sys/vm/vm_fault.c

   1 /*
   2  * (MPSAFE)
   3  *
   4  * Copyright (c) 1991, 1993
   5  *      The Regents of the University of California.  All rights reserved.
   6  * Copyright (c) 1994 John S. Dyson
   7  * All rights reserved.
   8  * Copyright (c) 1994 David Greenman
   9  * All rights reserved.
  10  *
  11  *
  12  * This code is derived from software contributed to Berkeley by
  13  * The Mach Operating System project at Carnegie-Mellon University.
  14  *
  15  * Redistribution and use in source and binary forms, with or without
  16  * modification, are permitted provided that the following conditions
  17  * are met:
  18  * 1. Redistributions of source code must retain the above copyright
  19  *    notice, this list of conditions and the following disclaimer.
  20  * 2. Redistributions in binary form must reproduce the above copyright
  21  *    notice, this list of conditions and the following disclaimer in the
  22  *    documentation and/or other materials provided with the distribution.
  23  * 3. All advertising materials mentioning features or use of this software
  24  *    must display the following acknowledgement:
  25  *      This product includes software developed by the University of
  26  *      California, Berkeley and its contributors.
  27  * 4. Neither the name of the University nor the names of its contributors
  28  *    may be used to endorse or promote products derived from this software
  29  *    without specific prior written permission.
  30  *
  31  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  32  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  34  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  35  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  36  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  37  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  38  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  39  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  40  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  41  * SUCH DAMAGE.
  42  *
  43  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
  44  *
  45  *
  46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  47  * All rights reserved.
  48  *
  49  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  50  *
  51  * Permission to use, copy, modify and distribute this software and
  52  * its documentation is hereby granted, provided that both the copyright
  53  * notice and this permission notice appear in all copies of the
  54  * software, derivative works or modified versions, and any portions
  55  * thereof, and that both notices appear in supporting documentation.
  56  *
  57  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  58  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  59  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  60  *
  61  * Carnegie Mellon requests users of this software to return to
  62  *
  63  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  64  *  School of Computer Science
  65  *  Carnegie Mellon University
  66  *  Pittsburgh PA 15213-3890
  67  *
  68  * any improvements or extensions that they make and grant Carnegie the
  69  * rights to redistribute these changes.
  70  *
  71  * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
  72  * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $
  73  */
  74
  75 /*
  76  *      Page fault handling module.
  77  */
  78
  79 #include <sys/param.h>
  80 #include <sys/systm.h>
  81 #include <sys/kernel.h>
  82 #include <sys/proc.h>
  83 #include <sys/vnode.h>
  84 #include <sys/resourcevar.h>
  85 #include <sys/vmmeter.h>
  86 #include <sys/vkernel.h>
  87 #include <sys/lock.h>
  88 #include <sys/sysctl.h>
  89
  90 #include <cpu/lwbuf.h>
  91
  92 #include <vm/vm.h>
  93 #include <vm/vm_param.h>
  94 #include <vm/pmap.h>
  95 #include <vm/vm_map.h>
  96 #include <vm/vm_object.h>
  97 #include <vm/vm_page.h>
  98 #include <vm/vm_pageout.h>
  99 #include <vm/vm_kern.h>
 100 #include <vm/vm_pager.h>
 101 #include <vm/vnode_pager.h>
 102 #include <vm/vm_extern.h>
 103
 104 #include <sys/thread2.h>
 105 #include <vm/vm_page2.h>
 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_fast_fault = 1;
 127 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
 128            "Burst fault zero-fill regions");
 129 static int debug_cluster = 0;
 130 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
 131
 132 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
 133 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
 134 #if 0
 135 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
 136 #endif
 137 static int vm_fault_ratelimit(struct vmspace *);
 138 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry);
 139 static void vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry,
 140                         int prot);
 141
 142 /*
 143  * The caller must hold vm_token.
 144  */
 145 static __inline void
 146 release_page(struct faultstate *fs)
 147 {
 148         vm_page_deactivate(fs->m);
 149         vm_page_wakeup(fs->m);
 150         fs->m = NULL;
 151 }
 152
 153 /*
 154  * The caller must hold vm_token.
 155  */
 156 static __inline void
 157 unlock_map(struct faultstate *fs)
 158 {
 159         if (fs->lookup_still_valid && fs->map) {
 160                 vm_map_lookup_done(fs->map, fs->entry, 0);
 161                 fs->lookup_still_valid = FALSE;
 162         }
 163 }
 164
 165 /*
 166  * Clean up after a successful call to vm_fault_object() so another call
 167  * to vm_fault_object() can be made.
 168  *
 169  * The caller must hold vm_token.
 170  */
 171 static void
 172 _cleanup_successful_fault(struct faultstate *fs, int relock)
 173 {
 174         if (fs->object != fs->first_object) {
 175                 vm_page_free(fs->first_m);
 176                 vm_object_pip_wakeup(fs->object);
 177                 fs->first_m = NULL;
 178         }
 179         fs->object = fs->first_object;
 180         if (relock && fs->lookup_still_valid == FALSE) {
 181                 if (fs->map)
 182                         vm_map_lock_read(fs->map);
 183                 fs->lookup_still_valid = TRUE;
 184         }
 185 }
 186
 187 /*
 188  * The caller must hold vm_token.
 189  */
 190 static void
 191 _unlock_things(struct faultstate *fs, int dealloc)
 192 {
 193         vm_object_pip_wakeup(fs->first_object);
 194         _cleanup_successful_fault(fs, 0);
 195         if (dealloc) {
 196                 vm_object_deallocate(fs->first_object);
 197                 fs->first_object = NULL;
 198         }
 199         unlock_map(fs);
 200         if (fs->vp != NULL) {
 201                 vput(fs->vp);
 202                 fs->vp = NULL;
 203         }
 204 }
 205
 206 #define unlock_things(fs) _unlock_things(fs, 0)
 207 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
 208 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
 209
 210 /*
 211  * TRYPAGER
 212  *
 213  * Determine if the pager for the current object *might* contain the page.
 214  *
 215  * We only need to try the pager if this is not a default object (default
 216  * objects are zero-fill and have no real pager), and if we are not taking
 217  * a wiring fault or if the FS entry is wired.
 218  */
 219 #define TRYPAGER(fs)    \
 220                 (fs->object->type != OBJT_DEFAULT && \
 221                 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
 222
 223 /*
 224  * vm_fault:
 225  *
 226  * Handle a page fault occuring at the given address, requiring the given
 227  * permissions, in the map specified.  If successful, the page is inserted
 228  * into the associated physical map.
 229  *
 230  * NOTE: The given address should be truncated to the proper page address.
 231  *
 232  * KERN_SUCCESS is returned if the page fault is handled; otherwise,
 233  * a standard error specifying why the fault is fatal is returned.
 234  *
 235  * The map in question must be referenced, and remains so.
 236  * The caller may hold no locks.
 237  * No other requirements.
 238  */
 239 int
 240 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
 241 {
 242         int result;
 243         vm_pindex_t first_pindex;
 244         struct faultstate fs;
 245         int growstack;
 246
 247         mycpu->gd_cnt.v_vm_faults++;
 248
 249         fs.didlimit = 0;
 250         fs.hardfault = 0;
 251         fs.fault_flags = fault_flags;
 252         growstack = 1;
 253
 254 RetryFault:
 255         /*
 256          * Find the vm_map_entry representing the backing store and resolve
 257          * the top level object and page index.  This may have the side
 258          * effect of executing a copy-on-write on the map entry and/or
 259          * creating a shadow object, but will not COW any actual VM pages.
 260          *
 261          * On success fs.map is left read-locked and various other fields
 262          * are initialized but not otherwise referenced or locked.
 263          *
 264          * NOTE!  vm_map_lookup will try to upgrade the fault_type to
 265          * VM_FAULT_WRITE if the map entry is a virtual page table and also
 266          * writable, so we can set the 'A'accessed bit in the virtual page
 267          * table entry.
 268          */
 269         fs.map = map;
 270         result = vm_map_lookup(&fs.map, vaddr, fault_type,
 271                                &fs.entry, &fs.first_object,
 272                                &first_pindex, &fs.first_prot, &fs.wired);
 273
 274         /*
 275          * If the lookup failed or the map protections are incompatible,
 276          * the fault generally fails.  However, if the caller is trying
 277          * to do a user wiring we have more work to do.
 278          */
 279         if (result != KERN_SUCCESS) {
 280                 if (result != KERN_PROTECTION_FAILURE ||
 281                     (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
 282                 {
 283                         if (result == KERN_INVALID_ADDRESS && growstack &&
 284                             map != &kernel_map && curproc != NULL) {
 285                                 result = vm_map_growstack(curproc, vaddr);
 286                                 if (result != KERN_SUCCESS)
 287                                         return (KERN_FAILURE);
 288                                 growstack = 0;
 289                                 goto RetryFault;
 290                         }
 291                         return (result);
 292                 }
 293
 294                 /*
 295                  * If we are user-wiring a r/w segment, and it is COW, then
 296                  * we need to do the COW operation.  Note that we don't
 297                  * currently COW RO sections now, because it is NOT desirable
 298                  * to COW .text.  We simply keep .text from ever being COW'ed
 299                  * and take the heat that one cannot debug wired .text sections.
 300                  */
 301                 result = vm_map_lookup(&fs.map, vaddr,
 302                                        VM_PROT_READ|VM_PROT_WRITE|
 303                                         VM_PROT_OVERRIDE_WRITE,
 304                                        &fs.entry, &fs.first_object,
 305                                        &first_pindex, &fs.first_prot,
 306                                        &fs.wired);
 307                 if (result != KERN_SUCCESS)
 308                         return result;
 309
 310                 /*
 311                  * If we don't COW now, on a user wire, the user will never
 312                  * be able to write to the mapping.  If we don't make this
 313                  * restriction, the bookkeeping would be nearly impossible.
 314                  */
 315                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 316                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 317         }
 318
 319         /*
 320          * fs.map is read-locked
 321          *
 322          * Misc checks.  Save the map generation number to detect races.
 323          */
 324         fs.map_generation = fs.map->timestamp;
 325
 326         if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) {
 327                 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 328                         panic("vm_fault: fault on nofault entry, addr: %p",
 329                               (void *)vaddr);
 330                 }
 331                 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) &&
 332                     vaddr >= fs.entry->start &&
 333                     vaddr < fs.entry->start + PAGE_SIZE) {
 334                         panic("vm_fault: fault on stack guard, addr: %p",
 335                               (void *)vaddr);
 336                 }
 337         }
 338
 339         /*
 340          * A system map entry may return a NULL object.  No object means
 341          * no pager means an unrecoverable kernel fault.
 342          */
 343         if (fs.first_object == NULL) {
 344                 panic("vm_fault: unrecoverable fault at %p in entry %p",
 345                         (void *)vaddr, fs.entry);
 346         }
 347
 348         /*
 349          * Make a reference to this object to prevent its disposal while we
 350          * are messing with it.  Once we have the reference, the map is free
 351          * to be diddled.  Since objects reference their shadows (and copies),
 352          * they will stay around as well.
 353          *
 354          * Bump the paging-in-progress count to prevent size changes (e.g.
 355          * truncation operations) during I/O.  This must be done after
 356          * obtaining the vnode lock in order to avoid possible deadlocks.
 357          *
 358          * The vm_token is needed to manipulate the vm_object
 359          */
 360         lwkt_gettoken(&vm_token);
 361         vm_object_reference(fs.first_object);
 362         fs.vp = vnode_pager_lock(fs.first_object);
 363         vm_object_pip_add(fs.first_object, 1);
 364         lwkt_reltoken(&vm_token);
 365
 366         fs.lookup_still_valid = TRUE;
 367         fs.first_m = NULL;
 368         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 369
 370         /*
 371          * If the entry is wired we cannot change the page protection.
 372          */
 373         if (fs.wired)
 374                 fault_type = fs.first_prot;
 375
 376         /*
 377          * The page we want is at (first_object, first_pindex), but if the
 378          * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
 379          * page table to figure out the actual pindex.
 380          *
 381          * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
 382          * ONLY
 383          */
 384         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 385                 result = vm_fault_vpagetable(&fs, &first_pindex,
 386                                              fs.entry->aux.master_pde,
 387                                              fault_type);
 388                 if (result == KERN_TRY_AGAIN)
 389                         goto RetryFault;
 390                 if (result != KERN_SUCCESS)
 391                         return (result);
 392         }
 393
 394         /*
 395          * Now we have the actual (object, pindex), fault in the page.  If
 396          * vm_fault_object() fails it will unlock and deallocate the FS
 397          * data.   If it succeeds everything remains locked and fs->object
 398          * will have an additional PIP count if it is not equal to
 399          * fs->first_object
 400          *
 401          * vm_fault_object will set fs->prot for the pmap operation.  It is
 402          * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
 403          * page can be safely written.  However, it will force a read-only
 404          * mapping for a read fault if the memory is managed by a virtual
 405          * page table.
 406          */
 407         result = vm_fault_object(&fs, first_pindex, fault_type);
 408
 409         if (result == KERN_TRY_AGAIN)
 410                 goto RetryFault;
 411         if (result != KERN_SUCCESS)
 412                 return (result);
 413
 414         /*
 415          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 416          * will contain a busied page.
 417          *
 418          * Enter the page into the pmap and do pmap-related adjustments.
 419          */
 420         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
 421
 422         /*
 423          * Burst in a few more pages if possible.  The fs.map should still
 424          * be locked.
 425          */
 426         if (fault_flags & VM_FAULT_BURST) {
 427                 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
 428                     fs.wired == 0) {
 429                         vm_prefault(fs.map->pmap, vaddr, fs.entry, fs.prot);
 430                 }
 431         }
 432         unlock_things(&fs);
 433
 434         vm_page_flag_clear(fs.m, PG_ZERO);
 435         vm_page_flag_set(fs.m, PG_REFERENCED);
 436
 437         /*
 438          * If the page is not wired down, then put it where the pageout daemon
 439          * can find it.
 440          *
 441          * We do not really need to get vm_token here but since all the
 442          * vm_*() calls have to doing it here improves efficiency.
 443          */
 444         lwkt_gettoken(&vm_token);
 445         if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
 446                 if (fs.wired)
 447                         vm_page_wire(fs.m);
 448                 else
 449                         vm_page_unwire(fs.m, 1);
 450         } else {
 451                 vm_page_activate(fs.m);
 452         }
 453
 454         if (curthread->td_lwp) {
 455                 if (fs.hardfault) {
 456                         curthread->td_lwp->lwp_ru.ru_majflt++;
 457                 } else {
 458                         curthread->td_lwp->lwp_ru.ru_minflt++;
 459                 }
 460         }
 461
 462         /*
 463          * Unlock everything, and return
 464          */
 465         vm_page_wakeup(fs.m);
 466         vm_object_deallocate(fs.first_object);
 467         lwkt_reltoken(&vm_token);
 468
 469         return (KERN_SUCCESS);
 470 }
 471
 472 /*
 473  * Fault in the specified virtual address in the current process map,
 474  * returning a held VM page or NULL.  See vm_fault_page() for more
 475  * information.
 476  *
 477  * No requirements.
 478  */
 479 vm_page_t
 480 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
 481 {
 482         struct lwp *lp = curthread->td_lwp;
 483         vm_page_t m;
 484
 485         m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
 486                           fault_type, VM_FAULT_NORMAL, errorp);
 487         return(m);
 488 }
 489
 490 /*
 491  * Fault in the specified virtual address in the specified map, doing all
 492  * necessary manipulation of the object store and all necessary I/O.  Return
 493  * a held VM page or NULL, and set *errorp.  The related pmap is not
 494  * updated.
 495  *
 496  * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
 497  * and marked PG_REFERENCED as well.
 498  *
 499  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
 500  * error will be returned.
 501  *
 502  * No requirements.
 503  */
 504 vm_page_t
 505 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 506               int fault_flags, int *errorp)
 507 {
 508         vm_pindex_t first_pindex;
 509         struct faultstate fs;
 510         int result;
 511         vm_prot_t orig_fault_type = fault_type;
 512
 513         mycpu->gd_cnt.v_vm_faults++;
 514
 515         fs.didlimit = 0;
 516         fs.hardfault = 0;
 517         fs.fault_flags = fault_flags;
 518         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
 519
 520 RetryFault:
 521         /*
 522          * Find the vm_map_entry representing the backing store and resolve
 523          * the top level object and page index.  This may have the side
 524          * effect of executing a copy-on-write on the map entry and/or
 525          * creating a shadow object, but will not COW any actual VM pages.
 526          *
 527          * On success fs.map is left read-locked and various other fields
 528          * are initialized but not otherwise referenced or locked.
 529          *
 530          * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
 531          * if the map entry is a virtual page table and also writable,
 532          * so we can set the 'A'accessed bit in the virtual page table entry.
 533          */
 534         fs.map = map;
 535         result = vm_map_lookup(&fs.map, vaddr, fault_type,
 536                                &fs.entry, &fs.first_object,
 537                                &first_pindex, &fs.first_prot, &fs.wired);
 538
 539         if (result != KERN_SUCCESS) {
 540                 *errorp = result;
 541                 return (NULL);
 542         }
 543
 544         /*
 545          * fs.map is read-locked
 546          *
 547          * Misc checks.  Save the map generation number to detect races.
 548          */
 549         fs.map_generation = fs.map->timestamp;
 550
 551         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 552                 panic("vm_fault: fault on nofault entry, addr: %lx",
 553                     (u_long)vaddr);
 554         }
 555
 556         /*
 557          * A system map entry may return a NULL object.  No object means
 558          * no pager means an unrecoverable kernel fault.
 559          */
 560         if (fs.first_object == NULL) {
 561                 panic("vm_fault: unrecoverable fault at %p in entry %p",
 562                         (void *)vaddr, fs.entry);
 563         }
 564
 565         /*
 566          * Make a reference to this object to prevent its disposal while we
 567          * are messing with it.  Once we have the reference, the map is free
 568          * to be diddled.  Since objects reference their shadows (and copies),
 569          * they will stay around as well.
 570          *
 571          * Bump the paging-in-progress count to prevent size changes (e.g.
 572          * truncation operations) during I/O.  This must be done after
 573          * obtaining the vnode lock in order to avoid possible deadlocks.
 574          *
 575          * The vm_token is needed to manipulate the vm_object
 576          */
 577         lwkt_gettoken(&vm_token);
 578         vm_object_reference(fs.first_object);
 579         fs.vp = vnode_pager_lock(fs.first_object);
 580         vm_object_pip_add(fs.first_object, 1);
 581         lwkt_reltoken(&vm_token);
 582
 583         fs.lookup_still_valid = TRUE;
 584         fs.first_m = NULL;
 585         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 586
 587         /*
 588          * If the entry is wired we cannot change the page protection.
 589          */
 590         if (fs.wired)
 591                 fault_type = fs.first_prot;
 592
 593         /*
 594          * The page we want is at (first_object, first_pindex), but if the
 595          * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
 596          * page table to figure out the actual pindex.
 597          *
 598          * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
 599          * ONLY
 600          */
 601         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 602                 result = vm_fault_vpagetable(&fs, &first_pindex,
 603                                              fs.entry->aux.master_pde,
 604                                              fault_type);
 605                 if (result == KERN_TRY_AGAIN)
 606                         goto RetryFault;
 607                 if (result != KERN_SUCCESS) {
 608                         *errorp = result;
 609                         return (NULL);
 610                 }
 611         }
 612
 613         /*
 614          * Now we have the actual (object, pindex), fault in the page.  If
 615          * vm_fault_object() fails it will unlock and deallocate the FS
 616          * data.   If it succeeds everything remains locked and fs->object
 617          * will have an additinal PIP count if it is not equal to
 618          * fs->first_object
 619          */
 620         result = vm_fault_object(&fs, first_pindex, fault_type);
 621
 622         if (result == KERN_TRY_AGAIN)
 623                 goto RetryFault;
 624         if (result != KERN_SUCCESS) {
 625                 *errorp = result;
 626                 return(NULL);
 627         }
 628
 629         if ((orig_fault_type & VM_PROT_WRITE) &&
 630             (fs.prot & VM_PROT_WRITE) == 0) {
 631                 *errorp = KERN_PROTECTION_FAILURE;
 632                 unlock_and_deallocate(&fs);
 633                 return(NULL);
 634         }
 635
 636         /*
 637          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 638          * will contain a busied page.
 639          */
 640         unlock_things(&fs);
 641
 642         /*
 643          * Return a held page.  We are not doing any pmap manipulation so do
 644          * not set PG_MAPPED.  However, adjust the page flags according to
 645          * the fault type because the caller may not use a managed pmapping
 646          * (so we don't want to lose the fact that the page will be dirtied
 647          * if a write fault was specified).
 648          */
 649         lwkt_gettoken(&vm_token);
 650         vm_page_hold(fs.m);
 651         vm_page_flag_clear(fs.m, PG_ZERO);
 652         if (fault_type & VM_PROT_WRITE)
 653                 vm_page_dirty(fs.m);
 654
 655         /*
 656          * Update the pmap.  We really only have to do this if a COW
 657          * occured to replace the read-only page with the new page.  For
 658          * now just do it unconditionally. XXX
 659          */
 660         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
 661         vm_page_flag_set(fs.m, PG_REFERENCED);
 662
 663         /*
 664          * Unbusy the page by activating it.  It remains held and will not
 665          * be reclaimed.
 666          */
 667         vm_page_activate(fs.m);
 668
 669         if (curthread->td_lwp) {
 670                 if (fs.hardfault) {
 671                         curthread->td_lwp->lwp_ru.ru_majflt++;
 672                 } else {
 673                         curthread->td_lwp->lwp_ru.ru_minflt++;
 674                 }
 675         }
 676
 677         /*
 678          * Unlock everything, and return the held page.
 679          */
 680         vm_page_wakeup(fs.m);
 681         vm_object_deallocate(fs.first_object);
 682         lwkt_reltoken(&vm_token);
 683
 684         *errorp = 0;
 685         return(fs.m);
 686 }
 687
 688 /*
 689  * Fault in the specified (object,offset), dirty the returned page as
 690  * needed.  If the requested fault_type cannot be done NULL and an
 691  * error is returned.
 692  *
 693  * A held (but not busied) page is returned.
 694  *
 695  * No requirements.
 696  */
 697 vm_page_t
 698 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
 699                      vm_prot_t fault_type, int fault_flags, int *errorp)
 700 {
 701         int result;
 702         vm_pindex_t first_pindex;
 703         struct faultstate fs;
 704         struct vm_map_entry entry;
 705
 706         bzero(&entry, sizeof(entry));
 707         entry.object.vm_object = object;
 708         entry.maptype = VM_MAPTYPE_NORMAL;
 709         entry.protection = entry.max_protection = fault_type;
 710
 711         fs.didlimit = 0;
 712         fs.hardfault = 0;
 713         fs.fault_flags = fault_flags;
 714         fs.map = NULL;
 715         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
 716
 717 RetryFault:
 718
 719         fs.first_object = object;
 720         first_pindex = OFF_TO_IDX(offset);
 721         fs.entry = &entry;
 722         fs.first_prot = fault_type;
 723         fs.wired = 0;
 724         /*fs.map_generation = 0; unused */
 725
 726         /*
 727          * Make a reference to this object to prevent its disposal while we
 728          * are messing with it.  Once we have the reference, the map is free
 729          * to be diddled.  Since objects reference their shadows (and copies),
 730          * they will stay around as well.
 731          *
 732          * Bump the paging-in-progress count to prevent size changes (e.g.
 733          * truncation operations) during I/O.  This must be done after
 734          * obtaining the vnode lock in order to avoid possible deadlocks.
 735          */
 736         lwkt_gettoken(&vm_token);
 737         vm_object_reference(fs.first_object);
 738         fs.vp = vnode_pager_lock(fs.first_object);
 739         vm_object_pip_add(fs.first_object, 1);
 740         lwkt_reltoken(&vm_token);
 741
 742         fs.lookup_still_valid = TRUE;
 743         fs.first_m = NULL;
 744         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 745
 746 #if 0
 747         /* XXX future - ability to operate on VM object using vpagetable */
 748         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 749                 result = vm_fault_vpagetable(&fs, &first_pindex,
 750                                              fs.entry->aux.master_pde,
 751                                              fault_type);
 752                 if (result == KERN_TRY_AGAIN)
 753                         goto RetryFault;
 754                 if (result != KERN_SUCCESS) {
 755                         *errorp = result;
 756                         return (NULL);
 757                 }
 758         }
 759 #endif
 760
 761         /*
 762          * Now we have the actual (object, pindex), fault in the page.  If
 763          * vm_fault_object() fails it will unlock and deallocate the FS
 764          * data.   If it succeeds everything remains locked and fs->object
 765          * will have an additinal PIP count if it is not equal to
 766          * fs->first_object
 767          */
 768         result = vm_fault_object(&fs, first_pindex, fault_type);
 769
 770         if (result == KERN_TRY_AGAIN)
 771                 goto RetryFault;
 772         if (result != KERN_SUCCESS) {
 773                 *errorp = result;
 774                 return(NULL);
 775         }
 776
 777         if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
 778                 *errorp = KERN_PROTECTION_FAILURE;
 779                 unlock_and_deallocate(&fs);
 780                 return(NULL);
 781         }
 782
 783         /*
 784          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 785          * will contain a busied page.
 786          */
 787         unlock_things(&fs);
 788
 789         /*
 790          * Return a held page.  We are not doing any pmap manipulation so do
 791          * not set PG_MAPPED.  However, adjust the page flags according to
 792          * the fault type because the caller may not use a managed pmapping
 793          * (so we don't want to lose the fact that the page will be dirtied
 794          * if a write fault was specified).
 795          */
 796         lwkt_gettoken(&vm_token);
 797         vm_page_hold(fs.m);
 798         vm_page_flag_clear(fs.m, PG_ZERO);
 799         if (fault_type & VM_PROT_WRITE)
 800                 vm_page_dirty(fs.m);
 801
 802         if (fault_flags & VM_FAULT_DIRTY)
 803                 vm_page_dirty(fs.m);
 804         if (fault_flags & VM_FAULT_UNSWAP)
 805                 swap_pager_unswapped(fs.m);
 806
 807         /*
 808          * Indicate that the page was accessed.
 809          */
 810         vm_page_flag_set(fs.m, PG_REFERENCED);
 811
 812         /*
 813          * Unbusy the page by activating it.  It remains held and will not
 814          * be reclaimed.
 815          */
 816         vm_page_activate(fs.m);
 817
 818         if (curthread->td_lwp) {
 819                 if (fs.hardfault) {
 820                         mycpu->gd_cnt.v_vm_faults++;
 821                         curthread->td_lwp->lwp_ru.ru_majflt++;
 822                 } else {
 823                         curthread->td_lwp->lwp_ru.ru_minflt++;
 824                 }
 825         }
 826
 827         /*
 828          * Unlock everything, and return the held page.
 829          */
 830         vm_page_wakeup(fs.m);
 831         vm_object_deallocate(fs.first_object);
 832         lwkt_reltoken(&vm_token);
 833
 834         *errorp = 0;
 835         return(fs.m);
 836 }
 837
 838 /*
 839  * Translate the virtual page number (first_pindex) that is relative
 840  * to the address space into a logical page number that is relative to the
 841  * backing object.  Use the virtual page table pointed to by (vpte).
 842  *
 843  * This implements an N-level page table.  Any level can terminate the
 844  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
 845  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
 846  *
 847  * No requirements (vm_token need not be held).
 848  */
 849 static
 850 int
 851 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
 852                     vpte_t vpte, int fault_type)
 853 {
 854         struct lwbuf *lwb;
 855         int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
 856         int result = KERN_SUCCESS;
 857         vpte_t *ptep;
 858
 859         for (;;) {
 860                 /*
 861                  * We cannot proceed if the vpte is not valid, not readable
 862                  * for a read fault, or not writable for a write fault.
 863                  */
 864                 if ((vpte & VPTE_V) == 0) {
 865                         unlock_and_deallocate(fs);
 866                         return (KERN_FAILURE);
 867                 }
 868                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
 869                         unlock_and_deallocate(fs);
 870                         return (KERN_FAILURE);
 871                 }
 872                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
 873                         unlock_and_deallocate(fs);
 874                         return (KERN_FAILURE);
 875                 }
 876                 if ((vpte & VPTE_PS) || vshift == 0)
 877                         break;
 878                 KKASSERT(vshift >= VPTE_PAGE_BITS);
 879
 880                 /*
 881                  * Get the page table page.  Nominally we only read the page
 882                  * table, but since we are actively setting VPTE_M and VPTE_A,
 883                  * tell vm_fault_object() that we are writing it.
 884                  *
 885                  * There is currently no real need to optimize this.
 886                  */
 887                 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
 888                                          VM_PROT_READ|VM_PROT_WRITE);
 889                 if (result != KERN_SUCCESS)
 890                         return (result);
 891
 892                 /*
 893                  * Process the returned fs.m and look up the page table
 894                  * entry in the page table page.
 895                  */
 896                 vshift -= VPTE_PAGE_BITS;
 897                 lwb = lwbuf_alloc(fs->m);
 898                 ptep = ((vpte_t *)lwbuf_kva(lwb) +
 899                         ((*pindex >> vshift) & VPTE_PAGE_MASK));
 900                 vpte = *ptep;
 901
 902                 /*
 903                  * Page table write-back.  If the vpte is valid for the
 904                  * requested operation, do a write-back to the page table.
 905                  *
 906                  * XXX VPTE_M is not set properly for page directory pages.
 907                  * It doesn't get set in the page directory if the page table
 908                  * is modified during a read access.
 909                  */
 910                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
 911                     (vpte & VPTE_W)) {
 912                         if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
 913                                 atomic_set_long(ptep, VPTE_M | VPTE_A);
 914                                 vm_page_dirty(fs->m);
 915                         }
 916                 }
 917                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
 918                     (vpte & VPTE_R)) {
 919                         if ((vpte & VPTE_A) == 0) {
 920                                 atomic_set_long(ptep, VPTE_A);
 921                                 vm_page_dirty(fs->m);
 922                         }
 923                 }
 924                 lwbuf_free(lwb);
 925                 vm_page_flag_set(fs->m, PG_REFERENCED);
 926                 vm_page_activate(fs->m);
 927                 vm_page_wakeup(fs->m);
 928                 cleanup_successful_fault(fs);
 929         }
 930         /*
 931          * Combine remaining address bits with the vpte.
 932          */
 933         /* JG how many bits from each? */
 934         *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
 935                   (*pindex & ((1L << vshift) - 1));
 936         return (KERN_SUCCESS);
 937 }
 938
 939
 940 /*
 941  * This is the core of the vm_fault code.
 942  *
 943  * Do all operations required to fault-in (fs.first_object, pindex).  Run
 944  * through the shadow chain as necessary and do required COW or virtual
 945  * copy operations.  The caller has already fully resolved the vm_map_entry
 946  * and, if appropriate, has created a copy-on-write layer.  All we need to
 947  * do is iterate the object chain.
 948  *
 949  * On failure (fs) is unlocked and deallocated and the caller may return or
 950  * retry depending on the failure code.  On success (fs) is NOT unlocked or
 951  * deallocated, fs.m will contained a resolved, busied page, and fs.object
 952  * will have an additional PIP count if it is not equal to fs.first_object.
 953  *
 954  * No requirements.
 955  */
 956 static
 957 int
 958 vm_fault_object(struct faultstate *fs,
 959                 vm_pindex_t first_pindex, vm_prot_t fault_type)
 960 {
 961         vm_object_t next_object;
 962         vm_pindex_t pindex;
 963
 964         fs->prot = fs->first_prot;
 965         fs->object = fs->first_object;
 966         pindex = first_pindex;
 967
 968         /*
 969          * If a read fault occurs we try to make the page writable if
 970          * possible.  There are three cases where we cannot make the
 971          * page mapping writable:
 972          *
 973          * (1) The mapping is read-only or the VM object is read-only,
 974          *     fs->prot above will simply not have VM_PROT_WRITE set.
 975          *
 976          * (2) If the mapping is a virtual page table we need to be able
 977          *     to detect writes so we can set VPTE_M in the virtual page
 978          *     table.
 979          *
 980          * (3) If the VM page is read-only or copy-on-write, upgrading would
 981          *     just result in an unnecessary COW fault.
 982          *
 983          * VM_PROT_VPAGED is set if faulting via a virtual page table and
 984          * causes adjustments to the 'M'odify bit to also turn off write
 985          * access to force a re-fault.
 986          */
 987         if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 988                 if ((fault_type & VM_PROT_WRITE) == 0)
 989                         fs->prot &= ~VM_PROT_WRITE;
 990         }
 991
 992         lwkt_gettoken(&vm_token);
 993
 994         for (;;) {
 995                 /*
 996                  * If the object is dead, we stop here
 997                  */
 998                 if (fs->object->flags & OBJ_DEAD) {
 999                         unlock_and_deallocate(fs);
1000                         lwkt_reltoken(&vm_token);
1001                         return (KERN_PROTECTION_FAILURE);
1002                 }
1003
1004                 /*
1005                  * See if the page is resident.
1006                  */
1007                 fs->m = vm_page_lookup(fs->object, pindex);
1008                 if (fs->m != NULL) {
1009                         int queue;
1010                         /*
1011                          * Wait/Retry if the page is busy.  We have to do this
1012                          * if the page is busy via either PG_BUSY or
1013                          * vm_page_t->busy because the vm_pager may be using
1014                          * vm_page_t->busy for pageouts ( and even pageins if
1015                          * it is the vnode pager ), and we could end up trying
1016                          * to pagein and pageout the same page simultaneously.
1017                          *
1018                          * We can theoretically allow the busy case on a read
1019                          * fault if the page is marked valid, but since such
1020                          * pages are typically already pmap'd, putting that
1021                          * special case in might be more effort then it is
1022                          * worth.  We cannot under any circumstances mess
1023                          * around with a vm_page_t->busy page except, perhaps,
1024                          * to pmap it.
1025                          */
1026                         if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
1027                                 unlock_things(fs);
1028                                 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1029                                 mycpu->gd_cnt.v_intrans++;
1030                                 vm_object_deallocate(fs->first_object);
1031                                 fs->first_object = NULL;
1032                                 lwkt_reltoken(&vm_token);
1033                                 return (KERN_TRY_AGAIN);
1034                         }
1035
1036                         /*
1037                          * If reactivating a page from PQ_CACHE we may have
1038                          * to rate-limit.
1039                          */
1040                         queue = fs->m->queue;
1041                         vm_page_unqueue_nowakeup(fs->m);
1042
1043                         if ((queue - fs->m->pc) == PQ_CACHE &&
1044                             vm_page_count_severe()) {
1045                                 vm_page_activate(fs->m);
1046                                 unlock_and_deallocate(fs);
1047                                 vm_waitpfault();
1048                                 lwkt_reltoken(&vm_token);
1049                                 return (KERN_TRY_AGAIN);
1050                         }
1051
1052                         /*
1053                          * Mark page busy for other processes, and the
1054                          * pagedaemon.  If it still isn't completely valid
1055                          * (readable), or if a read-ahead-mark is set on
1056                          * the VM page, jump to readrest, else we found the
1057                          * page and can return.
1058                          *
1059                          * We can release the spl once we have marked the
1060                          * page busy.
1061                          */
1062                         vm_page_busy(fs->m);
1063
1064                         if (fs->m->object != &kernel_object) {
1065                                 if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1066                                     VM_PAGE_BITS_ALL) {
1067                                         goto readrest;
1068                                 }
1069                                 if (fs->m->flags & PG_RAM) {
1070                                         if (debug_cluster)
1071                                                 kprintf("R");
1072                                         vm_page_flag_clear(fs->m, PG_RAM);
1073                                         goto readrest;
1074                                 }
1075                         }
1076                         break; /* break to PAGE HAS BEEN FOUND */
1077                 }
1078
1079                 /*
1080                  * Page is not resident, If this is the search termination
1081                  * or the pager might contain the page, allocate a new page.
1082                  */
1083                 if (TRYPAGER(fs) || fs->object == fs->first_object) {
1084                         /*
1085                          * If the page is beyond the object size we fail
1086                          */
1087                         if (pindex >= fs->object->size) {
1088                                 lwkt_reltoken(&vm_token);
1089                                 unlock_and_deallocate(fs);
1090                                 return (KERN_PROTECTION_FAILURE);
1091                         }
1092
1093                         /*
1094                          * Ratelimit.
1095                          */
1096                         if (fs->didlimit == 0 && curproc != NULL) {
1097                                 int limticks;
1098
1099                                 limticks = vm_fault_ratelimit(curproc->p_vmspace);
1100                                 if (limticks) {
1101                                         lwkt_reltoken(&vm_token);
1102                                         unlock_and_deallocate(fs);
1103                                         tsleep(curproc, 0, "vmrate", limticks);
1104                                         fs->didlimit = 1;
1105                                         return (KERN_TRY_AGAIN);
1106                                 }
1107                         }
1108
1109                         /*
1110                          * Allocate a new page for this object/offset pair.
1111                          */
1112                         fs->m = NULL;
1113                         if (!vm_page_count_severe()) {
1114                                 fs->m = vm_page_alloc(fs->object, pindex,
1115                                     (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
1116                         }
1117                         if (fs->m == NULL) {
1118                                 lwkt_reltoken(&vm_token);
1119                                 unlock_and_deallocate(fs);
1120                                 vm_waitpfault();
1121                                 return (KERN_TRY_AGAIN);
1122                         }
1123                 }
1124
1125 readrest:
1126                 /*
1127                  * We have found an invalid or partially valid page, a
1128                  * page with a read-ahead mark which might be partially or
1129                  * fully valid (and maybe dirty too), or we have allocated
1130                  * a new page.
1131                  *
1132                  * Attempt to fault-in the page if there is a chance that the
1133                  * pager has it, and potentially fault in additional pages
1134                  * at the same time.
1135                  *
1136                  * We are NOT in splvm here and if TRYPAGER is true then
1137                  * fs.m will be non-NULL and will be PG_BUSY for us.
1138                  */
1139                 if (TRYPAGER(fs)) {
1140                         int rv;
1141                         int seqaccess;
1142                         u_char behavior = vm_map_entry_behavior(fs->entry);
1143
1144                         if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1145                                 seqaccess = 0;
1146                         else
1147                                 seqaccess = -1;
1148
1149                         /*
1150                          * If sequential access is detected then attempt
1151                          * to deactivate/cache pages behind the scan to
1152                          * prevent resource hogging.
1153                          *
1154                          * Use of PG_RAM to detect sequential access
1155                          * also simulates multi-zone sequential access
1156                          * detection for free.
1157                          *
1158                          * NOTE: Partially valid dirty pages cannot be
1159                          *       deactivated without causing NFS picemeal
1160                          *       writes to barf.
1161                          */
1162                         if ((fs->first_object->type != OBJT_DEVICE) &&
1163                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1164                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1165                                  (fs->m->flags & PG_RAM)))
1166                         ) {
1167                                 vm_pindex_t scan_pindex;
1168                                 int scan_count = 16;
1169
1170                                 if (first_pindex < 16) {
1171                                         scan_pindex = 0;
1172                                         scan_count = 0;
1173                                 } else {
1174                                         scan_pindex = first_pindex - 16;
1175                                         if (scan_pindex < 16)
1176                                                 scan_count = scan_pindex;
1177                                         else
1178                                                 scan_count = 16;
1179                                 }
1180
1181                                 while (scan_count) {
1182                                         vm_page_t mt;
1183
1184                                         mt = vm_page_lookup(fs->first_object,
1185                                                             scan_pindex);
1186                                         if (mt == NULL ||
1187                                             (mt->valid != VM_PAGE_BITS_ALL)) {
1188                                                 break;
1189                                         }
1190                                         if (mt->busy ||
1191                                             (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
1192                                             mt->hold_count ||
1193                                             mt->wire_count)  {
1194                                                 goto skip;
1195                                         }
1196                                         if (mt->dirty == 0)
1197                                                 vm_page_test_dirty(mt);
1198                                         if (mt->dirty) {
1199                                                 vm_page_busy(mt);
1200                                                 vm_page_protect(mt,
1201                                                                 VM_PROT_NONE);
1202                                                 vm_page_deactivate(mt);
1203                                                 vm_page_wakeup(mt);
1204                                         } else {
1205                                                 vm_page_cache(mt);
1206                                         }
1207 skip:
1208                                         --scan_count;
1209                                         --scan_pindex;
1210                                 }
1211
1212                                 seqaccess = 1;
1213                         }
1214
1215                         /*
1216                          * Avoid deadlocking against the map when doing I/O.
1217                          * fs.object and the page is PG_BUSY'd.
1218                          */
1219                         unlock_map(fs);
1220
1221                         /*
1222                          * Acquire the page data.  We still hold a ref on
1223                          * fs.object and the page has been PG_BUSY's.
1224                          *
1225                          * The pager may replace the page (for example, in
1226                          * order to enter a fictitious page into the
1227                          * object).  If it does so it is responsible for
1228                          * cleaning up the passed page and properly setting
1229                          * the new page PG_BUSY.
1230                          *
1231                          * If we got here through a PG_RAM read-ahead
1232                          * mark the page may be partially dirty and thus
1233                          * not freeable.  Don't bother checking to see
1234                          * if the pager has the page because we can't free
1235                          * it anyway.  We have to depend on the get_page
1236                          * operation filling in any gaps whether there is
1237                          * backing store or not.
1238                          */
1239                         rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1240
1241                         if (rv == VM_PAGER_OK) {
1242                                 /*
1243                                  * Relookup in case pager changed page. Pager
1244                                  * is responsible for disposition of old page
1245                                  * if moved.
1246                                  *
1247                                  * XXX other code segments do relookups too.
1248                                  * It's a bad abstraction that needs to be
1249                                  * fixed/removed.
1250                                  */
1251                                 fs->m = vm_page_lookup(fs->object, pindex);
1252                                 if (fs->m == NULL) {
1253                                         lwkt_reltoken(&vm_token);
1254                                         unlock_and_deallocate(fs);
1255                                         return (KERN_TRY_AGAIN);
1256                                 }
1257
1258                                 ++fs->hardfault;
1259                                 break; /* break to PAGE HAS BEEN FOUND */
1260                         }
1261
1262                         /*
1263                          * Remove the bogus page (which does not exist at this
1264                          * object/offset); before doing so, we must get back
1265                          * our object lock to preserve our invariant.
1266                          *
1267                          * Also wake up any other process that may want to bring
1268                          * in this page.
1269                          *
1270                          * If this is the top-level object, we must leave the
1271                          * busy page to prevent another process from rushing
1272                          * past us, and inserting the page in that object at
1273                          * the same time that we are.
1274                          */
1275                         if (rv == VM_PAGER_ERROR) {
1276                                 if (curproc)
1277                                         kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
1278                                 else
1279                                         kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
1280                         }
1281
1282                         /*
1283                          * Data outside the range of the pager or an I/O error
1284                          *
1285                          * The page may have been wired during the pagein,
1286                          * e.g. by the buffer cache, and cannot simply be
1287                          * freed.  Call vnode_pager_freepage() to deal with it.
1288                          */
1289                         /*
1290                          * XXX - the check for kernel_map is a kludge to work
1291                          * around having the machine panic on a kernel space
1292                          * fault w/ I/O error.
1293                          */
1294                         if (((fs->map != &kernel_map) &&
1295                             (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1296                                 vnode_pager_freepage(fs->m);
1297                                 lwkt_reltoken(&vm_token);
1298                                 fs->m = NULL;
1299                                 unlock_and_deallocate(fs);
1300                                 if (rv == VM_PAGER_ERROR)
1301                                         return (KERN_FAILURE);
1302                                 else
1303                                         return (KERN_PROTECTION_FAILURE);
1304                                 /* NOT REACHED */
1305                         }
1306                         if (fs->object != fs->first_object) {
1307                                 vnode_pager_freepage(fs->m);
1308                                 fs->m = NULL;
1309                                 /*
1310                                  * XXX - we cannot just fall out at this
1311                                  * point, m has been freed and is invalid!
1312                                  */
1313                         }
1314                 }
1315
1316                 /*
1317                  * We get here if the object has a default pager (or unwiring)
1318                  * or the pager doesn't have the page.
1319                  */
1320                 if (fs->object == fs->first_object)
1321                         fs->first_m = fs->m;
1322
1323                 /*
1324                  * Move on to the next object.  Lock the next object before
1325                  * unlocking the current one.
1326                  */
1327                 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1328                 next_object = fs->object->backing_object;
1329                 if (next_object == NULL) {
1330                         /*
1331                          * If there's no object left, fill the page in the top
1332                          * object with zeros.
1333                          */
1334                         if (fs->object != fs->first_object) {
1335                                 vm_object_pip_wakeup(fs->object);
1336
1337                                 fs->object = fs->first_object;
1338                                 pindex = first_pindex;
1339                                 fs->m = fs->first_m;
1340                         }
1341                         fs->first_m = NULL;
1342
1343                         /*
1344                          * Zero the page if necessary and mark it valid.
1345                          */
1346                         if ((fs->m->flags & PG_ZERO) == 0) {
1347                                 vm_page_zero_fill(fs->m);
1348                         } else {
1349 #ifdef PMAP_DEBUG
1350                                 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m));
1351 #endif
1352                                 vm_page_flag_clear(fs->m, PG_ZERO);
1353                                 mycpu->gd_cnt.v_ozfod++;
1354                         }
1355                         mycpu->gd_cnt.v_zfod++;
1356                         fs->m->valid = VM_PAGE_BITS_ALL;
1357                         break;  /* break to PAGE HAS BEEN FOUND */
1358                 }
1359                 if (fs->object != fs->first_object) {
1360                         vm_object_pip_wakeup(fs->object);
1361                 }
1362                 KASSERT(fs->object != next_object,
1363                         ("object loop %p", next_object));
1364                 fs->object = next_object;
1365                 vm_object_pip_add(fs->object, 1);
1366         }
1367
1368         /*
1369          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1370          * is held.]
1371          *
1372          * vm_token is still held
1373          *
1374          * If the page is being written, but isn't already owned by the
1375          * top-level object, we have to copy it into a new page owned by the
1376          * top-level object.
1377          */
1378         KASSERT((fs->m->flags & PG_BUSY) != 0,
1379                 ("vm_fault: not busy after main loop"));
1380
1381         if (fs->object != fs->first_object) {
1382                 /*
1383                  * We only really need to copy if we want to write it.
1384                  */
1385                 if (fault_type & VM_PROT_WRITE) {
1386                         /*
1387                          * This allows pages to be virtually copied from a
1388                          * backing_object into the first_object, where the
1389                          * backing object has no other refs to it, and cannot
1390                          * gain any more refs.  Instead of a bcopy, we just
1391                          * move the page from the backing object to the
1392                          * first object.  Note that we must mark the page
1393                          * dirty in the first object so that it will go out
1394                          * to swap when needed.
1395                          */
1396                         if (
1397                                 /*
1398                                  * Map, if present, has not changed
1399                                  */
1400                                 (fs->map == NULL ||
1401                                 fs->map_generation == fs->map->timestamp) &&
1402                                 /*
1403                                  * Only one shadow object
1404                                  */
1405                                 (fs->object->shadow_count == 1) &&
1406                                 /*
1407                                  * No COW refs, except us
1408                                  */
1409                                 (fs->object->ref_count == 1) &&
1410                                 /*
1411                                  * No one else can look this object up
1412                                  */
1413                                 (fs->object->handle == NULL) &&
1414                                 /*
1415                                  * No other ways to look the object up
1416                                  */
1417                                 ((fs->object->type == OBJT_DEFAULT) ||
1418                                  (fs->object->type == OBJT_SWAP)) &&
1419                                 /*
1420                                  * We don't chase down the shadow chain
1421                                  */
1422                                 (fs->object == fs->first_object->backing_object) &&
1423
1424                                 /*
1425                                  * grab the lock if we need to
1426                                  */
1427                                 (fs->lookup_still_valid ||
1428                                  fs->map == NULL ||
1429                                  lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1430                             ) {
1431
1432                                 fs->lookup_still_valid = 1;
1433                                 /*
1434                                  * get rid of the unnecessary page
1435                                  */
1436                                 vm_page_protect(fs->first_m, VM_PROT_NONE);
1437                                 vm_page_free(fs->first_m);
1438                                 fs->first_m = NULL;
1439
1440                                 /*
1441                                  * grab the page and put it into the
1442                                  * process'es object.  The page is
1443                                  * automatically made dirty.
1444                                  */
1445                                 vm_page_rename(fs->m, fs->first_object, first_pindex);
1446                                 fs->first_m = fs->m;
1447                                 vm_page_busy(fs->first_m);
1448                                 fs->m = NULL;
1449                                 mycpu->gd_cnt.v_cow_optim++;
1450                         } else {
1451                                 /*
1452                                  * Oh, well, lets copy it.
1453                                  */
1454                                 vm_page_copy(fs->m, fs->first_m);
1455                                 vm_page_event(fs->m, VMEVENT_COW);
1456                         }
1457
1458                         if (fs->m) {
1459                                 /*
1460                                  * We no longer need the old page or object.
1461                                  */
1462                                 release_page(fs);
1463                         }
1464
1465                         /*
1466                          * fs->object != fs->first_object due to above
1467                          * conditional
1468                          */
1469                         vm_object_pip_wakeup(fs->object);
1470
1471                         /*
1472                          * Only use the new page below...
1473                          */
1474
1475                         mycpu->gd_cnt.v_cow_faults++;
1476                         fs->m = fs->first_m;
1477                         fs->object = fs->first_object;
1478                         pindex = first_pindex;
1479                 } else {
1480                         /*
1481                          * If it wasn't a write fault avoid having to copy
1482                          * the page by mapping it read-only.
1483                          */
1484                         fs->prot &= ~VM_PROT_WRITE;
1485                 }
1486         }
1487
1488         /*
1489          * We may have had to unlock a map to do I/O.  If we did then
1490          * lookup_still_valid will be FALSE.  If the map generation count
1491          * also changed then all sorts of things could have happened while
1492          * we were doing the I/O and we need to retry.
1493          */
1494
1495         if (!fs->lookup_still_valid &&
1496             fs->map != NULL &&
1497             (fs->map->timestamp != fs->map_generation)) {
1498                 release_page(fs);
1499                 lwkt_reltoken(&vm_token);
1500                 unlock_and_deallocate(fs);
1501                 return (KERN_TRY_AGAIN);
1502         }
1503
1504         /*
1505          * If the fault is a write, we know that this page is being
1506          * written NOW so dirty it explicitly to save on pmap_is_modified()
1507          * calls later.
1508          *
1509          * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1510          * if the page is already dirty to prevent data written with
1511          * the expectation of being synced from not being synced.
1512          * Likewise if this entry does not request NOSYNC then make
1513          * sure the page isn't marked NOSYNC.  Applications sharing
1514          * data should use the same flags to avoid ping ponging.
1515          *
1516          * Also tell the backing pager, if any, that it should remove
1517          * any swap backing since the page is now dirty.
1518          */
1519         if (fs->prot & VM_PROT_WRITE) {
1520                 vm_object_set_writeable_dirty(fs->m->object);
1521                 vm_set_nosync(fs->m, fs->entry);
1522                 if (fs->fault_flags & VM_FAULT_DIRTY) {
1523                         vm_page_dirty(fs->m);
1524                         swap_pager_unswapped(fs->m);
1525                 }
1526         }
1527
1528         lwkt_reltoken(&vm_token);
1529
1530         /*
1531          * Page had better still be busy.  We are still locked up and
1532          * fs->object will have another PIP reference if it is not equal
1533          * to fs->first_object.
1534          */
1535         KASSERT(fs->m->flags & PG_BUSY,
1536                 ("vm_fault: page %p not busy!", fs->m));
1537
1538         /*
1539          * Sanity check: page must be completely valid or it is not fit to
1540          * map into user space.  vm_pager_get_pages() ensures this.
1541          */
1542         if (fs->m->valid != VM_PAGE_BITS_ALL) {
1543                 vm_page_zero_invalid(fs->m, TRUE);
1544                 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1545         }
1546
1547         return (KERN_SUCCESS);
1548 }
1549
1550 /*
1551  * Wire down a range of virtual addresses in a map.  The entry in question
1552  * should be marked in-transition and the map must be locked.  We must
1553  * release the map temporarily while faulting-in the page to avoid a
1554  * deadlock.  Note that the entry may be clipped while we are blocked but
1555  * will never be freed.
1556  *
1557  * No requirements.
1558  */
1559 int
1560 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1561 {
1562         boolean_t fictitious;
1563         vm_offset_t start;
1564         vm_offset_t end;
1565         vm_offset_t va;
1566         vm_paddr_t pa;
1567         pmap_t pmap;
1568         int rv;
1569
1570         pmap = vm_map_pmap(map);
1571         start = entry->start;
1572         end = entry->end;
1573         fictitious = entry->object.vm_object &&
1574                         (entry->object.vm_object->type == OBJT_DEVICE);
1575         if (entry->eflags & MAP_ENTRY_KSTACK)
1576                 start += PAGE_SIZE;
1577         lwkt_gettoken(&vm_token);
1578         vm_map_unlock(map);
1579         map->timestamp++;
1580
1581         /*
1582          * We simulate a fault to get the page and enter it in the physical
1583          * map.
1584          */
1585         for (va = start; va < end; va += PAGE_SIZE) {
1586                 if (user_wire) {
1587                         rv = vm_fault(map, va, VM_PROT_READ,
1588                                         VM_FAULT_USER_WIRE);
1589                 } else {
1590                         rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1591                                         VM_FAULT_CHANGE_WIRING);
1592                 }
1593                 if (rv) {
1594                         while (va > start) {
1595                                 va -= PAGE_SIZE;
1596                                 if ((pa = pmap_extract(pmap, va)) == 0)
1597                                         continue;
1598                                 pmap_change_wiring(pmap, va, FALSE);
1599                                 if (!fictitious)
1600                                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1601                         }
1602                         vm_map_lock(map);
1603                         lwkt_reltoken(&vm_token);
1604                         return (rv);
1605                 }
1606         }
1607         vm_map_lock(map);
1608         lwkt_reltoken(&vm_token);
1609         return (KERN_SUCCESS);
1610 }
1611
1612 /*
1613  * Unwire a range of virtual addresses in a map.  The map should be
1614  * locked.
1615  */
1616 void
1617 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1618 {
1619         boolean_t fictitious;
1620         vm_offset_t start;
1621         vm_offset_t end;
1622         vm_offset_t va;
1623         vm_paddr_t pa;
1624         pmap_t pmap;
1625
1626         pmap = vm_map_pmap(map);
1627         start = entry->start;
1628         end = entry->end;
1629         fictitious = entry->object.vm_object &&
1630                         (entry->object.vm_object->type == OBJT_DEVICE);
1631         if (entry->eflags & MAP_ENTRY_KSTACK)
1632                 start += PAGE_SIZE;
1633
1634         /*
1635          * Since the pages are wired down, we must be able to get their
1636          * mappings from the physical map system.
1637          */
1638         lwkt_gettoken(&vm_token);
1639         for (va = start; va < end; va += PAGE_SIZE) {
1640                 pa = pmap_extract(pmap, va);
1641                 if (pa != 0) {
1642                         pmap_change_wiring(pmap, va, FALSE);
1643                         if (!fictitious)
1644                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1645                 }
1646         }
1647         lwkt_reltoken(&vm_token);
1648 }
1649
1650 /*
1651  * Reduce the rate at which memory is allocated to a process based
1652  * on the perceived load on the VM system. As the load increases
1653  * the allocation burst rate goes down and the delay increases.
1654  *
1655  * Rate limiting does not apply when faulting active or inactive
1656  * pages.  When faulting 'cache' pages, rate limiting only applies
1657  * if the system currently has a severe page deficit.
1658  *
1659  * XXX vm_pagesupply should be increased when a page is freed.
1660  *
1661  * We sleep up to 1/10 of a second.
1662  */
1663 static int
1664 vm_fault_ratelimit(struct vmspace *vmspace)
1665 {
1666         if (vm_load_enable == 0)
1667                 return(0);
1668         if (vmspace->vm_pagesupply > 0) {
1669                 --vmspace->vm_pagesupply;       /* SMP race ok */
1670                 return(0);
1671         }
1672 #ifdef INVARIANTS
1673         if (vm_load_debug) {
1674                 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1675                         vm_load,
1676                         (1000 - vm_load ) / 10, vm_load * hz / 10000,
1677                         curproc->p_pid, curproc->p_comm);
1678         }
1679 #endif
1680         vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1681         return(vm_load * hz / 10000);
1682 }
1683
1684 /*
1685  * Copy all of the pages from a wired-down map entry to another.
1686  *
1687  * The source and destination maps must be locked for write.
1688  * The source map entry must be wired down (or be a sharing map
1689  * entry corresponding to a main map entry that is wired down).
1690  *
1691  * No other requirements.
1692  */
1693 void
1694 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1695                     vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1696 {
1697         vm_object_t dst_object;
1698         vm_object_t src_object;
1699         vm_ooffset_t dst_offset;
1700         vm_ooffset_t src_offset;
1701         vm_prot_t prot;
1702         vm_offset_t vaddr;
1703         vm_page_t dst_m;
1704         vm_page_t src_m;
1705
1706 #ifdef  lint
1707         src_map++;
1708 #endif  /* lint */
1709
1710         src_object = src_entry->object.vm_object;
1711         src_offset = src_entry->offset;
1712
1713         /*
1714          * Create the top-level object for the destination entry. (Doesn't
1715          * actually shadow anything - we copy the pages directly.)
1716          */
1717         vm_map_entry_allocate_object(dst_entry);
1718         dst_object = dst_entry->object.vm_object;
1719
1720         prot = dst_entry->max_protection;
1721
1722         /*
1723          * Loop through all of the pages in the entry's range, copying each
1724          * one from the source object (it should be there) to the destination
1725          * object.
1726          */
1727         for (vaddr = dst_entry->start, dst_offset = 0;
1728             vaddr < dst_entry->end;
1729             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1730
1731                 /*
1732                  * Allocate a page in the destination object
1733                  */
1734                 do {
1735                         dst_m = vm_page_alloc(dst_object,
1736                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1737                         if (dst_m == NULL) {
1738                                 vm_wait(0);
1739                         }
1740                 } while (dst_m == NULL);
1741
1742                 /*
1743                  * Find the page in the source object, and copy it in.
1744                  * (Because the source is wired down, the page will be in
1745                  * memory.)
1746                  */
1747                 src_m = vm_page_lookup(src_object,
1748                                        OFF_TO_IDX(dst_offset + src_offset));
1749                 if (src_m == NULL)
1750                         panic("vm_fault_copy_wired: page missing");
1751
1752                 vm_page_copy(src_m, dst_m);
1753                 vm_page_event(src_m, VMEVENT_COW);
1754
1755                 /*
1756                  * Enter it in the pmap...
1757                  */
1758
1759                 vm_page_flag_clear(dst_m, PG_ZERO);
1760                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1761
1762                 /*
1763                  * Mark it no longer busy, and put it on the active list.
1764                  */
1765                 vm_page_activate(dst_m);
1766                 vm_page_wakeup(dst_m);
1767         }
1768 }
1769
1770 #if 0
1771
1772 /*
1773  * This routine checks around the requested page for other pages that
1774  * might be able to be faulted in.  This routine brackets the viable
1775  * pages for the pages to be paged in.
1776  *
1777  * Inputs:
1778  *      m, rbehind, rahead
1779  *
1780  * Outputs:
1781  *  marray (array of vm_page_t), reqpage (index of requested page)
1782  *
1783  * Return value:
1784  *  number of pages in marray
1785  */
1786 static int
1787 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1788                           vm_page_t *marray, int *reqpage)
1789 {
1790         int i,j;
1791         vm_object_t object;
1792         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1793         vm_page_t rtm;
1794         int cbehind, cahead;
1795
1796         object = m->object;
1797         pindex = m->pindex;
1798
1799         /*
1800          * we don't fault-ahead for device pager
1801          */
1802         if (object->type == OBJT_DEVICE) {
1803                 *reqpage = 0;
1804                 marray[0] = m;
1805                 return 1;
1806         }
1807
1808         /*
1809          * if the requested page is not available, then give up now
1810          */
1811         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1812                 *reqpage = 0;   /* not used by caller, fix compiler warn */
1813                 return 0;
1814         }
1815
1816         if ((cbehind == 0) && (cahead == 0)) {
1817                 *reqpage = 0;
1818                 marray[0] = m;
1819                 return 1;
1820         }
1821
1822         if (rahead > cahead) {
1823                 rahead = cahead;
1824         }
1825
1826         if (rbehind > cbehind) {
1827                 rbehind = cbehind;
1828         }
1829
1830         /*
1831          * Do not do any readahead if we have insufficient free memory.
1832          *
1833          * XXX code was broken disabled before and has instability
1834          * with this conditonal fixed, so shortcut for now.
1835          */
1836         if (burst_fault == 0 || vm_page_count_severe()) {
1837                 marray[0] = m;
1838                 *reqpage = 0;
1839                 return 1;
1840         }
1841
1842         /*
1843          * scan backward for the read behind pages -- in memory
1844          *
1845          * Assume that if the page is not found an interrupt will not
1846          * create it.  Theoretically interrupts can only remove (busy)
1847          * pages, not create new associations.
1848          */
1849         if (pindex > 0) {
1850                 if (rbehind > pindex) {
1851                         rbehind = pindex;
1852                         startpindex = 0;
1853                 } else {
1854                         startpindex = pindex - rbehind;
1855                 }
1856
1857                 lwkt_gettoken(&vm_token);
1858                 for (tpindex = pindex; tpindex > startpindex; --tpindex) {
1859                         if (vm_page_lookup(object, tpindex - 1))
1860                                 break;
1861                 }
1862
1863                 i = 0;
1864                 while (tpindex < pindex) {
1865                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1866                         if (rtm == NULL) {
1867                                 lwkt_reltoken(&vm_token);
1868                                 for (j = 0; j < i; j++) {
1869                                         vm_page_free(marray[j]);
1870                                 }
1871                                 marray[0] = m;
1872                                 *reqpage = 0;
1873                                 return 1;
1874                         }
1875                         marray[i] = rtm;
1876                         ++i;
1877                         ++tpindex;
1878                 }
1879                 lwkt_reltoken(&vm_token);
1880         } else {
1881                 i = 0;
1882         }
1883
1884         /*
1885          * Assign requested page
1886          */
1887         marray[i] = m;
1888         *reqpage = i;
1889         ++i;
1890
1891         /*
1892          * Scan forwards for read-ahead pages
1893          */
1894         tpindex = pindex + 1;
1895         endpindex = tpindex + rahead;
1896         if (endpindex > object->size)
1897                 endpindex = object->size;
1898
1899         lwkt_gettoken(&vm_token);
1900         while (tpindex < endpindex) {
1901                 if (vm_page_lookup(object, tpindex))
1902                         break;
1903                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1904                 if (rtm == NULL)
1905                         break;
1906                 marray[i] = rtm;
1907                 ++i;
1908                 ++tpindex;
1909         }
1910         lwkt_reltoken(&vm_token);
1911
1912         return (i);
1913 }
1914
1915 #endif
1916
1917 /*
1918  * vm_prefault() provides a quick way of clustering pagefaults into a
1919  * processes address space.  It is a "cousin" of pmap_object_init_pt,
1920  * except it runs at page fault time instead of mmap time.
1921  *
1922  * This code used to be per-platform pmap_prefault().  It is now
1923  * machine-independent and enhanced to also pre-fault zero-fill pages
1924  * (see vm.fast_fault) as well as make them writable, which greatly
1925  * reduces the number of page faults programs incur.
1926  *
1927  * Application performance when pre-faulting zero-fill pages is heavily
1928  * dependent on the application.  Very tiny applications like /bin/echo
1929  * lose a little performance while applications of any appreciable size
1930  * gain performance.  Prefaulting multiple pages also reduces SMP
1931  * congestion and can improve SMP performance significantly.
1932  *
1933  * NOTE!  prot may allow writing but this only applies to the top level
1934  *        object.  If we wind up mapping a page extracted from a backing
1935  *        object we have to make sure it is read-only.
1936  *
1937  * NOTE!  The caller has already handled any COW operations on the
1938  *        vm_map_entry via the normal fault code.  Do NOT call this
1939  *        shortcut unless the normal fault code has run on this entry.
1940  *
1941  * No other requirements.
1942  */
1943 #define PFBAK 4
1944 #define PFFOR 4
1945 #define PAGEORDER_SIZE (PFBAK+PFFOR)
1946
1947 static int vm_prefault_pageorder[] = {
1948         -PAGE_SIZE, PAGE_SIZE,
1949         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
1950         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
1951         -4 * PAGE_SIZE, 4 * PAGE_SIZE
1952 };
1953
1954 /*
1955  * Set PG_NOSYNC if the map entry indicates so, but only if the page
1956  * is not already dirty by other means.  This will prevent passive
1957  * filesystem syncing as well as 'sync' from writing out the page.
1958  */
1959 static void
1960 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
1961 {
1962         if (entry->eflags & MAP_ENTRY_NOSYNC) {
1963                 if (m->dirty == 0)
1964                         vm_page_flag_set(m, PG_NOSYNC);
1965         } else {
1966                 vm_page_flag_clear(m, PG_NOSYNC);
1967         }
1968 }
1969
1970 static void
1971 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot)
1972 {
1973         struct lwp *lp;
1974         vm_page_t m;
1975         vm_offset_t starta;
1976         vm_offset_t addr;
1977         vm_pindex_t index;
1978         vm_pindex_t pindex;
1979         vm_object_t object;
1980         int pprot;
1981         int i;
1982
1983         /*
1984          * We do not currently prefault mappings that use virtual page
1985          * tables.  We do not prefault foreign pmaps.
1986          */
1987         if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
1988                 return;
1989         lp = curthread->td_lwp;
1990         if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
1991                 return;
1992
1993         object = entry->object.vm_object;
1994
1995         starta = addra - PFBAK * PAGE_SIZE;
1996         if (starta < entry->start)
1997                 starta = entry->start;
1998         else if (starta > addra)
1999                 starta = 0;
2000
2001         lwkt_gettoken(&vm_token);
2002         for (i = 0; i < PAGEORDER_SIZE; i++) {
2003                 vm_object_t lobject;
2004                 int allocated = 0;
2005
2006                 addr = addra + vm_prefault_pageorder[i];
2007                 if (addr > addra + (PFFOR * PAGE_SIZE))
2008                         addr = 0;
2009
2010                 if (addr < starta || addr >= entry->end)
2011                         continue;
2012
2013                 if (pmap_prefault_ok(pmap, addr) == 0)
2014                         continue;
2015
2016                 /*
2017                  * Follow the VM object chain to obtain the page to be mapped
2018                  * into the pmap.
2019                  *
2020                  * If we reach the terminal object without finding a page
2021                  * and we determine it would be advantageous, then allocate
2022                  * a zero-fill page for the base object.  The base object
2023                  * is guaranteed to be OBJT_DEFAULT for this case.
2024                  *
2025                  * In order to not have to check the pager via *haspage*()
2026                  * we stop if any non-default object is encountered.  e.g.
2027                  * a vnode or swap object would stop the loop.
2028                  */
2029                 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2030                 lobject = object;
2031                 pindex = index;
2032                 pprot = prot;
2033
2034                 while ((m = vm_page_lookup(lobject, pindex)) == NULL) {
2035                         if (lobject->type != OBJT_DEFAULT)
2036                                 break;
2037                         if (lobject->backing_object == NULL) {
2038                                 if (vm_fast_fault == 0)
2039                                         break;
2040                                 if (vm_prefault_pageorder[i] < 0 ||
2041                                     (prot & VM_PROT_WRITE) == 0 ||
2042                                     vm_page_count_min(0)) {
2043                                         break;
2044                                 }
2045                                 /* note: allocate from base object */
2046                                 m = vm_page_alloc(object, index,
2047                                               VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
2048
2049                                 if ((m->flags & PG_ZERO) == 0) {
2050                                         vm_page_zero_fill(m);
2051                                 } else {
2052 #ifdef PMAP_DEBUG
2053                                         pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
2054 #endif
2055                                         vm_page_flag_clear(m, PG_ZERO);
2056                                         mycpu->gd_cnt.v_ozfod++;
2057                                 }
2058                                 mycpu->gd_cnt.v_zfod++;
2059                                 m->valid = VM_PAGE_BITS_ALL;
2060                                 allocated = 1;
2061                                 pprot = prot;
2062                                 /* lobject = object .. not needed */
2063                                 break;
2064                         }
2065                         if (lobject->backing_object_offset & PAGE_MASK)
2066                                 break;
2067                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2068                         lobject = lobject->backing_object;
2069                         pprot &= ~VM_PROT_WRITE;
2070                 }
2071                 /*
2072                  * NOTE: lobject now invalid (if we did a zero-fill we didn't
2073                  *       bother assigning lobject = object).
2074                  *
2075                  * Give-up if the page is not available.
2076                  */
2077                 if (m == NULL)
2078                         break;
2079
2080                 /*
2081                  * Do not conditionalize on PG_RAM.  If pages are present in
2082                  * the VM system we assume optimal caching.  If caching is
2083                  * not optimal the I/O gravy train will be restarted when we
2084                  * hit an unavailable page.  We do not want to try to restart
2085                  * the gravy train now because we really don't know how much
2086                  * of the object has been cached.  The cost for restarting
2087                  * the gravy train should be low (since accesses will likely
2088                  * be I/O bound anyway).
2089                  *
2090                  * The object must be marked dirty if we are mapping a
2091                  * writable page.
2092                  */
2093                 if (pprot & VM_PROT_WRITE)
2094                         vm_object_set_writeable_dirty(m->object);
2095
2096                 /*
2097                  * Enter the page into the pmap if appropriate.  If we had
2098                  * allocated the page we have to place it on a queue.  If not
2099                  * we just have to make sure it isn't on the cache queue
2100                  * (pages on the cache queue are not allowed to be mapped).
2101                  */
2102                 if (allocated) {
2103                         if (pprot & VM_PROT_WRITE)
2104                                 vm_set_nosync(m, entry);
2105                         pmap_enter(pmap, addr, m, pprot, 0);
2106                         vm_page_deactivate(m);
2107                         vm_page_wakeup(m);
2108                 } else if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2109                     (m->busy == 0) &&
2110                     (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2111
2112                         if ((m->queue - m->pc) == PQ_CACHE) {
2113                                 vm_page_deactivate(m);
2114                         }
2115                         vm_page_busy(m);
2116                         if (pprot & VM_PROT_WRITE)
2117                                 vm_set_nosync(m, entry);
2118                         pmap_enter(pmap, addr, m, pprot, 0);
2119                         vm_page_wakeup(m);
2120                 }
2121         }
2122         lwkt_reltoken(&vm_token);
2123 }