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_getpages(vm_object_t, vm_page_t *, int, int *, int *);
  88 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
  89     int *, vop_getpages_iodone_t, void *);
  90 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
  91 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
  92 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
  93     vm_ooffset_t, struct ucred *cred);
  94 static int vnode_pager_generic_getpages_done(struct buf *);
  95 static void vnode_pager_generic_getpages_done_async(struct buf *);
  96
  97 struct pagerops vnodepagerops = {
  98         .pgo_alloc =    vnode_pager_alloc,
  99         .pgo_dealloc =  vnode_pager_dealloc,
 100         .pgo_getpages = vnode_pager_getpages,
 101         .pgo_getpages_async = vnode_pager_getpages_async,
 102         .pgo_putpages = vnode_pager_putpages,
 103         .pgo_haspage =  vnode_pager_haspage,
 104 };
 105
 106 int vnode_pbuf_freecnt;
 107 int vnode_async_pbuf_freecnt;
 108
 109 /* Create the VM system backing object for this vnode */
 110 int
 111 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
 112 {
 113         vm_object_t object;
 114         vm_ooffset_t size = isize;
 115         struct vattr va;
 116
 117         if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
 118                 return (0);
 119
 120         while ((object = vp->v_object) != NULL) {
 121                 VM_OBJECT_WLOCK(object);
 122                 if (!(object->flags & OBJ_DEAD)) {
 123                         VM_OBJECT_WUNLOCK(object);
 124                         return (0);
 125                 }
 126                 VOP_UNLOCK(vp, 0);
 127                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 128                 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0);
 129                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 130         }
 131
 132         if (size == 0) {
 133                 if (vn_isdisk(vp, NULL)) {
 134                         size = IDX_TO_OFF(INT_MAX);
 135                 } else {
 136                         if (VOP_GETATTR(vp, &va, td->td_ucred))
 137                                 return (0);
 138                         size = va.va_size;
 139                 }
 140         }
 141
 142         object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
 143         /*
 144          * Dereference the reference we just created.  This assumes
 145          * that the object is associated with the vp.
 146          */
 147         VM_OBJECT_WLOCK(object);
 148         object->ref_count--;
 149         VM_OBJECT_WUNLOCK(object);
 150         vrele(vp);
 151
 152         KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
 153
 154         return (0);
 155 }
 156
 157 void
 158 vnode_destroy_vobject(struct vnode *vp)
 159 {
 160         struct vm_object *obj;
 161
 162         obj = vp->v_object;
 163         if (obj == NULL)
 164                 return;
 165         ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
 166         VM_OBJECT_WLOCK(obj);
 167         umtx_shm_object_terminated(obj);
 168         if (obj->ref_count == 0) {
 169                 /*
 170                  * don't double-terminate the object
 171                  */
 172                 if ((obj->flags & OBJ_DEAD) == 0) {
 173                         vm_object_terminate(obj);
 174                 } else {
 175                         /*
 176                          * Waiters were already handled during object
 177                          * termination.  The exclusive vnode lock hopefully
 178                          * prevented new waiters from referencing the dying
 179                          * object.
 180                          */
 181                         KASSERT((obj->flags & OBJ_DISCONNECTWNT) == 0,
 182                             ("OBJ_DISCONNECTWNT set obj %p flags %x",
 183                             obj, obj->flags));
 184                         vp->v_object = NULL;
 185                         VM_OBJECT_WUNLOCK(obj);
 186                 }
 187         } else {
 188                 /*
 189                  * Woe to the process that tries to page now :-).
 190                  */
 191                 vm_pager_deallocate(obj);
 192                 VM_OBJECT_WUNLOCK(obj);
 193         }
 194         KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object));
 195 }
 196
 197
 198 /*
 199  * Allocate (or lookup) pager for a vnode.
 200  * Handle is a vnode pointer.
 201  *
 202  * MPSAFE
 203  */
 204 vm_object_t
 205 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
 206     vm_ooffset_t offset, struct ucred *cred)
 207 {
 208         vm_object_t object;
 209         struct vnode *vp;
 210
 211         /*
 212          * Pageout to vnode, no can do yet.
 213          */
 214         if (handle == NULL)
 215                 return (NULL);
 216
 217         vp = (struct vnode *) handle;
 218
 219         /*
 220          * If the object is being terminated, wait for it to
 221          * go away.
 222          */
 223 retry:
 224         while ((object = vp->v_object) != NULL) {
 225                 VM_OBJECT_WLOCK(object);
 226                 if ((object->flags & OBJ_DEAD) == 0)
 227                         break;
 228                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 229                 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
 230         }
 231
 232         KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
 233
 234         if (object == NULL) {
 235                 /*
 236                  * Add an object of the appropriate size
 237                  */
 238                 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
 239
 240                 object->un_pager.vnp.vnp_size = size;
 241                 object->un_pager.vnp.writemappings = 0;
 242
 243                 object->handle = handle;
 244                 VI_LOCK(vp);
 245                 if (vp->v_object != NULL) {
 246                         /*
 247                          * Object has been created while we were sleeping
 248                          */
 249                         VI_UNLOCK(vp);
 250                         VM_OBJECT_WLOCK(object);
 251                         KASSERT(object->ref_count == 1,
 252                             ("leaked ref %p %d", object, object->ref_count));
 253                         object->type = OBJT_DEAD;
 254                         object->ref_count = 0;
 255                         VM_OBJECT_WUNLOCK(object);
 256                         vm_object_destroy(object);
 257                         goto retry;
 258                 }
 259                 vp->v_object = object;
 260                 VI_UNLOCK(vp);
 261         } else {
 262                 object->ref_count++;
 263 #if VM_NRESERVLEVEL > 0
 264                 vm_object_color(object, 0);
 265 #endif
 266                 VM_OBJECT_WUNLOCK(object);
 267         }
 268         vref(vp);
 269         return (object);
 270 }
 271
 272 /*
 273  *      The object must be locked.
 274  */
 275 static void
 276 vnode_pager_dealloc(vm_object_t object)
 277 {
 278         struct vnode *vp;
 279         int refs;
 280
 281         vp = object->handle;
 282         if (vp == NULL)
 283                 panic("vnode_pager_dealloc: pager already dealloced");
 284
 285         VM_OBJECT_ASSERT_WLOCKED(object);
 286         vm_object_pip_wait(object, "vnpdea");
 287         refs = object->ref_count;
 288
 289         object->handle = NULL;
 290         object->type = OBJT_DEAD;
 291         if (object->flags & OBJ_DISCONNECTWNT) {
 292                 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
 293                 wakeup(object);
 294         }
 295         ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
 296         if (object->un_pager.vnp.writemappings > 0) {
 297                 object->un_pager.vnp.writemappings = 0;
 298                 VOP_ADD_WRITECOUNT(vp, -1);
 299                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
 300                     __func__, vp, vp->v_writecount);
 301         }
 302         vp->v_object = NULL;
 303         VOP_UNSET_TEXT(vp);
 304         VM_OBJECT_WUNLOCK(object);
 305         while (refs-- > 0)
 306                 vunref(vp);
 307         VM_OBJECT_WLOCK(object);
 308 }
 309
 310 static boolean_t
 311 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
 312     int *after)
 313 {
 314         struct vnode *vp = object->handle;
 315         daddr_t bn;
 316         int err;
 317         daddr_t reqblock;
 318         int poff;
 319         int bsize;
 320         int pagesperblock, blocksperpage;
 321
 322         VM_OBJECT_ASSERT_WLOCKED(object);
 323         /*
 324          * If no vp or vp is doomed or marked transparent to VM, we do not
 325          * have the page.
 326          */
 327         if (vp == NULL || vp->v_iflag & VI_DOOMED)
 328                 return FALSE;
 329         /*
 330          * If the offset is beyond end of file we do
 331          * not have the page.
 332          */
 333         if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
 334                 return FALSE;
 335
 336         bsize = vp->v_mount->mnt_stat.f_iosize;
 337         pagesperblock = bsize / PAGE_SIZE;
 338         blocksperpage = 0;
 339         if (pagesperblock > 0) {
 340                 reqblock = pindex / pagesperblock;
 341         } else {
 342                 blocksperpage = (PAGE_SIZE / bsize);
 343                 reqblock = pindex * blocksperpage;
 344         }
 345         VM_OBJECT_WUNLOCK(object);
 346         err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
 347         VM_OBJECT_WLOCK(object);
 348         if (err)
 349                 return TRUE;
 350         if (bn == -1)
 351                 return FALSE;
 352         if (pagesperblock > 0) {
 353                 poff = pindex - (reqblock * pagesperblock);
 354                 if (before) {
 355                         *before *= pagesperblock;
 356                         *before += poff;
 357                 }
 358                 if (after) {
 359                         /*
 360                          * The BMAP vop can report a partial block in the
 361                          * 'after', but must not report blocks after EOF.
 362                          * Assert the latter, and truncate 'after' in case
 363                          * of the former.
 364                          */
 365                         KASSERT((reqblock + *after) * pagesperblock <
 366                             roundup2(object->size, pagesperblock),
 367                             ("%s: reqblock %jd after %d size %ju", __func__,
 368                             (intmax_t )reqblock, *after,
 369                             (uintmax_t )object->size));
 370                         *after *= pagesperblock;
 371                         *after += pagesperblock - (poff + 1);
 372                         if (pindex + *after >= object->size)
 373                                 *after = object->size - 1 - pindex;
 374                 }
 375         } else {
 376                 if (before) {
 377                         *before /= blocksperpage;
 378                 }
 379
 380                 if (after) {
 381                         *after /= blocksperpage;
 382                 }
 383         }
 384         return TRUE;
 385 }
 386
 387 /*
 388  * Lets the VM system know about a change in size for a file.
 389  * We adjust our own internal size and flush any cached pages in
 390  * the associated object that are affected by the size change.
 391  *
 392  * Note: this routine may be invoked as a result of a pager put
 393  * operation (possibly at object termination time), so we must be careful.
 394  */
 395 void
 396 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
 397 {
 398         vm_object_t object;
 399         vm_page_t m;
 400         vm_pindex_t nobjsize;
 401
 402         if ((object = vp->v_object) == NULL)
 403                 return;
 404 /*      ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
 405         VM_OBJECT_WLOCK(object);
 406         if (object->type == OBJT_DEAD) {
 407                 VM_OBJECT_WUNLOCK(object);
 408                 return;
 409         }
 410         KASSERT(object->type == OBJT_VNODE,
 411             ("not vnode-backed object %p", object));
 412         if (nsize == object->un_pager.vnp.vnp_size) {
 413                 /*
 414                  * Hasn't changed size
 415                  */
 416                 VM_OBJECT_WUNLOCK(object);
 417                 return;
 418         }
 419         nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
 420         if (nsize < object->un_pager.vnp.vnp_size) {
 421                 /*
 422                  * File has shrunk. Toss any cached pages beyond the new EOF.
 423                  */
 424                 if (nobjsize < object->size)
 425                         vm_object_page_remove(object, nobjsize, object->size,
 426                             0);
 427                 /*
 428                  * this gets rid of garbage at the end of a page that is now
 429                  * only partially backed by the vnode.
 430                  *
 431                  * XXX for some reason (I don't know yet), if we take a
 432                  * completely invalid page and mark it partially valid
 433                  * it can screw up NFS reads, so we don't allow the case.
 434                  */
 435                 if ((nsize & PAGE_MASK) &&
 436                     (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
 437                     m->valid != 0) {
 438                         int base = (int)nsize & PAGE_MASK;
 439                         int size = PAGE_SIZE - base;
 440
 441                         /*
 442                          * Clear out partial-page garbage in case
 443                          * the page has been mapped.
 444                          */
 445                         pmap_zero_page_area(m, base, size);
 446
 447                         /*
 448                          * Update the valid bits to reflect the blocks that
 449                          * have been zeroed.  Some of these valid bits may
 450                          * have already been set.
 451                          */
 452                         vm_page_set_valid_range(m, base, size);
 453
 454                         /*
 455                          * Round "base" to the next block boundary so that the
 456                          * dirty bit for a partially zeroed block is not
 457                          * cleared.
 458                          */
 459                         base = roundup2(base, DEV_BSIZE);
 460
 461                         /*
 462                          * Clear out partial-page dirty bits.
 463                          *
 464                          * note that we do not clear out the valid
 465                          * bits.  This would prevent bogus_page
 466                          * replacement from working properly.
 467                          */
 468                         vm_page_clear_dirty(m, base, PAGE_SIZE - base);
 469                 }
 470         }
 471         object->un_pager.vnp.vnp_size = nsize;
 472         object->size = nobjsize;
 473         VM_OBJECT_WUNLOCK(object);
 474 }
 475
 476 /*
 477  * calculate the linear (byte) disk address of specified virtual
 478  * file address
 479  */
 480 static int
 481 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
 482     int *run)
 483 {
 484         int bsize;
 485         int err;
 486         daddr_t vblock;
 487         daddr_t voffset;
 488
 489         if (address < 0)
 490                 return -1;
 491
 492         if (vp->v_iflag & VI_DOOMED)
 493                 return -1;
 494
 495         bsize = vp->v_mount->mnt_stat.f_iosize;
 496         vblock = address / bsize;
 497         voffset = address % bsize;
 498
 499         err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
 500         if (err == 0) {
 501                 if (*rtaddress != -1)
 502                         *rtaddress += voffset / DEV_BSIZE;
 503                 if (run) {
 504                         *run += 1;
 505                         *run *= bsize/PAGE_SIZE;
 506                         *run -= voffset/PAGE_SIZE;
 507                 }
 508         }
 509
 510         return (err);
 511 }
 512
 513 /*
 514  * small block filesystem vnode pager input
 515  */
 516 static int
 517 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
 518 {
 519         struct vnode *vp;
 520         struct bufobj *bo;
 521         struct buf *bp;
 522         struct sf_buf *sf;
 523         daddr_t fileaddr;
 524         vm_offset_t bsize;
 525         vm_page_bits_t bits;
 526         int error, i;
 527
 528         error = 0;
 529         vp = object->handle;
 530         if (vp->v_iflag & VI_DOOMED)
 531                 return VM_PAGER_BAD;
 532
 533         bsize = vp->v_mount->mnt_stat.f_iosize;
 534
 535         VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
 536
 537         sf = sf_buf_alloc(m, 0);
 538
 539         for (i = 0; i < PAGE_SIZE / bsize; i++) {
 540                 vm_ooffset_t address;
 541
 542                 bits = vm_page_bits(i * bsize, bsize);
 543                 if (m->valid & bits)
 544                         continue;
 545
 546                 address = IDX_TO_OFF(m->pindex) + i * bsize;
 547                 if (address >= object->un_pager.vnp.vnp_size) {
 548                         fileaddr = -1;
 549                 } else {
 550                         error = vnode_pager_addr(vp, address, &fileaddr, NULL);
 551                         if (error)
 552                                 break;
 553                 }
 554                 if (fileaddr != -1) {
 555                         bp = getpbuf(&vnode_pbuf_freecnt);
 556
 557                         /* build a minimal buffer header */
 558                         bp->b_iocmd = BIO_READ;
 559                         bp->b_iodone = bdone;
 560                         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 561                         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 562                         bp->b_rcred = crhold(curthread->td_ucred);
 563                         bp->b_wcred = crhold(curthread->td_ucred);
 564                         bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
 565                         bp->b_blkno = fileaddr;
 566                         pbgetbo(bo, bp);
 567                         bp->b_vp = vp;
 568                         bp->b_bcount = bsize;
 569                         bp->b_bufsize = bsize;
 570                         bp->b_runningbufspace = bp->b_bufsize;
 571                         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 572
 573                         /* do the input */
 574                         bp->b_iooffset = dbtob(bp->b_blkno);
 575                         bstrategy(bp);
 576
 577                         bwait(bp, PVM, "vnsrd");
 578
 579                         if ((bp->b_ioflags & BIO_ERROR) != 0)
 580                                 error = EIO;
 581
 582                         /*
 583                          * free the buffer header back to the swap buffer pool
 584                          */
 585                         bp->b_vp = NULL;
 586                         pbrelbo(bp);
 587                         relpbuf(bp, &vnode_pbuf_freecnt);
 588                         if (error)
 589                                 break;
 590                 } else
 591                         bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
 592                 KASSERT((m->dirty & bits) == 0,
 593                     ("vnode_pager_input_smlfs: page %p is dirty", m));
 594                 VM_OBJECT_WLOCK(object);
 595                 m->valid |= bits;
 596                 VM_OBJECT_WUNLOCK(object);
 597         }
 598         sf_buf_free(sf);
 599         if (error) {
 600                 return VM_PAGER_ERROR;
 601         }
 602         return VM_PAGER_OK;
 603 }
 604
 605 /*
 606  * old style vnode pager input routine
 607  */
 608 static int
 609 vnode_pager_input_old(vm_object_t object, vm_page_t m)
 610 {
 611         struct uio auio;
 612         struct iovec aiov;
 613         int error;
 614         int size;
 615         struct sf_buf *sf;
 616         struct vnode *vp;
 617
 618         VM_OBJECT_ASSERT_WLOCKED(object);
 619         error = 0;
 620
 621         /*
 622          * Return failure if beyond current EOF
 623          */
 624         if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
 625                 return VM_PAGER_BAD;
 626         } else {
 627                 size = PAGE_SIZE;
 628                 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
 629                         size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
 630                 vp = object->handle;
 631                 VM_OBJECT_WUNLOCK(object);
 632
 633                 /*
 634                  * Allocate a kernel virtual address and initialize so that
 635                  * we can use VOP_READ/WRITE routines.
 636                  */
 637                 sf = sf_buf_alloc(m, 0);
 638
 639                 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
 640                 aiov.iov_len = size;
 641                 auio.uio_iov = &aiov;
 642                 auio.uio_iovcnt = 1;
 643                 auio.uio_offset = IDX_TO_OFF(m->pindex);
 644                 auio.uio_segflg = UIO_SYSSPACE;
 645                 auio.uio_rw = UIO_READ;
 646                 auio.uio_resid = size;
 647                 auio.uio_td = curthread;
 648
 649                 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
 650                 if (!error) {
 651                         int count = size - auio.uio_resid;
 652
 653                         if (count == 0)
 654                                 error = EINVAL;
 655                         else if (count != PAGE_SIZE)
 656                                 bzero((caddr_t)sf_buf_kva(sf) + count,
 657                                     PAGE_SIZE - count);
 658                 }
 659                 sf_buf_free(sf);
 660
 661                 VM_OBJECT_WLOCK(object);
 662         }
 663         KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
 664         if (!error)
 665                 m->valid = VM_PAGE_BITS_ALL;
 666         return error ? VM_PAGER_ERROR : VM_PAGER_OK;
 667 }
 668
 669 /*
 670  * generic vnode pager input routine
 671  */
 672
 673 /*
 674  * Local media VFS's that do not implement their own VOP_GETPAGES
 675  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
 676  * to implement the previous behaviour.
 677  *
 678  * All other FS's should use the bypass to get to the local media
 679  * backing vp's VOP_GETPAGES.
 680  */
 681 static int
 682 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
 683     int *rahead)
 684 {
 685         struct vnode *vp;
 686         int rtval;
 687
 688         vp = object->handle;
 689         VM_OBJECT_WUNLOCK(object);
 690         rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
 691         KASSERT(rtval != EOPNOTSUPP,
 692             ("vnode_pager: FS getpages not implemented\n"));
 693         VM_OBJECT_WLOCK(object);
 694         return rtval;
 695 }
 696
 697 static int
 698 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
 699     int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
 700 {
 701         struct vnode *vp;
 702         int rtval;
 703
 704         vp = object->handle;
 705         VM_OBJECT_WUNLOCK(object);
 706         rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
 707         KASSERT(rtval != EOPNOTSUPP,
 708             ("vnode_pager: FS getpages_async not implemented\n"));
 709         VM_OBJECT_WLOCK(object);
 710         return (rtval);
 711 }
 712
 713 /*
 714  * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
 715  * local filesystems, where partially valid pages can only occur at
 716  * the end of file.
 717  */
 718 int
 719 vnode_pager_local_getpages(struct vop_getpages_args *ap)
 720 {
 721
 722         return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
 723             ap->a_rbehind, ap->a_rahead, NULL, NULL));
 724 }
 725
 726 int
 727 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
 728 {
 729
 730         return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
 731             ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg));
 732 }
 733
 734 /*
 735  * This is now called from local media FS's to operate against their
 736  * own vnodes if they fail to implement VOP_GETPAGES.
 737  */
 738 int
 739 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
 740     int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
 741 {
 742         vm_object_t object;
 743         struct bufobj *bo;
 744         struct buf *bp;
 745         off_t foff;
 746 #ifdef INVARIANTS
 747         off_t blkno0;
 748 #endif
 749         int bsize, pagesperblock, *freecnt;
 750         int error, before, after, rbehind, rahead, poff, i;
 751         int bytecount, secmask;
 752
 753         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 754             ("%s does not support devices", __func__));
 755
 756         if (vp->v_iflag & VI_DOOMED)
 757                 return (VM_PAGER_BAD);
 758
 759         object = vp->v_object;
 760         foff = IDX_TO_OFF(m[0]->pindex);
 761         bsize = vp->v_mount->mnt_stat.f_iosize;
 762         pagesperblock = bsize / PAGE_SIZE;
 763
 764         KASSERT(foff < object->un_pager.vnp.vnp_size,
 765             ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
 766         KASSERT(count <= sizeof(bp->b_pages),
 767             ("%s: requested %d pages", __func__, count));
 768
 769         /*
 770          * The last page has valid blocks.  Invalid part can only
 771          * exist at the end of file, and the page is made fully valid
 772          * by zeroing in vm_pager_get_pages().
 773          */
 774         if (m[count - 1]->valid != 0 && --count == 0) {
 775                 if (iodone != NULL)
 776                         iodone(arg, m, 1, 0);
 777                 return (VM_PAGER_OK);
 778         }
 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, &bp->b_blkno, &after, &before);
 798         if (error == EOPNOTSUPP) {
 799                 relpbuf(bp, freecnt);
 800                 VM_OBJECT_WLOCK(object);
 801                 for (i = 0; i < count; i++) {
 802                         PCPU_INC(cnt.v_vnodein);
 803                         PCPU_INC(cnt.v_vnodepgsin);
 804                         error = vnode_pager_input_old(object, m[i]);
 805                         if (error)
 806                                 break;
 807                 }
 808                 VM_OBJECT_WUNLOCK(object);
 809                 return (error);
 810         } else if (error != 0) {
 811                 relpbuf(bp, freecnt);
 812                 return (VM_PAGER_ERROR);
 813         }
 814
 815         /*
 816          * If the file system supports BMAP, but blocksize is smaller
 817          * than a page size, then use special small filesystem code.
 818          */
 819         if (pagesperblock == 0) {
 820                 relpbuf(bp, freecnt);
 821                 for (i = 0; i < count; i++) {
 822                         PCPU_INC(cnt.v_vnodein);
 823                         PCPU_INC(cnt.v_vnodepgsin);
 824                         error = vnode_pager_input_smlfs(object, m[i]);
 825                         if (error)
 826                                 break;
 827                 }
 828                 return (error);
 829         }
 830
 831         /*
 832          * A sparse file can be encountered only for a single page request,
 833          * which may not be preceded by call to vm_pager_haspage().
 834          */
 835         if (bp->b_blkno == -1) {
 836                 KASSERT(count == 1,
 837                     ("%s: array[%d] request to a sparse file %p", __func__,
 838                     count, vp));
 839                 relpbuf(bp, freecnt);
 840                 pmap_zero_page(m[0]);
 841                 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
 842                     __func__, m[0]));
 843                 VM_OBJECT_WLOCK(object);
 844                 m[0]->valid = VM_PAGE_BITS_ALL;
 845                 VM_OBJECT_WUNLOCK(object);
 846                 return (VM_PAGER_OK);
 847         }
 848
 849 #ifdef INVARIANTS
 850         blkno0 = bp->b_blkno;
 851 #endif
 852         bp->b_blkno += (foff % bsize) / DEV_BSIZE;
 853
 854         /* Recalculate blocks available after/before to pages. */
 855         poff = (foff % bsize) / PAGE_SIZE;
 856         before *= pagesperblock;
 857         before += poff;
 858         after *= pagesperblock;
 859         after += pagesperblock - (poff + 1);
 860         if (m[0]->pindex + after >= object->size)
 861                 after = object->size - 1 - m[0]->pindex;
 862         KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
 863             __func__, count, after + 1));
 864         after -= count - 1;
 865
 866         /* Trim requested rbehind/rahead to possible values. */
 867         rbehind = a_rbehind ? *a_rbehind : 0;
 868         rahead = a_rahead ? *a_rahead : 0;
 869         rbehind = min(rbehind, before);
 870         rbehind = min(rbehind, m[0]->pindex);
 871         rahead = min(rahead, after);
 872         rahead = min(rahead, object->size - m[count - 1]->pindex);
 873         /*
 874          * Check that total amount of pages fit into buf.  Trim rbehind and
 875          * rahead evenly if not.
 876          */
 877         if (rbehind + rahead + count > nitems(bp->b_pages)) {
 878                 int trim, sum;
 879
 880                 trim = rbehind + rahead + count - nitems(bp->b_pages) + 1;
 881                 sum = rbehind + rahead;
 882                 if (rbehind == before) {
 883                         /* Roundup rbehind trim to block size. */
 884                         rbehind -= roundup(trim * rbehind / sum, pagesperblock);
 885                         if (rbehind < 0)
 886                                 rbehind = 0;
 887                 } else
 888                         rbehind -= trim * rbehind / sum;
 889                 rahead -= trim * rahead / sum;
 890         }
 891         KASSERT(rbehind + rahead + count <= nitems(bp->b_pages),
 892             ("%s: behind %d ahead %d count %d", __func__,
 893             rbehind, rahead, count));
 894
 895         /*
 896          * Fill in the bp->b_pages[] array with requested and optional
 897          * read behind or read ahead pages.  Read behind pages are looked
 898          * up in a backward direction, down to a first cached page.  Same
 899          * for read ahead pages, but there is no need to shift the array
 900          * in case of encountering a cached page.
 901          */
 902         i = bp->b_npages = 0;
 903         if (rbehind) {
 904                 vm_pindex_t startpindex, tpindex;
 905                 vm_page_t p;
 906
 907                 VM_OBJECT_WLOCK(object);
 908                 startpindex = m[0]->pindex - rbehind;
 909                 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
 910                     p->pindex >= startpindex)
 911                         startpindex = p->pindex + 1;
 912
 913                 /* tpindex is unsigned; beware of numeric underflow. */
 914                 for (tpindex = m[0]->pindex - 1;
 915                     tpindex >= startpindex && tpindex < m[0]->pindex;
 916                     tpindex--, i++) {
 917                         p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
 918                         if (p == NULL) {
 919                                 /* Shift the array. */
 920                                 for (int j = 0; j < i; j++)
 921                                         bp->b_pages[j] = bp->b_pages[j +
 922                                             tpindex + 1 - startpindex];
 923                                 break;
 924                         }
 925                         bp->b_pages[tpindex - startpindex] = p;
 926                 }
 927
 928                 bp->b_pgbefore = i;
 929                 bp->b_npages += i;
 930                 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
 931         } else
 932                 bp->b_pgbefore = 0;
 933
 934         /* Requested pages. */
 935         for (int j = 0; j < count; j++, i++)
 936                 bp->b_pages[i] = m[j];
 937         bp->b_npages += count;
 938
 939         if (rahead) {
 940                 vm_pindex_t endpindex, tpindex;
 941                 vm_page_t p;
 942
 943                 if (!VM_OBJECT_WOWNED(object))
 944                         VM_OBJECT_WLOCK(object);
 945                 endpindex = m[count - 1]->pindex + rahead + 1;
 946                 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
 947                     p->pindex < endpindex)
 948                         endpindex = p->pindex;
 949                 if (endpindex > object->size)
 950                         endpindex = object->size;
 951
 952                 for (tpindex = m[count - 1]->pindex + 1;
 953                     tpindex < endpindex; i++, tpindex++) {
 954                         p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
 955                         if (p == NULL)
 956                                 break;
 957                         bp->b_pages[i] = p;
 958                 }
 959
 960                 bp->b_pgafter = i - bp->b_npages;
 961                 bp->b_npages = i;
 962         } else
 963                 bp->b_pgafter = 0;
 964
 965         if (VM_OBJECT_WOWNED(object))
 966                 VM_OBJECT_WUNLOCK(object);
 967
 968         /* Report back actual behind/ahead read. */
 969         if (a_rbehind)
 970                 *a_rbehind = bp->b_pgbefore;
 971         if (a_rahead)
 972                 *a_rahead = bp->b_pgafter;
 973
 974 #ifdef INVARIANTS
 975         KASSERT(bp->b_npages <= nitems(bp->b_pages),
 976             ("%s: buf %p overflowed", __func__, bp));
 977         for (int j = 1; j < bp->b_npages; j++)
 978                 KASSERT(bp->b_pages[j]->pindex - 1 ==
 979                     bp->b_pages[j - 1]->pindex,
 980                     ("%s: pages array not consecutive, bp %p", __func__, bp));
 981 #endif
 982
 983         /*
 984          * Recalculate first offset and bytecount with regards to read behind.
 985          * Truncate bytecount to vnode real size and round up physical size
 986          * for real devices.
 987          */
 988         foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
 989         bytecount = bp->b_npages << PAGE_SHIFT;
 990         if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
 991                 bytecount = object->un_pager.vnp.vnp_size - foff;
 992         secmask = bo->bo_bsize - 1;
 993         KASSERT(secmask < PAGE_SIZE && secmask > 0,
 994             ("%s: sector size %d too large", __func__, secmask + 1));
 995         bytecount = (bytecount + secmask) & ~secmask;
 996
 997         /*
 998          * And map the pages to be read into the kva, if the filesystem
 999          * requires mapped buffers.
1000          */
1001         if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
1002             unmapped_buf_allowed) {
1003                 bp->b_data = unmapped_buf;
1004                 bp->b_offset = 0;
1005         } else {
1006                 bp->b_data = bp->b_kvabase;
1007                 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1008         }
1009
1010         /* Build a minimal buffer header. */
1011         bp->b_iocmd = BIO_READ;
1012         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
1013         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
1014         bp->b_rcred = crhold(curthread->td_ucred);
1015         bp->b_wcred = crhold(curthread->td_ucred);
1016         pbgetbo(bo, bp);
1017         bp->b_vp = vp;
1018         bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
1019         bp->b_iooffset = dbtob(bp->b_blkno);
1020         KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) ==
1021             (blkno0 - bp->b_blkno) * DEV_BSIZE +
1022             IDX_TO_OFF(m[0]->pindex) % bsize,
1023             ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju "
1024             "blkno0 %ju b_blkno %ju", bsize,
1025             (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex,
1026             (uintmax_t)blkno0, (uintmax_t)bp->b_blkno));
1027
1028         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
1029         PCPU_INC(cnt.v_vnodein);
1030         PCPU_ADD(cnt.v_vnodepgsin, bp->b_npages);
1031
1032         if (iodone != NULL) { /* async */
1033                 bp->b_pgiodone = iodone;
1034                 bp->b_caller1 = arg;
1035                 bp->b_iodone = vnode_pager_generic_getpages_done_async;
1036                 bp->b_flags |= B_ASYNC;
1037                 BUF_KERNPROC(bp);
1038                 bstrategy(bp);
1039                 return (VM_PAGER_OK);
1040         } else {
1041                 bp->b_iodone = bdone;
1042                 bstrategy(bp);
1043                 bwait(bp, PVM, "vnread");
1044                 error = vnode_pager_generic_getpages_done(bp);
1045                 for (i = 0; i < bp->b_npages; i++)
1046                         bp->b_pages[i] = NULL;
1047                 bp->b_vp = NULL;
1048                 pbrelbo(bp);
1049                 relpbuf(bp, &vnode_pbuf_freecnt);
1050                 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1051         }
1052 }
1053
1054 static void
1055 vnode_pager_generic_getpages_done_async(struct buf *bp)
1056 {
1057         int error;
1058
1059         error = vnode_pager_generic_getpages_done(bp);
1060         /* Run the iodone upon the requested range. */
1061         bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1062             bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1063         for (int i = 0; i < bp->b_npages; i++)
1064                 bp->b_pages[i] = NULL;
1065         bp->b_vp = NULL;
1066         pbrelbo(bp);
1067         relpbuf(bp, &vnode_async_pbuf_freecnt);
1068 }
1069
1070 static int
1071 vnode_pager_generic_getpages_done(struct buf *bp)
1072 {
1073         vm_object_t object;
1074         off_t tfoff, nextoff;
1075         int i, error;
1076
1077         error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1078         object = bp->b_vp->v_object;
1079
1080         if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1081                 if (!buf_mapped(bp)) {
1082                         bp->b_data = bp->b_kvabase;
1083                         pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1084                             bp->b_npages);
1085                 }
1086                 bzero(bp->b_data + bp->b_bcount,
1087                     PAGE_SIZE * bp->b_npages - bp->b_bcount);
1088         }
1089         if (buf_mapped(bp)) {
1090                 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1091                 bp->b_data = unmapped_buf;
1092         }
1093
1094         VM_OBJECT_WLOCK(object);
1095         for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1096             i < bp->b_npages; i++, tfoff = nextoff) {
1097                 vm_page_t mt;
1098
1099                 nextoff = tfoff + PAGE_SIZE;
1100                 mt = bp->b_pages[i];
1101
1102                 if (nextoff <= object->un_pager.vnp.vnp_size) {
1103                         /*
1104                          * Read filled up entire page.
1105                          */
1106                         mt->valid = VM_PAGE_BITS_ALL;
1107                         KASSERT(mt->dirty == 0,
1108                             ("%s: page %p is dirty", __func__, mt));
1109                         KASSERT(!pmap_page_is_mapped(mt),
1110                             ("%s: page %p is mapped", __func__, mt));
1111                 } else {
1112                         /*
1113                          * Read did not fill up entire page.
1114                          *
1115                          * Currently we do not set the entire page valid,
1116                          * we just try to clear the piece that we couldn't
1117                          * read.
1118                          */
1119                         vm_page_set_valid_range(mt, 0,
1120                             object->un_pager.vnp.vnp_size - tfoff);
1121                         KASSERT((mt->dirty & vm_page_bits(0,
1122                             object->un_pager.vnp.vnp_size - tfoff)) == 0,
1123                             ("%s: page %p is dirty", __func__, mt));
1124                 }
1125
1126                 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1127                         vm_page_readahead_finish(mt);
1128         }
1129         VM_OBJECT_WUNLOCK(object);
1130         if (error != 0)
1131                 printf("%s: I/O read error %d\n", __func__, error);
1132
1133         return (error);
1134 }
1135
1136 /*
1137  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1138  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1139  * vnode_pager_generic_putpages() to implement the previous behaviour.
1140  *
1141  * All other FS's should use the bypass to get to the local media
1142  * backing vp's VOP_PUTPAGES.
1143  */
1144 static void
1145 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1146     int flags, int *rtvals)
1147 {
1148         int rtval;
1149         struct vnode *vp;
1150         int bytes = count * PAGE_SIZE;
1151
1152         /*
1153          * Force synchronous operation if we are extremely low on memory
1154          * to prevent a low-memory deadlock.  VOP operations often need to
1155          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1156          * operation ).  The swapper handles the case by limiting the amount
1157          * of asynchronous I/O, but that sort of solution doesn't scale well
1158          * for the vnode pager without a lot of work.
1159          *
1160          * Also, the backing vnode's iodone routine may not wake the pageout
1161          * daemon up.  This should be probably be addressed XXX.
1162          */
1163
1164         if (vm_cnt.v_free_count < vm_cnt.v_pageout_free_min)
1165                 flags |= VM_PAGER_PUT_SYNC;
1166
1167         /*
1168          * Call device-specific putpages function
1169          */
1170         vp = object->handle;
1171         VM_OBJECT_WUNLOCK(object);
1172         rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1173         KASSERT(rtval != EOPNOTSUPP,
1174             ("vnode_pager: stale FS putpages\n"));
1175         VM_OBJECT_WLOCK(object);
1176 }
1177
1178
1179 /*
1180  * This is now called from local media FS's to operate against their
1181  * own vnodes if they fail to implement VOP_PUTPAGES.
1182  *
1183  * This is typically called indirectly via the pageout daemon and
1184  * clustering has already typically occurred, so in general we ask the
1185  * underlying filesystem to write the data out asynchronously rather
1186  * then delayed.
1187  */
1188 int
1189 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1190     int flags, int *rtvals)
1191 {
1192         int i;
1193         vm_object_t object;
1194         vm_page_t m;
1195         int count;
1196
1197         int maxsize, ncount;
1198         vm_ooffset_t poffset;
1199         struct uio auio;
1200         struct iovec aiov;
1201         int error;
1202         int ioflags;
1203         int ppscheck = 0;
1204         static struct timeval lastfail;
1205         static int curfail;
1206
1207         object = vp->v_object;
1208         count = bytecount / PAGE_SIZE;
1209
1210         for (i = 0; i < count; i++)
1211                 rtvals[i] = VM_PAGER_ERROR;
1212
1213         if ((int64_t)ma[0]->pindex < 0) {
1214                 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1215                     (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1216                 rtvals[0] = VM_PAGER_BAD;
1217                 return VM_PAGER_BAD;
1218         }
1219
1220         maxsize = count * PAGE_SIZE;
1221         ncount = count;
1222
1223         poffset = IDX_TO_OFF(ma[0]->pindex);
1224
1225         /*
1226          * If the page-aligned write is larger then the actual file we
1227          * have to invalidate pages occurring beyond the file EOF.  However,
1228          * there is an edge case where a file may not be page-aligned where
1229          * the last page is partially invalid.  In this case the filesystem
1230          * may not properly clear the dirty bits for the entire page (which
1231          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1232          * With the page locked we are free to fix-up the dirty bits here.
1233          *
1234          * We do not under any circumstances truncate the valid bits, as
1235          * this will screw up bogus page replacement.
1236          */
1237         VM_OBJECT_WLOCK(object);
1238         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1239                 if (object->un_pager.vnp.vnp_size > poffset) {
1240                         int pgoff;
1241
1242                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1243                         ncount = btoc(maxsize);
1244                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1245                                 /*
1246                                  * If the object is locked and the following
1247                                  * conditions hold, then the page's dirty
1248                                  * field cannot be concurrently changed by a
1249                                  * pmap operation.
1250                                  */
1251                                 m = ma[ncount - 1];
1252                                 vm_page_assert_sbusied(m);
1253                                 KASSERT(!pmap_page_is_write_mapped(m),
1254                 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1255                                 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1256                                     pgoff);
1257                         }
1258                 } else {
1259                         maxsize = 0;
1260                         ncount = 0;
1261                 }
1262                 if (ncount < count) {
1263                         for (i = ncount; i < count; i++) {
1264                                 rtvals[i] = VM_PAGER_BAD;
1265                         }
1266                 }
1267         }
1268         VM_OBJECT_WUNLOCK(object);
1269
1270         /*
1271          * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1272          * rather then a bdwrite() to prevent paging I/O from saturating
1273          * the buffer cache.  Dummy-up the sequential heuristic to cause
1274          * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1275          * the system decides how to cluster.
1276          */
1277         ioflags = IO_VMIO;
1278         if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1279                 ioflags |= IO_SYNC;
1280         else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1281                 ioflags |= IO_ASYNC;
1282         ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1283         ioflags |= (flags & VM_PAGER_PUT_NOREUSE) ? IO_NOREUSE : 0;
1284         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1285
1286         aiov.iov_base = (caddr_t) 0;
1287         aiov.iov_len = maxsize;
1288         auio.uio_iov = &aiov;
1289         auio.uio_iovcnt = 1;
1290         auio.uio_offset = poffset;
1291         auio.uio_segflg = UIO_NOCOPY;
1292         auio.uio_rw = UIO_WRITE;
1293         auio.uio_resid = maxsize;
1294         auio.uio_td = (struct thread *) 0;
1295         error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1296         PCPU_INC(cnt.v_vnodeout);
1297         PCPU_ADD(cnt.v_vnodepgsout, ncount);
1298
1299         if (error) {
1300                 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1301                         printf("vnode_pager_putpages: I/O error %d\n", error);
1302         }
1303         if (auio.uio_resid) {
1304                 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1305                         printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1306                             auio.uio_resid, (u_long)ma[0]->pindex);
1307         }
1308         for (i = 0; i < ncount; i++) {
1309                 rtvals[i] = VM_PAGER_OK;
1310         }
1311         return rtvals[0];
1312 }
1313
1314 void
1315 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1316 {
1317         vm_object_t obj;
1318         int i, pos;
1319
1320         if (written == 0)
1321                 return;
1322         obj = ma[0]->object;
1323         VM_OBJECT_WLOCK(obj);
1324         for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1325                 if (pos < trunc_page(written)) {
1326                         rtvals[i] = VM_PAGER_OK;
1327                         vm_page_undirty(ma[i]);
1328                 } else {
1329                         /* Partially written page. */
1330                         rtvals[i] = VM_PAGER_AGAIN;
1331                         vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1332                 }
1333         }
1334         VM_OBJECT_WUNLOCK(obj);
1335 }
1336
1337 void
1338 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1339     vm_offset_t end)
1340 {
1341         struct vnode *vp;
1342         vm_ooffset_t old_wm;
1343
1344         VM_OBJECT_WLOCK(object);
1345         if (object->type != OBJT_VNODE) {
1346                 VM_OBJECT_WUNLOCK(object);
1347                 return;
1348         }
1349         old_wm = object->un_pager.vnp.writemappings;
1350         object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1351         vp = object->handle;
1352         if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1353                 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1354                 VOP_ADD_WRITECOUNT(vp, 1);
1355                 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1356                     __func__, vp, vp->v_writecount);
1357         } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1358                 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1359                 VOP_ADD_WRITECOUNT(vp, -1);
1360                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1361                     __func__, vp, vp->v_writecount);
1362         }
1363         VM_OBJECT_WUNLOCK(object);
1364 }
1365
1366 void
1367 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1368     vm_offset_t end)
1369 {
1370         struct vnode *vp;
1371         struct mount *mp;
1372         vm_offset_t inc;
1373
1374         VM_OBJECT_WLOCK(object);
1375
1376         /*
1377          * First, recheck the object type to account for the race when
1378          * the vnode is reclaimed.
1379          */
1380         if (object->type != OBJT_VNODE) {
1381                 VM_OBJECT_WUNLOCK(object);
1382                 return;
1383         }
1384
1385         /*
1386          * Optimize for the case when writemappings is not going to
1387          * zero.
1388          */
1389         inc = end - start;
1390         if (object->un_pager.vnp.writemappings != inc) {
1391                 object->un_pager.vnp.writemappings -= inc;
1392                 VM_OBJECT_WUNLOCK(object);
1393                 return;
1394         }
1395
1396         vp = object->handle;
1397         vhold(vp);
1398         VM_OBJECT_WUNLOCK(object);
1399         mp = NULL;
1400         vn_start_write(vp, &mp, V_WAIT);
1401         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1402
1403         /*
1404          * Decrement the object's writemappings, by swapping the start
1405          * and end arguments for vnode_pager_update_writecount().  If
1406          * there was not a race with vnode reclaimation, then the
1407          * vnode's v_writecount is decremented.
1408          */
1409         vnode_pager_update_writecount(object, end, start);
1410         VOP_UNLOCK(vp, 0);
1411         vdrop(vp);
1412         if (mp != NULL)
1413                 vn_finished_write(mp);
1414 }