = NULL;
 655         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 656
 657 #if 0
 658         /* XXX future - ability to operate on VM object using vpagetable */
 659         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 660                 result = vm_fault_vpagetable(&fs, &first_pindex,
 661                                              fs.entry->aux.master_pde,
 662                                              fault_type);
 663                 if (result == KERN_TRY_AGAIN)
 664                         goto RetryFault;
 665                 if (result != KERN_SUCCESS) {
 666                         *errorp = result;
 667                         return (NULL);
 668                 }
 669         }
 670 #endif
 671
 672         /*
 673          * Now we have the actual (object, pindex), fault in the page.  If
 674          * vm_fault_object() fails it will unlock and deallocate the FS
 675          * data.   If it succeeds everything remains locked and fs->object
 676          * will have an additinal PIP count if it is not equal to
 677          * fs->first_object
 678          */
 679         result = vm_fault_object(&fs, first_pindex, fault_type);
 680
 681         if (result == KERN_TRY_AGAIN)
 682                 goto RetryFault;
 683         if (result != KERN_SUCCESS) {
 684                 *errorp = result;
 685                 return(NULL);
 686         }
 687
 688         /*
 689          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 690          * will contain a busied page.
 691          */
 692         unlock_things(&fs);
 693
 694         /*
 695          * Return a held page.  We are not doing any pmap manipulation so do
 696          * not set PG_MAPPED.  However, adjust the page flags according to
 697          * the fault type because the caller may not use a managed pmapping
 698          * (so we don't want to lose the fact that the page will be dirtied
 699          * if a write fault was specified).
 700          */
 701         vm_page_hold(fs.m);
 702         vm_page_flag_clear(fs.m, PG_ZERO);
 703         if (fault_type & VM_PROT_WRITE)
 704                 vm_page_dirty(fs.m);
 705
 706         /*
 707          * Indicate that the page was accessed.
 708          */
 709         vm_page_flag_set(fs.m, PG_REFERENCED);
 710
 711         /*
 712          * Unbusy the page by activating it.  It remains held and will not
 713          * be reclaimed.
 714          */
 715         vm_page_activate(fs.m);
 716
 717         if (curthread->td_lwp) {
 718                 if (fs.hardfault) {
 719                         mycpu->gd_cnt.v_vm_faults++;
 720                         curthread->td_lwp->lwp_ru.ru_majflt++;
 721                 } else {
 722                         curthread->td_lwp->lwp_ru.ru_minflt++;
 723                 }
 724         }
 725
 726         /*
 727          * Unlock everything, and return the held page.
 728          */
 729         vm_page_wakeup(fs.m);
 730         vm_object_deallocate(fs.first_object);
 731
 732         *errorp = 0;
 733         return(fs.m);
 734 }
 735
 736 /*
 737  * Translate the virtual page number (first_pindex) that is relative
 738  * to the address space into a logical page number that is relative to the
 739  * backing object.  Use the virtual page table pointed to by (vpte).
 740  *
 741  * This implements an N-level page table.  Any level can terminate the
 742  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
 743  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
 744  */
 745 static
 746 int
 747 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
 748                     vpte_t vpte, int fault_type)
 749 {
 750         struct sf_buf *sf;
 751         int vshift = 32 - PAGE_SHIFT;   /* page index bits remaining */
 752         int result = KERN_SUCCESS;
 753         vpte_t *ptep;
 754
 755         for (;;) {
 756                 /*
 757                  * We cannot proceed if the vpte is not valid, not readable
 758                  * for a read fault, or not writable for a write fault.
 759                  */
 760                 if ((vpte & VPTE_V) == 0) {
 761                         unlock_and_deallocate(fs);
 762                         return (KERN_FAILURE);
 763                 }
 764                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
 765                         unlock_and_deallocate(fs);
 766                         return (KERN_FAILURE);
 767                 }
 768                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
 769                         unlock_and_deallocate(fs);
 770                         return (KERN_FAILURE);
 771                 }
 772                 if ((vpte & VPTE_PS) || vshift == 0)
 773                         break;
 774                 KKASSERT(vshift >= VPTE_PAGE_BITS);
 775
 776                 /*
 777                  * Get the page table page.  Nominally we only read the page
 778                  * table, but since we are actively setting VPTE_M and VPTE_A,
 779                  * tell vm_fault_object() that we are writing it.
 780                  *
 781                  * There is currently no real need to optimize this.
 782                  */
 783                 result = vm_fault_object(fs, vpte >> PAGE_SHIFT,
 784                                          VM_PROT_READ|VM_PROT_WRITE);
 785                 if (result != KERN_SUCCESS)
 786                         return (result);
 787
 788                 /*
 789                  * Process the returned fs.m and look up the page table
 790                  * entry in the page table page.
 791                  */
 792                 vshift -= VPTE_PAGE_BITS;
 793                 sf = sf_buf_alloc(fs->m, SFB_CPUPRIVATE);
 794                 ptep = ((vpte_t *)sf_buf_kva(sf) +
 795                         ((*pindex >> vshift) & VPTE_PAGE_MASK));
 796                 vpte = *ptep;
 797
 798                 /*
 799                  * Page table write-back.  If the vpte is valid for the
 800                  * requested operation, do a write-back to the page table.
 801                  *
 802                  * XXX VPTE_M is not set properly for page directory pages.
 803                  * It doesn't get set in the page directory if the page table
 804                  * is modified during a read access.
 805                  */
 806                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
 807                     (vpte & VPTE_W)) {
 808                         if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
 809                                 atomic_set_int(ptep, VPTE_M|VPTE_A);
 810                                 vm_page_dirty(fs->m);
 811                         }
 812                 }
 813                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
 814                     (vpte & VPTE_R)) {
 815                         if ((vpte & VPTE_A) == 0) {
 816                                 atomic_set_int(ptep, VPTE_A);
 817                                 vm_page_dirty(fs->m);
 818                         }
 819                 }
 820                 sf_buf_free(sf);
 821                 vm_page_flag_set(fs->m, PG_REFERENCED);
 822                 vm_page_activate(fs->m);
 823                 vm_page_wakeup(fs->m);
 824                 cleanup_successful_fault(fs);
 825         }
 826         /*
 827          * Combine remaining address bits with the vpte.
 828          */
 829         *pindex = (vpte >> PAGE_SHIFT) +
 830                   (*pindex & ((1 << vshift) - 1));
 831         return (KERN_SUCCESS);
 832 }
 833
 834
 835 /*
 836  * Do all operations required to fault-in (fs.first_object, pindex).  Run
 837  * through the shadow chain as necessary and do required COW or virtual
 838  * copy operations.  The caller has already fully resolved the vm_map_entry
 839  * and, if appropriate, has created a copy-on-write layer.  All we need to
 840  * do is iterate the object chain.
 841  *
 842  * On failure (fs) is unlocked and deallocated and the caller may return or
 843  * retry depending on the failure code.  On success (fs) is NOT unlocked or
 844  * deallocated, fs.m will contained a resolved, busied page, and fs.object
 845  * will have an additional PIP count if it is not equal to fs.first_object.
 846  */
 847 static
 848 int
 849 vm_fault_object(struct faultstate *fs,
 850                 vm_pindex_t first_pindex, vm_prot_t fault_type)
 851 {
 852         vm_object_t next_object;
 853         vm_page_t marray[VM_FAULT_READ];
 854         vm_pindex_t pindex;
 855         int faultcount;
 856
 857         fs->prot = fs->first_prot;
 858         fs->object = fs->first_object;
 859         pindex = first_pindex;
 860
 861         /*
 862          * If a read fault occurs we try to make the page writable if
 863          * possible.  There are three cases where we cannot make the
 864          * page mapping writable:
 865          *
 866          * (1) The mapping is read-only or the VM object is read-only,
 867          *     fs->prot above will simply not have VM_PROT_WRITE set.
 868          *
 869          * (2) If the mapping is a virtual page table we need to be able
 870          *     to detect writes so we can set VPTE_M in the virtual page
 871          *     table.
 872          *
 873          * (3) If the VM page is read-only or copy-on-write, upgrading would
 874          *     just result in an unnecessary COW fault.
 875          *
 876          * VM_PROT_VPAGED is set if faulting via a virtual page table and
 877          * causes adjustments to the 'M'odify bit to also turn off write
 878          * access to force a re-fault.
 879          */
 880         if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 881                 if ((fault_type & VM_PROT_WRITE) == 0)
 882                         fs->prot &= ~VM_PROT_WRITE;
 883         }
 884
 885         for (;;) {
 886                 /*
 887                  * If the object is dead, we stop here
 888                  */
 889                 if (fs->object->flags & OBJ_DEAD) {
 890                         unlock_and_deallocate(fs);
 891                         return (KERN_PROTECTION_FAILURE);
 892                 }
 893
 894                 /*
 895                  * See if page is resident.  spl protection is required
 896                  * to avoid an interrupt unbusy/free race against our
 897                  * lookup.  We must hold the protection through a page
 898                  * allocation or busy.
 899                  */
 900                 crit_enter();
 901                 fs->m = vm_page_lookup(fs->object, pindex);
 902                 if (fs->m != NULL) {
 903                         int queue;
 904                         /*
 905                          * Wait/Retry if the page is busy.  We have to do this
 906                          * if the page is busy via either PG_BUSY or
 907                          * vm_page_t->busy because the vm_pager may be using
 908                          * vm_page_t->busy for pageouts ( and even pageins if
 909                          * it is the vnode pager ), and we could end up trying
 910                          * to pagein and pageout the same page simultaneously.
 911                          *
 912                          * We can theoretically allow the busy case on a read
 913                          * fault if the page is marked valid, but since such
 914                          * pages are typically already pmap'd, putting that
 915                          * special case in might be more effort then it is
 916                          * worth.  We cannot under any circumstances mess
 917                          * around with a vm_page_t->busy page except, perhaps,
 918                          * to pmap it.
 919                          */
 920                         if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
 921                                 unlock_things(fs);
 922                                 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
 923                                 mycpu->gd_cnt.v_intrans++;
 924                                 vm_object_deallocate(fs->first_object);
 925                                 crit_exit();
 926                                 return (KERN_TRY_AGAIN);
 927                         }
 928
 929                         /*
 930                          * If reactivating a page from PQ_CACHE we may have
 931                          * to rate-limit.
 932                          */
 933                         queue = fs->m->queue;
 934                         vm_page_unqueue_nowakeup(fs->m);
 935
 936                         if ((queue - fs->m->pc) == PQ_CACHE &&
 937                             vm_page_count_severe()) {
 938                                 vm_page_activate(fs->m);
 939                                 unlock_and_deallocate(fs);
 940                                 vm_waitpfault();
 941                                 crit_exit();
 942                                 return (KERN_TRY_AGAIN);
 943                         }
 944
 945                         /*
 946                          * Mark page busy for other processes, and the
 947                          * pagedaemon.  If it still isn't completely valid
 948                          * (readable), jump to readrest, else we found the
 949                          * page and can return.
 950                          *
 951                          * We can release the spl once we have marked the
 952                          * page busy.
 953                          */
 954                         vm_page_busy(fs->m);
 955                         crit_exit();
 956
 957                         if (((fs->m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
 958                             fs->m->object != &kernel_object) {
 959                                 goto readrest;
 960                         }
 961                         break; /* break to PAGE HAS BEEN FOUND */
 962                 }
 963
 964                 /*
 965                  * Page is not resident, If this is the search termination
 966                  * or the pager might contain the page, allocate a new page.
 967                  *
 968                  * NOTE: We are still in a critical section.
 969                  */
 970                 if (TRYPAGER(fs) || fs->object == fs->first_object) {
 971                         /*
 972                          * If the page is beyond the object size we fail
 973                          */
 974                         if (pindex >= fs->object->size) {
 975                                 crit_exit();
 976                                 unlock_and_deallocate(fs);
 977                                 return (KERN_PROTECTION_FAILURE);
 978                         }
 979
 980                         /*
 981                          * Ratelimit.
 982                          */
 983                         if (fs->didlimit == 0 && curproc != NULL) {
 984                                 int limticks;
 985
 986                                 limticks = vm_fault_ratelimit(curproc->p_vmspace);
 987                                 if (limticks) {
 988                                         crit_exit();
 989                                         unlock_and_deallocate(fs);
 990                                         tsleep(curproc, 0, "vmrate", limticks);
 991                                         fs->didlimit = 1;
 992                                         return (KERN_TRY_AGAIN);
 993                                 }
 994                         }
 995
 996                         /*
 997                          * Allocate a new page for this object/offset pair.
 998                          */
 999                         fs->m = NULL;
1000                         if (!vm_page_count_severe()) {
1001                                 fs->m = vm_page_alloc(fs->object, pindex,
1002                                     (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
1003                         }
1004                         if (fs->m == NULL) {
1005                                 crit_exit();
1006                                 unlock_and_deallocate(fs);
1007                                 vm_waitpfault();
1008                                 return (KERN_TRY_AGAIN);
1009                         }
1010                 }
1011                 crit_exit();
1012
1013 readrest:
1014                 /*
1015                  * We have found a valid page or we have allocated a new page.
1016                  * The page thus may not be valid or may not be entirely
1017                  * valid.
1018                  *
1019                  * Attempt to fault-in the page if there is a chance that the
1020                  * pager has it, and potentially fault in additional pages
1021                  * at the same time.
1022                  *
1023                  * We are NOT in splvm here and if TRYPAGER is true then
1024                  * fs.m will be non-NULL and will be PG_BUSY for us.
1025                  */
1026
1027                 if (TRYPAGER(fs)) {
1028                         int rv;
1029                         int reqpage;
1030                         int ahead, behind;
1031                         u_char behavior = vm_map_entry_behavior(fs->entry);
1032
1033                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
1034                                 ahead = 0;
1035                                 behind = 0;
1036                         } else {
1037                                 behind = pindex;
1038                                 if (behind > VM_FAULT_READ_BEHIND)
1039                                         behind = VM_FAULT_READ_BEHIND;
1040
1041                                 ahead = fs->object->size - pindex;
1042                                 if (ahead < 1)
1043                                         ahead = 1;
1044                                 if (ahead > VM_FAULT_READ_AHEAD)
1045                                         ahead = VM_FAULT_READ_AHEAD;
1046                         }
1047
1048                         if ((fs->first_object->type != OBJT_DEVICE) &&
1049                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1050                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1051                                 pindex >= fs->entry->lastr &&
1052                                 pindex < fs->entry->lastr + VM_FAULT_READ))
1053                         ) {
1054                                 vm_pindex_t firstpindex, tmppindex;
1055
1056                                 if (first_pindex < 2 * VM_FAULT_READ)
1057                                         firstpindex = 0;
1058                                 else
1059                                         firstpindex = first_pindex - 2 * VM_FAULT_READ;
1060
1061                                 /*
1062                                  * note: partially valid pages cannot be
1063                                  * included in the lookahead - NFS piecemeal
1064                                  * writes will barf on it badly.
1065                                  *
1066                                  * spl protection is required to avoid races
1067                                  * between the lookup and an interrupt
1068                                  * unbusy/free sequence occuring prior to
1069                                  * our busy check.
1070                                  */
1071                                 crit_enter();
1072                                 for (tmppindex = first_pindex - 1;
1073                                     tmppindex >= firstpindex;
1074                                     --tmppindex
1075                                 ) {
1076                                         vm_page_t mt;
1077
1078                                         mt = vm_page_lookup(fs->first_object, tmppindex);
1079                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
1080                                                 break;
1081                                         if (mt->busy ||
1082                                                 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
1083                                                 mt->hold_count ||
1084                                                 mt->wire_count)
1085                                                 continue;
1086                                         if (mt->dirty == 0)
1087                                                 vm_page_test_dirty(mt);
1088                                         if (mt->dirty) {
1089                                                 vm_page_protect(mt, VM_PROT_NONE);
1090                                                 vm_page_deactivate(mt);
1091                                         } else {
1092                                                 vm_page_cache(mt);
1093                                         }
1094                                 }
1095                                 crit_exit();
1096
1097                                 ahead += behind;
1098                                 behind = 0;
1099                         }
1100
1101                         /*
1102                          * now we find out if any other pages should be paged
1103                          * in at this time this routine checks to see if the
1104                          * pages surrounding this fault reside in the same
1105                          * object as the page for this fault.  If they do,
1106                          * then they are faulted in also into the object.  The
1107                          * array "marray" returned contains an array of
1108                          * vm_page_t structs where one of them is the
1109                          * vm_page_t passed to the routine.  The reqpage
1110                          * return value is the index into the marray for the
1111                          * vm_page_t passed to the routine.
1112                          *
1113                          * fs.m plus the additional pages are PG_BUSY'd.
1114                          */
1115                         faultcount = vm_fault_additional_pages(
1116                             fs->m, behind, ahead, marray, &reqpage);
1117
1118                         /*
1119                          * update lastr imperfectly (we do not know how much
1120                          * getpages will actually read), but good enough.
1121                          */
1122                         fs->entry->lastr = pindex + faultcount - behind;
1123
1124                         /*
1125                          * Call the pager to retrieve the data, if any, after
1126                          * releasing the lock on the map.  We hold a ref on
1127                          * fs.object and the pages are PG_BUSY'd.
1128                          */
1129                         unlock_map(fs);
1130
1131                         if (faultcount) {
1132                                 rv = vm_pager_get_pages(fs->object, marray,
1133                                                         faultcount, reqpage);
1134                         } else {
1135                                 rv = VM_PAGER_FAIL;
1136                         }
1137
1138                         if (rv == VM_PAGER_OK) {
1139                                 /*
1140                                  * Found the page. Leave it busy while we play
1141                                  * with it.
1142                                  */
1143
1144                                 /*
1145                                  * Relookup in case pager changed page. Pager
1146                                  * is responsible for disposition of old page
1147                                  * if moved.
1148                                  *
1149                                  * XXX other code segments do relookups too.
1150                                  * It's a bad abstraction that needs to be
1151                                  * fixed/removed.
1152                                  */
1153                                 fs->m = vm_page_lookup(fs->object, pindex);
1154                                 if (fs->m == NULL) {
1155                                         unlock_and_deallocate(fs);
1156                                         return (KERN_TRY_AGAIN);
1157                                 }
1158
1159                                 ++fs->hardfault;
1160                                 break; /* break to PAGE HAS BEEN FOUND */
1161                         }
1162
1163                         /*
1164                          * Remove the bogus page (which does not exist at this
1165                          * object/offset); before doing so, we must get back
1166                          * our object lock to preserve our invariant.
1167                          *
1168                          * Also wake up any other process that may want to bring
1169                          * in this page.
1170                          *
1171                          * If this is the top-level object, we must leave the
1172                          * busy page to prevent another process from rushing
1173                          * past us, and inserting the page in that object at
1174                          * the same time that we are.
1175                          */
1176                         if (rv == VM_PAGER_ERROR) {
1177                                 if (curproc)
1178                                         kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
1179                                 else
1180                                         kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
1181                         }
1182                         /*
1183                          * Data outside the range of the pager or an I/O error
1184                          */
1185                         /*
1186                          * XXX - the check for kernel_map is a kludge to work
1187                          * around having the machine panic on a kernel space
1188                          * fault w/ I/O error.
1189                          */
1190                         if (((fs->map != &kernel_map) && (rv == VM_PAGER_ERROR)) ||
1191                                 (rv == VM_PAGER_BAD)) {
1192                                 vm_page_free(fs->m);
1193                                 fs->m = NULL;
1194                                 unlock_and_deallocate(fs);
1195                                 if (rv == VM_PAGER_ERROR)
1196                                         return (KERN_FAILURE);
1197                                 else
1198                                         return (KERN_PROTECTION_FAILURE);
1199                                 /* NOT REACHED */
1200                         }
1201                         if (fs->object != fs->first_object) {
1202                                 vm_page_free(fs->m);
1203                                 fs->m = NULL;
1204                                 /*
1205                                  * XXX - we cannot just fall out at this
1206                                  * point, m has been freed and is invalid!
1207                                  */
1208                         }
1209                 }
1210
1211                 /*
1212                  * We get here if the object has a default pager (or unwiring)
1213                  * or the pager doesn't have the page.
1214                  */
1215                 if (fs->object == fs->first_object)
1216                         fs->first_m = fs->m;
1217
1218                 /*
1219                  * Move on to the next object.  Lock the next object before
1220                  * unlocking the current one.
1221                  */
1222                 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1223                 next_object = fs->object->backing_object;
1224                 if (next_object == NULL) {
1225                         /*
1226                          * If there's no object left, fill the page in the top
1227                          * object with zeros.
1228                          */
1229                         if (fs->object != fs->first_object) {
1230                                 vm_object_pip_wakeup(fs->object);
1231
1232                                 fs->object = fs->first_object;
1233                                 pindex = first_pindex;
1234                                 fs->m = fs->first_m;
1235                         }
1236                         fs->first_m = NULL;
1237
1238                         /*
1239                          * Zero the page if necessary and mark it valid.
1240                          */
1241                         if ((fs->m->flags & PG_ZERO) == 0) {
1242                                 vm_page_zero_fill(fs->m);
1243                         } else {
1244                                 mycpu->gd_cnt.v_ozfod++;
1245                         }
1246                         mycpu->gd_cnt.v_zfod++;
1247                         fs->m->valid = VM_PAGE_BITS_ALL;
1248                         break;  /* break to PAGE HAS BEEN FOUND */
1249                 } else {
1250                         if (fs->object != fs->first_object) {
1251                                 vm_object_pip_wakeup(fs->object);
1252                         }
1253                         KASSERT(fs->object != next_object, ("object loop %p", next_object));
1254                         fs->object = next_object;
1255                         vm_object_pip_add(fs->object, 1);
1256                 }
1257         }
1258
1259         KASSERT((fs->m->flags & PG_BUSY) != 0,
1260                 ("vm_fault: not busy after main loop"));
1261
1262         /*
1263          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1264          * is held.]
1265          */
1266
1267         /*
1268          * If the page is being written, but isn't already owned by the
1269          * top-level object, we have to copy it into a new page owned by the
1270          * top-level object.
1271          */
1272         if (fs->object != fs->first_object) {
1273                 /*
1274                  * We only really need to copy if we want to write it.
1275                  */
1276                 if (fault_type & VM_PROT_WRITE) {
1277                         /*
1278                          * This allows pages to be virtually copied from a
1279                          * backing_object into the first_object, where the
1280                          * backing object has no other refs to it, and cannot
1281                          * gain any more refs.  Instead of a bcopy, we just
1282                          * move the page from the backing object to the
1283                          * first object.  Note that we must mark the page
1284                          * dirty in the first object so that it will go out
1285                          * to swap when needed.
1286                          */
1287                         if (
1288                                 /*
1289                                  * Map, if present, has not changed
1290                                  */
1291                                 (fs->map == NULL ||
1292                                 fs->map_generation == fs->map->timestamp) &&
1293                                 /*
1294                                  * Only one shadow object
1295                                  */
1296                                 (fs->object->shadow_count == 1) &&
1297                                 /*
1298                                  * No COW refs, except us
1299                                  */
1300                                 (fs->object->ref_count == 1) &&
1301                                 /*
1302                                  * No one else can look this object up
1303                                  */
1304                                 (fs->object->handle == NULL) &&
1305                                 /*
1306                                  * No other ways to look the object up
1307                                  */
1308                                 ((fs->object->type == OBJT_DEFAULT) ||
1309                                  (fs->object->type == OBJT_SWAP)) &&
1310                                 /*
1311                                  * We don't chase down the shadow chain
1312                                  */
1313                                 (fs->object == fs->first_object->backing_object) &&
1314
1315                                 /*
1316                                  * grab the lock if we need to
1317                                  */
1318                                 (fs->lookup_still_valid ||
1319                                  fs->map == NULL ||
1320                                  lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1321                             ) {
1322
1323                                 fs->lookup_still_valid = 1;
1324                                 /*
1325                                  * get rid of the unnecessary page
1326                                  */
1327                                 vm_page_protect(fs->first_m, VM_PROT_NONE);
1328                                 vm_page_free(fs->first_m);
1329                                 fs->first_m = NULL;
1330
1331                                 /*
1332                                  * grab the page and put it into the
1333                                  * process'es object.  The page is
1334                                  * automatically made dirty.
1335                                  */
1336                                 vm_page_rename(fs->m, fs->first_object, first_pindex);
1337                                 fs->first_m = fs->m;
1338                                 vm_page_busy(fs->first_m);
1339                                 fs->m = NULL;
1340                                 mycpu->gd_cnt.v_cow_optim++;
1341                         } else {
1342                                 /*
1343                                  * Oh, well, lets copy it.
1344                                  */
1345                                 vm_page_copy(fs->m, fs->first_m);
1346                         }
1347
1348                         if (fs->m) {
1349                                 /*
1350                                  * We no longer need the old page or object.
1351                                  */
1352                                 release_page(fs);
1353                         }
1354
1355                         /*
1356                          * fs->object != fs->first_object due to above
1357                          * conditional
1358                          */
1359                         vm_object_pip_wakeup(fs->object);
1360
1361                         /*
1362                          * Only use the new page below...
1363                          */
1364
1365                         mycpu->gd_cnt.v_cow_faults++;
1366                         fs->m = fs->first_m;
1367                         fs->object = fs->first_object;
1368                         pindex = first_pindex;
1369                 } else {
1370                         /*
1371                          * If it wasn't a write fault avoid having to copy
1372                          * the page by mapping it read-only.
1373                          */
1374                         fs->prot &= ~VM_PROT_WRITE;
1375                 }
1376         }
1377
1378         /*
1379          * We may have had to unlock a map to do I/O.  If we did then
1380          * lookup_still_valid will be FALSE.  If the map generation count
1381          * also changed then all sorts of things could have happened while
1382          * we were doing the I/O and we need to retry.
1383          */
1384
1385         if (!fs->lookup_still_valid &&
1386             fs->map != NULL &&
1387             (fs->map->timestamp != fs->map_generation)) {
1388                 release_page(fs);
1389                 unlock_and_deallocate(fs);
1390                 return (KERN_TRY_AGAIN);
1391         }
1392
1393         /*
1394          * Put this page into the physical map. We had to do the unlock above
1395          * because pmap_enter may cause other faults.   We don't put the page
1396          * back on the active queue until later so that the page-out daemon
1397          * won't find us (yet).
1398          */
1399         if (fs->prot & VM_PROT_WRITE) {
1400                 vm_page_flag_set(fs->m, PG_WRITEABLE);
1401                 vm_object_set_writeable_dirty(fs->m->object);
1402
1403                 /*
1404                  * If the fault is a write, we know that this page is being
1405                  * written NOW so dirty it explicitly to save on
1406                  * pmap_is_modified() calls later.
1407                  *
1408                  * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1409                  * if the page is already dirty to prevent data written with
1410                  * the expectation of being synced from not being synced.
1411                  * Likewise if this entry does not request NOSYNC then make
1412                  * sure the page isn't marked NOSYNC.  Applications sharing
1413                  * data should use the same flags to avoid ping ponging.
1414                  *
1415                  * Also tell the backing pager, if any, that it should remove
1416                  * any swap backing since the page is now dirty.
1417                  */
1418                 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) {
1419                         if (fs->m->dirty == 0)
1420                                 vm_page_flag_set(fs->m, PG_NOSYNC);
1421                 } else {
1422                         vm_page_flag_clear(fs->m, PG_NOSYNC);
1423                 }
1424                 if (fs->fault_flags & VM_FAULT_DIRTY) {
1425                         crit_enter();
1426                         vm_page_dirty(fs->m);
1427                         vm_pager_page_unswapped(fs->m);
1428                         crit_exit();
1429                 }
1430         }
1431
1432         /*
1433          * Page had better still be busy.  We are still locked up and
1434          * fs->object will have another PIP reference if it is not equal
1435          * to fs->first_object.
1436          */
1437         KASSERT(fs->m->flags & PG_BUSY,
1438                 ("vm_fault: page %p not busy!", fs->m));
1439
1440         /*
1441          * Sanity check: page must be completely valid or it is not fit to
1442          * map into user space.  vm_pager_get_pages() ensures this.
1443          */
1444         if (fs->m->valid != VM_PAGE_BITS_ALL) {
1445                 vm_page_zero_invalid(fs->m, TRUE);
1446                 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1447         }
1448
1449         return (KERN_SUCCESS);
1450 }
1451
1452 /*
1453  * Wire down a range of virtual addresses in a map.  The entry in question
1454  * should be marked in-transition and the map must be locked.  We must
1455  * release the map temporarily while faulting-in the page to avoid a
1456  * deadlock.  Note that the entry may be clipped while we are blocked but
1457  * will never be freed.
1458  */
1459 int
1460 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1461 {
1462         boolean_t fictitious;
1463         vm_offset_t start;
1464         vm_offset_t end;
1465         vm_offset_t va;
1466         vm_paddr_t pa;
1467         pmap_t pmap;
1468         int rv;
1469
1470         pmap = vm_map_pmap(map);
1471         start = entry->start;
1472         end = entry->end;
1473         fictitious = entry->object.vm_object &&
1474                         (entry->object.vm_object->type == OBJT_DEVICE);
1475
1476         vm_map_unlock(map);
1477         map->timestamp++;
1478
1479         /*
1480          * We simulate a fault to get the page and enter it in the physical
1481          * map.
1482          */
1483         for (va = start; va < end; va += PAGE_SIZE) {
1484                 if (user_wire) {
1485                         rv = vm_fault(map, va, VM_PROT_READ,
1486                                         VM_FAULT_USER_WIRE);
1487                 } else {
1488                         rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1489                                         VM_FAULT_CHANGE_WIRING);
1490                 }
1491                 if (rv) {
1492                         while (va > start) {
1493                                 va -= PAGE_SIZE;
1494                                 if ((pa = pmap_extract(pmap, va)) == 0)
1495                                         continue;
1496                                 pmap_change_wiring(pmap, va, FALSE);
1497                                 if (!fictitious)
1498                                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1499                         }
1500                         vm_map_lock(map);
1501                         return (rv);
1502                 }
1503         }
1504         vm_map_lock(map);
1505         return (KERN_SUCCESS);
1506 }
1507
1508 /*
1509  * Unwire a range of virtual addresses in a map.  The map should be
1510  * locked.
1511  */
1512 void
1513 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1514 {
1515         boolean_t fictitious;
1516         vm_offset_t start;
1517         vm_offset_t end;
1518         vm_offset_t va;
1519         vm_paddr_t pa;
1520         pmap_t pmap;
1521
1522         pmap = vm_map_pmap(map);
1523         start = entry->start;
1524         end = entry->end;
1525         fictitious = entry->object.vm_object &&
1526                         (entry->object.vm_object->type == OBJT_DEVICE);
1527
1528         /*
1529          * Since the pages are wired down, we must be able to get their
1530          * mappings from the physical map system.
1531          */
1532         for (va = start; va < end; va += PAGE_SIZE) {
1533                 pa = pmap_extract(pmap, va);
1534                 if (pa != 0) {
1535                         pmap_change_wiring(pmap, va, FALSE);
1536                         if (!fictitious)
1537                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1538                 }
1539         }
1540 }
1541
1542 /*
1543  * Reduce the rate at which memory is allocated to a process based
1544  * on the perceived load on the VM system. As the load increases
1545  * the allocation burst rate goes down and the delay increases.
1546  *
1547  * Rate limiting does not apply when faulting active or inactive
1548  * pages.  When faulting 'cache' pages, rate limiting only applies
1549  * if the system currently has a severe page deficit.
1550  *
1551  * XXX vm_pagesupply should be increased when a page is freed.
1552  *
1553  * We sleep up to 1/10 of a second.
1554  */
1555 static int
1556 vm_fault_ratelimit(struct vmspace *vmspace)
1557 {
1558         if (vm_load_enable == 0)
1559                 return(0);
1560         if (vmspace->vm_pagesupply > 0) {
1561                 --vmspace->vm_pagesupply;
1562                 return(0);
1563         }
1564 #ifdef INVARIANTS
1565         if (vm_load_debug) {
1566                 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1567                         vm_load,
1568                         (1000 - vm_load ) / 10, vm_load * hz / 10000,
1569                         curproc->p_pid, curproc->p_comm);
1570         }
1571 #endif
1572         vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1573         return(vm_load * hz / 10000);
1574 }
1575
1576 /*
1577  *      Routine:
1578  *              vm_fault_copy_entry
1579  *      Function:
1580  *              Copy all of the pages from a wired-down map entry to another.
1581  *
1582  *      In/out conditions:
1583  *              The source and destination maps must be locked for write.
1584  *              The source map entry must be wired down (or be a sharing map
1585  *              entry corresponding to a main map entry that is wired down).
1586  */
1587
1588 void
1589 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1590     vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1591 {
1592         vm_object_t dst_object;
1593         vm_object_t src_object;
1594         vm_ooffset_t dst_offset;
1595         vm_ooffset_t src_offset;
1596         vm_prot_t prot;
1597         vm_offset_t vaddr;
1598         vm_page_t dst_m;
1599         vm_page_t src_m;
1600
1601 #ifdef  lint
1602         src_map++;
1603 #endif  /* lint */
1604
1605         src_object = src_entry->object.vm_object;
1606         src_offset = src_entry->offset;
1607
1608         /*
1609          * Create the top-level object for the destination entry. (Doesn't
1610          * actually shadow anything - we copy the pages directly.)
1611          */
1612         vm_map_entry_allocate_object(dst_entry);
1613         dst_object = dst_entry->object.vm_object;
1614
1615         prot = dst_entry->max_protection;
1616
1617         /*
1618          * Loop through all of the pages in the entry's range, copying each
1619          * one from the source object (it should be there) to the destination
1620          * object.
1621          */
1622         for (vaddr = dst_entry->start, dst_offset = 0;
1623             vaddr < dst_entry->end;
1624             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1625
1626                 /*
1627                  * Allocate a page in the destination object
1628                  */
1629                 do {
1630                         dst_m = vm_page_alloc(dst_object,
1631                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1632                         if (dst_m == NULL) {
1633                                 vm_wait();
1634                         }
1635                 } while (dst_m == NULL);
1636
1637                 /*
1638                  * Find the page in the source object, and copy it in.
1639                  * (Because the source is wired down, the page will be in
1640                  * memory.)
1641                  */
1642                 src_m = vm_page_lookup(src_object,
1643                         OFF_TO_IDX(dst_offset + src_offset));
1644                 if (src_m == NULL)
1645                         panic("vm_fault_copy_wired: page missing");
1646
1647                 vm_page_copy(src_m, dst_m);
1648
1649                 /*
1650                  * Enter it in the pmap...
1651                  */
1652
1653                 vm_page_flag_clear(dst_m, PG_ZERO);
1654                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1655                 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1656
1657                 /*
1658                  * Mark it no longer busy, and put it on the active list.
1659                  */
1660                 vm_page_activate(dst_m);
1661                 vm_page_wakeup(dst_m);
1662         }
1663 }
1664
1665
1666 /*
1667  * This routine checks around the requested page for other pages that
1668  * might be able to be faulted in.  This routine brackets the viable
1669  * pages for the pages to be paged in.
1670  *
1671  * Inputs:
1672  *      m, rbehind, rahead
1673  *
1674  * Outputs:
1675  *  marray (array of vm_page_t), reqpage (index of requested page)
1676  *
1677  * Return value:
1678  *  number of pages in marray
1679  */
1680 static int
1681 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1682     vm_page_t *marray, int *reqpage)
1683 {
1684         int i,j;
1685         vm_object_t object;
1686         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1687         vm_page_t rtm;
1688         int cbehind, cahead;
1689
1690         object = m->object;
1691         pindex = m->pindex;
1692
1693         /*
1694          * we don't fault-ahead for device pager
1695          */
1696         if (object->type == OBJT_DEVICE) {
1697                 *reqpage = 0;
1698                 marray[0] = m;
1699                 return 1;
1700         }
1701
1702         /*
1703          * if the requested page is not available, then give up now
1704          */
1705
1706         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1707                 return 0;
1708         }
1709
1710         if ((cbehind == 0) && (cahead == 0)) {
1711                 *reqpage = 0;
1712                 marray[0] = m;
1713                 return 1;
1714         }
1715
1716         if (rahead > cahead) {
1717                 rahead = cahead;
1718         }
1719
1720         if (rbehind > cbehind) {
1721                 rbehind = cbehind;
1722         }
1723
1724         /*
1725          * try to do any readahead that we might have free pages for.
1726          */
1727         if ((rahead + rbehind) >
1728                 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) {
1729                 pagedaemon_wakeup();
1730                 marray[0] = m;
1731                 *reqpage = 0;
1732                 return 1;
1733         }
1734
1735         /*
1736          * scan backward for the read behind pages -- in memory
1737          *
1738          * Assume that if the page is not found an interrupt will not
1739          * create it.  Theoretically interrupts can only remove (busy)
1740          * pages, not create new associations.
1741          */
1742         if (pindex > 0) {
1743                 if (rbehind > pindex) {
1744                         rbehind = pindex;
1745                         startpindex = 0;
1746                 } else {
1747                         startpindex = pindex - rbehind;
1748                 }
1749
1750                 crit_enter();
1751                 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1752                         if (vm_page_lookup( object, tpindex)) {
1753                                 startpindex = tpindex + 1;
1754                                 break;
1755                         }
1756                         if (tpindex == 0)
1757                                 break;
1758                 }
1759
1760                 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1761
1762                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1763                         if (rtm == NULL) {
1764                                 crit_exit();
1765                                 for (j = 0; j < i; j++) {
1766                                         vm_page_free(marray[j]);
1767                                 }
1768                                 marray[0] = m;
1769                                 *reqpage = 0;
1770                                 return 1;
1771                         }
1772
1773                         marray[i] = rtm;
1774                 }
1775                 crit_exit();
1776         } else {
1777                 startpindex = 0;
1778                 i = 0;
1779         }
1780
1781         marray[i] = m;
1782         /* page offset of the required page */
1783         *reqpage = i;
1784
1785         tpindex = pindex + 1;
1786         i++;
1787
1788         /*
1789          * scan forward for the read ahead pages
1790          */
1791         endpindex = tpindex + rahead;
1792         if (endpindex > object->size)
1793                 endpindex = object->size;
1794
1795         crit_enter();
1796         for( ; tpindex < endpindex; i++, tpindex++) {
1797
1798                 if (vm_page_lookup(object, tpindex)) {
1799                         break;
1800                 }
1801
1802                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1803                 if (rtm == NULL) {
1804                         break;
1805                 }
1806
1807                 marray[i] = rtm;
1808         }
1809         crit_exit();
1810
1811         /* return number of bytes of pages */
1812         return i;
1813 }