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 "opt_vm.h"
  57
  58 #include <sys/param.h>
  59 #include <sys/systm.h>
  60 #include <sys/proc.h>
  61 #include <sys/vnode.h>
  62 #include <sys/mount.h>
  63 #include <sys/bio.h>
  64 #include <sys/buf.h>
  65 #include <sys/vmmeter.h>
  66 #include <sys/limits.h>
  67 #include <sys/conf.h>
  68 #include <sys/rwlock.h>
  69 #include <sys/sf_buf.h>
  70
  71 #include <machine/atomic.h>
  72
  73 #include <vm/vm.h>
  74 #include <vm/vm_param.h>
  75 #include <vm/vm_object.h>
  76 #include <vm/vm_page.h>
  77 #include <vm/vm_pager.h>
  78 #include <vm/vm_map.h>
  79 #include <vm/vnode_pager.h>
  80 #include <vm/vm_extern.h>
  81
  82 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
  83     daddr_t *rtaddress, int *run);
  84 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
  85 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
  86 static void vnode_pager_dealloc(vm_object_t);
  87 static int vnode_pager_local_getpages0(struct vnode *, vm_page_t *, int, int,
  88     vop_getpages_iodone_t, void *);
  89 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
  90 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int,
  91     vop_getpages_iodone_t, void *);
  92 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
  93 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
  94 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
  95     vm_ooffset_t, struct ucred *cred);
  96 static int vnode_pager_generic_getpages_done(struct buf *);
  97 static void vnode_pager_generic_getpages_done_async(struct buf *);
  98
  99 struct pagerops vnodepagerops = {
 100         .pgo_alloc =    vnode_pager_alloc,
 101         .pgo_dealloc =  vnode_pager_dealloc,
 102         .pgo_getpages = vnode_pager_getpages,
 103         .pgo_getpages_async = vnode_pager_getpages_async,
 104         .pgo_putpages = vnode_pager_putpages,
 105         .pgo_haspage =  vnode_pager_haspage,
 106 };
 107
 108 int vnode_pbuf_freecnt;
 109 int vnode_async_pbuf_freecnt;
 110
 111 /* Create the VM system backing object for this vnode */
 112 int
 113 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
 114 {
 115         vm_object_t object;
 116         vm_ooffset_t size = isize;
 117         struct vattr va;
 118
 119         if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
 120                 return (0);
 121
 122         while ((object = vp->v_object) != NULL) {
 123                 VM_OBJECT_WLOCK(object);
 124                 if (!(object->flags & OBJ_DEAD)) {
 125                         VM_OBJECT_WUNLOCK(object);
 126                         return (0);
 127                 }
 128                 VOP_UNLOCK(vp, 0);
 129                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 130                 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0);
 131                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 132         }
 133
 134         if (size == 0) {
 135                 if (vn_isdisk(vp, NULL)) {
 136                         size = IDX_TO_OFF(INT_MAX);
 137                 } else {
 138                         if (VOP_GETATTR(vp, &va, td->td_ucred))
 139                                 return (0);
 140                         size = va.va_size;
 141                 }
 142         }
 143
 144         object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
 145         /*
 146          * Dereference the reference we just created.  This assumes
 147          * that the object is associated with the vp.
 148          */
 149         VM_OBJECT_WLOCK(object);
 150         object->ref_count--;
 151         VM_OBJECT_WUNLOCK(object);
 152         vrele(vp);
 153
 154         KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
 155
 156         return (0);
 157 }
 158
 159 void
 160 vnode_destroy_vobject(struct vnode *vp)
 161 {
 162         struct vm_object *obj;
 163
 164         obj = vp->v_object;
 165         if (obj == NULL)
 166                 return;
 167         ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
 168         VM_OBJECT_WLOCK(obj);
 169         if (obj->ref_count == 0) {
 170                 /*
 171                  * don't double-terminate the object
 172                  */
 173                 if ((obj->flags & OBJ_DEAD) == 0)
 174                         vm_object_terminate(obj);
 175                 else
 176                         VM_OBJECT_WUNLOCK(obj);
 177         } else {
 178                 /*
 179                  * Woe to the process that tries to page now :-).
 180                  */
 181                 vm_pager_deallocate(obj);
 182                 VM_OBJECT_WUNLOCK(obj);
 183         }
 184         vp->v_object = NULL;
 185 }
 186
 187
 188 /*
 189  * Allocate (or lookup) pager for a vnode.
 190  * Handle is a vnode pointer.
 191  *
 192  * MPSAFE
 193  */
 194 vm_object_t
 195 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
 196     vm_ooffset_t offset, struct ucred *cred)
 197 {
 198         vm_object_t object;
 199         struct vnode *vp;
 200
 201         /*
 202          * Pageout to vnode, no can do yet.
 203          */
 204         if (handle == NULL)
 205                 return (NULL);
 206
 207         vp = (struct vnode *) handle;
 208
 209         /*
 210          * If the object is being terminated, wait for it to
 211          * go away.
 212          */
 213 retry:
 214         while ((object = vp->v_object) != NULL) {
 215                 VM_OBJECT_WLOCK(object);
 216                 if ((object->flags & OBJ_DEAD) == 0)
 217                         break;
 218                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 219                 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
 220         }
 221
 222         KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
 223
 224         if (object == NULL) {
 225                 /*
 226                  * Add an object of the appropriate size
 227                  */
 228                 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
 229
 230                 object->un_pager.vnp.vnp_size = size;
 231                 object->un_pager.vnp.writemappings = 0;
 232
 233                 object->handle = handle;
 234                 VI_LOCK(vp);
 235                 if (vp->v_object != NULL) {
 236                         /*
 237                          * Object has been created while we were sleeping
 238                          */
 239                         VI_UNLOCK(vp);
 240                         vm_object_destroy(object);
 241                         goto retry;
 242                 }
 243                 vp->v_object = object;
 244                 VI_UNLOCK(vp);
 245         } else {
 246                 object->ref_count++;
 247 #if VM_NRESERVLEVEL > 0
 248                 vm_object_color(object, 0);
 249 #endif
 250                 VM_OBJECT_WUNLOCK(object);
 251         }
 252         vref(vp);
 253         return (object);
 254 }
 255
 256 /*
 257  *      The object must be locked.
 258  */
 259 static void
 260 vnode_pager_dealloc(vm_object_t object)
 261 {
 262         struct vnode *vp;
 263         int refs;
 264
 265         vp = object->handle;
 266         if (vp == NULL)
 267                 panic("vnode_pager_dealloc: pager already dealloced");
 268
 269         VM_OBJECT_ASSERT_WLOCKED(object);
 270         vm_object_pip_wait(object, "vnpdea");
 271         refs = object->ref_count;
 272
 273         object->handle = NULL;
 274         object->type = OBJT_DEAD;
 275         if (object->flags & OBJ_DISCONNECTWNT) {
 276                 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
 277                 wakeup(object);
 278         }
 279         ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
 280         if (object->un_pager.vnp.writemappings > 0) {
 281                 object->un_pager.vnp.writemappings = 0;
 282                 VOP_ADD_WRITECOUNT(vp, -1);
 283                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
 284                     __func__, vp, vp->v_writecount);
 285         }
 286         vp->v_object = NULL;
 287         VOP_UNSET_TEXT(vp);
 288         VM_OBJECT_WUNLOCK(object);
 289         while (refs-- > 0)
 290                 vunref(vp);
 291         VM_OBJECT_WLOCK(object);
 292 }
 293
 294 static boolean_t
 295 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
 296     int *after)
 297 {
 298         struct vnode *vp = object->handle;
 299         daddr_t bn;
 300         int err;
 301         daddr_t reqblock;
 302         int poff;
 303         int bsize;
 304         int pagesperblock, blocksperpage;
 305
 306         VM_OBJECT_ASSERT_WLOCKED(object);
 307         /*
 308          * If no vp or vp is doomed or marked transparent to VM, we do not
 309          * have the page.
 310          */
 311         if (vp == NULL || vp->v_iflag & VI_DOOMED)
 312                 return FALSE;
 313         /*
 314          * If the offset is beyond end of file we do
 315          * not have the page.
 316          */
 317         if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
 318                 return FALSE;
 319
 320         bsize = vp->v_mount->mnt_stat.f_iosize;
 321         pagesperblock = bsize / PAGE_SIZE;
 322         blocksperpage = 0;
 323         if (pagesperblock > 0) {
 324                 reqblock = pindex / pagesperblock;
 325         } else {
 326                 blocksperpage = (PAGE_SIZE / bsize);
 327                 reqblock = pindex * blocksperpage;
 328         }
 329         VM_OBJECT_WUNLOCK(object);
 330         err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
 331         VM_OBJECT_WLOCK(object);
 332         if (err)
 333                 return TRUE;
 334         if (bn == -1)
 335                 return FALSE;
 336         if (pagesperblock > 0) {
 337                 poff = pindex - (reqblock * pagesperblock);
 338                 if (before) {
 339                         *before *= pagesperblock;
 340                         *before += poff;
 341                 }
 342                 if (after) {
 343                         int numafter;
 344                         *after *= pagesperblock;
 345                         numafter = pagesperblock - (poff + 1);
 346                         if (IDX_TO_OFF(pindex + numafter) >
 347                             object->un_pager.vnp.vnp_size) {
 348                                 numafter =
 349                                     OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
 350                                     pindex;
 351                         }
 352                         *after += numafter;
 353                 }
 354         } else {
 355                 if (before) {
 356                         *before /= blocksperpage;
 357                 }
 358
 359                 if (after) {
 360                         *after /= blocksperpage;
 361                 }
 362         }
 363         return TRUE;
 364 }
 365
 366 /*
 367  * Lets the VM system know about a change in size for a file.
 368  * We adjust our own internal size and flush any cached pages in
 369  * the associated object that are affected by the size change.
 370  *
 371  * Note: this routine may be invoked as a result of a pager put
 372  * operation (possibly at object termination time), so we must be careful.
 373  */
 374 void
 375 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
 376 {
 377         vm_object_t object;
 378         vm_page_t m;
 379         vm_pindex_t nobjsize;
 380
 381         if ((object = vp->v_object) == NULL)
 382                 return;
 383 /*      ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
 384         VM_OBJECT_WLOCK(object);
 385         if (object->type == OBJT_DEAD) {
 386                 VM_OBJECT_WUNLOCK(object);
 387                 return;
 388         }
 389         KASSERT(object->type == OBJT_VNODE,
 390             ("not vnode-backed object %p", object));
 391         if (nsize == object->un_pager.vnp.vnp_size) {
 392                 /*
 393                  * Hasn't changed size
 394                  */
 395                 VM_OBJECT_WUNLOCK(object);
 396                 return;
 397         }
 398         nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
 399         if (nsize < object->un_pager.vnp.vnp_size) {
 400                 /*
 401                  * File has shrunk. Toss any cached pages beyond the new EOF.
 402                  */
 403                 if (nobjsize < object->size)
 404                         vm_object_page_remove(object, nobjsize, object->size,
 405                             0);
 406                 /*
 407                  * this gets rid of garbage at the end of a page that is now
 408                  * only partially backed by the vnode.
 409                  *
 410                  * XXX for some reason (I don't know yet), if we take a
 411                  * completely invalid page and mark it partially valid
 412                  * it can screw up NFS reads, so we don't allow the case.
 413                  */
 414                 if ((nsize & PAGE_MASK) &&
 415                     (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
 416                     m->valid != 0) {
 417                         int base = (int)nsize & PAGE_MASK;
 418                         int size = PAGE_SIZE - base;
 419
 420                         /*
 421                          * Clear out partial-page garbage in case
 422                          * the page has been mapped.
 423                          */
 424                         pmap_zero_page_area(m, base, size);
 425
 426                         /*
 427                          * Update the valid bits to reflect the blocks that
 428                          * have been zeroed.  Some of these valid bits may
 429                          * have already been set.
 430                          */
 431                         vm_page_set_valid_range(m, base, size);
 432
 433                         /*
 434                          * Round "base" to the next block boundary so that the
 435                          * dirty bit for a partially zeroed block is not
 436                          * cleared.
 437                          */
 438                         base = roundup2(base, DEV_BSIZE);
 439
 440                         /*
 441                          * Clear out partial-page dirty bits.
 442                          *
 443                          * note that we do not clear out the valid
 444                          * bits.  This would prevent bogus_page
 445                          * replacement from working properly.
 446                          */
 447                         vm_page_clear_dirty(m, base, PAGE_SIZE - base);
 448                 } else if ((nsize & PAGE_MASK) &&
 449                     vm_page_is_cached(object, OFF_TO_IDX(nsize))) {
 450                         vm_page_cache_free(object, OFF_TO_IDX(nsize),
 451                             nobjsize);
 452                 }
 453         }
 454         object->un_pager.vnp.vnp_size = nsize;
 455         object->size = nobjsize;
 456         VM_OBJECT_WUNLOCK(object);
 457 }
 458
 459 /*
 460  * calculate the linear (byte) disk address of specified virtual
 461  * file address
 462  */
 463 static int
 464 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
 465     int *run)
 466 {
 467         int bsize;
 468         int err;
 469         daddr_t vblock;
 470         daddr_t voffset;
 471
 472         if (address < 0)
 473                 return -1;
 474
 475         if (vp->v_iflag & VI_DOOMED)
 476                 return -1;
 477
 478         bsize = vp->v_mount->mnt_stat.f_iosize;
 479         vblock = address / bsize;
 480         voffset = address % bsize;
 481
 482         err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
 483         if (err == 0) {
 484                 if (*rtaddress != -1)
 485                         *rtaddress += voffset / DEV_BSIZE;
 486                 if (run) {
 487                         *run += 1;
 488                         *run *= bsize/PAGE_SIZE;
 489                         *run -= voffset/PAGE_SIZE;
 490                 }
 491         }
 492
 493         return (err);
 494 }
 495
 496 /*
 497  * small block filesystem vnode pager input
 498  */
 499 static int
 500 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
 501 {
 502         struct vnode *vp;
 503         struct bufobj *bo;
 504         struct buf *bp;
 505         struct sf_buf *sf;
 506         daddr_t fileaddr;
 507         vm_offset_t bsize;
 508         vm_page_bits_t bits;
 509         int error, i;
 510
 511         error = 0;
 512         vp = object->handle;
 513         if (vp->v_iflag & VI_DOOMED)
 514                 return VM_PAGER_BAD;
 515
 516         bsize = vp->v_mount->mnt_stat.f_iosize;
 517
 518         VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
 519
 520         sf = sf_buf_alloc(m, 0);
 521
 522         for (i = 0; i < PAGE_SIZE / bsize; i++) {
 523                 vm_ooffset_t address;
 524
 525                 bits = vm_page_bits(i * bsize, bsize);
 526                 if (m->valid & bits)
 527                         continue;
 528
 529                 address = IDX_TO_OFF(m->pindex) + i * bsize;
 530                 if (address >= object->un_pager.vnp.vnp_size) {
 531                         fileaddr = -1;
 532                 } else {
 533                         error = vnode_pager_addr(vp, address, &fileaddr, NULL);
 534                         if (error)
 535                                 break;
 536                 }
 537                 if (fileaddr != -1) {
 538                         bp = getpbuf(&vnode_pbuf_freecnt);
 539
 540                         /* build a minimal buffer header */
 541                         bp->b_iocmd = BIO_READ;
 542                         bp->b_iodone = bdone;
 543                         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 544                         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 545                         bp->b_rcred = crhold(curthread->td_ucred);
 546                         bp->b_wcred = crhold(curthread->td_ucred);
 547                         bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
 548                         bp->b_blkno = fileaddr;
 549                         pbgetbo(bo, bp);
 550                         bp->b_vp = vp;
 551                         bp->b_bcount = bsize;
 552                         bp->b_bufsize = bsize;
 553                         bp->b_runningbufspace = bp->b_bufsize;
 554                         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 555
 556                         /* do the input */
 557                         bp->b_iooffset = dbtob(bp->b_blkno);
 558                         bstrategy(bp);
 559
 560                         bwait(bp, PVM, "vnsrd");
 561
 562                         if ((bp->b_ioflags & BIO_ERROR) != 0)
 563                                 error = EIO;
 564
 565                         /*
 566                          * free the buffer header back to the swap buffer pool
 567                          */
 568                         bp->b_vp = NULL;
 569                         pbrelbo(bp);
 570                         relpbuf(bp, &vnode_pbuf_freecnt);
 571                         if (error)
 572                                 break;
 573                 } else
 574                         bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
 575                 KASSERT((m->dirty & bits) == 0,
 576                     ("vnode_pager_input_smlfs: page %p is dirty", m));
 577                 VM_OBJECT_WLOCK(object);
 578                 m->valid |= bits;
 579                 VM_OBJECT_WUNLOCK(object);
 580         }
 581         sf_buf_free(sf);
 582         if (error) {
 583                 return VM_PAGER_ERROR;
 584         }
 585         return VM_PAGER_OK;
 586 }
 587
 588 /*
 589  * old style vnode pager input routine
 590  */
 591 static int
 592 vnode_pager_input_old(vm_object_t object, vm_page_t m)
 593 {
 594         struct uio auio;
 595         struct iovec aiov;
 596         int error;
 597         int size;
 598         struct sf_buf *sf;
 599         struct vnode *vp;
 600
 601         VM_OBJECT_ASSERT_WLOCKED(object);
 602         error = 0;
 603
 604         /*
 605          * Return failure if beyond current EOF
 606          */
 607         if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
 608                 return VM_PAGER_BAD;
 609         } else {
 610                 size = PAGE_SIZE;
 611                 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
 612                         size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
 613                 vp = object->handle;
 614                 VM_OBJECT_WUNLOCK(object);
 615
 616                 /*
 617                  * Allocate a kernel virtual address and initialize so that
 618                  * we can use VOP_READ/WRITE routines.
 619                  */
 620                 sf = sf_buf_alloc(m, 0);
 621
 622                 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
 623                 aiov.iov_len = size;
 624                 auio.uio_iov = &aiov;
 625                 auio.uio_iovcnt = 1;
 626                 auio.uio_offset = IDX_TO_OFF(m->pindex);
 627                 auio.uio_segflg = UIO_SYSSPACE;
 628                 auio.uio_rw = UIO_READ;
 629                 auio.uio_resid = size;
 630                 auio.uio_td = curthread;
 631
 632                 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
 633                 if (!error) {
 634                         int count = size - auio.uio_resid;
 635
 636                         if (count == 0)
 637                                 error = EINVAL;
 638                         else if (count != PAGE_SIZE)
 639                                 bzero((caddr_t)sf_buf_kva(sf) + count,
 640                                     PAGE_SIZE - count);
 641                 }
 642                 sf_buf_free(sf);
 643
 644                 VM_OBJECT_WLOCK(object);
 645         }
 646         KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
 647         if (!error)
 648                 m->valid = VM_PAGE_BITS_ALL;
 649         return error ? VM_PAGER_ERROR : VM_PAGER_OK;
 650 }
 651
 652 /*
 653  * generic vnode pager input routine
 654  */
 655
 656 /*
 657  * Local media VFS's that do not implement their own VOP_GETPAGES
 658  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
 659  * to implement the previous behaviour.
 660  *
 661  * All other FS's should use the bypass to get to the local media
 662  * backing vp's VOP_GETPAGES.
 663  */
 664 static int
 665 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, 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);
 674         KASSERT(rtval != EOPNOTSUPP,
 675             ("vnode_pager: FS getpages not implemented\n"));
 676         VM_OBJECT_WLOCK(object);
 677         return rtval;
 678 }
 679
 680 static int
 681 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
 682     int reqpage, vop_getpages_iodone_t iodone, void *arg)
 683 {
 684         struct vnode *vp;
 685         int rtval;
 686
 687         vp = object->handle;
 688         VM_OBJECT_WUNLOCK(object);
 689         rtval = VOP_GETPAGES_ASYNC(vp, m, count * PAGE_SIZE, reqpage,
 690             iodone, arg);
 691         KASSERT(rtval != EOPNOTSUPP,
 692             ("vnode_pager: FS getpages_async not implemented\n"));
 693         VM_OBJECT_WLOCK(object);
 694         return (rtval);
 695 }
 696
 697 /*
 698  * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
 699  * local filesystems, where partially valid pages can only occur at
 700  * the end of file.
 701  */
 702 int
 703 vnode_pager_local_getpages(struct vop_getpages_args *ap)
 704 {
 705
 706         return (vnode_pager_local_getpages0(ap->a_vp, ap->a_m, ap->a_count,
 707             ap->a_reqpage, NULL, NULL));
 708 }
 709
 710 int
 711 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
 712 {
 713
 714         return (vnode_pager_local_getpages0(ap->a_vp, ap->a_m, ap->a_count,
 715             ap->a_reqpage, ap->a_iodone, ap->a_arg));
 716 }
 717
 718 static int
 719 vnode_pager_local_getpages0(struct vnode *vp, vm_page_t *m, int bytecount,
 720     int reqpage, vop_getpages_iodone_t iodone, void *arg)
 721 {
 722         vm_page_t mreq;
 723
 724         mreq = m[reqpage];
 725
 726         /*
 727          * Since the caller has busied the requested page, that page's valid
 728          * field will not be changed by other threads.
 729          */
 730         vm_page_assert_xbusied(mreq);
 731
 732         /*
 733          * The requested page has valid blocks.  Invalid part can only
 734          * exist at the end of file, and the page is made fully valid
 735          * by zeroing in vm_pager_get_pages().  Free non-requested
 736          * pages, since no i/o is done to read its content.
 737          */
 738         if (mreq->valid != 0) {
 739                 vm_pager_free_nonreq(mreq->object, m, reqpage,
 740                     round_page(bytecount) / PAGE_SIZE, FALSE);
 741                 if (iodone != NULL)
 742                         iodone(arg, m, reqpage, 0);
 743                 return (VM_PAGER_OK);
 744         }
 745
 746         return (vnode_pager_generic_getpages(vp, m, bytecount, reqpage,
 747             iodone, arg));
 748 }
 749
 750 /*
 751  * This is now called from local media FS's to operate against their
 752  * own vnodes if they fail to implement VOP_GETPAGES.
 753  */
 754 int
 755 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount,
 756     int reqpage, vop_getpages_iodone_t iodone, void *arg)
 757 {
 758         vm_object_t object;
 759         off_t foff;
 760         int i, j, size, bsize, first, *freecnt;
 761         daddr_t firstaddr, reqblock;
 762         struct bufobj *bo;
 763         int runpg;
 764         int runend;
 765         struct buf *bp;
 766         int count;
 767         int error;
 768
 769         object = vp->v_object;
 770         count = bytecount / PAGE_SIZE;
 771
 772         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 773             ("vnode_pager_generic_getpages does not support devices"));
 774         if (vp->v_iflag & VI_DOOMED)
 775                 return VM_PAGER_BAD;
 776
 777         bsize = vp->v_mount->mnt_stat.f_iosize;
 778         foff = IDX_TO_OFF(m[reqpage]->pindex);
 779
 780         /*
 781          * Synchronous and asynchronous paging operations use different
 782          * free pbuf counters.  This is done to avoid asynchronous requests
 783          * to consume all pbufs.
 784          * Allocate the pbuf at the very beginning of the function, so that
 785          * if we are low on certain kind of pbufs don't even proceed to BMAP,
 786          * but sleep.
 787          */
 788         freecnt = iodone != NULL ?
 789             &vnode_async_pbuf_freecnt : &vnode_pbuf_freecnt;
 790         bp = getpbuf(freecnt);
 791
 792         /*
 793          * Get the underlying device blocks for the file with VOP_BMAP().
 794          * If the file system doesn't support VOP_BMAP, use old way of
 795          * getting pages via VOP_READ.
 796          */
 797         error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL);
 798         if (error == EOPNOTSUPP) {
 799                 relpbuf(bp, freecnt);
 800                 VM_OBJECT_WLOCK(object);
 801                 for (i = 0; i < count; i++)
 802                         if (i != reqpage) {
 803                                 vm_page_lock(m[i]);
 804                                 vm_page_free(m[i]);
 805                                 vm_page_unlock(m[i]);
 806                         }
 807                 PCPU_INC(cnt.v_vnodein);
 808                 PCPU_INC(cnt.v_vnodepgsin);
 809                 error = vnode_pager_input_old(object, m[reqpage]);
 810                 VM_OBJECT_WUNLOCK(object);
 811                 return (error);
 812         } else if (error != 0) {
 813                 relpbuf(bp, freecnt);
 814                 vm_pager_free_nonreq(object, m, reqpage, count, FALSE);
 815                 return (VM_PAGER_ERROR);
 816
 817                 /*
 818                  * if the blocksize is smaller than a page size, then use
 819                  * special small filesystem code.  NFS sometimes has a small
 820                  * blocksize, but it can handle large reads itself.
 821                  */
 822         } else if ((PAGE_SIZE / bsize) > 1 &&
 823             (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
 824                 relpbuf(bp, freecnt);
 825                 vm_pager_free_nonreq(object, m, reqpage, count, FALSE);
 826                 PCPU_INC(cnt.v_vnodein);
 827                 PCPU_INC(cnt.v_vnodepgsin);
 828                 return vnode_pager_input_smlfs(object, m[reqpage]);
 829         }
 830
 831         /*
 832          * Since the caller has busied the requested page, that page's valid
 833          * field will not be changed by other threads.
 834          */
 835         vm_page_assert_xbusied(m[reqpage]);
 836
 837         /*
 838          * If we have a completely valid page available to us, we can
 839          * clean up and return.  Otherwise we have to re-read the
 840          * media.
 841          */
 842         if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
 843                 relpbuf(bp, freecnt);
 844                 vm_pager_free_nonreq(object, m, reqpage, count, FALSE);
 845                 return (VM_PAGER_OK);
 846         } else if (reqblock == -1) {
 847                 relpbuf(bp, freecnt);
 848                 pmap_zero_page(m[reqpage]);
 849                 KASSERT(m[reqpage]->dirty == 0,
 850                     ("vnode_pager_generic_getpages: page %p is dirty", m));
 851                 VM_OBJECT_WLOCK(object);
 852                 m[reqpage]->valid = VM_PAGE_BITS_ALL;
 853                 vm_pager_free_nonreq(object, m, reqpage, count, TRUE);
 854                 VM_OBJECT_WUNLOCK(object);
 855                 return (VM_PAGER_OK);
 856         } else if (m[reqpage]->valid != 0) {
 857                 VM_OBJECT_WLOCK(object);
 858                 m[reqpage]->valid = 0;
 859                 VM_OBJECT_WUNLOCK(object);
 860         }
 861
 862         /*
 863          * here on direct device I/O
 864          */
 865         firstaddr = -1;
 866
 867         /*
 868          * calculate the run that includes the required page
 869          */
 870         for (first = 0, i = 0; i < count; i = runend) {
 871                 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr,
 872                     &runpg) != 0) {
 873                         relpbuf(bp, freecnt);
 874                         /* The requested page may be out of range. */
 875                         vm_pager_free_nonreq(object, m + i, reqpage - i,
 876                             count - i, FALSE);
 877                         return (VM_PAGER_ERROR);
 878                 }
 879                 if (firstaddr == -1) {
 880                         VM_OBJECT_WLOCK(object);
 881                         if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
 882                                 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
 883                                     (intmax_t)firstaddr, (uintmax_t)(foff >> 32),
 884                                     (uintmax_t)foff,
 885                                     (uintmax_t)
 886                                     (object->un_pager.vnp.vnp_size >> 32),
 887                                     (uintmax_t)object->un_pager.vnp.vnp_size);
 888                         }
 889                         vm_page_lock(m[i]);
 890                         vm_page_free(m[i]);
 891                         vm_page_unlock(m[i]);
 892                         VM_OBJECT_WUNLOCK(object);
 893                         runend = i + 1;
 894                         first = runend;
 895                         continue;
 896                 }
 897                 runend = i + runpg;
 898                 if (runend <= reqpage) {
 899                         VM_OBJECT_WLOCK(object);
 900                         for (j = i; j < runend; j++) {
 901                                 vm_page_lock(m[j]);
 902                                 vm_page_free(m[j]);
 903                                 vm_page_unlock(m[j]);
 904                         }
 905                         VM_OBJECT_WUNLOCK(object);
 906                 } else {
 907                         if (runpg < (count - first)) {
 908                                 VM_OBJECT_WLOCK(object);
 909                                 for (i = first + runpg; i < count; i++) {
 910                                         vm_page_lock(m[i]);
 911                                         vm_page_free(m[i]);
 912                                         vm_page_unlock(m[i]);
 913                                 }
 914                                 VM_OBJECT_WUNLOCK(object);
 915                                 count = first + runpg;
 916                         }
 917                         break;
 918                 }
 919                 first = runend;
 920         }
 921
 922         /*
 923          * the first and last page have been calculated now, move input pages
 924          * to be zero based...
 925          */
 926         if (first != 0) {
 927                 m += first;
 928                 count -= first;
 929                 reqpage -= first;
 930         }
 931
 932         /*
 933          * calculate the file virtual address for the transfer
 934          */
 935         foff = IDX_TO_OFF(m[0]->pindex);
 936
 937         /*
 938          * calculate the size of the transfer
 939          */
 940         size = count * PAGE_SIZE;
 941         KASSERT(count > 0, ("zero count"));
 942         if ((foff + size) > object->un_pager.vnp.vnp_size)
 943                 size = object->un_pager.vnp.vnp_size - foff;
 944         KASSERT(size > 0, ("zero size"));
 945
 946         /*
 947          * round up physical size for real devices.
 948          */
 949         if (1) {
 950                 int secmask = bo->bo_bsize - 1;
 951                 KASSERT(secmask < PAGE_SIZE && secmask > 0,
 952                     ("vnode_pager_generic_getpages: sector size %d too large",
 953                     secmask + 1));
 954                 size = (size + secmask) & ~secmask;
 955         }
 956
 957         bp->b_kvaalloc = bp->b_data;
 958
 959         /*
 960          * and map the pages to be read into the kva, if the filesystem
 961          * requires mapped buffers.
 962          */
 963         if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
 964             unmapped_buf_allowed) {
 965                 bp->b_data = unmapped_buf;
 966                 bp->b_kvabase = unmapped_buf;
 967                 bp->b_offset = 0;
 968                 bp->b_flags |= B_UNMAPPED;
 969         } else
 970                 pmap_qenter((vm_offset_t)bp->b_kvaalloc, m, count);
 971
 972         /* build a minimal buffer header */
 973         bp->b_iocmd = BIO_READ;
 974         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 975         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 976         bp->b_rcred = crhold(curthread->td_ucred);
 977         bp->b_wcred = crhold(curthread->td_ucred);
 978         bp->b_blkno = firstaddr;
 979         pbgetbo(bo, bp);
 980         bp->b_vp = vp;
 981         bp->b_bcount = size;
 982         bp->b_bufsize = size;
 983         bp->b_runningbufspace = bp->b_bufsize;
 984         for (i = 0; i < count; i++)
 985                 bp->b_pages[i] = m[i];
 986         bp->b_npages = count;
 987         bp->b_pager.pg_reqpage = reqpage;
 988         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 989
 990         PCPU_INC(cnt.v_vnodein);
 991         PCPU_ADD(cnt.v_vnodepgsin, count);
 992
 993         /* do the input */
 994         bp->b_iooffset = dbtob(bp->b_blkno);
 995
 996         if (iodone != NULL) { /* async */
 997                 bp->b_pager.pg_iodone = iodone;
 998                 bp->b_caller1 = arg;
 999                 bp->b_iodone = vnode_pager_generic_getpages_done_async;
1000                 bp->b_flags |= B_ASYNC;
1001                 BUF_KERNPROC(bp);
1002                 bstrategy(bp);
1003                 /* Good bye! */
1004         } else {
1005                 bp->b_iodone = bdone;
1006                 bstrategy(bp);
1007                 bwait(bp, PVM, "vnread");
1008                 error = vnode_pager_generic_getpages_done(bp);
1009                 for (i = 0; i < bp->b_npages; i++)
1010                         bp->b_pages[i] = NULL;
1011                 bp->b_vp = NULL;
1012                 pbrelbo(bp);
1013                 relpbuf(bp, &vnode_pbuf_freecnt);
1014         }
1015
1016         return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1017 }
1018
1019 static void
1020 vnode_pager_generic_getpages_done_async(struct buf *bp)
1021 {
1022         int error;
1023
1024         error = vnode_pager_generic_getpages_done(bp);
1025         bp->b_pager.pg_iodone(bp->b_caller1, bp->b_pages,
1026           bp->b_pager.pg_reqpage, error);
1027         for (int i = 0; i < bp->b_npages; i++)
1028                 bp->b_pages[i] = NULL;
1029         bp->b_vp = NULL;
1030         pbrelbo(bp);
1031         relpbuf(bp, &vnode_async_pbuf_freecnt);
1032 }
1033
1034 static int
1035 vnode_pager_generic_getpages_done(struct buf *bp)
1036 {
1037         vm_object_t object;
1038         off_t tfoff, nextoff;
1039         int i, error;
1040
1041         error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1042         object = bp->b_vp->v_object;
1043
1044         if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1045                 if ((bp->b_flags & B_UNMAPPED) != 0) {
1046                         bp->b_flags &= ~B_UNMAPPED;
1047                         pmap_qenter((vm_offset_t)bp->b_kvaalloc, bp->b_pages,
1048                             bp->b_npages);
1049                 }
1050                 bzero(bp->b_kvaalloc + bp->b_bcount,
1051                     PAGE_SIZE * bp->b_npages - bp->b_bcount);
1052         }
1053         if ((bp->b_flags & B_UNMAPPED) == 0)
1054                 pmap_qremove((vm_offset_t)bp->b_kvaalloc, bp->b_npages);
1055         if ((bp->b_vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0) {
1056                 bp->b_data = bp->b_kvaalloc;
1057                 bp->b_kvabase = bp->b_kvaalloc;
1058                 bp->b_flags &= ~B_UNMAPPED;
1059         }
1060
1061         VM_OBJECT_WLOCK(object);
1062         for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1063             i < bp->b_npages; i++, tfoff = nextoff) {
1064                 vm_page_t mt;
1065
1066                 nextoff = tfoff + PAGE_SIZE;
1067                 mt = bp->b_pages[i];
1068
1069                 if (nextoff <= object->un_pager.vnp.vnp_size) {
1070                         /*
1071                          * Read filled up entire page.
1072                          */
1073                         mt->valid = VM_PAGE_BITS_ALL;
1074                         KASSERT(mt->dirty == 0,
1075                             ("%s: page %p is dirty", __func__, mt));
1076                         KASSERT(!pmap_page_is_mapped(mt),
1077                             ("%s: page %p is mapped", __func__, mt));
1078                 } else {
1079                         /*
1080                          * Read did not fill up entire page.
1081                          *
1082                          * Currently we do not set the entire page valid,
1083                          * we just try to clear the piece that we couldn't
1084                          * read.
1085                          */
1086                         vm_page_set_valid_range(mt, 0,
1087                             object->un_pager.vnp.vnp_size - tfoff);
1088                         KASSERT((mt->dirty & vm_page_bits(0,
1089                             object->un_pager.vnp.vnp_size - tfoff)) == 0,
1090                             ("%s: page %p is dirty", __func__, mt));
1091                 }
1092
1093                 if (i != bp->b_pager.pg_reqpage)
1094                         vm_page_readahead_finish(mt);
1095         }
1096         VM_OBJECT_WUNLOCK(object);
1097         if (error != 0)
1098                 printf("%s: I/O read error %d\n", __func__, error);
1099
1100         return (error);
1101 }
1102
1103 /*
1104  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1105  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1106  * vnode_pager_generic_putpages() to implement the previous behaviour.
1107  *
1108  * All other FS's should use the bypass to get to the local media
1109  * backing vp's VOP_PUTPAGES.
1110  */
1111 static void
1112 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1113     int flags, int *rtvals)
1114 {
1115         int rtval;
1116         struct vnode *vp;
1117         int bytes = count * PAGE_SIZE;
1118
1119         /*
1120          * Force synchronous operation if we are extremely low on memory
1121          * to prevent a low-memory deadlock.  VOP operations often need to
1122          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1123          * operation ).  The swapper handles the case by limiting the amount
1124          * of asynchronous I/O, but that sort of solution doesn't scale well
1125          * for the vnode pager without a lot of work.
1126          *
1127          * Also, the backing vnode's iodone routine may not wake the pageout
1128          * daemon up.  This should be probably be addressed XXX.
1129          */
1130
1131         if (vm_cnt.v_free_count + vm_cnt.v_cache_count <
1132             vm_cnt.v_pageout_free_min)
1133                 flags |= VM_PAGER_PUT_SYNC;
1134
1135         /*
1136          * Call device-specific putpages function
1137          */
1138         vp = object->handle;
1139         VM_OBJECT_WUNLOCK(object);
1140         rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1141         KASSERT(rtval != EOPNOTSUPP,
1142             ("vnode_pager: stale FS putpages\n"));
1143         VM_OBJECT_WLOCK(object);
1144 }
1145
1146
1147 /*
1148  * This is now called from local media FS's to operate against their
1149  * own vnodes if they fail to implement VOP_PUTPAGES.
1150  *
1151  * This is typically called indirectly via the pageout daemon and
1152  * clustering has already typically occured, so in general we ask the
1153  * underlying filesystem to write the data out asynchronously rather
1154  * then delayed.
1155  */
1156 int
1157 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1158     int flags, int *rtvals)
1159 {
1160         int i;
1161         vm_object_t object;
1162         vm_page_t m;
1163         int count;
1164
1165         int maxsize, ncount;
1166         vm_ooffset_t poffset;
1167         struct uio auio;
1168         struct iovec aiov;
1169         int error;
1170         int ioflags;
1171         int ppscheck = 0;
1172         static struct timeval lastfail;
1173         static int curfail;
1174
1175         object = vp->v_object;
1176         count = bytecount / PAGE_SIZE;
1177
1178         for (i = 0; i < count; i++)
1179                 rtvals[i] = VM_PAGER_ERROR;
1180
1181         if ((int64_t)ma[0]->pindex < 0) {
1182                 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1183                     (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1184                 rtvals[0] = VM_PAGER_BAD;
1185                 return VM_PAGER_BAD;
1186         }
1187
1188         maxsize = count * PAGE_SIZE;
1189         ncount = count;
1190
1191         poffset = IDX_TO_OFF(ma[0]->pindex);
1192
1193         /*
1194          * If the page-aligned write is larger then the actual file we
1195          * have to invalidate pages occuring beyond the file EOF.  However,
1196          * there is an edge case where a file may not be page-aligned where
1197          * the last page is partially invalid.  In this case the filesystem
1198          * may not properly clear the dirty bits for the entire page (which
1199          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1200          * With the page locked we are free to fix-up the dirty bits here.
1201          *
1202          * We do not under any circumstances truncate the valid bits, as
1203          * this will screw up bogus page replacement.
1204          */
1205         VM_OBJECT_WLOCK(object);
1206         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1207                 if (object->un_pager.vnp.vnp_size > poffset) {
1208                         int pgoff;
1209
1210                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1211                         ncount = btoc(maxsize);
1212                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1213                                 /*
1214                                  * If the object is locked and the following
1215                                  * conditions hold, then the page's dirty
1216                                  * field cannot be concurrently changed by a
1217                                  * pmap operation.
1218                                  */
1219                                 m = ma[ncount - 1];
1220                                 vm_page_assert_sbusied(m);
1221                                 KASSERT(!pmap_page_is_write_mapped(m),
1222                 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1223                                 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1224                                     pgoff);
1225                         }
1226                 } else {
1227                         maxsize = 0;
1228                         ncount = 0;
1229                 }
1230                 if (ncount < count) {
1231                         for (i = ncount; i < count; i++) {
1232                                 rtvals[i] = VM_PAGER_BAD;
1233                         }
1234                 }
1235         }
1236         VM_OBJECT_WUNLOCK(object);
1237
1238         /*
1239          * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1240          * rather then a bdwrite() to prevent paging I/O from saturating
1241          * the buffer cache.  Dummy-up the sequential heuristic to cause
1242          * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1243          * the system decides how to cluster.
1244          */
1245         ioflags = IO_VMIO;
1246         if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1247                 ioflags |= IO_SYNC;
1248         else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1249                 ioflags |= IO_ASYNC;
1250         ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1251         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1252
1253         aiov.iov_base = (caddr_t) 0;
1254         aiov.iov_len = maxsize;
1255         auio.uio_iov = &aiov;
1256         auio.uio_iovcnt = 1;
1257         auio.uio_offset = poffset;
1258         auio.uio_segflg = UIO_NOCOPY;
1259         auio.uio_rw = UIO_WRITE;
1260         auio.uio_resid = maxsize;
1261         auio.uio_td = (struct thread *) 0;
1262         error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1263         PCPU_INC(cnt.v_vnodeout);
1264         PCPU_ADD(cnt.v_vnodepgsout, ncount);
1265
1266         if (error) {
1267                 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1268                         printf("vnode_pager_putpages: I/O error %d\n", error);
1269         }
1270         if (auio.uio_resid) {
1271                 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1272                         printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1273                             auio.uio_resid, (u_long)ma[0]->pindex);
1274         }
1275         for (i = 0; i < ncount; i++) {
1276                 rtvals[i] = VM_PAGER_OK;
1277         }
1278         return rtvals[0];
1279 }
1280
1281 void
1282 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1283 {
1284         vm_object_t obj;
1285         int i, pos;
1286
1287         if (written == 0)
1288                 return;
1289         obj = ma[0]->object;
1290         VM_OBJECT_WLOCK(obj);
1291         for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1292                 if (pos < trunc_page(written)) {
1293                         rtvals[i] = VM_PAGER_OK;
1294                         vm_page_undirty(ma[i]);
1295                 } else {
1296                         /* Partially written page. */
1297                         rtvals[i] = VM_PAGER_AGAIN;
1298                         vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1299                 }
1300         }
1301         VM_OBJECT_WUNLOCK(obj);
1302 }
1303
1304 void
1305 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1306     vm_offset_t end)
1307 {
1308         struct vnode *vp;
1309         vm_ooffset_t old_wm;
1310
1311         VM_OBJECT_WLOCK(object);
1312         if (object->type != OBJT_VNODE) {
1313                 VM_OBJECT_WUNLOCK(object);
1314                 return;
1315         }
1316         old_wm = object->un_pager.vnp.writemappings;
1317         object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1318         vp = object->handle;
1319         if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1320                 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1321                 VOP_ADD_WRITECOUNT(vp, 1);
1322                 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1323                     __func__, vp, vp->v_writecount);
1324         } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1325                 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1326                 VOP_ADD_WRITECOUNT(vp, -1);
1327                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1328                     __func__, vp, vp->v_writecount);
1329         }
1330         VM_OBJECT_WUNLOCK(object);
1331 }
1332
1333 void
1334 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1335     vm_offset_t end)
1336 {
1337         struct vnode *vp;
1338         struct mount *mp;
1339         vm_offset_t inc;
1340
1341         VM_OBJECT_WLOCK(object);
1342
1343         /*
1344          * First, recheck the object type to account for the race when
1345          * the vnode is reclaimed.
1346          */
1347         if (object->type != OBJT_VNODE) {
1348                 VM_OBJECT_WUNLOCK(object);
1349                 return;
1350         }
1351
1352         /*
1353          * Optimize for the case when writemappings is not going to
1354          * zero.
1355          */
1356         inc = end - start;
1357         if (object->un_pager.vnp.writemappings != inc) {
1358                 object->un_pager.vnp.writemappings -= inc;
1359                 VM_OBJECT_WUNLOCK(object);
1360                 return;
1361         }
1362
1363         vp = object->handle;
1364         vhold(vp);
1365         VM_OBJECT_WUNLOCK(object);
1366         mp = NULL;
1367         vn_start_write(vp, &mp, V_WAIT);
1368         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1369
1370         /*
1371          * Decrement the object's writemappings, by swapping the start
1372          * and end arguments for vnode_pager_update_writecount().  If
1373          * there was not a race with vnode reclaimation, then the
1374          * vnode's v_writecount is decremented.
1375          */
1376         vnode_pager_update_writecount(object, end, start);
1377         VOP_UNLOCK(vp, 0);
1378         vdrop(vp);
1379         if (mp != NULL)
1380                 vn_finished_write(mp);
1381 }