sys/vm/vnode_pager.c

   1 /*-
   2  * Copyright (c) 1990 University of Utah.
   3  * Copyright (c) 1991 The Regents of the University of California.
   4  * All rights reserved.
   5  * Copyright (c) 1993, 1994 John S. Dyson
   6  * Copyright (c) 1995, David Greenman
   7  *
   8  * This code is derived from software contributed to Berkeley by
   9  * the Systems Programming Group of the University of Utah Computer
  10  * Science Department.
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  * 1. Redistributions of source code must retain the above copyright
  16  *    notice, this list of conditions and the following disclaimer.
  17  * 2. Redistributions in binary form must reproduce the above copyright
  18  *    notice, this list of conditions and the following disclaimer in the
  19  *    documentation and/or other materials provided with the distribution.
  20  * 3. All advertising materials mentioning features or use of this software
  21  *    must display the following acknowledgement:
  22  *      This product includes software developed by the University of
  23  *      California, Berkeley and its contributors.
  24  * 4. Neither the name of the University nor the names of its contributors
  25  *    may be used to endorse or promote products derived from this software
  26  *    without specific prior written permission.
  27  *
  28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  38  * SUCH DAMAGE.
  39  *
  40  *      from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91
  41  */
  42
  43 /*
  44  * Page to/from files (vnodes).
  45  */
  46
  47 /*
  48  * TODO:
  49  *      Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
  50  *      greatly re-simplify the vnode_pager.
  51  */
  52
  53 #include <sys/cdefs.h>
  54 __FBSDID("$FreeBSD$");
  55
  56 #include <sys/param.h>
  57 #include <sys/systm.h>
  58 #include <sys/proc.h>
  59 #include <sys/vnode.h>
  60 #include <sys/mount.h>
  61 #include <sys/bio.h>
  62 #include <sys/buf.h>
  63 #include <sys/vmmeter.h>
  64 #include <sys/limits.h>
  65 #include <sys/conf.h>
  66 #include <sys/rwlock.h>
  67 #include <sys/sf_buf.h>
  68
  69 #include <machine/atomic.h>
  70
  71 #include <vm/vm.h>
  72 #include <vm/vm_param.h>
  73 #include <vm/vm_object.h>
  74 #include <vm/vm_page.h>
  75 #include <vm/vm_pager.h>
  76 #include <vm/vm_map.h>
  77 #include <vm/vnode_pager.h>
  78 #include <vm/vm_extern.h>
  79
  80 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
  81     daddr_t *rtaddress, int *run);
  82 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
  83 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
  84 static void vnode_pager_dealloc(vm_object_t);
  85 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
  86 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
  87 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
  88 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
  89     vm_ooffset_t, struct ucred *cred);
  90
  91 struct pagerops vnodepagerops = {
  92         .pgo_alloc =    vnode_pager_alloc,
  93         .pgo_dealloc =  vnode_pager_dealloc,
  94         .pgo_getpages = vnode_pager_getpages,
  95         .pgo_putpages = vnode_pager_putpages,
  96         .pgo_haspage =  vnode_pager_haspage,
  97 };
  98
  99 int vnode_pbuf_freecnt;
 100
 101 /* Create the VM system backing object for this vnode */
 102 int
 103 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
 104 {
 105         vm_object_t object;
 106         vm_ooffset_t size = isize;
 107         struct vattr va;
 108
 109         if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
 110                 return (0);
 111
 112         while ((object = vp->v_object) != NULL) {
 113                 VM_OBJECT_WLOCK(object);
 114                 if (!(object->flags & OBJ_DEAD)) {
 115                         VM_OBJECT_WUNLOCK(object);
 116                         return (0);
 117                 }
 118                 VOP_UNLOCK(vp, 0);
 119                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 120                 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0);
 121                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 122         }
 123
 124         if (size == 0) {
 125                 if (vn_isdisk(vp, NULL)) {
 126                         size = IDX_TO_OFF(INT_MAX);
 127                 } else {
 128                         if (VOP_GETATTR(vp, &va, td->td_ucred))
 129                                 return (0);
 130                         size = va.va_size;
 131                 }
 132         }
 133
 134         object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
 135         /*
 136          * Dereference the reference we just created.  This assumes
 137          * that the object is associated with the vp.
 138          */
 139         VM_OBJECT_WLOCK(object);
 140         object->ref_count--;
 141         VM_OBJECT_WUNLOCK(object);
 142         vrele(vp);
 143
 144         KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
 145
 146         return (0);
 147 }
 148
 149 void
 150 vnode_destroy_vobject(struct vnode *vp)
 151 {
 152         struct vm_object *obj;
 153
 154         obj = vp->v_object;
 155         if (obj == NULL)
 156                 return;
 157         ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
 158         VM_OBJECT_WLOCK(obj);
 159         if (obj->ref_count == 0) {
 160                 /*
 161                  * vclean() may be called twice. The first time
 162                  * removes the primary reference to the object,
 163                  * the second time goes one further and is a
 164                  * special-case to terminate the object.
 165                  *
 166                  * don't double-terminate the object
 167                  */
 168                 if ((obj->flags & OBJ_DEAD) == 0)
 169                         vm_object_terminate(obj);
 170                 else
 171                         VM_OBJECT_WUNLOCK(obj);
 172         } else {
 173                 /*
 174                  * Woe to the process that tries to page now :-).
 175                  */
 176                 vm_pager_deallocate(obj);
 177                 VM_OBJECT_WUNLOCK(obj);
 178         }
 179         vp->v_object = NULL;
 180 }
 181
 182
 183 /*
 184  * Allocate (or lookup) pager for a vnode.
 185  * Handle is a vnode pointer.
 186  *
 187  * MPSAFE
 188  */
 189 vm_object_t
 190 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
 191     vm_ooffset_t offset, struct ucred *cred)
 192 {
 193         vm_object_t object;
 194         struct vnode *vp;
 195
 196         /*
 197          * Pageout to vnode, no can do yet.
 198          */
 199         if (handle == NULL)
 200                 return (NULL);
 201
 202         vp = (struct vnode *) handle;
 203
 204         /*
 205          * If the object is being terminated, wait for it to
 206          * go away.
 207          */
 208 retry:
 209         while ((object = vp->v_object) != NULL) {
 210                 VM_OBJECT_WLOCK(object);
 211                 if ((object->flags & OBJ_DEAD) == 0)
 212                         break;
 213                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 214                 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
 215         }
 216
 217         KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
 218
 219         if (object == NULL) {
 220                 /*
 221                  * Add an object of the appropriate size
 222                  */
 223                 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
 224
 225                 object->un_pager.vnp.vnp_size = size;
 226                 object->un_pager.vnp.writemappings = 0;
 227
 228                 object->handle = handle;
 229                 VI_LOCK(vp);
 230                 if (vp->v_object != NULL) {
 231                         /*
 232                          * Object has been created while we were sleeping
 233                          */
 234                         VI_UNLOCK(vp);
 235                         vm_object_destroy(object);
 236                         goto retry;
 237                 }
 238                 vp->v_object = object;
 239                 VI_UNLOCK(vp);
 240         } else {
 241                 object->ref_count++;
 242                 VM_OBJECT_WUNLOCK(object);
 243         }
 244         vref(vp);
 245         return (object);
 246 }
 247
 248 /*
 249  *      The object must be locked.
 250  */
 251 static void
 252 vnode_pager_dealloc(object)
 253         vm_object_t object;
 254 {
 255         struct vnode *vp;
 256         int refs;
 257
 258         vp = object->handle;
 259         if (vp == NULL)
 260                 panic("vnode_pager_dealloc: pager already dealloced");
 261
 262         VM_OBJECT_ASSERT_WLOCKED(object);
 263         vm_object_pip_wait(object, "vnpdea");
 264         refs = object->ref_count;
 265
 266         object->handle = NULL;
 267         object->type = OBJT_DEAD;
 268         if (object->flags & OBJ_DISCONNECTWNT) {
 269                 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
 270                 wakeup(object);
 271         }
 272         ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
 273         if (object->un_pager.vnp.writemappings > 0) {
 274                 object->un_pager.vnp.writemappings = 0;
 275                 VOP_ADD_WRITECOUNT(vp, -1);
 276                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
 277                     __func__, vp, vp->v_writecount);
 278         }
 279         vp->v_object = NULL;
 280         VOP_UNSET_TEXT(vp);
 281         VM_OBJECT_WUNLOCK(object);
 282         while (refs-- > 0)
 283                 vunref(vp);
 284         VM_OBJECT_WLOCK(object);
 285 }
 286
 287 static boolean_t
 288 vnode_pager_haspage(object, pindex, before, after)
 289         vm_object_t object;
 290         vm_pindex_t pindex;
 291         int *before;
 292         int *after;
 293 {
 294         struct vnode *vp = object->handle;
 295         daddr_t bn;
 296         int err;
 297         daddr_t reqblock;
 298         int poff;
 299         int bsize;
 300         int pagesperblock, blocksperpage;
 301
 302         VM_OBJECT_ASSERT_WLOCKED(object);
 303         /*
 304          * If no vp or vp is doomed or marked transparent to VM, we do not
 305          * have the page.
 306          */
 307         if (vp == NULL || vp->v_iflag & VI_DOOMED)
 308                 return FALSE;
 309         /*
 310          * If the offset is beyond end of file we do
 311          * not have the page.
 312          */
 313         if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
 314                 return FALSE;
 315
 316         bsize = vp->v_mount->mnt_stat.f_iosize;
 317         pagesperblock = bsize / PAGE_SIZE;
 318         blocksperpage = 0;
 319         if (pagesperblock > 0) {
 320                 reqblock = pindex / pagesperblock;
 321         } else {
 322                 blocksperpage = (PAGE_SIZE / bsize);
 323                 reqblock = pindex * blocksperpage;
 324         }
 325         VM_OBJECT_WUNLOCK(object);
 326         err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
 327         VM_OBJECT_WLOCK(object);
 328         if (err)
 329                 return TRUE;
 330         if (bn == -1)
 331                 return FALSE;
 332         if (pagesperblock > 0) {
 333                 poff = pindex - (reqblock * pagesperblock);
 334                 if (before) {
 335                         *before *= pagesperblock;
 336                         *before += poff;
 337                 }
 338                 if (after) {
 339                         int numafter;
 340                         *after *= pagesperblock;
 341                         numafter = pagesperblock - (poff + 1);
 342                         if (IDX_TO_OFF(pindex + numafter) >
 343                             object->un_pager.vnp.vnp_size) {
 344                                 numafter =
 345                                     OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
 346                                     pindex;
 347                         }
 348                         *after += numafter;
 349                 }
 350         } else {
 351                 if (before) {
 352                         *before /= blocksperpage;
 353                 }
 354
 355                 if (after) {
 356                         *after /= blocksperpage;
 357                 }
 358         }
 359         return TRUE;
 360 }
 361
 362 /*
 363  * Lets the VM system know about a change in size for a file.
 364  * We adjust our own internal size and flush any cached pages in
 365  * the associated object that are affected by the size change.
 366  *
 367  * Note: this routine may be invoked as a result of a pager put
 368  * operation (possibly at object termination time), so we must be careful.
 369  */
 370 void
 371 vnode_pager_setsize(vp, nsize)
 372         struct vnode *vp;
 373         vm_ooffset_t nsize;
 374 {
 375         vm_object_t object;
 376         vm_page_t m;
 377         vm_pindex_t nobjsize;
 378
 379         if ((object = vp->v_object) == NULL)
 380                 return;
 381 /*      ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
 382         VM_OBJECT_WLOCK(object);
 383         if (nsize == object->un_pager.vnp.vnp_size) {
 384                 /*
 385                  * Hasn't changed size
 386                  */
 387                 VM_OBJECT_WUNLOCK(object);
 388                 return;
 389         }
 390         nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
 391         if (nsize < object->un_pager.vnp.vnp_size) {
 392                 /*
 393                  * File has shrunk. Toss any cached pages beyond the new EOF.
 394                  */
 395                 if (nobjsize < object->size)
 396                         vm_object_page_remove(object, nobjsize, object->size,
 397                             0);
 398                 /*
 399                  * this gets rid of garbage at the end of a page that is now
 400                  * only partially backed by the vnode.
 401                  *
 402                  * XXX for some reason (I don't know yet), if we take a
 403                  * completely invalid page and mark it partially valid
 404                  * it can screw up NFS reads, so we don't allow the case.
 405                  */
 406                 if ((nsize & PAGE_MASK) &&
 407                     (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
 408                     m->valid != 0) {
 409                         int base = (int)nsize & PAGE_MASK;
 410                         int size = PAGE_SIZE - base;
 411
 412                         /*
 413                          * Clear out partial-page garbage in case
 414                          * the page has been mapped.
 415                          */
 416                         pmap_zero_page_area(m, base, size);
 417
 418                         /*
 419                          * Update the valid bits to reflect the blocks that
 420                          * have been zeroed.  Some of these valid bits may
 421                          * have already been set.
 422                          */
 423                         vm_page_set_valid_range(m, base, size);
 424
 425                         /*
 426                          * Round "base" to the next block boundary so that the
 427                          * dirty bit for a partially zeroed block is not
 428                          * cleared.
 429                          */
 430                         base = roundup2(base, DEV_BSIZE);
 431
 432                         /*
 433                          * Clear out partial-page dirty bits.
 434                          *
 435                          * note that we do not clear out the valid
 436                          * bits.  This would prevent bogus_page
 437                          * replacement from working properly.
 438                          */
 439                         vm_page_clear_dirty(m, base, PAGE_SIZE - base);
 440                 } else if ((nsize & PAGE_MASK) &&
 441                     vm_page_is_cached(object, OFF_TO_IDX(nsize))) {
 442                         vm_page_cache_free(object, OFF_TO_IDX(nsize),
 443                             nobjsize);
 444                 }
 445         }
 446         object->un_pager.vnp.vnp_size = nsize;
 447         object->size = nobjsize;
 448         VM_OBJECT_WUNLOCK(object);
 449 }
 450
 451 /*
 452  * calculate the linear (byte) disk address of specified virtual
 453  * file address
 454  */
 455 static int
 456 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
 457     int *run)
 458 {
 459         int bsize;
 460         int err;
 461         daddr_t vblock;
 462         daddr_t voffset;
 463
 464         if (address < 0)
 465                 return -1;
 466
 467         if (vp->v_iflag & VI_DOOMED)
 468                 return -1;
 469
 470         bsize = vp->v_mount->mnt_stat.f_iosize;
 471         vblock = address / bsize;
 472         voffset = address % bsize;
 473
 474         err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
 475         if (err == 0) {
 476                 if (*rtaddress != -1)
 477                         *rtaddress += voffset / DEV_BSIZE;
 478                 if (run) {
 479                         *run += 1;
 480                         *run *= bsize/PAGE_SIZE;
 481                         *run -= voffset/PAGE_SIZE;
 482                 }
 483         }
 484
 485         return (err);
 486 }
 487
 488 /*
 489  * small block filesystem vnode pager input
 490  */
 491 static int
 492 vnode_pager_input_smlfs(object, m)
 493         vm_object_t object;
 494         vm_page_t m;
 495 {
 496         struct vnode *vp;
 497         struct bufobj *bo;
 498         struct buf *bp;
 499         struct sf_buf *sf;
 500         daddr_t fileaddr;
 501         vm_offset_t bsize;
 502         vm_page_bits_t bits;
 503         int error, i;
 504
 505         error = 0;
 506         vp = object->handle;
 507         if (vp->v_iflag & VI_DOOMED)
 508                 return VM_PAGER_BAD;
 509
 510         bsize = vp->v_mount->mnt_stat.f_iosize;
 511
 512         VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
 513
 514         sf = sf_buf_alloc(m, 0);
 515
 516         for (i = 0; i < PAGE_SIZE / bsize; i++) {
 517                 vm_ooffset_t address;
 518
 519                 bits = vm_page_bits(i * bsize, bsize);
 520                 if (m->valid & bits)
 521                         continue;
 522
 523                 address = IDX_TO_OFF(m->pindex) + i * bsize;
 524                 if (address >= object->un_pager.vnp.vnp_size) {
 525                         fileaddr = -1;
 526                 } else {
 527                         error = vnode_pager_addr(vp, address, &fileaddr, NULL);
 528                         if (error)
 529                                 break;
 530                 }
 531                 if (fileaddr != -1) {
 532                         bp = getpbuf(&vnode_pbuf_freecnt);
 533
 534                         /* build a minimal buffer header */
 535                         bp->b_iocmd = BIO_READ;
 536                         bp->b_iodone = bdone;
 537                         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 538                         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 539                         bp->b_rcred = crhold(curthread->td_ucred);
 540                         bp->b_wcred = crhold(curthread->td_ucred);
 541                         bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
 542                         bp->b_blkno = fileaddr;
 543                         pbgetbo(bo, bp);
 544                         bp->b_vp = vp;
 545                         bp->b_bcount = bsize;
 546                         bp->b_bufsize = bsize;
 547                         bp->b_runningbufspace = bp->b_bufsize;
 548                         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 549
 550                         /* do the input */
 551                         bp->b_iooffset = dbtob(bp->b_blkno);
 552                         bstrategy(bp);
 553
 554                         bwait(bp, PVM, "vnsrd");
 555
 556                         if ((bp->b_ioflags & BIO_ERROR) != 0)
 557                                 error = EIO;
 558
 559                         /*
 560                          * free the buffer header back to the swap buffer pool
 561                          */
 562                         bp->b_vp = NULL;
 563                         pbrelbo(bp);
 564                         relpbuf(bp, &vnode_pbuf_freecnt);
 565                         if (error)
 566                                 break;
 567                 } else
 568                         bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
 569                 KASSERT((m->dirty & bits) == 0,
 570                     ("vnode_pager_input_smlfs: page %p is dirty", m));
 571                 VM_OBJECT_WLOCK(object);
 572                 m->valid |= bits;
 573                 VM_OBJECT_WUNLOCK(object);
 574         }
 575         sf_buf_free(sf);
 576         if (error) {
 577                 return VM_PAGER_ERROR;
 578         }
 579         return VM_PAGER_OK;
 580 }
 581
 582 /*
 583  * old style vnode pager input routine
 584  */
 585 static int
 586 vnode_pager_input_old(object, m)
 587         vm_object_t object;
 588         vm_page_t m;
 589 {
 590         struct uio auio;
 591         struct iovec aiov;
 592         int error;
 593         int size;
 594         struct sf_buf *sf;
 595         struct vnode *vp;
 596
 597         VM_OBJECT_ASSERT_WLOCKED(object);
 598         error = 0;
 599
 600         /*
 601          * Return failure if beyond current EOF
 602          */
 603         if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
 604                 return VM_PAGER_BAD;
 605         } else {
 606                 size = PAGE_SIZE;
 607                 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
 608                         size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
 609                 vp = object->handle;
 610                 VM_OBJECT_WUNLOCK(object);
 611
 612                 /*
 613                  * Allocate a kernel virtual address and initialize so that
 614                  * we can use VOP_READ/WRITE routines.
 615                  */
 616                 sf = sf_buf_alloc(m, 0);
 617
 618                 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
 619                 aiov.iov_len = size;
 620                 auio.uio_iov = &aiov;
 621                 auio.uio_iovcnt = 1;
 622                 auio.uio_offset = IDX_TO_OFF(m->pindex);
 623                 auio.uio_segflg = UIO_SYSSPACE;
 624                 auio.uio_rw = UIO_READ;
 625                 auio.uio_resid = size;
 626                 auio.uio_td = curthread;
 627
 628                 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
 629                 if (!error) {
 630                         int count = size - auio.uio_resid;
 631
 632                         if (count == 0)
 633                                 error = EINVAL;
 634                         else if (count != PAGE_SIZE)
 635                                 bzero((caddr_t)sf_buf_kva(sf) + count,
 636                                     PAGE_SIZE - count);
 637                 }
 638                 sf_buf_free(sf);
 639
 640                 VM_OBJECT_WLOCK(object);
 641         }
 642         KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
 643         if (!error)
 644                 m->valid = VM_PAGE_BITS_ALL;
 645         return error ? VM_PAGER_ERROR : VM_PAGER_OK;
 646 }
 647
 648 /*
 649  * generic vnode pager input routine
 650  */
 651
 652 /*
 653  * Local media VFS's that do not implement their own VOP_GETPAGES
 654  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
 655  * to implement the previous behaviour.
 656  *
 657  * All other FS's should use the bypass to get to the local media
 658  * backing vp's VOP_GETPAGES.
 659  */
 660 static int
 661 vnode_pager_getpages(object, m, count, reqpage)
 662         vm_object_t object;
 663         vm_page_t *m;
 664         int count;
 665         int reqpage;
 666 {
 667         int rtval;
 668         struct vnode *vp;
 669         int bytes = count * PAGE_SIZE;
 670
 671         vp = object->handle;
 672         VM_OBJECT_WUNLOCK(object);
 673         rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
 674         KASSERT(rtval != EOPNOTSUPP,
 675             ("vnode_pager: FS getpages not implemented\n"));
 676         VM_OBJECT_WLOCK(object);
 677         return rtval;
 678 }
 679
 680 /*
 681  * This is now called from local media FS's to operate against their
 682  * own vnodes if they fail to implement VOP_GETPAGES.
 683  */
 684 int
 685 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
 686         struct vnode *vp;
 687         vm_page_t *m;
 688         int bytecount;
 689         int reqpage;
 690 {
 691         vm_object_t object;
 692         vm_offset_t kva;
 693         off_t foff, tfoff, nextoff;
 694         int i, j, size, bsize, first;
 695         daddr_t firstaddr, reqblock;
 696         struct bufobj *bo;
 697         int runpg;
 698         int runend;
 699         struct buf *bp;
 700         struct mount *mp;
 701         int count;
 702         int error;
 703
 704         object = vp->v_object;
 705         count = bytecount / PAGE_SIZE;
 706
 707         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 708             ("vnode_pager_generic_getpages does not support devices"));
 709         if (vp->v_iflag & VI_DOOMED)
 710                 return VM_PAGER_BAD;
 711
 712         bsize = vp->v_mount->mnt_stat.f_iosize;
 713
 714         /* get the UNDERLYING device for the file with VOP_BMAP() */
 715
 716         /*
 717          * originally, we did not check for an error return value -- assuming
 718          * an fs always has a bmap entry point -- that assumption is wrong!!!
 719          */
 720         foff = IDX_TO_OFF(m[reqpage]->pindex);
 721
 722         /*
 723          * if we can't bmap, use old VOP code
 724          */
 725         error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL);
 726         if (error == EOPNOTSUPP) {
 727                 VM_OBJECT_WLOCK(object);
 728
 729                 for (i = 0; i < count; i++)
 730                         if (i != reqpage) {
 731                                 vm_page_lock(m[i]);
 732                                 vm_page_free(m[i]);
 733                                 vm_page_unlock(m[i]);
 734                         }
 735                 PCPU_INC(cnt.v_vnodein);
 736                 PCPU_INC(cnt.v_vnodepgsin);
 737                 error = vnode_pager_input_old(object, m[reqpage]);
 738                 VM_OBJECT_WUNLOCK(object);
 739                 return (error);
 740         } else if (error != 0) {
 741                 VM_OBJECT_WLOCK(object);
 742                 for (i = 0; i < count; i++)
 743                         if (i != reqpage) {
 744                                 vm_page_lock(m[i]);
 745                                 vm_page_free(m[i]);
 746                                 vm_page_unlock(m[i]);
 747                         }
 748                 VM_OBJECT_WUNLOCK(object);
 749                 return (VM_PAGER_ERROR);
 750
 751                 /*
 752                  * if the blocksize is smaller than a page size, then use
 753                  * special small filesystem code.  NFS sometimes has a small
 754                  * blocksize, but it can handle large reads itself.
 755                  */
 756         } else if ((PAGE_SIZE / bsize) > 1 &&
 757             (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
 758                 VM_OBJECT_WLOCK(object);
 759                 for (i = 0; i < count; i++)
 760                         if (i != reqpage) {
 761                                 vm_page_lock(m[i]);
 762                                 vm_page_free(m[i]);
 763                                 vm_page_unlock(m[i]);
 764                         }
 765                 VM_OBJECT_WUNLOCK(object);
 766                 PCPU_INC(cnt.v_vnodein);
 767                 PCPU_INC(cnt.v_vnodepgsin);
 768                 return vnode_pager_input_smlfs(object, m[reqpage]);
 769         }
 770
 771         /*
 772          * If we have a completely valid page available to us, we can
 773          * clean up and return.  Otherwise we have to re-read the
 774          * media.
 775          */
 776         VM_OBJECT_WLOCK(object);
 777         if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
 778                 for (i = 0; i < count; i++)
 779                         if (i != reqpage) {
 780                                 vm_page_lock(m[i]);
 781                                 vm_page_free(m[i]);
 782                                 vm_page_unlock(m[i]);
 783                         }
 784                 VM_OBJECT_WUNLOCK(object);
 785                 return VM_PAGER_OK;
 786         } else if (reqblock == -1) {
 787                 pmap_zero_page(m[reqpage]);
 788                 KASSERT(m[reqpage]->dirty == 0,
 789                     ("vnode_pager_generic_getpages: page %p is dirty", m));
 790                 m[reqpage]->valid = VM_PAGE_BITS_ALL;
 791                 for (i = 0; i < count; i++)
 792                         if (i != reqpage) {
 793                                 vm_page_lock(m[i]);
 794                                 vm_page_free(m[i]);
 795                                 vm_page_unlock(m[i]);
 796                         }
 797                 VM_OBJECT_WUNLOCK(object);
 798                 return (VM_PAGER_OK);
 799         }
 800         m[reqpage]->valid = 0;
 801         VM_OBJECT_WUNLOCK(object);
 802
 803         /*
 804          * here on direct device I/O
 805          */
 806         firstaddr = -1;
 807
 808         /*
 809          * calculate the run that includes the required page
 810          */
 811         for (first = 0, i = 0; i < count; i = runend) {
 812                 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr,
 813                     &runpg) != 0) {
 814                         VM_OBJECT_WLOCK(object);
 815                         for (; i < count; i++)
 816                                 if (i != reqpage) {
 817                                         vm_page_lock(m[i]);
 818                                         vm_page_free(m[i]);
 819                                         vm_page_unlock(m[i]);
 820                                 }
 821                         VM_OBJECT_WUNLOCK(object);
 822                         return (VM_PAGER_ERROR);
 823                 }
 824                 if (firstaddr == -1) {
 825                         VM_OBJECT_WLOCK(object);
 826                         if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
 827                                 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
 828                                     (intmax_t)firstaddr, (uintmax_t)(foff >> 32),
 829                                     (uintmax_t)foff,
 830                                     (uintmax_t)
 831                                     (object->un_pager.vnp.vnp_size >> 32),
 832                                     (uintmax_t)object->un_pager.vnp.vnp_size);
 833                         }
 834                         vm_page_lock(m[i]);
 835                         vm_page_free(m[i]);
 836                         vm_page_unlock(m[i]);
 837                         VM_OBJECT_WUNLOCK(object);
 838                         runend = i + 1;
 839                         first = runend;
 840                         continue;
 841                 }
 842                 runend = i + runpg;
 843                 if (runend <= reqpage) {
 844                         VM_OBJECT_WLOCK(object);
 845                         for (j = i; j < runend; j++) {
 846                                 vm_page_lock(m[j]);
 847                                 vm_page_free(m[j]);
 848                                 vm_page_unlock(m[j]);
 849                         }
 850                         VM_OBJECT_WUNLOCK(object);
 851                 } else {
 852                         if (runpg < (count - first)) {
 853                                 VM_OBJECT_WLOCK(object);
 854                                 for (i = first + runpg; i < count; i++) {
 855                                         vm_page_lock(m[i]);
 856                                         vm_page_free(m[i]);
 857                                         vm_page_unlock(m[i]);
 858                                 }
 859                                 VM_OBJECT_WUNLOCK(object);
 860                                 count = first + runpg;
 861                         }
 862                         break;
 863                 }
 864                 first = runend;
 865         }
 866
 867         /*
 868          * the first and last page have been calculated now, move input pages
 869          * to be zero based...
 870          */
 871         if (first != 0) {
 872                 m += first;
 873                 count -= first;
 874                 reqpage -= first;
 875         }
 876
 877         /*
 878          * calculate the file virtual address for the transfer
 879          */
 880         foff = IDX_TO_OFF(m[0]->pindex);
 881
 882         /*
 883          * calculate the size of the transfer
 884          */
 885         size = count * PAGE_SIZE;
 886         KASSERT(count > 0, ("zero count"));
 887         if ((foff + size) > object->un_pager.vnp.vnp_size)
 888                 size = object->un_pager.vnp.vnp_size - foff;
 889         KASSERT(size > 0, ("zero size"));
 890
 891         /*
 892          * round up physical size for real devices.
 893          */
 894         if (1) {
 895                 int secmask = bo->bo_bsize - 1;
 896                 KASSERT(secmask < PAGE_SIZE && secmask > 0,
 897                     ("vnode_pager_generic_getpages: sector size %d too large",
 898                     secmask + 1));
 899                 size = (size + secmask) & ~secmask;
 900         }
 901
 902         bp = getpbuf(&vnode_pbuf_freecnt);
 903         kva = (vm_offset_t)bp->b_data;
 904
 905         /*
 906          * and map the pages to be read into the kva, if the filesystem
 907          * requires mapped buffers.
 908          */
 909         mp = vp->v_mount;
 910         if (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
 911             unmapped_buf_allowed) {
 912                 bp->b_data = unmapped_buf;
 913                 bp->b_kvabase = unmapped_buf;
 914                 bp->b_offset = 0;
 915                 bp->b_flags |= B_UNMAPPED;
 916                 bp->b_npages = count;
 917                 for (i = 0; i < count; i++)
 918                         bp->b_pages[i] = m[i];
 919         } else
 920                 pmap_qenter(kva, m, count);
 921
 922         /* build a minimal buffer header */
 923         bp->b_iocmd = BIO_READ;
 924         bp->b_iodone = bdone;
 925         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 926         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 927         bp->b_rcred = crhold(curthread->td_ucred);
 928         bp->b_wcred = crhold(curthread->td_ucred);
 929         bp->b_blkno = firstaddr;
 930         pbgetbo(bo, bp);
 931         bp->b_vp = vp;
 932         bp->b_bcount = size;
 933         bp->b_bufsize = size;
 934         bp->b_runningbufspace = bp->b_bufsize;
 935         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 936
 937         PCPU_INC(cnt.v_vnodein);
 938         PCPU_ADD(cnt.v_vnodepgsin, count);
 939
 940         /* do the input */
 941         bp->b_iooffset = dbtob(bp->b_blkno);
 942         bstrategy(bp);
 943
 944         bwait(bp, PVM, "vnread");
 945
 946         if ((bp->b_ioflags & BIO_ERROR) != 0)
 947                 error = EIO;
 948
 949         if (error == 0 && size != count * PAGE_SIZE) {
 950                 if ((bp->b_flags & B_UNMAPPED) != 0) {
 951                         bp->b_flags &= ~B_UNMAPPED;
 952                         pmap_qenter(kva, m, count);
 953                 }
 954                 bzero((caddr_t)kva + size, PAGE_SIZE * count - size);
 955         }
 956         if ((bp->b_flags & B_UNMAPPED) == 0)
 957                 pmap_qremove(kva, count);
 958         if (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0) {
 959                 bp->b_data = (caddr_t)kva;
 960                 bp->b_kvabase = (caddr_t)kva;
 961                 bp->b_flags &= ~B_UNMAPPED;
 962                 for (i = 0; i < count; i++)
 963                         bp->b_pages[i] = NULL;
 964         }
 965
 966         /*
 967          * free the buffer header back to the swap buffer pool
 968          */
 969         bp->b_vp = NULL;
 970         pbrelbo(bp);
 971         relpbuf(bp, &vnode_pbuf_freecnt);
 972
 973         VM_OBJECT_WLOCK(object);
 974         for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
 975                 vm_page_t mt;
 976
 977                 nextoff = tfoff + PAGE_SIZE;
 978                 mt = m[i];
 979
 980                 if (nextoff <= object->un_pager.vnp.vnp_size) {
 981                         /*
 982                          * Read filled up entire page.
 983                          */
 984                         mt->valid = VM_PAGE_BITS_ALL;
 985                         KASSERT(mt->dirty == 0,
 986                             ("vnode_pager_generic_getpages: page %p is dirty",
 987                             mt));
 988                         KASSERT(!pmap_page_is_mapped(mt),
 989                             ("vnode_pager_generic_getpages: page %p is mapped",
 990                             mt));
 991                 } else {
 992                         /*
 993                          * Read did not fill up entire page.
 994                          *
 995                          * Currently we do not set the entire page valid,
 996                          * we just try to clear the piece that we couldn't
 997                          * read.
 998                          */
 999                         vm_page_set_valid_range(mt, 0,
1000                             object->un_pager.vnp.vnp_size - tfoff);
1001                         KASSERT((mt->dirty & vm_page_bits(0,
1002                             object->un_pager.vnp.vnp_size - tfoff)) == 0,
1003                             ("vnode_pager_generic_getpages: page %p is dirty",
1004                             mt));
1005                 }
1006
1007                 if (i != reqpage)
1008                         vm_page_readahead_finish(mt);
1009         }
1010         VM_OBJECT_WUNLOCK(object);
1011         if (error) {
1012                 printf("vnode_pager_getpages: I/O read error\n");
1013         }
1014         return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
1015 }
1016
1017 /*
1018  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1019  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1020  * vnode_pager_generic_putpages() to implement the previous behaviour.
1021  *
1022  * All other FS's should use the bypass to get to the local media
1023  * backing vp's VOP_PUTPAGES.
1024  */
1025 static void
1026 vnode_pager_putpages(object, m, count, sync, rtvals)
1027         vm_object_t object;
1028         vm_page_t *m;
1029         int count;
1030         boolean_t sync;
1031         int *rtvals;
1032 {
1033         int rtval;
1034         struct vnode *vp;
1035         int bytes = count * PAGE_SIZE;
1036
1037         /*
1038          * Force synchronous operation if we are extremely low on memory
1039          * to prevent a low-memory deadlock.  VOP operations often need to
1040          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1041          * operation ).  The swapper handles the case by limiting the amount
1042          * of asynchronous I/O, but that sort of solution doesn't scale well
1043          * for the vnode pager without a lot of work.
1044          *
1045          * Also, the backing vnode's iodone routine may not wake the pageout
1046          * daemon up.  This should be probably be addressed XXX.
1047          */
1048
1049         if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
1050                 sync |= OBJPC_SYNC;
1051
1052         /*
1053          * Call device-specific putpages function
1054          */
1055         vp = object->handle;
1056         VM_OBJECT_WUNLOCK(object);
1057         rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
1058         KASSERT(rtval != EOPNOTSUPP,
1059             ("vnode_pager: stale FS putpages\n"));
1060         VM_OBJECT_WLOCK(object);
1061 }
1062
1063
1064 /*
1065  * This is now called from local media FS's to operate against their
1066  * own vnodes if they fail to implement VOP_PUTPAGES.
1067  *
1068  * This is typically called indirectly via the pageout daemon and
1069  * clustering has already typically occured, so in general we ask the
1070  * underlying filesystem to write the data out asynchronously rather
1071  * then delayed.
1072  */
1073 int
1074 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1075     int flags, int *rtvals)
1076 {
1077         int i;
1078         vm_object_t object;
1079         vm_page_t m;
1080         int count;
1081
1082         int maxsize, ncount;
1083         vm_ooffset_t poffset;
1084         struct uio auio;
1085         struct iovec aiov;
1086         int error;
1087         int ioflags;
1088         int ppscheck = 0;
1089         static struct timeval lastfail;
1090         static int curfail;
1091
1092         object = vp->v_object;
1093         count = bytecount / PAGE_SIZE;
1094
1095         for (i = 0; i < count; i++)
1096                 rtvals[i] = VM_PAGER_ERROR;
1097
1098         if ((int64_t)ma[0]->pindex < 0) {
1099                 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1100                     (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1101                 rtvals[0] = VM_PAGER_BAD;
1102                 return VM_PAGER_BAD;
1103         }
1104
1105         maxsize = count * PAGE_SIZE;
1106         ncount = count;
1107
1108         poffset = IDX_TO_OFF(ma[0]->pindex);
1109
1110         /*
1111          * If the page-aligned write is larger then the actual file we
1112          * have to invalidate pages occuring beyond the file EOF.  However,
1113          * there is an edge case where a file may not be page-aligned where
1114          * the last page is partially invalid.  In this case the filesystem
1115          * may not properly clear the dirty bits for the entire page (which
1116          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1117          * With the page locked we are free to fix-up the dirty bits here.
1118          *
1119          * We do not under any circumstances truncate the valid bits, as
1120          * this will screw up bogus page replacement.
1121          */
1122         VM_OBJECT_WLOCK(object);
1123         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1124                 if (object->un_pager.vnp.vnp_size > poffset) {
1125                         int pgoff;
1126
1127                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1128                         ncount = btoc(maxsize);
1129                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1130                                 /*
1131                                  * If the object is locked and the following
1132                                  * conditions hold, then the page's dirty
1133                                  * field cannot be concurrently changed by a
1134                                  * pmap operation.
1135                                  */
1136                                 m = ma[ncount - 1];
1137                                 KASSERT(m->busy > 0,
1138                 ("vnode_pager_generic_putpages: page %p is not busy", m));
1139                                 KASSERT(!pmap_page_is_write_mapped(m),
1140                 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1141                                 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1142                                     pgoff);
1143                         }
1144                 } else {
1145                         maxsize = 0;
1146                         ncount = 0;
1147                 }
1148                 if (ncount < count) {
1149                         for (i = ncount; i < count; i++) {
1150                                 rtvals[i] = VM_PAGER_BAD;
1151                         }
1152                 }
1153         }
1154         VM_OBJECT_WUNLOCK(object);
1155
1156         /*
1157          * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1158          * rather then a bdwrite() to prevent paging I/O from saturating
1159          * the buffer cache.  Dummy-up the sequential heuristic to cause
1160          * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1161          * the system decides how to cluster.
1162          */
1163         ioflags = IO_VMIO;
1164         if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1165                 ioflags |= IO_SYNC;
1166         else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1167                 ioflags |= IO_ASYNC;
1168         ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1169         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1170
1171         aiov.iov_base = (caddr_t) 0;
1172         aiov.iov_len = maxsize;
1173         auio.uio_iov = &aiov;
1174         auio.uio_iovcnt = 1;
1175         auio.uio_offset = poffset;
1176         auio.uio_segflg = UIO_NOCOPY;
1177         auio.uio_rw = UIO_WRITE;
1178         auio.uio_resid = maxsize;
1179         auio.uio_td = (struct thread *) 0;
1180         error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1181         PCPU_INC(cnt.v_vnodeout);
1182         PCPU_ADD(cnt.v_vnodepgsout, ncount);
1183
1184         if (error) {
1185                 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1186                         printf("vnode_pager_putpages: I/O error %d\n", error);
1187         }
1188         if (auio.uio_resid) {
1189                 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1190                         printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1191                             auio.uio_resid, (u_long)ma[0]->pindex);
1192         }
1193         for (i = 0; i < ncount; i++) {
1194                 rtvals[i] = VM_PAGER_OK;
1195         }
1196         return rtvals[0];
1197 }
1198
1199 void
1200 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1201 {
1202         vm_object_t obj;
1203         int i, pos;
1204
1205         if (written == 0)
1206                 return;
1207         obj = ma[0]->object;
1208         VM_OBJECT_WLOCK(obj);
1209         for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1210                 if (pos < trunc_page(written)) {
1211                         rtvals[i] = VM_PAGER_OK;
1212                         vm_page_undirty(ma[i]);
1213                 } else {
1214                         /* Partially written page. */
1215                         rtvals[i] = VM_PAGER_AGAIN;
1216                         vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1217                 }
1218         }
1219         VM_OBJECT_WUNLOCK(obj);
1220 }
1221
1222 void
1223 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1224     vm_offset_t end)
1225 {
1226         struct vnode *vp;
1227         vm_ooffset_t old_wm;
1228
1229         VM_OBJECT_WLOCK(object);
1230         if (object->type != OBJT_VNODE) {
1231                 VM_OBJECT_WUNLOCK(object);
1232                 return;
1233         }
1234         old_wm = object->un_pager.vnp.writemappings;
1235         object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1236         vp = object->handle;
1237         if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1238                 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1239                 VOP_ADD_WRITECOUNT(vp, 1);
1240                 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1241                     __func__, vp, vp->v_writecount);
1242         } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1243                 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1244                 VOP_ADD_WRITECOUNT(vp, -1);
1245                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1246                     __func__, vp, vp->v_writecount);
1247         }
1248         VM_OBJECT_WUNLOCK(object);
1249 }
1250
1251 void
1252 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1253     vm_offset_t end)
1254 {
1255         struct vnode *vp;
1256         struct mount *mp;
1257         vm_offset_t inc;
1258
1259         VM_OBJECT_WLOCK(object);
1260
1261         /*
1262          * First, recheck the object type to account for the race when
1263          * the vnode is reclaimed.
1264          */
1265         if (object->type != OBJT_VNODE) {
1266                 VM_OBJECT_WUNLOCK(object);
1267                 return;
1268         }
1269
1270         /*
1271          * Optimize for the case when writemappings is not going to
1272          * zero.
1273          */
1274         inc = end - start;
1275         if (object->un_pager.vnp.writemappings != inc) {
1276                 object->un_pager.vnp.writemappings -= inc;
1277                 VM_OBJECT_WUNLOCK(object);
1278                 return;
1279         }
1280
1281         vp = object->handle;
1282         vhold(vp);
1283         VM_OBJECT_WUNLOCK(object);
1284         mp = NULL;
1285         vn_start_write(vp, &mp, V_WAIT);
1286         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1287
1288         /*
1289          * Decrement the object's writemappings, by swapping the start
1290          * and end arguments for vnode_pager_update_writecount().  If
1291          * there was not a race with vnode reclaimation, then the
1292          * vnode's v_writecount is decremented.
1293          */
1294         vnode_pager_update_writecount(object, end, start);
1295         VOP_UNLOCK(vp, 0);
1296         vdrop(vp);
1297         if (mp != NULL)
1298                 vn_finished_write(mp);
1299 }