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 1.46 2008/05/09 07:24:48 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_deactivate(fs->m);
 135         vm_page_wakeup(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_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         struct lwp *lp = curthread->td_lwp;
 422         vm_page_t m;
 423
 424         m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
 425                           fault_type, VM_FAULT_NORMAL, errorp);
 426         return(m);
 427 }
 428
 429 /*
 430  * Fault in the specified virtual address in the specified map, doing all
 431  * necessary manipulation of the object store and all necessary I/O.  Return
 432  * a held VM page or NULL, and set *errorp.  The related pmap is not
 433  * updated.
 434  *
 435  * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
 436  * and marked PG_REFERENCED as well.
 437  *
 438  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
 439  * error will be returned.
 440  */
 441 vm_page_t
 442 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 443               int fault_flags, int *errorp)
 444 {
 445         vm_pindex_t first_pindex;
 446         struct faultstate fs;
 447         int result;
 448         vm_prot_t orig_fault_type = fault_type;
 449
 450         mycpu->gd_cnt.v_vm_faults++;
 451
 452         fs.didlimit = 0;
 453         fs.hardfault = 0;
 454         fs.fault_flags = fault_flags;
 455         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
 456
 457 RetryFault:
 458         /*
 459          * Find the vm_map_entry representing the backing store and resolve
 460          * the top level object and page index.  This may have the side
 461          * effect of executing a copy-on-write on the map entry and/or
 462          * creating a shadow object, but will not COW any actual VM pages.
 463          *
 464          * On success fs.map is left read-locked and various other fields
 465          * are initialized but not otherwise referenced or locked.
 466          *
 467          * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
 468          * if the map entry is a virtual page table and also writable,
 469          * so we can set the 'A'accessed bit in the virtual page table entry.
 470          */
 471         fs.map = map;
 472         result = vm_map_lookup(&fs.map, vaddr, fault_type,
 473                                &fs.entry, &fs.first_object,
 474                                &first_pindex, &fs.first_prot, &fs.wired);
 475
 476         if (result != KERN_SUCCESS) {
 477                 *errorp = result;
 478                 return (NULL);
 479         }
 480
 481         /*
 482          * fs.map is read-locked
 483          *
 484          * Misc checks.  Save the map generation number to detect races.
 485          */
 486         fs.map_generation = fs.map->timestamp;
 487
 488         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 489                 panic("vm_fault: fault on nofault entry, addr: %lx",
 490                     (u_long)vaddr);
 491         }
 492
 493         /*
 494          * A system map entry may return a NULL object.  No object means
 495          * no pager means an unrecoverable kernel fault.
 496          */
 497         if (fs.first_object == NULL) {
 498                 panic("vm_fault: unrecoverable fault at %p in entry %p",
 499                         (void *)vaddr, fs.entry);
 500         }
 501
 502         /*
 503          * Make a reference to this object to prevent its disposal while we
 504          * are messing with it.  Once we have the reference, the map is free
 505          * to be diddled.  Since objects reference their shadows (and copies),
 506          * they will stay around as well.
 507          *
 508          * Bump the paging-in-progress count to prevent size changes (e.g.
 509          * truncation operations) during I/O.  This must be done after
 510          * obtaining the vnode lock in order to avoid possible deadlocks.
 511          */
 512         vm_object_reference(fs.first_object);
 513         fs.vp = vnode_pager_lock(fs.first_object);
 514         vm_object_pip_add(fs.first_object, 1);
 515
 516         fs.lookup_still_valid = TRUE;
 517         fs.first_m = NULL;
 518         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 519
 520         /*
 521          * If the entry is wired we cannot change the page protection.
 522          */
 523         if (fs.wired)
 524                 fault_type = fs.first_prot;
 525
 526         /*
 527          * The page we want is at (first_object, first_pindex), but if the
 528          * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
 529          * page table to figure out the actual pindex.
 530          *
 531          * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
 532          * ONLY
 533          */
 534         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 535                 result = vm_fault_vpagetable(&fs, &first_pindex,
 536                                              fs.entry->aux.master_pde,
 537                                              fault_type);
 538                 if (result == KERN_TRY_AGAIN)
 539                         goto RetryFault;
 540                 if (result != KERN_SUCCESS) {
 541                         *errorp = result;
 542                         return (NULL);
 543                 }
 544         }
 545
 546         /*
 547          * Now we have the actual (object, pindex), fault in the page.  If
 548          * vm_fault_object() fails it will unlock and deallocate the FS
 549          * data.   If it succeeds everything remains locked and fs->object
 550          * will have an additinal PIP count if it is not equal to
 551          * fs->first_object
 552          */
 553         result = vm_fault_object(&fs, first_pindex, fault_type);
 554
 555         if (result == KERN_TRY_AGAIN)
 556                 goto RetryFault;
 557         if (result != KERN_SUCCESS) {
 558                 *errorp = result;
 559                 return(NULL);
 560         }
 561
 562         if ((orig_fault_type & VM_PROT_WRITE) &&
 563             (fs.prot & VM_PROT_WRITE) == 0) {
 564                 *errorp = KERN_PROTECTION_FAILURE;
 565                 unlock_and_deallocate(&fs);
 566                 return(NULL);
 567         }
 568
 569         /*
 570          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 571          * will contain a busied page.
 572          */
 573         unlock_things(&fs);
 574
 575         /*
 576          * Return a held page.  We are not doing any pmap manipulation so do
 577          * not set PG_MAPPED.  However, adjust the page flags according to
 578          * the fault type because the caller may not use a managed pmapping
 579          * (so we don't want to lose the fact that the page will be dirtied
 580          * if a write fault was specified).
 581          */
 582         vm_page_hold(fs.m);
 583         vm_page_flag_clear(fs.m, PG_ZERO);
 584         if (fault_type & VM_PROT_WRITE)
 585                 vm_page_dirty(fs.m);
 586
 587         /*
 588          * Update the pmap.  We really only have to do this if a COW
 589          * occured to replace the read-only page with the new page.  For
 590          * now just do it unconditionally. XXX
 591          */
 592         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
 593         vm_page_flag_set(fs.m, PG_REFERENCED);
 594
 595         /*
 596          * Unbusy the page by activating it.  It remains held and will not
 597          * be reclaimed.
 598          */
 599         vm_page_activate(fs.m);
 600
 601         if (curthread->td_lwp) {
 602                 if (fs.hardfault) {
 603                         curthread->td_lwp->lwp_ru.ru_majflt++;
 604                 } else {
 605                         curthread->td_lwp->lwp_ru.ru_minflt++;
 606                 }
 607         }
 608
 609         /*
 610          * Unlock everything, and return the held page.
 611          */
 612         vm_page_wakeup(fs.m);
 613         vm_object_deallocate(fs.first_object);
 614
 615         *errorp = 0;
 616         return(fs.m);
 617 }
 618
 619 /*
 620  * Fault in the specified (object,offset), dirty the returned page as
 621  * needed.  If the requested fault_type cannot be done NULL and an
 622  * error is returned.
 623  */
 624 vm_page_t
 625 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
 626                      vm_prot_t fault_type, int fault_flags, int *errorp)
 627 {
 628         int result;
 629         vm_pindex_t first_pindex;
 630         struct faultstate fs;
 631         struct vm_map_entry entry;
 632
 633         bzero(&entry, sizeof(entry));
 634         entry.object.vm_object = object;
 635         entry.maptype = VM_MAPTYPE_NORMAL;
 636         entry.protection = entry.max_protection = fault_type;
 637
 638         fs.didlimit = 0;
 639         fs.hardfault = 0;
 640         fs.fault_flags = fault_flags;
 641         fs.map = NULL;
 642         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
 643
 644 RetryFault:
 645
 646         fs.first_object = object;
 647         first_pindex = OFF_TO_IDX(offset);
 648         fs.entry = &entry;
 649         fs.first_prot = fault_type;
 650         fs.wired = 0;
 651         /*fs.map_generation = 0; unused */
 652
 653         /*
 654          * Make a reference to this object to prevent its disposal while we
 655          * are messing with it.  Once we have the reference, the map is free
 656          * to be diddled.  Since objects reference their shadows (and copies),
 657          * they will stay around as well.
 658          *
 659          * Bump the paging-in-progress count to prevent size changes (e.g.
 660          * truncation operations) during I/O.  This must be done after
 661          * obtaining the vnode lock in order to avoid possible deadlocks.
 662          */
 663         vm_object_reference(fs.first_object);
 664         fs.vp = vnode_pager_lock(fs.first_object);
 665         vm_object_pip_add(fs.first_object, 1);
 666
 667         fs.lookup_still_valid = TRUE;
 668         fs.first_m = NULL;
 669         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 670
 671 #if 0
 672         /* XXX future - ability to operate on VM object using vpagetable */
 673         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 674                 result = vm_fault_vpagetable(&fs, &first_pindex,
 675                                              fs.entry->aux.master_pde,
 676                                              fault_type);
 677                 if (result == KERN_TRY_AGAIN)
 678                         goto RetryFault;
 679                 if (result != KERN_SUCCESS) {
 680                         *errorp = result;
 681                         return (NULL);
 682                 }
 683         }
 684 #endif
 685
 686         /*
 687          * Now we have the actual (object, pindex), fault in the page.  If
 688          * vm_fault_object() fails it will unlock and deallocate the FS
 689          * data.   If it succeeds everything remains locked and fs->object
 690          * will have an additinal PIP count if it is not equal to
 691          * fs->first_object
 692          */
 693         result = vm_fault_object(&fs, first_pindex, fault_type);
 694
 695         if (result == KERN_TRY_AGAIN)
 696                 goto RetryFault;
 697         if (result != KERN_SUCCESS) {
 698                 *errorp = result;
 699                 return(NULL);
 700         }
 701
 702         if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
 703                 *errorp = KERN_PROTECTION_FAILURE;
 704                 unlock_and_deallocate(&fs);
 705                 return(NULL);
 706         }
 707
 708         /*
 709          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 710          * will contain a busied page.
 711          */
 712         unlock_things(&fs);
 713
 714         /*
 715          * Return a held page.  We are not doing any pmap manipulation so do
 716          * not set PG_MAPPED.  However, adjust the page flags according to
 717          * the fault type because the caller may not use a managed pmapping
 718          * (so we don't want to lose the fact that the page will be dirtied
 719          * if a write fault was specified).
 720          */
 721         vm_page_hold(fs.m);
 722         vm_page_flag_clear(fs.m, PG_ZERO);
 723         if (fault_type & VM_PROT_WRITE)
 724                 vm_page_dirty(fs.m);
 725
 726         /*
 727          * Indicate that the page was accessed.
 728          */
 729         vm_page_flag_set(fs.m, PG_REFERENCED);
 730
 731         /*
 732          * Unbusy the page by activating it.  It remains held and will not
 733          * be reclaimed.
 734          */
 735         vm_page_activate(fs.m);
 736
 737         if (curthread->td_lwp) {
 738                 if (fs.hardfault) {
 739                         mycpu->gd_cnt.v_vm_faults++;
 740                         curthread->td_lwp->lwp_ru.ru_majflt++;
 741                 } else {
 742                         curthread->td_lwp->lwp_ru.ru_minflt++;
 743                 }
 744         }
 745
 746         /*
 747          * Unlock everything, and return the held page.
 748          */
 749         vm_page_wakeup(fs.m);
 750         vm_object_deallocate(fs.first_object);
 751
 752         *errorp = 0;
 753         return(fs.m);
 754 }
 755
 756 /*
 757  * Translate the virtual page number (first_pindex) that is relative
 758  * to the address space into a logical page number that is relative to the
 759  * backing object.  Use the virtual page table pointed to by (vpte).
 760  *
 761  * This implements an N-level page table.  Any level can terminate the
 762  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
 763  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
 764  */
 765 static
 766 int
 767 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
 768                     vpte_t vpte, int fault_type)
 769 {
 770         struct sf_buf *sf;
 771         int vshift = 32 - PAGE_SHIFT;   /* page index bits remaining */
 772         int result = KERN_SUCCESS;
 773         vpte_t *ptep;
 774
 775         for (;;) {
 776                 /*
 777                  * We cannot proceed if the vpte is not valid, not readable
 778                  * for a read fault, or not writable for a write fault.
 779                  */
 780                 if ((vpte & VPTE_V) == 0) {
 781                         unlock_and_deallocate(fs);
 782                         return (KERN_FAILURE);
 783                 }
 784                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
 785                         unlock_and_deallocate(fs);
 786                         return (KERN_FAILURE);
 787                 }
 788                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
 789                         unlock_and_deallocate(fs);
 790                         return (KERN_FAILURE);
 791                 }
 792                 if ((vpte & VPTE_PS) || vshift == 0)
 793                         break;
 794                 KKASSERT(vshift >= VPTE_PAGE_BITS);
 795
 796                 /*
 797                  * Get the page table page.  Nominally we only read the page
 798                  * table, but since we are actively setting VPTE_M and VPTE_A,
 799                  * tell vm_fault_object() that we are writing it.
 800                  *
 801                  * There is currently no real need to optimize this.
 802                  */
 803                 result = vm_fault_object(fs, vpte >> PAGE_SHIFT,
 804                                          VM_PROT_READ|VM_PROT_WRITE);
 805                 if (result != KERN_SUCCESS)
 806                         return (result);
 807
 808                 /*
 809                  * Process the returned fs.m and look up the page table
 810                  * entry in the page table page.
 811                  */
 812                 vshift -= VPTE_PAGE_BITS;
 813                 sf = sf_buf_alloc(fs->m, SFB_CPUPRIVATE);
 814                 ptep = ((vpte_t *)sf_buf_kva(sf) +
 815                         ((*pindex >> vshift) & VPTE_PAGE_MASK));
 816                 vpte = *ptep;
 817
 818                 /*
 819                  * Page table write-back.  If the vpte is valid for the
 820                  * requested operation, do a write-back to the page table.
 821                  *
 822                  * XXX VPTE_M is not set properly for page directory pages.
 823                  * It doesn't get set in the page directory if the page table
 824                  * is modified during a read access.
 825                  */
 826                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
 827                     (vpte & VPTE_W)) {
 828                         if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
 829                                 atomic_set_int(ptep, VPTE_M|VPTE_A);
 830                                 vm_page_dirty(fs->m);
 831                         }
 832                 }
 833                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
 834                     (vpte & VPTE_R)) {
 835                         if ((vpte & VPTE_A) == 0) {
 836                                 atomic_set_int(ptep, VPTE_A);
 837                                 vm_page_dirty(fs->m);
 838                         }
 839                 }
 840                 sf_buf_free(sf);
 841                 vm_page_flag_set(fs->m, PG_REFERENCED);
 842                 vm_page_activate(fs->m);
 843                 vm_page_wakeup(fs->m);
 844                 cleanup_successful_fault(fs);
 845         }
 846         /*
 847          * Combine remaining address bits with the vpte.
 848          */
 849         *pindex = (vpte >> PAGE_SHIFT) +
 850                   (*pindex & ((1 << vshift) - 1));
 851         return (KERN_SUCCESS);
 852 }
 853
 854
 855 /*
 856  * Do all operations required to fault-in (fs.first_object, pindex).  Run
 857  * through the shadow chain as necessary and do required COW or virtual
 858  * copy operations.  The caller has already fully resolved the vm_map_entry
 859  * and, if appropriate, has created a copy-on-write layer.  All we need to
 860  * do is iterate the object chain.
 861  *
 862  * On failure (fs) is unlocked and deallocated and the caller may return or
 863  * retry depending on the failure code.  On success (fs) is NOT unlocked or
 864  * deallocated, fs.m will contained a resolved, busied page, and fs.object
 865  * will have an additional PIP count if it is not equal to fs.first_object.
 866  */
 867 static
 868 int
 869 vm_fault_object(struct faultstate *fs,
 870                 vm_pindex_t first_pindex, vm_prot_t fault_type)
 871 {
 872         vm_object_t next_object;
 873         vm_page_t marray[VM_FAULT_READ];
 874         vm_pindex_t pindex;
 875         int faultcount;
 876
 877         fs->prot = fs->first_prot;
 878         fs->object = fs->first_object;
 879         pindex = first_pindex;
 880
 881         /*
 882          * If a read fault occurs we try to make the page writable if
 883          * possible.  There are three cases where we cannot make the
 884          * page mapping writable:
 885          *
 886          * (1) The mapping is read-only or the VM object is read-only,
 887          *     fs->prot above will simply not have VM_PROT_WRITE set.
 888          *
 889          * (2) If the mapping is a virtual page table we need to be able
 890          *     to detect writes so we can set VPTE_M in the virtual page
 891          *     table.
 892          *
 893          * (3) If the VM page is read-only or copy-on-write, upgrading would
 894          *     just result in an unnecessary COW fault.
 895          *
 896          * VM_PROT_VPAGED is set if faulting via a virtual page table and
 897          * causes adjustments to the 'M'odify bit to also turn off write
 898          * access to force a re-fault.
 899          */
 900         if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 901                 if ((fault_type & VM_PROT_WRITE) == 0)
 902                         fs->prot &= ~VM_PROT_WRITE;
 903         }
 904
 905         for (;;) {
 906                 /*
 907                  * If the object is dead, we stop here
 908                  */
 909                 if (fs->object->flags & OBJ_DEAD) {
 910                         unlock_and_deallocate(fs);
 911                         return (KERN_PROTECTION_FAILURE);
 912                 }
 913
 914                 /*
 915                  * See if page is resident.  spl protection is required
 916                  * to avoid an interrupt unbusy/free race against our
 917                  * lookup.  We must hold the protection through a page
 918                  * allocation or busy.
 919                  */
 920                 crit_enter();
 921                 fs->m = vm_page_lookup(fs->object, pindex);
 922                 if (fs->m != NULL) {
 923                         int queue;
 924                         /*
 925                          * Wait/Retry if the page is busy.  We have to do this
 926                          * if the page is busy via either PG_BUSY or
 927                          * vm_page_t->busy because the vm_pager may be using
 928                          * vm_page_t->busy for pageouts ( and even pageins if
 929                          * it is the vnode pager ), and we could end up trying
 930                          * to pagein and pageout the same page simultaneously.
 931                          *
 932                          * We can theoretically allow the busy case on a read
 933                          * fault if the page is marked valid, but since such
 934                          * pages are typically already pmap'd, putting that
 935                          * special case in might be more effort then it is
 936                          * worth.  We cannot under any circumstances mess
 937                          * around with a vm_page_t->busy page except, perhaps,
 938                          * to pmap it.
 939                          */
 940                         if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
 941                                 unlock_things(fs);
 942                                 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
 943                                 mycpu->gd_cnt.v_intrans++;
 944                                 vm_object_deallocate(fs->first_object);
 945                                 crit_exit();
 946                                 return (KERN_TRY_AGAIN);
 947                         }
 948
 949                         /*
 950                          * If reactivating a page from PQ_CACHE we may have
 951                          * to rate-limit.
 952                          */
 953                         queue = fs->m->queue;
 954                         vm_page_unqueue_nowakeup(fs->m);
 955
 956                         if ((queue - fs->m->pc) == PQ_CACHE &&
 957                             vm_page_count_severe()) {
 958                                 vm_page_activate(fs->m);
 959                                 unlock_and_deallocate(fs);
 960                                 vm_waitpfault();
 961                                 crit_exit();
 962                                 return (KERN_TRY_AGAIN);
 963                         }
 964
 965                         /*
 966                          * Mark page busy for other processes, and the
 967                          * pagedaemon.  If it still isn't completely valid
 968                          * (readable), jump to readrest, else we found the
 969                          * page and can return.
 970                          *
 971                          * We can release the spl once we have marked the
 972                          * page busy.
 973                          */
 974                         vm_page_busy(fs->m);
 975                         crit_exit();
 976
 977                         if (((fs->m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
 978                             fs->m->object != &kernel_object) {
 979                                 goto readrest;
 980                         }
 981                         break; /* break to PAGE HAS BEEN FOUND */
 982                 }
 983
 984                 /*
 985                  * Page is not resident, If this is the search termination
 986                  * or the pager might contain the page, allocate a new page.
 987                  *
 988                  * NOTE: We are still in a critical section.
 989                  */
 990                 if (TRYPAGER(fs) || fs->object == fs->first_object) {
 991                         /*
 992                          * If the page is beyond the object size we fail
 993                          */
 994                         if (pindex >= fs->object->size) {
 995                                 crit_exit();
 996                                 unlock_and_deallocate(fs);
 997                                 return (KERN_PROTECTION_FAILURE);
 998                         }
 999
1000                         /*
1001                          * Ratelimit.
1002                          */
1003                         if (fs->didlimit == 0 && curproc != NULL) {
1004                                 int limticks;
1005
1006                                 limticks = vm_fault_ratelimit(curproc->p_vmspace);
1007                                 if (limticks) {
1008                                         crit_exit();
1009                                         unlock_and_deallocate(fs);
1010                                         tsleep(curproc, 0, "vmrate", limticks);
1011                                         fs->didlimit = 1;
1012                                         return (KERN_TRY_AGAIN);
1013                                 }
1014                         }
1015
1016                         /*
1017                          * Allocate a new page for this object/offset pair.
1018                          */
1019                         fs->m = NULL;
1020                         if (!vm_page_count_severe()) {
1021                                 fs->m = vm_page_alloc(fs->object, pindex,
1022                                     (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
1023                         }
1024                         if (fs->m == NULL) {
1025                                 crit_exit();
1026                                 unlock_and_deallocate(fs);
1027                                 vm_waitpfault();
1028                                 return (KERN_TRY_AGAIN);
1029                         }
1030                 }
1031                 crit_exit();
1032
1033 readrest:
1034                 /*
1035                  * We have found a valid page or we have allocated a new page.
1036                  * The page thus may not be valid or may not be entirely
1037                  * valid.
1038                  *
1039                  * Attempt to fault-in the page if there is a chance that the
1040                  * pager has it, and potentially fault in additional pages
1041                  * at the same time.
1042                  *
1043                  * We are NOT in splvm here and if TRYPAGER is true then
1044                  * fs.m will be non-NULL and will be PG_BUSY for us.
1045                  */
1046
1047                 if (TRYPAGER(fs)) {
1048                         int rv;
1049                         int reqpage;
1050                         int ahead, behind;
1051                         u_char behavior = vm_map_entry_behavior(fs->entry);
1052
1053                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
1054                                 ahead = 0;
1055                                 behind = 0;
1056                         } else {
1057                                 behind = pindex;
1058                                 if (behind > VM_FAULT_READ_BEHIND)
1059                                         behind = VM_FAULT_READ_BEHIND;
1060
1061                                 ahead = fs->object->size - pindex;
1062                                 if (ahead < 1)
1063                                         ahead = 1;
1064                                 if (ahead > VM_FAULT_READ_AHEAD)
1065                                         ahead = VM_FAULT_READ_AHEAD;
1066                         }
1067
1068                         if ((fs->first_object->type != OBJT_DEVICE) &&
1069                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1070                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1071                                 pindex >= fs->entry->lastr &&
1072                                 pindex < fs->entry->lastr + VM_FAULT_READ))
1073                         ) {
1074                                 vm_pindex_t firstpindex, tmppindex;
1075
1076                                 if (first_pindex < 2 * VM_FAULT_READ)
1077                                         firstpindex = 0;
1078                                 else
1079                                         firstpindex = first_pindex - 2 * VM_FAULT_READ;
1080
1081                                 /*
1082                                  * note: partially valid pages cannot be
1083                                  * included in the lookahead - NFS piecemeal
1084                                  * writes will barf on it badly.
1085                                  *
1086                                  * spl protection is required to avoid races
1087                                  * between the lookup and an interrupt
1088                                  * unbusy/free sequence occuring prior to
1089                                  * our busy check.
1090                                  */
1091                                 crit_enter();
1092                                 for (tmppindex = first_pindex - 1;
1093                                     tmppindex >= firstpindex;
1094                                     --tmppindex
1095                                 ) {
1096                                         vm_page_t mt;
1097
1098                                         mt = vm_page_lookup(fs->first_object, tmppindex);
1099                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
1100                                                 break;
1101                                         if (mt->busy ||
1102                                                 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
1103                                                 mt->hold_count ||
1104                                                 mt->wire_count)
1105                                                 continue;
1106                                         if (mt->dirty == 0)
1107                                                 vm_page_test_dirty(mt);
1108                                         if (mt->dirty) {
1109                                                 vm_page_busy(mt);
1110                                                 vm_page_protect(mt, VM_PROT_NONE);
1111                                                 vm_page_deactivate(mt);
1112                                                 vm_page_wakeup(mt);
1113                                         } else {
1114                                                 vm_page_cache(mt);
1115                                         }
1116                                 }
1117                                 crit_exit();
1118
1119                                 ahead += behind;
1120                                 behind = 0;
1121                         }
1122
1123                         /*
1124                          * now we find out if any other pages should be paged
1125                          * in at this time this routine checks to see if the
1126                          * pages surrounding this fault reside in the same
1127                          * object as the page for this fault.  If they do,
1128                          * then they are faulted in also into the object.  The
1129                          * array "marray" returned contains an array of
1130                          * vm_page_t structs where one of them is the
1131                          * vm_page_t passed to the routine.  The reqpage
1132                          * return value is the index into the marray for the
1133                          * vm_page_t passed to the routine.
1134                          *
1135                          * fs.m plus the additional pages are PG_BUSY'd.
1136                          */
1137                         faultcount = vm_fault_additional_pages(
1138                             fs->m, behind, ahead, marray, &reqpage);
1139
1140                         /*
1141                          * update lastr imperfectly (we do not know how much
1142                          * getpages will actually read), but good enough.
1143                          */
1144                         fs->entry->lastr = pindex + faultcount - behind;
1145
1146                         /*
1147                          * Call the pager to retrieve the data, if any, after
1148                          * releasing the lock on the map.  We hold a ref on
1149                          * fs.object and the pages are PG_BUSY'd.
1150                          */
1151                         unlock_map(fs);
1152
1153                         if (faultcount) {
1154                                 rv = vm_pager_get_pages(fs->object, marray,
1155                                                         faultcount, reqpage);
1156                         } else {
1157                                 rv = VM_PAGER_FAIL;
1158                         }
1159
1160                         if (rv == VM_PAGER_OK) {
1161                                 /*
1162                                  * Found the page. Leave it busy while we play
1163                                  * with it.
1164                                  */
1165
1166                                 /*
1167                                  * Relookup in case pager changed page. Pager
1168                                  * is responsible for disposition of old page
1169                                  * if moved.
1170                                  *
1171                                  * XXX other code segments do relookups too.
1172                                  * It's a bad abstraction that needs to be
1173                                  * fixed/removed.
1174                                  */
1175                                 fs->m = vm_page_lookup(fs->object, pindex);
1176                                 if (fs->m == NULL) {
1177                                         unlock_and_deallocate(fs);
1178                                         return (KERN_TRY_AGAIN);
1179                                 }
1180
1181                                 ++fs->hardfault;
1182                                 break; /* break to PAGE HAS BEEN FOUND */
1183                         }
1184
1185                         /*
1186                          * Remove the bogus page (which does not exist at this
1187                          * object/offset); before doing so, we must get back
1188                          * our object lock to preserve our invariant.
1189                          *
1190                          * Also wake up any other process that may want to bring
1191                          * in this page.
1192                          *
1193                          * If this is the top-level object, we must leave the
1194                          * busy page to prevent another process from rushing
1195                          * past us, and inserting the page in that object at
1196                          * the same time that we are.
1197                          */
1198                         if (rv == VM_PAGER_ERROR) {
1199                                 if (curproc)
1200                                         kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
1201                                 else
1202                                         kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
1203                         }
1204                         /*
1205                          * Data outside the range of the pager or an I/O error
1206                          *
1207                          * The page may have been wired during the pagein,
1208                          * e.g. by the buffer cache, and cannot simply be
1209                          * freed.  Call vnode_pager_freepag() to deal with it.
1210                          */
1211                         /*
1212                          * XXX - the check for kernel_map is a kludge to work
1213                          * around having the machine panic on a kernel space
1214                          * fault w/ I/O error.
1215                          */
1216                         if (((fs->map != &kernel_map) && (rv == VM_PAGER_ERROR)) ||
1217                                 (rv == VM_PAGER_BAD)) {
1218                                 vnode_pager_freepage(fs->m);
1219                                 fs->m = NULL;
1220                                 unlock_and_deallocate(fs);
1221                                 if (rv == VM_PAGER_ERROR)
1222                                         return (KERN_FAILURE);
1223                                 else
1224                                         return (KERN_PROTECTION_FAILURE);
1225                                 /* NOT REACHED */
1226                         }
1227                         if (fs->object != fs->first_object) {
1228                                 vnode_pager_freepage(fs->m);
1229                                 fs->m = NULL;
1230                                 /*
1231                                  * XXX - we cannot just fall out at this
1232                                  * point, m has been freed and is invalid!
1233                                  */
1234                         }
1235                 }
1236
1237                 /*
1238                  * We get here if the object has a default pager (or unwiring)
1239                  * or the pager doesn't have the page.
1240                  */
1241                 if (fs->object == fs->first_object)
1242                         fs->first_m = fs->m;
1243
1244                 /*
1245                  * Move on to the next object.  Lock the next object before
1246                  * unlocking the current one.
1247                  */
1248                 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1249                 next_object = fs->object->backing_object;
1250                 if (next_object == NULL) {
1251                         /*
1252                          * If there's no object left, fill the page in the top
1253                          * object with zeros.
1254                          */
1255                         if (fs->object != fs->first_object) {
1256                                 vm_object_pip_wakeup(fs->object);
1257
1258                                 fs->object = fs->first_object;
1259                                 pindex = first_pindex;
1260                                 fs->m = fs->first_m;
1261                         }
1262                         fs->first_m = NULL;
1263
1264                         /*
1265                          * Zero the page if necessary and mark it valid.
1266                          */
1267                         if ((fs->m->flags & PG_ZERO) == 0) {
1268                                 vm_page_zero_fill(fs->m);
1269                         } else {
1270                                 mycpu->gd_cnt.v_ozfod++;
1271                         }
1272                         mycpu->gd_cnt.v_zfod++;
1273                         fs->m->valid = VM_PAGE_BITS_ALL;
1274                         break;  /* break to PAGE HAS BEEN FOUND */
1275                 } else {
1276                         if (fs->object != fs->first_object) {
1277                                 vm_object_pip_wakeup(fs->object);
1278                         }
1279                         KASSERT(fs->object != next_object, ("object loop %p", next_object));
1280                         fs->object = next_object;
1281                         vm_object_pip_add(fs->object, 1);
1282                 }
1283         }
1284
1285         KASSERT((fs->m->flags & PG_BUSY) != 0,
1286                 ("vm_fault: not busy after main loop"));
1287
1288         /*
1289          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1290          * is held.]
1291          */
1292
1293         /*
1294          * If the page is being written, but isn't already owned by the
1295          * top-level object, we have to copy it into a new page owned by the
1296          * top-level object.
1297          */
1298         if (fs->object != fs->first_object) {
1299                 /*
1300                  * We only really need to copy if we want to write it.
1301                  */
1302                 if (fault_type & VM_PROT_WRITE) {
1303                         /*
1304                          * This allows pages to be virtually copied from a
1305                          * backing_object into the first_object, where the
1306                          * backing object has no other refs to it, and cannot
1307                          * gain any more refs.  Instead of a bcopy, we just
1308                          * move the page from the backing object to the
1309                          * first object.  Note that we must mark the page
1310                          * dirty in the first object so that it will go out
1311                          * to swap when needed.
1312                          */
1313                         if (
1314                                 /*
1315                                  * Map, if present, has not changed
1316                                  */
1317                                 (fs->map == NULL ||
1318                                 fs->map_generation == fs->map->timestamp) &&
1319                                 /*
1320                                  * Only one shadow object
1321                                  */
1322                                 (fs->object->shadow_count == 1) &&
1323                                 /*
1324                                  * No COW refs, except us
1325                                  */
1326                                 (fs->object->ref_count == 1) &&
1327                                 /*
1328                                  * No one else can look this object up
1329                                  */
1330                                 (fs->object->handle == NULL) &&
1331                                 /*
1332                                  * No other ways to look the object up
1333                                  */
1334                                 ((fs->object->type == OBJT_DEFAULT) ||
1335                                  (fs->object->type == OBJT_SWAP)) &&
1336                                 /*
1337                                  * We don't chase down the shadow chain
1338                                  */
1339                                 (fs->object == fs->first_object->backing_object) &&
1340
1341                                 /*
1342                                  * grab the lock if we need to
1343                                  */
1344                                 (fs->lookup_still_valid ||
1345                                  fs->map == NULL ||
1346                                  lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1347                             ) {
1348
1349                                 fs->lookup_still_valid = 1;
1350                                 /*
1351                                  * get rid of the unnecessary page
1352                                  */
1353                                 vm_page_protect(fs->first_m, VM_PROT_NONE);
1354                                 vm_page_free(fs->first_m);
1355                                 fs->first_m = NULL;
1356
1357                                 /*
1358                                  * grab the page and put it into the
1359                                  * process'es object.  The page is
1360                                  * automatically made dirty.
1361                                  */
1362                                 vm_page_rename(fs->m, fs->first_object, first_pindex);
1363                                 fs->first_m = fs->m;
1364                                 vm_page_busy(fs->first_m);
1365                                 fs->m = NULL;
1366                                 mycpu->gd_cnt.v_cow_optim++;
1367                         } else {
1368                                 /*
1369                                  * Oh, well, lets copy it.
1370                                  */
1371                                 vm_page_copy(fs->m, fs->first_m);
1372                                 vm_page_event(fs->m, VMEVENT_COW);
1373                         }
1374
1375                         if (fs->m) {
1376                                 /*
1377                                  * We no longer need the old page or object.
1378                                  */
1379                                 release_page(fs);
1380                         }
1381
1382                         /*
1383                          * fs->object != fs->first_object due to above
1384                          * conditional
1385                          */
1386                         vm_object_pip_wakeup(fs->object);
1387
1388                         /*
1389                          * Only use the new page below...
1390                          */
1391
1392                         mycpu->gd_cnt.v_cow_faults++;
1393                         fs->m = fs->first_m;
1394                         fs->object = fs->first_object;
1395                         pindex = first_pindex;
1396                 } else {
1397                         /*
1398                          * If it wasn't a write fault avoid having to copy
1399                          * the page by mapping it read-only.
1400                          */
1401                         fs->prot &= ~VM_PROT_WRITE;
1402                 }
1403         }
1404
1405         /*
1406          * We may have had to unlock a map to do I/O.  If we did then
1407          * lookup_still_valid will be FALSE.  If the map generation count
1408          * also changed then all sorts of things could have happened while
1409          * we were doing the I/O and we need to retry.
1410          */
1411
1412         if (!fs->lookup_still_valid &&
1413             fs->map != NULL &&
1414             (fs->map->timestamp != fs->map_generation)) {
1415                 release_page(fs);
1416                 unlock_and_deallocate(fs);
1417                 return (KERN_TRY_AGAIN);
1418         }
1419
1420         /*
1421          * If the fault is a write, we know that this page is being
1422          * written NOW so dirty it explicitly to save on pmap_is_modified()
1423          * calls later.
1424          *
1425          * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1426          * if the page is already dirty to prevent data written with
1427          * the expectation of being synced from not being synced.
1428          * Likewise if this entry does not request NOSYNC then make
1429          * sure the page isn't marked NOSYNC.  Applications sharing
1430          * data should use the same flags to avoid ping ponging.
1431          *
1432          * Also tell the backing pager, if any, that it should remove
1433          * any swap backing since the page is now dirty.
1434          */
1435         if (fs->prot & VM_PROT_WRITE) {
1436                 vm_object_set_writeable_dirty(fs->m->object);
1437                 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) {
1438                         if (fs->m->dirty == 0)
1439                                 vm_page_flag_set(fs->m, PG_NOSYNC);
1440                 } else {
1441                         vm_page_flag_clear(fs->m, PG_NOSYNC);
1442                 }
1443                 if (fs->fault_flags & VM_FAULT_DIRTY) {
1444                         crit_enter();
1445                         vm_page_dirty(fs->m);
1446                         vm_pager_page_unswapped(fs->m);
1447                         crit_exit();
1448                 }
1449         }
1450
1451         /*
1452          * Page had better still be busy.  We are still locked up and
1453          * fs->object will have another PIP reference if it is not equal
1454          * to fs->first_object.
1455          */
1456         KASSERT(fs->m->flags & PG_BUSY,
1457                 ("vm_fault: page %p not busy!", fs->m));
1458
1459         /*
1460          * Sanity check: page must be completely valid or it is not fit to
1461          * map into user space.  vm_pager_get_pages() ensures this.
1462          */
1463         if (fs->m->valid != VM_PAGE_BITS_ALL) {
1464                 vm_page_zero_invalid(fs->m, TRUE);
1465                 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1466         }
1467
1468         return (KERN_SUCCESS);
1469 }
1470
1471 /*
1472  * Wire down a range of virtual addresses in a map.  The entry in question
1473  * should be marked in-transition and the map must be locked.  We must
1474  * release the map temporarily while faulting-in the page to avoid a
1475  * deadlock.  Note that the entry may be clipped while we are blocked but
1476  * will never be freed.
1477  */
1478 int
1479 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1480 {
1481         boolean_t fictitious;
1482         vm_offset_t start;
1483         vm_offset_t end;
1484         vm_offset_t va;
1485         vm_paddr_t pa;
1486         pmap_t pmap;
1487         int rv;
1488
1489         pmap = vm_map_pmap(map);
1490         start = entry->start;
1491         end = entry->end;
1492         fictitious = entry->object.vm_object &&
1493                         (entry->object.vm_object->type == OBJT_DEVICE);
1494
1495         vm_map_unlock(map);
1496         map->timestamp++;
1497
1498         /*
1499          * We simulate a fault to get the page and enter it in the physical
1500          * map.
1501          */
1502         for (va = start; va < end; va += PAGE_SIZE) {
1503                 if (user_wire) {
1504                         rv = vm_fault(map, va, VM_PROT_READ,
1505                                         VM_FAULT_USER_WIRE);
1506                 } else {
1507                         rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1508                                         VM_FAULT_CHANGE_WIRING);
1509                 }
1510                 if (rv) {
1511                         while (va > start) {
1512                                 va -= PAGE_SIZE;
1513                                 if ((pa = pmap_extract(pmap, va)) == 0)
1514                                         continue;
1515                                 pmap_change_wiring(pmap, va, FALSE);
1516                                 if (!fictitious)
1517                                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1518                         }
1519                         vm_map_lock(map);
1520                         return (rv);
1521                 }
1522         }
1523         vm_map_lock(map);
1524         return (KERN_SUCCESS);
1525 }
1526
1527 /*
1528  * Unwire a range of virtual addresses in a map.  The map should be
1529  * locked.
1530  */
1531 void
1532 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1533 {
1534         boolean_t fictitious;
1535         vm_offset_t start;
1536         vm_offset_t end;
1537         vm_offset_t va;
1538         vm_paddr_t pa;
1539         pmap_t pmap;
1540
1541         pmap = vm_map_pmap(map);
1542         start = entry->start;
1543         end = entry->end;
1544         fictitious = entry->object.vm_object &&
1545                         (entry->object.vm_object->type == OBJT_DEVICE);
1546
1547         /*
1548          * Since the pages are wired down, we must be able to get their
1549          * mappings from the physical map system.
1550          */
1551         for (va = start; va < end; va += PAGE_SIZE) {
1552                 pa = pmap_extract(pmap, va);
1553                 if (pa != 0) {
1554                         pmap_change_wiring(pmap, va, FALSE);
1555                         if (!fictitious)
1556                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1557                 }
1558         }
1559 }
1560
1561 /*
1562  * Reduce the rate at which memory is allocated to a process based
1563  * on the perceived load on the VM system. As the load increases
1564  * the allocation burst rate goes down and the delay increases.
1565  *
1566  * Rate limiting does not apply when faulting active or inactive
1567  * pages.  When faulting 'cache' pages, rate limiting only applies
1568  * if the system currently has a severe page deficit.
1569  *
1570  * XXX vm_pagesupply should be increased when a page is freed.
1571  *
1572  * We sleep up to 1/10 of a second.
1573  */
1574 static int
1575 vm_fault_ratelimit(struct vmspace *vmspace)
1576 {
1577         if (vm_load_enable == 0)
1578                 return(0);
1579         if (vmspace->vm_pagesupply > 0) {
1580                 --vmspace->vm_pagesupply;
1581                 return(0);
1582         }
1583 #ifdef INVARIANTS
1584         if (vm_load_debug) {
1585                 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1586                         vm_load,
1587                         (1000 - vm_load ) / 10, vm_load * hz / 10000,
1588                         curproc->p_pid, curproc->p_comm);
1589         }
1590 #endif
1591         vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1592         return(vm_load * hz / 10000);
1593 }
1594
1595 /*
1596  *      Routine:
1597  *              vm_fault_copy_entry
1598  *      Function:
1599  *              Copy all of the pages from a wired-down map entry to another.
1600  *
1601  *      In/out conditions:
1602  *              The source and destination maps must be locked for write.
1603  *              The source map entry must be wired down (or be a sharing map
1604  *              entry corresponding to a main map entry that is wired down).
1605  */
1606
1607 void
1608 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1609     vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1610 {
1611         vm_object_t dst_object;
1612         vm_object_t src_object;
1613         vm_ooffset_t dst_offset;
1614         vm_ooffset_t src_offset;
1615         vm_prot_t prot;
1616         vm_offset_t vaddr;
1617         vm_page_t dst_m;
1618         vm_page_t src_m;
1619
1620 #ifdef  lint
1621         src_map++;
1622 #endif  /* lint */
1623
1624         src_object = src_entry->object.vm_object;
1625         src_offset = src_entry->offset;
1626
1627         /*
1628          * Create the top-level object for the destination entry. (Doesn't
1629          * actually shadow anything - we copy the pages directly.)
1630          */
1631         vm_map_entry_allocate_object(dst_entry);
1632         dst_object = dst_entry->object.vm_object;
1633
1634         prot = dst_entry->max_protection;
1635
1636         /*
1637          * Loop through all of the pages in the entry's range, copying each
1638          * one from the source object (it should be there) to the destination
1639          * object.
1640          */
1641         for (vaddr = dst_entry->start, dst_offset = 0;
1642             vaddr < dst_entry->end;
1643             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1644
1645                 /*
1646                  * Allocate a page in the destination object
1647                  */
1648                 do {
1649                         dst_m = vm_page_alloc(dst_object,
1650                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1651                         if (dst_m == NULL) {
1652                                 vm_wait();
1653                         }
1654                 } while (dst_m == NULL);
1655
1656                 /*
1657                  * Find the page in the source object, and copy it in.
1658                  * (Because the source is wired down, the page will be in
1659                  * memory.)
1660                  */
1661                 src_m = vm_page_lookup(src_object,
1662                         OFF_TO_IDX(dst_offset + src_offset));
1663                 if (src_m == NULL)
1664                         panic("vm_fault_copy_wired: page missing");
1665
1666                 vm_page_copy(src_m, dst_m);
1667                 vm_page_event(src_m, VMEVENT_COW);
1668
1669                 /*
1670                  * Enter it in the pmap...
1671                  */
1672
1673                 vm_page_flag_clear(dst_m, PG_ZERO);
1674                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1675
1676                 /*
1677                  * Mark it no longer busy, and put it on the active list.
1678                  */
1679                 vm_page_activate(dst_m);
1680                 vm_page_wakeup(dst_m);
1681         }
1682 }
1683
1684
1685 /*
1686  * This routine checks around the requested page for other pages that
1687  * might be able to be faulted in.  This routine brackets the viable
1688  * pages for the pages to be paged in.
1689  *
1690  * Inputs:
1691  *      m, rbehind, rahead
1692  *
1693  * Outputs:
1694  *  marray (array of vm_page_t), reqpage (index of requested page)
1695  *
1696  * Return value:
1697  *  number of pages in marray
1698  */
1699 static int
1700 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1701     vm_page_t *marray, int *reqpage)
1702 {
1703         int i,j;
1704         vm_object_t object;
1705         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1706         vm_page_t rtm;
1707         int cbehind, cahead;
1708
1709         object = m->object;
1710         pindex = m->pindex;
1711
1712         /*
1713          * we don't fault-ahead for device pager
1714          */
1715         if (object->type == OBJT_DEVICE) {
1716                 *reqpage = 0;
1717                 marray[0] = m;
1718                 return 1;
1719         }
1720
1721         /*
1722          * if the requested page is not available, then give up now
1723          */
1724         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1725                 *reqpage = 0;   /* not used by caller, fix compiler warn */
1726                 return 0;
1727         }
1728
1729         if ((cbehind == 0) && (cahead == 0)) {
1730                 *reqpage = 0;
1731                 marray[0] = m;
1732                 return 1;
1733         }
1734
1735         if (rahead > cahead) {
1736                 rahead = cahead;
1737         }
1738
1739         if (rbehind > cbehind) {
1740                 rbehind = cbehind;
1741         }
1742
1743         /*
1744          * try to do any readahead that we might have free pages for.
1745          */
1746         if ((rahead + rbehind) >
1747                 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) {
1748                 pagedaemon_wakeup();
1749                 marray[0] = m;
1750                 *reqpage = 0;
1751                 return 1;
1752         }
1753
1754         /*
1755          * scan backward for the read behind pages -- in memory
1756          *
1757          * Assume that if the page is not found an interrupt will not
1758          * create it.  Theoretically interrupts can only remove (busy)
1759          * pages, not create new associations.
1760          */
1761         if (pindex > 0) {
1762                 if (rbehind > pindex) {
1763                         rbehind = pindex;
1764                         startpindex = 0;
1765                 } else {
1766                         startpindex = pindex - rbehind;
1767                 }
1768
1769                 crit_enter();
1770                 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1771                         if (vm_page_lookup( object, tpindex)) {
1772                                 startpindex = tpindex + 1;
1773                                 break;
1774                         }
1775                         if (tpindex == 0)
1776                                 break;
1777                 }
1778
1779                 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1780
1781                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1782                         if (rtm == NULL) {
1783                                 crit_exit();
1784                                 for (j = 0; j < i; j++) {
1785                                         vm_page_free(marray[j]);
1786                                 }
1787                                 marray[0] = m;
1788                                 *reqpage = 0;
1789                                 return 1;
1790                         }
1791
1792                         marray[i] = rtm;
1793                 }
1794                 crit_exit();
1795         } else {
1796                 startpindex = 0;
1797                 i = 0;
1798         }
1799
1800         marray[i] = m;
1801         /* page offset of the required page */
1802         *reqpage = i;
1803
1804         tpindex = pindex + 1;
1805         i++;
1806
1807         /*
1808          * scan forward for the read ahead pages
1809          */
1810         endpindex = tpindex + rahead;
1811         if (endpindex > object->size)
1812                 endpindex = object->size;
1813
1814         crit_enter();
1815         for( ; tpindex < endpindex; i++, tpindex++) {
1816
1817                 if (vm_page_lookup(object, tpindex)) {
1818                         break;
1819                 }
1820
1821                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1822                 if (rtm == NULL) {
1823                         break;
1824                 }
1825
1826                 marray[i] = rtm;
1827         }
1828         crit_exit();
1829
1830         /* return number of bytes of pages */
1831         return i;
1832 }