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         /* BEFORE */
 408         result = vm_fault_object(&fs, first_pindex, fault_type);
 409
 410         if (result == KERN_TRY_AGAIN) {
 411                 /*lwkt_reltoken(&vm_token);*/
 412                 goto RetryFault;
 413         }
 414         if (result != KERN_SUCCESS) {
 415                 /*lwkt_reltoken(&vm_token);*/
 416                 return (result);
 417         }
 418
 419         /*
 420          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 421          * will contain a busied page.
 422          *
 423          * Enter the page into the pmap and do pmap-related adjustments.
 424          */
 425         vm_page_flag_set(fs.m, PG_REFERENCED);
 426         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
 427
 428         /*
 429          * Burst in a few more pages if possible.  The fs.map should still
 430          * be locked.
 431          */
 432         if (fault_flags & VM_FAULT_BURST) {
 433                 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
 434                     fs.wired == 0) {
 435                         vm_prefault(fs.map->pmap, vaddr, fs.entry, fs.prot);
 436                 }
 437         }
 438         lwkt_gettoken(&vm_token);
 439         unlock_things(&fs);
 440
 441         /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
 442         KKASSERT(fs.m->flags & PG_BUSY);
 443
 444         /*
 445          * If the page is not wired down, then put it where the pageout daemon
 446          * can find it.
 447          *
 448          * We do not really need to get vm_token here but since all the
 449          * vm_*() calls have to doing it here improves efficiency.
 450          */
 451         /*lwkt_gettoken(&vm_token);*/
 452
 453         if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
 454                 lwkt_reltoken(&vm_token); /* before wire activate does not */
 455                 if (fs.wired)
 456                         vm_page_wire(fs.m);
 457                 else
 458                         vm_page_unwire(fs.m, 1);
 459         } else {
 460                 vm_page_activate(fs.m);
 461                 lwkt_reltoken(&vm_token); /* before wire activate does not */
 462         }
 463         /*lwkt_reltoken(&vm_token); after wire/activate works */
 464
 465         if (curthread->td_lwp) {
 466                 if (fs.hardfault) {
 467                         curthread->td_lwp->lwp_ru.ru_majflt++;
 468                 } else {
 469                         curthread->td_lwp->lwp_ru.ru_minflt++;
 470                 }
 471         }
 472
 473         /*
 474          * Unlock everything, and return
 475          */
 476         vm_page_wakeup(fs.m);
 477         vm_object_deallocate(fs.first_object);
 478         /*fs.m = NULL; */
 479         /*fs.first_object = NULL; */
 480         /*lwkt_reltoken(&vm_token);*/
 481
 482         return (KERN_SUCCESS);
 483 }
 484
 485 /*
 486  * Fault in the specified virtual address in the current process map,
 487  * returning a held VM page or NULL.  See vm_fault_page() for more
 488  * information.
 489  *
 490  * No requirements.
 491  */
 492 vm_page_t
 493 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
 494 {
 495         struct lwp *lp = curthread->td_lwp;
 496         vm_page_t m;
 497
 498         m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
 499                           fault_type, VM_FAULT_NORMAL, errorp);
 500         return(m);
 501 }
 502
 503 /*
 504  * Fault in the specified virtual address in the specified map, doing all
 505  * necessary manipulation of the object store and all necessary I/O.  Return
 506  * a held VM page or NULL, and set *errorp.  The related pmap is not
 507  * updated.
 508  *
 509  * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
 510  * and marked PG_REFERENCED as well.
 511  *
 512  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
 513  * error will be returned.
 514  *
 515  * No requirements.
 516  */
 517 vm_page_t
 518 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 519               int fault_flags, int *errorp)
 520 {
 521         vm_pindex_t first_pindex;
 522         struct faultstate fs;
 523         int result;
 524         vm_prot_t orig_fault_type = fault_type;
 525
 526         mycpu->gd_cnt.v_vm_faults++;
 527
 528         fs.didlimit = 0;
 529         fs.hardfault = 0;
 530         fs.fault_flags = fault_flags;
 531         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
 532
 533 RetryFault:
 534         /*
 535          * Find the vm_map_entry representing the backing store and resolve
 536          * the top level object and page index.  This may have the side
 537          * effect of executing a copy-on-write on the map entry and/or
 538          * creating a shadow object, but will not COW any actual VM pages.
 539          *
 540          * On success fs.map is left read-locked and various other fields
 541          * are initialized but not otherwise referenced or locked.
 542          *
 543          * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
 544          * if the map entry is a virtual page table and also writable,
 545          * so we can set the 'A'accessed bit in the virtual page table entry.
 546          */
 547         fs.map = map;
 548         result = vm_map_lookup(&fs.map, vaddr, fault_type,
 549                                &fs.entry, &fs.first_object,
 550                                &first_pindex, &fs.first_prot, &fs.wired);
 551
 552         if (result != KERN_SUCCESS) {
 553                 *errorp = result;
 554                 return (NULL);
 555         }
 556
 557         /*
 558          * fs.map is read-locked
 559          *
 560          * Misc checks.  Save the map generation number to detect races.
 561          */
 562         fs.map_generation = fs.map->timestamp;
 563
 564         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 565                 panic("vm_fault: fault on nofault entry, addr: %lx",
 566                     (u_long)vaddr);
 567         }
 568
 569         /*
 570          * A system map entry may return a NULL object.  No object means
 571          * no pager means an unrecoverable kernel fault.
 572          */
 573         if (fs.first_object == NULL) {
 574                 panic("vm_fault: unrecoverable fault at %p in entry %p",
 575                         (void *)vaddr, fs.entry);
 576         }
 577
 578         /*
 579          * Make a reference to this object to prevent its disposal while we
 580          * are messing with it.  Once we have the reference, the map is free
 581          * to be diddled.  Since objects reference their shadows (and copies),
 582          * they will stay around as well.
 583          *
 584          * Bump the paging-in-progress count to prevent size changes (e.g.
 585          * truncation operations) during I/O.  This must be done after
 586          * obtaining the vnode lock in order to avoid possible deadlocks.
 587          *
 588          * The vm_token is needed to manipulate the vm_object
 589          */
 590         lwkt_gettoken(&vm_token);
 591         vm_object_reference(fs.first_object);
 592         fs.vp = vnode_pager_lock(fs.first_object);
 593         vm_object_pip_add(fs.first_object, 1);
 594         lwkt_reltoken(&vm_token);
 595
 596         fs.lookup_still_valid = TRUE;
 597         fs.first_m = NULL;
 598         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 599
 600         /*
 601          * If the entry is wired we cannot change the page protection.
 602          */
 603         if (fs.wired)
 604                 fault_type = fs.first_prot;
 605
 606         /*
 607          * The page we want is at (first_object, first_pindex), but if the
 608          * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
 609          * page table to figure out the actual pindex.
 610          *
 611          * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
 612          * ONLY
 613          */
 614         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 615                 result = vm_fault_vpagetable(&fs, &first_pindex,
 616                                              fs.entry->aux.master_pde,
 617                                              fault_type);
 618                 if (result == KERN_TRY_AGAIN)
 619                         goto RetryFault;
 620                 if (result != KERN_SUCCESS) {
 621                         *errorp = result;
 622                         return (NULL);
 623                 }
 624         }
 625
 626         /*
 627          * Now we have the actual (object, pindex), fault in the page.  If
 628          * vm_fault_object() fails it will unlock and deallocate the FS
 629          * data.   If it succeeds everything remains locked and fs->object
 630          * will have an additinal PIP count if it is not equal to
 631          * fs->first_object
 632          */
 633         result = vm_fault_object(&fs, first_pindex, fault_type);
 634
 635         if (result == KERN_TRY_AGAIN)
 636                 goto RetryFault;
 637         if (result != KERN_SUCCESS) {
 638                 *errorp = result;
 639                 return(NULL);
 640         }
 641
 642         if ((orig_fault_type & VM_PROT_WRITE) &&
 643             (fs.prot & VM_PROT_WRITE) == 0) {
 644                 *errorp = KERN_PROTECTION_FAILURE;
 645                 unlock_and_deallocate(&fs);
 646                 return(NULL);
 647         }
 648
 649         /*
 650          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 651          * will contain a busied page.
 652          */
 653         unlock_things(&fs);
 654
 655         /*
 656          * Return a held page.  We are not doing any pmap manipulation so do
 657          * not set PG_MAPPED.  However, adjust the page flags according to
 658          * the fault type because the caller may not use a managed pmapping
 659          * (so we don't want to lose the fact that the page will be dirtied
 660          * if a write fault was specified).
 661          */
 662         lwkt_gettoken(&vm_token);
 663         vm_page_hold(fs.m);
 664         if (fault_type & VM_PROT_WRITE)
 665                 vm_page_dirty(fs.m);
 666
 667         /*
 668          * Update the pmap.  We really only have to do this if a COW
 669          * occured to replace the read-only page with the new page.  For
 670          * now just do it unconditionally. XXX
 671          */
 672         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
 673         vm_page_flag_set(fs.m, PG_REFERENCED);
 674
 675         /*
 676          * Unbusy the page by activating it.  It remains held and will not
 677          * be reclaimed.
 678          */
 679         vm_page_activate(fs.m);
 680
 681         if (curthread->td_lwp) {
 682                 if (fs.hardfault) {
 683                         curthread->td_lwp->lwp_ru.ru_majflt++;
 684                 } else {
 685                         curthread->td_lwp->lwp_ru.ru_minflt++;
 686                 }
 687         }
 688
 689         /*
 690          * Unlock everything, and return the held page.
 691          */
 692         vm_page_wakeup(fs.m);
 693         vm_object_deallocate(fs.first_object);
 694         /*fs.first_object = NULL; */
 695         lwkt_reltoken(&vm_token);
 696
 697         *errorp = 0;
 698         return(fs.m);
 699 }
 700
 701 /*
 702  * Fault in the specified (object,offset), dirty the returned page as
 703  * needed.  If the requested fault_type cannot be done NULL and an
 704  * error is returned.
 705  *
 706  * A held (but not busied) page is returned.
 707  *
 708  * No requirements.
 709  */
 710 vm_page_t
 711 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
 712                      vm_prot_t fault_type, int fault_flags, int *errorp)
 713 {
 714         int result;
 715         vm_pindex_t first_pindex;
 716         struct faultstate fs;
 717         struct vm_map_entry entry;
 718
 719         bzero(&entry, sizeof(entry));
 720         entry.object.vm_object = object;
 721         entry.maptype = VM_MAPTYPE_NORMAL;
 722         entry.protection = entry.max_protection = fault_type;
 723
 724         fs.didlimit = 0;
 725         fs.hardfault = 0;
 726         fs.fault_flags = fault_flags;
 727         fs.map = NULL;
 728         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
 729
 730 RetryFault:
 731
 732         fs.first_object = object;
 733         first_pindex = OFF_TO_IDX(offset);
 734         fs.entry = &entry;
 735         fs.first_prot = fault_type;
 736         fs.wired = 0;
 737         /*fs.map_generation = 0; unused */
 738
 739         /*
 740          * Make a reference to this object to prevent its disposal while we
 741          * are messing with it.  Once we have the reference, the map is free
 742          * to be diddled.  Since objects reference their shadows (and copies),
 743          * they will stay around as well.
 744          *
 745          * Bump the paging-in-progress count to prevent size changes (e.g.
 746          * truncation operations) during I/O.  This must be done after
 747          * obtaining the vnode lock in order to avoid possible deadlocks.
 748          */
 749         lwkt_gettoken(&vm_token);
 750         vm_object_reference(fs.first_object);
 751         fs.vp = vnode_pager_lock(fs.first_object);
 752         vm_object_pip_add(fs.first_object, 1);
 753         lwkt_reltoken(&vm_token);
 754
 755         fs.lookup_still_valid = TRUE;
 756         fs.first_m = NULL;
 757         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
 758
 759 #if 0
 760         /* XXX future - ability to operate on VM object using vpagetable */
 761         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
 762                 result = vm_fault_vpagetable(&fs, &first_pindex,
 763                                              fs.entry->aux.master_pde,
 764                                              fault_type);
 765                 if (result == KERN_TRY_AGAIN)
 766                         goto RetryFault;
 767                 if (result != KERN_SUCCESS) {
 768                         *errorp = result;
 769                         return (NULL);
 770                 }
 771         }
 772 #endif
 773
 774         /*
 775          * Now we have the actual (object, pindex), fault in the page.  If
 776          * vm_fault_object() fails it will unlock and deallocate the FS
 777          * data.   If it succeeds everything remains locked and fs->object
 778          * will have an additinal PIP count if it is not equal to
 779          * fs->first_object
 780          */
 781         result = vm_fault_object(&fs, first_pindex, fault_type);
 782
 783         if (result == KERN_TRY_AGAIN)
 784                 goto RetryFault;
 785         if (result != KERN_SUCCESS) {
 786                 *errorp = result;
 787                 return(NULL);
 788         }
 789
 790         if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
 791                 *errorp = KERN_PROTECTION_FAILURE;
 792                 unlock_and_deallocate(&fs);
 793                 return(NULL);
 794         }
 795
 796         /*
 797          * On success vm_fault_object() does not unlock or deallocate, and fs.m
 798          * will contain a busied page.
 799          */
 800         unlock_things(&fs);
 801
 802         /*
 803          * Return a held page.  We are not doing any pmap manipulation so do
 804          * not set PG_MAPPED.  However, adjust the page flags according to
 805          * the fault type because the caller may not use a managed pmapping
 806          * (so we don't want to lose the fact that the page will be dirtied
 807          * if a write fault was specified).
 808          */
 809         lwkt_gettoken(&vm_token);
 810         vm_page_hold(fs.m);
 811         if (fault_type & VM_PROT_WRITE)
 812                 vm_page_dirty(fs.m);
 813
 814         if (fault_flags & VM_FAULT_DIRTY)
 815                 vm_page_dirty(fs.m);
 816         if (fault_flags & VM_FAULT_UNSWAP)
 817                 swap_pager_unswapped(fs.m);
 818
 819         /*
 820          * Indicate that the page was accessed.
 821          */
 822         vm_page_flag_set(fs.m, PG_REFERENCED);
 823
 824         /*
 825          * Unbusy the page by activating it.  It remains held and will not
 826          * be reclaimed.
 827          */
 828         vm_page_activate(fs.m);
 829
 830         if (curthread->td_lwp) {
 831                 if (fs.hardfault) {
 832                         mycpu->gd_cnt.v_vm_faults++;
 833                         curthread->td_lwp->lwp_ru.ru_majflt++;
 834                 } else {
 835                         curthread->td_lwp->lwp_ru.ru_minflt++;
 836                 }
 837         }
 838
 839         /*
 840          * Unlock everything, and return the held page.
 841          */
 842         vm_page_wakeup(fs.m);
 843         vm_object_deallocate(fs.first_object);
 844         /*fs.first_object = NULL; */
 845         lwkt_reltoken(&vm_token);
 846
 847         *errorp = 0;
 848         return(fs.m);
 849 }
 850
 851 /*
 852  * Translate the virtual page number (first_pindex) that is relative
 853  * to the address space into a logical page number that is relative to the
 854  * backing object.  Use the virtual page table pointed to by (vpte).
 855  *
 856  * This implements an N-level page table.  Any level can terminate the
 857  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
 858  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
 859  *
 860  * No requirements (vm_token need not be held).
 861  */
 862 static
 863 int
 864 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
 865                     vpte_t vpte, int fault_type)
 866 {
 867         struct lwbuf *lwb;
 868         struct lwbuf lwb_cache;
 869         int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
 870         int result = KERN_SUCCESS;
 871         vpte_t *ptep;
 872
 873         for (;;) {
 874                 /*
 875                  * We cannot proceed if the vpte is not valid, not readable
 876                  * for a read fault, or not writable for a write fault.
 877                  */
 878                 if ((vpte & VPTE_V) == 0) {
 879                         unlock_and_deallocate(fs);
 880                         return (KERN_FAILURE);
 881                 }
 882                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
 883                         unlock_and_deallocate(fs);
 884                         return (KERN_FAILURE);
 885                 }
 886                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
 887                         unlock_and_deallocate(fs);
 888                         return (KERN_FAILURE);
 889                 }
 890                 if ((vpte & VPTE_PS) || vshift == 0)
 891                         break;
 892                 KKASSERT(vshift >= VPTE_PAGE_BITS);
 893
 894                 /*
 895                  * Get the page table page.  Nominally we only read the page
 896                  * table, but since we are actively setting VPTE_M and VPTE_A,
 897                  * tell vm_fault_object() that we are writing it.
 898                  *
 899                  * There is currently no real need to optimize this.
 900                  */
 901                 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
 902                                          VM_PROT_READ|VM_PROT_WRITE);
 903                 if (result != KERN_SUCCESS)
 904                         return (result);
 905
 906                 /*
 907                  * Process the returned fs.m and look up the page table
 908                  * entry in the page table page.
 909                  */
 910                 vshift -= VPTE_PAGE_BITS;
 911                 lwb = lwbuf_alloc(fs->m, &lwb_cache);
 912                 ptep = ((vpte_t *)lwbuf_kva(lwb) +
 913                         ((*pindex >> vshift) & VPTE_PAGE_MASK));
 914                 vpte = *ptep;
 915
 916                 /*
 917                  * Page table write-back.  If the vpte is valid for the
 918                  * requested operation, do a write-back to the page table.
 919                  *
 920                  * XXX VPTE_M is not set properly for page directory pages.
 921                  * It doesn't get set in the page directory if the page table
 922                  * is modified during a read access.
 923                  */
 924                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
 925                     (vpte & VPTE_W)) {
 926                         if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
 927                                 atomic_set_long(ptep, VPTE_M | VPTE_A);
 928                                 vm_page_dirty(fs->m);
 929                         }
 930                 }
 931                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
 932                     (vpte & VPTE_R)) {
 933                         if ((vpte & VPTE_A) == 0) {
 934                                 atomic_set_long(ptep, VPTE_A);
 935                                 vm_page_dirty(fs->m);
 936                         }
 937                 }
 938                 lwbuf_free(lwb);
 939                 vm_page_flag_set(fs->m, PG_REFERENCED);
 940                 vm_page_activate(fs->m);
 941                 vm_page_wakeup(fs->m);
 942                 fs->m = NULL;
 943                 cleanup_successful_fault(fs);
 944         }
 945         /*
 946          * Combine remaining address bits with the vpte.
 947          */
 948         /* JG how many bits from each? */
 949         *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
 950                   (*pindex & ((1L << vshift) - 1));
 951         return (KERN_SUCCESS);
 952 }
 953
 954
 955 /*
 956  * This is the core of the vm_fault code.
 957  *
 958  * Do all operations required to fault-in (fs.first_object, pindex).  Run
 959  * through the shadow chain as necessary and do required COW or virtual
 960  * copy operations.  The caller has already fully resolved the vm_map_entry
 961  * and, if appropriate, has created a copy-on-write layer.  All we need to
 962  * do is iterate the object chain.
 963  *
 964  * On failure (fs) is unlocked and deallocated and the caller may return or
 965  * retry depending on the failure code.  On success (fs) is NOT unlocked or
 966  * deallocated, fs.m will contained a resolved, busied page, and fs.object
 967  * will have an additional PIP count if it is not equal to fs.first_object.
 968  *
 969  * No requirements.
 970  */
 971 static
 972 int
 973 vm_fault_object(struct faultstate *fs,
 974                 vm_pindex_t first_pindex, vm_prot_t fault_type)
 975 {
 976         vm_object_t next_object;
 977         vm_pindex_t pindex;
 978
 979         fs->prot = fs->first_prot;
 980         fs->object = fs->first_object;
 981         pindex = first_pindex;
 982
 983         /*
 984          * If a read fault occurs we try to make the page writable if
 985          * possible.  There are three cases where we cannot make the
 986          * page mapping writable:
 987          *
 988          * (1) The mapping is read-only or the VM object is read-only,
 989          *     fs->prot above will simply not have VM_PROT_WRITE set.
 990          *
 991          * (2) If the mapping is a virtual page table we need to be able
 992          *     to detect writes so we can set VPTE_M in the virtual page
 993          *     table.
 994          *
 995          * (3) If the VM page is read-only or copy-on-write, upgrading would
 996          *     just result in an unnecessary COW fault.
 997          *
 998          * VM_PROT_VPAGED is set if faulting via a virtual page table and
 999          * causes adjustments to the 'M'odify bit to also turn off write
1000          * access to force a re-fault.
1001          */
1002         if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1003                 if ((fault_type & VM_PROT_WRITE) == 0)
1004                         fs->prot &= ~VM_PROT_WRITE;
1005         }
1006
1007         lwkt_gettoken(&vm_token);
1008
1009         for (;;) {
1010                 /*
1011                  * If the object is dead, we stop here
1012                  */
1013                 if (fs->object->flags & OBJ_DEAD) {
1014                         unlock_and_deallocate(fs);
1015                         lwkt_reltoken(&vm_token);
1016                         return (KERN_PROTECTION_FAILURE);
1017                 }
1018
1019                 /*
1020                  * See if the page is resident.
1021                  */
1022                 fs->m = vm_page_lookup(fs->object, pindex);
1023                 if (fs->m != NULL) {
1024                         int queue;
1025                         /*
1026                          * Wait/Retry if the page is busy.  We have to do this
1027                          * if the page is busy via either PG_BUSY or
1028                          * vm_page_t->busy because the vm_pager may be using
1029                          * vm_page_t->busy for pageouts ( and even pageins if
1030                          * it is the vnode pager ), and we could end up trying
1031                          * to pagein and pageout the same page simultaneously.
1032                          *
1033                          * We can theoretically allow the busy case on a read
1034                          * fault if the page is marked valid, but since such
1035                          * pages are typically already pmap'd, putting that
1036                          * special case in might be more effort then it is
1037                          * worth.  We cannot under any circumstances mess
1038                          * around with a vm_page_t->busy page except, perhaps,
1039                          * to pmap it.
1040                          */
1041                         if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
1042                                 unlock_things(fs);
1043                                 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1044                                 mycpu->gd_cnt.v_intrans++;
1045                                 vm_object_deallocate(fs->first_object);
1046                                 fs->first_object = NULL;
1047                                 lwkt_reltoken(&vm_token);
1048                                 return (KERN_TRY_AGAIN);
1049                         }
1050
1051                         /*
1052                          * If reactivating a page from PQ_CACHE we may have
1053                          * to rate-limit.
1054                          */
1055                         queue = fs->m->queue;
1056                         vm_page_unqueue_nowakeup(fs->m);
1057
1058                         if ((queue - fs->m->pc) == PQ_CACHE &&
1059                             vm_page_count_severe()) {
1060                                 vm_page_activate(fs->m);
1061                                 unlock_and_deallocate(fs);
1062                                 vm_waitpfault();
1063                                 lwkt_reltoken(&vm_token);
1064                                 return (KERN_TRY_AGAIN);
1065                         }
1066
1067                         /*
1068                          * Mark page busy for other processes, and the
1069                          * pagedaemon.  If it still isn't completely valid
1070                          * (readable), or if a read-ahead-mark is set on
1071                          * the VM page, jump to readrest, else we found the
1072                          * page and can return.
1073                          *
1074                          * We can release the spl once we have marked the
1075                          * page busy.
1076                          */
1077                         vm_page_busy(fs->m);
1078
1079                         if (fs->m->object != &kernel_object) {
1080                                 if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1081                                     VM_PAGE_BITS_ALL) {
1082                                         goto readrest;
1083                                 }
1084                                 if (fs->m->flags & PG_RAM) {
1085                                         if (debug_cluster)
1086                                                 kprintf("R");
1087                                         vm_page_flag_clear(fs->m, PG_RAM);
1088                                         goto readrest;
1089                                 }
1090                         }
1091                         break; /* break to PAGE HAS BEEN FOUND */
1092                 }
1093
1094                 /*
1095                  * Page is not resident, If this is the search termination
1096                  * or the pager might contain the page, allocate a new page.
1097                  */
1098                 if (TRYPAGER(fs) || fs->object == fs->first_object) {
1099                         /*
1100                          * If the page is beyond the object size we fail
1101                          */
1102                         if (pindex >= fs->object->size) {
1103                                 lwkt_reltoken(&vm_token);
1104                                 unlock_and_deallocate(fs);
1105                                 return (KERN_PROTECTION_FAILURE);
1106                         }
1107
1108                         /*
1109                          * Ratelimit.
1110                          */
1111                         if (fs->didlimit == 0 && curproc != NULL) {
1112                                 int limticks;
1113
1114                                 limticks = vm_fault_ratelimit(curproc->p_vmspace);
1115                                 if (limticks) {
1116                                         lwkt_reltoken(&vm_token);
1117                                         unlock_and_deallocate(fs);
1118                                         tsleep(curproc, 0, "vmrate", limticks);
1119                                         fs->didlimit = 1;
1120                                         return (KERN_TRY_AGAIN);
1121                                 }
1122                         }
1123
1124                         /*
1125                          * Allocate a new page for this object/offset pair.
1126                          */
1127                         fs->m = NULL;
1128                         if (!vm_page_count_severe()) {
1129                                 fs->m = vm_page_alloc(fs->object, pindex,
1130                                     (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
1131                         }
1132                         if (fs->m == NULL) {
1133                                 lwkt_reltoken(&vm_token);
1134                                 unlock_and_deallocate(fs);
1135                                 vm_waitpfault();
1136                                 return (KERN_TRY_AGAIN);
1137                         }
1138                 }
1139
1140 readrest:
1141                 /*
1142                  * We have found an invalid or partially valid page, a
1143                  * page with a read-ahead mark which might be partially or
1144                  * fully valid (and maybe dirty too), or we have allocated
1145                  * a new page.
1146                  *
1147                  * Attempt to fault-in the page if there is a chance that the
1148                  * pager has it, and potentially fault in additional pages
1149                  * at the same time.
1150                  *
1151                  * We are NOT in splvm here and if TRYPAGER is true then
1152                  * fs.m will be non-NULL and will be PG_BUSY for us.
1153                  */
1154                 if (TRYPAGER(fs)) {
1155                         int rv;
1156                         int seqaccess;
1157                         u_char behavior = vm_map_entry_behavior(fs->entry);
1158
1159                         if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1160                                 seqaccess = 0;
1161                         else
1162                                 seqaccess = -1;
1163
1164                         /*
1165                          * If sequential access is detected then attempt
1166                          * to deactivate/cache pages behind the scan to
1167                          * prevent resource hogging.
1168                          *
1169                          * Use of PG_RAM to detect sequential access
1170                          * also simulates multi-zone sequential access
1171                          * detection for free.
1172                          *
1173                          * NOTE: Partially valid dirty pages cannot be
1174                          *       deactivated without causing NFS picemeal
1175                          *       writes to barf.
1176                          */
1177                         if ((fs->first_object->type != OBJT_DEVICE) &&
1178                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1179                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1180                                  (fs->m->flags & PG_RAM)))
1181                         ) {
1182                                 vm_pindex_t scan_pindex;
1183                                 int scan_count = 16;
1184
1185                                 if (first_pindex < 16) {
1186                                         scan_pindex = 0;
1187                                         scan_count = 0;
1188                                 } else {
1189                                         scan_pindex = first_pindex - 16;
1190                                         if (scan_pindex < 16)
1191                                                 scan_count = scan_pindex;
1192                                         else
1193                                                 scan_count = 16;
1194                                 }
1195
1196                                 while (scan_count) {
1197                                         vm_page_t mt;
1198
1199                                         mt = vm_page_lookup(fs->first_object,
1200                                                             scan_pindex);
1201                                         if (mt == NULL ||
1202                                             (mt->valid != VM_PAGE_BITS_ALL)) {
1203                                                 break;
1204                                         }
1205                                         if (mt->busy ||
1206                                             (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
1207                                             mt->hold_count ||
1208                                             mt->wire_count)  {
1209                                                 goto skip;
1210                                         }
1211                                         vm_page_busy(mt);
1212                                         if (mt->dirty == 0)
1213                                                 vm_page_test_dirty(mt);
1214                                         if (mt->dirty) {
1215                                                 vm_page_protect(mt,
1216                                                                 VM_PROT_NONE);
1217                                                 vm_page_deactivate(mt);
1218                                                 vm_page_wakeup(mt);
1219                                         } else {
1220                                                 vm_page_cache(mt);
1221                                         }
1222 skip:
1223                                         --scan_count;
1224                                         --scan_pindex;
1225                                 }
1226
1227                                 seqaccess = 1;
1228                         }
1229
1230                         /*
1231                          * Avoid deadlocking against the map when doing I/O.
1232                          * fs.object and the page is PG_BUSY'd.
1233                          */
1234                         unlock_map(fs);
1235
1236                         /*
1237                          * Acquire the page data.  We still hold a ref on
1238                          * fs.object and the page has been PG_BUSY's.
1239                          *
1240                          * The pager may replace the page (for example, in
1241                          * order to enter a fictitious page into the
1242                          * object).  If it does so it is responsible for
1243                          * cleaning up the passed page and properly setting
1244                          * the new page PG_BUSY.
1245                          *
1246                          * If we got here through a PG_RAM read-ahead
1247                          * mark the page may be partially dirty and thus
1248                          * not freeable.  Don't bother checking to see
1249                          * if the pager has the page because we can't free
1250                          * it anyway.  We have to depend on the get_page
1251                          * operation filling in any gaps whether there is
1252                          * backing store or not.
1253                          */
1254                         rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1255
1256                         if (rv == VM_PAGER_OK) {
1257                                 /*
1258                                  * Relookup in case pager changed page. Pager
1259                                  * is responsible for disposition of old page
1260                                  * if moved.
1261                                  *
1262                                  * XXX other code segments do relookups too.
1263                                  * It's a bad abstraction that needs to be
1264                                  * fixed/removed.
1265                                  */
1266                                 fs->m = vm_page_lookup(fs->object, pindex);
1267                                 if (fs->m == NULL) {
1268                                         lwkt_reltoken(&vm_token);
1269                                         unlock_and_deallocate(fs);
1270                                         return (KERN_TRY_AGAIN);
1271                                 }
1272
1273                                 ++fs->hardfault;
1274                                 break; /* break to PAGE HAS BEEN FOUND */
1275                         }
1276
1277                         /*
1278                          * Remove the bogus page (which does not exist at this
1279                          * object/offset); before doing so, we must get back
1280                          * our object lock to preserve our invariant.
1281                          *
1282                          * Also wake up any other process that may want to bring
1283                          * in this page.
1284                          *
1285                          * If this is the top-level object, we must leave the
1286                          * busy page to prevent another process from rushing
1287                          * past us, and inserting the page in that object at
1288                          * the same time that we are.
1289                          */
1290                         if (rv == VM_PAGER_ERROR) {
1291                                 if (curproc)
1292                                         kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
1293                                 else
1294                                         kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
1295                         }
1296
1297                         /*
1298                          * Data outside the range of the pager or an I/O error
1299                          *
1300                          * The page may have been wired during the pagein,
1301                          * e.g. by the buffer cache, and cannot simply be
1302                          * freed.  Call vnode_pager_freepage() to deal with it.
1303                          */
1304                         /*
1305                          * XXX - the check for kernel_map is a kludge to work
1306                          * around having the machine panic on a kernel space
1307                          * fault w/ I/O error.
1308                          */
1309                         if (((fs->map != &kernel_map) &&
1310                             (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1311                                 vnode_pager_freepage(fs->m);
1312                                 lwkt_reltoken(&vm_token);
1313                                 fs->m = NULL;
1314                                 unlock_and_deallocate(fs);
1315                                 if (rv == VM_PAGER_ERROR)
1316                                         return (KERN_FAILURE);
1317                                 else
1318                                         return (KERN_PROTECTION_FAILURE);
1319                                 /* NOT REACHED */
1320                         }
1321                         if (fs->object != fs->first_object) {
1322                                 vnode_pager_freepage(fs->m);
1323                                 fs->m = NULL;
1324                                 /*
1325                                  * XXX - we cannot just fall out at this
1326                                  * point, m has been freed and is invalid!
1327                                  */
1328                         }
1329                 }
1330
1331                 /*
1332                  * We get here if the object has a default pager (or unwiring)
1333                  * or the pager doesn't have the page.
1334                  */
1335                 if (fs->object == fs->first_object)
1336                         fs->first_m = fs->m;
1337
1338                 /*
1339                  * Move on to the next object.  Lock the next object before
1340                  * unlocking the current one.
1341                  */
1342                 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1343                 next_object = fs->object->backing_object;
1344                 if (next_object == NULL) {
1345                         /*
1346                          * If there's no object left, fill the page in the top
1347                          * object with zeros.
1348                          */
1349                         if (fs->object != fs->first_object) {
1350                                 vm_object_pip_wakeup(fs->object);
1351
1352                                 fs->object = fs->first_object;
1353                                 pindex = first_pindex;
1354                                 fs->m = fs->first_m;
1355                         }
1356                         fs->first_m = NULL;
1357
1358                         /*
1359                          * Zero the page if necessary and mark it valid.
1360                          */
1361                         if ((fs->m->flags & PG_ZERO) == 0) {
1362                                 vm_page_zero_fill(fs->m);
1363                         } else {
1364 #ifdef PMAP_DEBUG
1365                                 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m));
1366 #endif
1367                                 vm_page_flag_clear(fs->m, PG_ZERO);
1368                                 mycpu->gd_cnt.v_ozfod++;
1369                         }
1370                         mycpu->gd_cnt.v_zfod++;
1371                         fs->m->valid = VM_PAGE_BITS_ALL;
1372                         break;  /* break to PAGE HAS BEEN FOUND */
1373                 }
1374                 if (fs->object != fs->first_object) {
1375                         vm_object_pip_wakeup(fs->object);
1376                 }
1377                 KASSERT(fs->object != next_object,
1378                         ("object loop %p", next_object));
1379                 fs->object = next_object;
1380                 vm_object_pip_add(fs->object, 1);
1381         }
1382
1383         /*
1384          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1385          * is held.]
1386          *
1387          * vm_token is still held
1388          *
1389          * If the page is being written, but isn't already owned by the
1390          * top-level object, we have to copy it into a new page owned by the
1391          * top-level object.
1392          */
1393         KASSERT((fs->m->flags & PG_BUSY) != 0,
1394                 ("vm_fault: not busy after main loop"));
1395
1396         if (fs->object != fs->first_object) {
1397                 /*
1398                  * We only really need to copy if we want to write it.
1399                  */
1400                 if (fault_type & VM_PROT_WRITE) {
1401                         /*
1402                          * This allows pages to be virtually copied from a
1403                          * backing_object into the first_object, where the
1404                          * backing object has no other refs to it, and cannot
1405                          * gain any more refs.  Instead of a bcopy, we just
1406                          * move the page from the backing object to the
1407                          * first object.  Note that we must mark the page
1408                          * dirty in the first object so that it will go out
1409                          * to swap when needed.
1410                          */
1411                         if (
1412                                 /*
1413                                  * Map, if present, has not changed
1414                                  */
1415                                 (fs->map == NULL ||
1416                                 fs->map_generation == fs->map->timestamp) &&
1417                                 /*
1418                                  * Only one shadow object
1419                                  */
1420                                 (fs->object->shadow_count == 1) &&
1421                                 /*
1422                                  * No COW refs, except us
1423                                  */
1424                                 (fs->object->ref_count == 1) &&
1425                                 /*
1426                                  * No one else can look this object up
1427                                  */
1428                                 (fs->object->handle == NULL) &&
1429                                 /*
1430                                  * No other ways to look the object up
1431                                  */
1432                                 ((fs->object->type == OBJT_DEFAULT) ||
1433                                  (fs->object->type == OBJT_SWAP)) &&
1434                                 /*
1435                                  * We don't chase down the shadow chain
1436                                  */
1437                                 (fs->object == fs->first_object->backing_object) &&
1438
1439                                 /*
1440                                  * grab the lock if we need to
1441                                  */
1442                                 (fs->lookup_still_valid ||
1443                                  fs->map == NULL ||
1444                                  lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1445                             ) {
1446
1447                                 fs->lookup_still_valid = 1;
1448                                 /*
1449                                  * get rid of the unnecessary page
1450                                  */
1451                                 vm_page_protect(fs->first_m, VM_PROT_NONE);
1452                                 vm_page_free(fs->first_m);
1453                                 fs->first_m = NULL;
1454
1455                                 /*
1456                                  * grab the page and put it into the
1457                                  * process'es object.  The page is
1458                                  * automatically made dirty.
1459                                  */
1460                                 vm_page_rename(fs->m, fs->first_object, first_pindex);
1461                                 fs->first_m = fs->m;
1462                                 vm_page_busy(fs->first_m);
1463                                 fs->m = NULL;
1464                                 mycpu->gd_cnt.v_cow_optim++;
1465                         } else {
1466                                 /*
1467                                  * Oh, well, lets copy it.
1468                                  */
1469                                 vm_page_copy(fs->m, fs->first_m);
1470                                 vm_page_event(fs->m, VMEVENT_COW);
1471                         }
1472
1473                         if (fs->m) {
1474                                 /*
1475                                  * We no longer need the old page or object.
1476                                  */
1477                                 release_page(fs);
1478                         }
1479
1480                         /*
1481                          * fs->object != fs->first_object due to above
1482                          * conditional
1483                          */
1484                         vm_object_pip_wakeup(fs->object);
1485
1486                         /*
1487                          * Only use the new page below...
1488                          */
1489
1490                         mycpu->gd_cnt.v_cow_faults++;
1491                         fs->m = fs->first_m;
1492                         fs->object = fs->first_object;
1493                         pindex = first_pindex;
1494                 } else {
1495                         /*
1496                          * If it wasn't a write fault avoid having to copy
1497                          * the page by mapping it read-only.
1498                          */
1499                         fs->prot &= ~VM_PROT_WRITE;
1500                 }
1501         }
1502
1503         /*
1504          * We may have had to unlock a map to do I/O.  If we did then
1505          * lookup_still_valid will be FALSE.  If the map generation count
1506          * also changed then all sorts of things could have happened while
1507          * we were doing the I/O and we need to retry.
1508          */
1509
1510         if (!fs->lookup_still_valid &&
1511             fs->map != NULL &&
1512             (fs->map->timestamp != fs->map_generation)) {
1513                 release_page(fs);
1514                 lwkt_reltoken(&vm_token);
1515                 unlock_and_deallocate(fs);
1516                 return (KERN_TRY_AGAIN);
1517         }
1518
1519         /*
1520          * If the fault is a write, we know that this page is being
1521          * written NOW so dirty it explicitly to save on pmap_is_modified()
1522          * calls later.
1523          *
1524          * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1525          * if the page is already dirty to prevent data written with
1526          * the expectation of being synced from not being synced.
1527          * Likewise if this entry does not request NOSYNC then make
1528          * sure the page isn't marked NOSYNC.  Applications sharing
1529          * data should use the same flags to avoid ping ponging.
1530          *
1531          * Also tell the backing pager, if any, that it should remove
1532          * any swap backing since the page is now dirty.
1533          */
1534         if (fs->prot & VM_PROT_WRITE) {
1535                 vm_object_set_writeable_dirty(fs->m->object);
1536                 vm_set_nosync(fs->m, fs->entry);
1537                 if (fs->fault_flags & VM_FAULT_DIRTY) {
1538                         vm_page_dirty(fs->m);
1539                         swap_pager_unswapped(fs->m);
1540                 }
1541         }
1542
1543         lwkt_reltoken(&vm_token);
1544
1545         /*
1546          * Page had better still be busy.  We are still locked up and
1547          * fs->object will have another PIP reference if it is not equal
1548          * to fs->first_object.
1549          */
1550         KASSERT(fs->m->flags & PG_BUSY,
1551                 ("vm_fault: page %p not busy!", fs->m));
1552
1553         /*
1554          * Sanity check: page must be completely valid or it is not fit to
1555          * map into user space.  vm_pager_get_pages() ensures this.
1556          */
1557         if (fs->m->valid != VM_PAGE_BITS_ALL) {
1558                 vm_page_zero_invalid(fs->m, TRUE);
1559                 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1560         }
1561         vm_page_flag_clear(fs->m, PG_ZERO);
1562
1563         return (KERN_SUCCESS);
1564 }
1565
1566 /*
1567  * Wire down a range of virtual addresses in a map.  The entry in question
1568  * should be marked in-transition and the map must be locked.  We must
1569  * release the map temporarily while faulting-in the page to avoid a
1570  * deadlock.  Note that the entry may be clipped while we are blocked but
1571  * will never be freed.
1572  *
1573  * No requirements.
1574  */
1575 int
1576 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1577 {
1578         boolean_t fictitious;
1579         vm_offset_t start;
1580         vm_offset_t end;
1581         vm_offset_t va;
1582         vm_paddr_t pa;
1583         pmap_t pmap;
1584         int rv;
1585
1586         pmap = vm_map_pmap(map);
1587         start = entry->start;
1588         end = entry->end;
1589         fictitious = entry->object.vm_object &&
1590                         (entry->object.vm_object->type == OBJT_DEVICE);
1591         if (entry->eflags & MAP_ENTRY_KSTACK)
1592                 start += PAGE_SIZE;
1593         lwkt_gettoken(&vm_token);
1594         vm_map_unlock(map);
1595         map->timestamp++;
1596
1597         /*
1598          * We simulate a fault to get the page and enter it in the physical
1599          * map.
1600          */
1601         for (va = start; va < end; va += PAGE_SIZE) {
1602                 if (user_wire) {
1603                         rv = vm_fault(map, va, VM_PROT_READ,
1604                                         VM_FAULT_USER_WIRE);
1605                 } else {
1606                         rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1607                                         VM_FAULT_CHANGE_WIRING);
1608                 }
1609                 if (rv) {
1610                         while (va > start) {
1611                                 va -= PAGE_SIZE;
1612                                 if ((pa = pmap_extract(pmap, va)) == 0)
1613                                         continue;
1614                                 pmap_change_wiring(pmap, va, FALSE);
1615                                 if (!fictitious)
1616                                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1617                         }
1618                         vm_map_lock(map);
1619                         lwkt_reltoken(&vm_token);
1620                         return (rv);
1621                 }
1622         }
1623         vm_map_lock(map);
1624         lwkt_reltoken(&vm_token);
1625         return (KERN_SUCCESS);
1626 }
1627
1628 /*
1629  * Unwire a range of virtual addresses in a map.  The map should be
1630  * locked.
1631  */
1632 void
1633 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1634 {
1635         boolean_t fictitious;
1636         vm_offset_t start;
1637         vm_offset_t end;
1638         vm_offset_t va;
1639         vm_paddr_t pa;
1640         pmap_t pmap;
1641
1642         pmap = vm_map_pmap(map);
1643         start = entry->start;
1644         end = entry->end;
1645         fictitious = entry->object.vm_object &&
1646                         (entry->object.vm_object->type == OBJT_DEVICE);
1647         if (entry->eflags & MAP_ENTRY_KSTACK)
1648                 start += PAGE_SIZE;
1649
1650         /*
1651          * Since the pages are wired down, we must be able to get their
1652          * mappings from the physical map system.
1653          */
1654         lwkt_gettoken(&vm_token);
1655         for (va = start; va < end; va += PAGE_SIZE) {
1656                 pa = pmap_extract(pmap, va);
1657                 if (pa != 0) {
1658                         pmap_change_wiring(pmap, va, FALSE);
1659                         if (!fictitious)
1660                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1661                 }
1662         }
1663         lwkt_reltoken(&vm_token);
1664 }
1665
1666 /*
1667  * Reduce the rate at which memory is allocated to a process based
1668  * on the perceived load on the VM system. As the load increases
1669  * the allocation burst rate goes down and the delay increases.
1670  *
1671  * Rate limiting does not apply when faulting active or inactive
1672  * pages.  When faulting 'cache' pages, rate limiting only applies
1673  * if the system currently has a severe page deficit.
1674  *
1675  * XXX vm_pagesupply should be increased when a page is freed.
1676  *
1677  * We sleep up to 1/10 of a second.
1678  */
1679 static int
1680 vm_fault_ratelimit(struct vmspace *vmspace)
1681 {
1682         if (vm_load_enable == 0)
1683                 return(0);
1684         if (vmspace->vm_pagesupply > 0) {
1685                 --vmspace->vm_pagesupply;       /* SMP race ok */
1686                 return(0);
1687         }
1688 #ifdef INVARIANTS
1689         if (vm_load_debug) {
1690                 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1691                         vm_load,
1692                         (1000 - vm_load ) / 10, vm_load * hz / 10000,
1693                         curproc->p_pid, curproc->p_comm);
1694         }
1695 #endif
1696         vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1697         return(vm_load * hz / 10000);
1698 }
1699
1700 /*
1701  * Copy all of the pages from a wired-down map entry to another.
1702  *
1703  * The source and destination maps must be locked for write.
1704  * The source map entry must be wired down (or be a sharing map
1705  * entry corresponding to a main map entry that is wired down).
1706  *
1707  * No other requirements.
1708  */
1709 void
1710 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1711                     vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1712 {
1713         vm_object_t dst_object;
1714         vm_object_t src_object;
1715         vm_ooffset_t dst_offset;
1716         vm_ooffset_t src_offset;
1717         vm_prot_t prot;
1718         vm_offset_t vaddr;
1719         vm_page_t dst_m;
1720         vm_page_t src_m;
1721
1722 #ifdef  lint
1723         src_map++;
1724 #endif  /* lint */
1725
1726         src_object = src_entry->object.vm_object;
1727         src_offset = src_entry->offset;
1728
1729         /*
1730          * Create the top-level object for the destination entry. (Doesn't
1731          * actually shadow anything - we copy the pages directly.)
1732          */
1733         vm_map_entry_allocate_object(dst_entry);
1734         dst_object = dst_entry->object.vm_object;
1735
1736         prot = dst_entry->max_protection;
1737
1738         /*
1739          * Loop through all of the pages in the entry's range, copying each
1740          * one from the source object (it should be there) to the destination
1741          * object.
1742          */
1743         for (vaddr = dst_entry->start, dst_offset = 0;
1744             vaddr < dst_entry->end;
1745             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1746
1747                 /*
1748                  * Allocate a page in the destination object
1749                  */
1750                 do {
1751                         dst_m = vm_page_alloc(dst_object,
1752                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1753                         if (dst_m == NULL) {
1754                                 vm_wait(0);
1755                         }
1756                 } while (dst_m == NULL);
1757
1758                 /*
1759                  * Find the page in the source object, and copy it in.
1760                  * (Because the source is wired down, the page will be in
1761                  * memory.)
1762                  */
1763                 src_m = vm_page_lookup(src_object,
1764                                        OFF_TO_IDX(dst_offset + src_offset));
1765                 if (src_m == NULL)
1766                         panic("vm_fault_copy_wired: page missing");
1767
1768                 vm_page_copy(src_m, dst_m);
1769                 vm_page_event(src_m, VMEVENT_COW);
1770
1771                 /*
1772                  * Enter it in the pmap...
1773                  */
1774
1775                 vm_page_flag_clear(dst_m, PG_ZERO);
1776                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1777
1778                 /*
1779                  * Mark it no longer busy, and put it on the active list.
1780                  */
1781                 vm_page_activate(dst_m);
1782                 vm_page_wakeup(dst_m);
1783         }
1784 }
1785
1786 #if 0
1787
1788 /*
1789  * This routine checks around the requested page for other pages that
1790  * might be able to be faulted in.  This routine brackets the viable
1791  * pages for the pages to be paged in.
1792  *
1793  * Inputs:
1794  *      m, rbehind, rahead
1795  *
1796  * Outputs:
1797  *  marray (array of vm_page_t), reqpage (index of requested page)
1798  *
1799  * Return value:
1800  *  number of pages in marray
1801  */
1802 static int
1803 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1804                           vm_page_t *marray, int *reqpage)
1805 {
1806         int i,j;
1807         vm_object_t object;
1808         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1809         vm_page_t rtm;
1810         int cbehind, cahead;
1811
1812         object = m->object;
1813         pindex = m->pindex;
1814
1815         /*
1816          * we don't fault-ahead for device pager
1817          */
1818         if (object->type == OBJT_DEVICE) {
1819                 *reqpage = 0;
1820                 marray[0] = m;
1821                 return 1;
1822         }
1823
1824         /*
1825          * if the requested page is not available, then give up now
1826          */
1827         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1828                 *reqpage = 0;   /* not used by caller, fix compiler warn */
1829                 return 0;
1830         }
1831
1832         if ((cbehind == 0) && (cahead == 0)) {
1833                 *reqpage = 0;
1834                 marray[0] = m;
1835                 return 1;
1836         }
1837
1838         if (rahead > cahead) {
1839                 rahead = cahead;
1840         }
1841
1842         if (rbehind > cbehind) {
1843                 rbehind = cbehind;
1844         }
1845
1846         /*
1847          * Do not do any readahead if we have insufficient free memory.
1848          *
1849          * XXX code was broken disabled before and has instability
1850          * with this conditonal fixed, so shortcut for now.
1851          */
1852         if (burst_fault == 0 || vm_page_count_severe()) {
1853                 marray[0] = m;
1854                 *reqpage = 0;
1855                 return 1;
1856         }
1857
1858         /*
1859          * scan backward for the read behind pages -- in memory
1860          *
1861          * Assume that if the page is not found an interrupt will not
1862          * create it.  Theoretically interrupts can only remove (busy)
1863          * pages, not create new associations.
1864          */
1865         if (pindex > 0) {
1866                 if (rbehind > pindex) {
1867                         rbehind = pindex;
1868                         startpindex = 0;
1869                 } else {
1870                         startpindex = pindex - rbehind;
1871                 }
1872
1873                 lwkt_gettoken(&vm_token);
1874                 for (tpindex = pindex; tpindex > startpindex; --tpindex) {
1875                         if (vm_page_lookup(object, tpindex - 1))
1876                                 break;
1877                 }
1878
1879                 i = 0;
1880                 while (tpindex < pindex) {
1881                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1882                         if (rtm == NULL) {
1883                                 lwkt_reltoken(&vm_token);
1884                                 for (j = 0; j < i; j++) {
1885                                         vm_page_free(marray[j]);
1886                                 }
1887                                 marray[0] = m;
1888                                 *reqpage = 0;
1889                                 return 1;
1890                         }
1891                         marray[i] = rtm;
1892                         ++i;
1893                         ++tpindex;
1894                 }
1895                 lwkt_reltoken(&vm_token);
1896         } else {
1897                 i = 0;
1898         }
1899
1900         /*
1901          * Assign requested page
1902          */
1903         marray[i] = m;
1904         *reqpage = i;
1905         ++i;
1906
1907         /*
1908          * Scan forwards for read-ahead pages
1909          */
1910         tpindex = pindex + 1;
1911         endpindex = tpindex + rahead;
1912         if (endpindex > object->size)
1913                 endpindex = object->size;
1914
1915         lwkt_gettoken(&vm_token);
1916         while (tpindex < endpindex) {
1917                 if (vm_page_lookup(object, tpindex))
1918                         break;
1919                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1920                 if (rtm == NULL)
1921                         break;
1922                 marray[i] = rtm;
1923                 ++i;
1924                 ++tpindex;
1925         }
1926         lwkt_reltoken(&vm_token);
1927
1928         return (i);
1929 }
1930
1931 #endif
1932
1933 /*
1934  * vm_prefault() provides a quick way of clustering pagefaults into a
1935  * processes address space.  It is a "cousin" of pmap_object_init_pt,
1936  * except it runs at page fault time instead of mmap time.
1937  *
1938  * This code used to be per-platform pmap_prefault().  It is now
1939  * machine-independent and enhanced to also pre-fault zero-fill pages
1940  * (see vm.fast_fault) as well as make them writable, which greatly
1941  * reduces the number of page faults programs incur.
1942  *
1943  * Application performance when pre-faulting zero-fill pages is heavily
1944  * dependent on the application.  Very tiny applications like /bin/echo
1945  * lose a little performance while applications of any appreciable size
1946  * gain performance.  Prefaulting multiple pages also reduces SMP
1947  * congestion and can improve SMP performance significantly.
1948  *
1949  * NOTE!  prot may allow writing but this only applies to the top level
1950  *        object.  If we wind up mapping a page extracted from a backing
1951  *        object we have to make sure it is read-only.
1952  *
1953  * NOTE!  The caller has already handled any COW operations on the
1954  *        vm_map_entry via the normal fault code.  Do NOT call this
1955  *        shortcut unless the normal fault code has run on this entry.
1956  *
1957  * No other requirements.
1958  */
1959 #define PFBAK 4
1960 #define PFFOR 4
1961 #define PAGEORDER_SIZE (PFBAK+PFFOR)
1962
1963 static int vm_prefault_pageorder[] = {
1964         -PAGE_SIZE, PAGE_SIZE,
1965         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
1966         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
1967         -4 * PAGE_SIZE, 4 * PAGE_SIZE
1968 };
1969
1970 /*
1971  * Set PG_NOSYNC if the map entry indicates so, but only if the page
1972  * is not already dirty by other means.  This will prevent passive
1973  * filesystem syncing as well as 'sync' from writing out the page.
1974  */
1975 static void
1976 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
1977 {
1978         if (entry->eflags & MAP_ENTRY_NOSYNC) {
1979                 if (m->dirty == 0)
1980                         vm_page_flag_set(m, PG_NOSYNC);
1981         } else {
1982                 vm_page_flag_clear(m, PG_NOSYNC);
1983         }
1984 }
1985
1986 static void
1987 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot)
1988 {
1989         struct lwp *lp;
1990         vm_page_t m;
1991         vm_offset_t starta;
1992         vm_offset_t addr;
1993         vm_pindex_t index;
1994         vm_pindex_t pindex;
1995         vm_object_t object;
1996         int pprot;
1997         int i;
1998
1999         /*
2000          * We do not currently prefault mappings that use virtual page
2001          * tables.  We do not prefault foreign pmaps.
2002          */
2003         if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2004                 return;
2005         lp = curthread->td_lwp;
2006         if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2007                 return;
2008
2009         object = entry->object.vm_object;
2010
2011         starta = addra - PFBAK * PAGE_SIZE;
2012         if (starta < entry->start)
2013                 starta = entry->start;
2014         else if (starta > addra)
2015                 starta = 0;
2016
2017         lwkt_gettoken(&vm_token);
2018         for (i = 0; i < PAGEORDER_SIZE; i++) {
2019                 vm_object_t lobject;
2020                 int allocated = 0;
2021
2022                 addr = addra + vm_prefault_pageorder[i];
2023                 if (addr > addra + (PFFOR * PAGE_SIZE))
2024                         addr = 0;
2025
2026                 if (addr < starta || addr >= entry->end)
2027                         continue;
2028
2029                 if (pmap_prefault_ok(pmap, addr) == 0)
2030                         continue;
2031
2032                 /*
2033                  * Follow the VM object chain to obtain the page to be mapped
2034                  * into the pmap.
2035                  *
2036                  * If we reach the terminal object without finding a page
2037                  * and we determine it would be advantageous, then allocate
2038                  * a zero-fill page for the base object.  The base object
2039                  * is guaranteed to be OBJT_DEFAULT for this case.
2040                  *
2041                  * In order to not have to check the pager via *haspage*()
2042                  * we stop if any non-default object is encountered.  e.g.
2043                  * a vnode or swap object would stop the loop.
2044                  */
2045                 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2046                 lobject = object;
2047                 pindex = index;
2048                 pprot = prot;
2049
2050                 while ((m = vm_page_lookup(lobject, pindex)) == NULL) {
2051                         if (lobject->type != OBJT_DEFAULT)
2052                                 break;
2053                         if (lobject->backing_object == NULL) {
2054                                 if (vm_fast_fault == 0)
2055                                         break;
2056                                 if (vm_prefault_pageorder[i] < 0 ||
2057                                     (prot & VM_PROT_WRITE) == 0 ||
2058                                     vm_page_count_min(0)) {
2059                                         break;
2060                                 }
2061                                 /* note: allocate from base object */
2062                                 m = vm_page_alloc(object, index,
2063                                               VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
2064
2065                                 if ((m->flags & PG_ZERO) == 0) {
2066                                         vm_page_zero_fill(m);
2067                                 } else {
2068 #ifdef PMAP_DEBUG
2069                                         pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
2070 #endif
2071                                         vm_page_flag_clear(m, PG_ZERO);
2072                                         mycpu->gd_cnt.v_ozfod++;
2073                                 }
2074                                 mycpu->gd_cnt.v_zfod++;
2075                                 m->valid = VM_PAGE_BITS_ALL;
2076                                 allocated = 1;
2077                                 pprot = prot;
2078                                 /* lobject = object .. not needed */
2079                                 break;
2080                         }
2081                         if (lobject->backing_object_offset & PAGE_MASK)
2082                                 break;
2083                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2084                         lobject = lobject->backing_object;
2085                         pprot &= ~VM_PROT_WRITE;
2086                 }
2087                 /*
2088                  * NOTE: lobject now invalid (if we did a zero-fill we didn't
2089                  *       bother assigning lobject = object).
2090                  *
2091                  * Give-up if the page is not available.
2092                  */
2093                 if (m == NULL)
2094                         break;
2095
2096                 /*
2097                  * Do not conditionalize on PG_RAM.  If pages are present in
2098                  * the VM system we assume optimal caching.  If caching is
2099                  * not optimal the I/O gravy train will be restarted when we
2100                  * hit an unavailable page.  We do not want to try to restart
2101                  * the gravy train now because we really don't know how much
2102                  * of the object has been cached.  The cost for restarting
2103                  * the gravy train should be low (since accesses will likely
2104                  * be I/O bound anyway).
2105                  *
2106                  * The object must be marked dirty if we are mapping a
2107                  * writable page.
2108                  */
2109                 if (pprot & VM_PROT_WRITE)
2110                         vm_object_set_writeable_dirty(m->object);
2111
2112                 /*
2113                  * Enter the page into the pmap if appropriate.  If we had
2114                  * allocated the page we have to place it on a queue.  If not
2115                  * we just have to make sure it isn't on the cache queue
2116                  * (pages on the cache queue are not allowed to be mapped).
2117                  */
2118                 if (allocated) {
2119                         if (pprot & VM_PROT_WRITE)
2120                                 vm_set_nosync(m, entry);
2121                         pmap_enter(pmap, addr, m, pprot, 0);
2122                         vm_page_deactivate(m);
2123                         vm_page_wakeup(m);
2124                 } else if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2125                     (m->busy == 0) &&
2126                     (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2127
2128                         vm_page_busy(m);
2129                         if ((m->queue - m->pc) == PQ_CACHE) {
2130                                 vm_page_deactivate(m);
2131                         }
2132                         if (pprot & VM_PROT_WRITE)
2133                                 vm_set_nosync(m, entry);
2134                         pmap_enter(pmap, addr, m, pprot, 0);
2135                         vm_page_wakeup(m);
2136                 }
2137         }
2138         lwkt_reltoken(&vm_token);
2139 }