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         vrefact(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         int bsize, pagesperblock, *freecnt;
 747         int error, before, after, rbehind, rahead, poff, i;
 748         int bytecount, secmask;
 749
 750         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 751             ("%s does not support devices", __func__));
 752
 753         if (vp->v_iflag & VI_DOOMED)
 754                 return (VM_PAGER_BAD);
 755
 756         object = vp->v_object;
 757         foff = IDX_TO_OFF(m[0]->pindex);
 758         bsize = vp->v_mount->mnt_stat.f_iosize;
 759         pagesperblock = bsize / PAGE_SIZE;
 760
 761         KASSERT(foff < object->un_pager.vnp.vnp_size,
 762             ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
 763         KASSERT(count <= sizeof(bp->b_pages),
 764             ("%s: requested %d pages", __func__, count));
 765
 766         /*
 767          * The last page has valid blocks.  Invalid part can only
 768          * exist at the end of file, and the page is made fully valid
 769          * by zeroing in vm_pager_get_pages().
 770          */
 771         if (m[count - 1]->valid != 0 && --count == 0) {
 772                 if (iodone != NULL)
 773                         iodone(arg, m, 1, 0);
 774                 return (VM_PAGER_OK);
 775         }
 776
 777         /*
 778          * Synchronous and asynchronous paging operations use different
 779          * free pbuf counters.  This is done to avoid asynchronous requests
 780          * to consume all pbufs.
 781          * Allocate the pbuf at the very beginning of the function, so that
 782          * if we are low on certain kind of pbufs don't even proceed to BMAP,
 783          * but sleep.
 784          */
 785         freecnt = iodone != NULL ?
 786             &vnode_async_pbuf_freecnt : &vnode_pbuf_freecnt;
 787         bp = getpbuf(freecnt);
 788
 789         /*
 790          * Get the underlying device blocks for the file with VOP_BMAP().
 791          * If the file system doesn't support VOP_BMAP, use old way of
 792          * getting pages via VOP_READ.
 793          */
 794         error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
 795         if (error == EOPNOTSUPP) {
 796                 relpbuf(bp, freecnt);
 797                 VM_OBJECT_WLOCK(object);
 798                 for (i = 0; i < count; i++) {
 799                         PCPU_INC(cnt.v_vnodein);
 800                         PCPU_INC(cnt.v_vnodepgsin);
 801                         error = vnode_pager_input_old(object, m[i]);
 802                         if (error)
 803                                 break;
 804                 }
 805                 VM_OBJECT_WUNLOCK(object);
 806                 return (error);
 807         } else if (error != 0) {
 808                 relpbuf(bp, freecnt);
 809                 return (VM_PAGER_ERROR);
 810         }
 811
 812         /*
 813          * If the file system supports BMAP, but blocksize is smaller
 814          * than a page size, then use special small filesystem code.
 815          */
 816         if (pagesperblock == 0) {
 817                 relpbuf(bp, freecnt);
 818                 for (i = 0; i < count; i++) {
 819                         PCPU_INC(cnt.v_vnodein);
 820                         PCPU_INC(cnt.v_vnodepgsin);
 821                         error = vnode_pager_input_smlfs(object, m[i]);
 822                         if (error)
 823                                 break;
 824                 }
 825                 return (error);
 826         }
 827
 828         /*
 829          * A sparse file can be encountered only for a single page request,
 830          * which may not be preceded by call to vm_pager_haspage().
 831          */
 832         if (bp->b_blkno == -1) {
 833                 KASSERT(count == 1,
 834                     ("%s: array[%d] request to a sparse file %p", __func__,
 835                     count, vp));
 836                 relpbuf(bp, freecnt);
 837                 pmap_zero_page(m[0]);
 838                 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
 839                     __func__, m[0]));
 840                 VM_OBJECT_WLOCK(object);
 841                 m[0]->valid = VM_PAGE_BITS_ALL;
 842                 VM_OBJECT_WUNLOCK(object);
 843                 return (VM_PAGER_OK);
 844         }
 845
 846         bp->b_blkno += (foff % bsize) / DEV_BSIZE;
 847
 848         /* Recalculate blocks available after/before to pages. */
 849         poff = (foff % bsize) / PAGE_SIZE;
 850         before *= pagesperblock;
 851         before += poff;
 852         after *= pagesperblock;
 853         after += pagesperblock - (poff + 1);
 854         if (m[0]->pindex + after >= object->size)
 855                 after = object->size - 1 - m[0]->pindex;
 856         KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
 857             __func__, count, after + 1));
 858         after -= count - 1;
 859
 860         /* Trim requested rbehind/rahead to possible values. */
 861         rbehind = a_rbehind ? *a_rbehind : 0;
 862         rahead = a_rahead ? *a_rahead : 0;
 863         rbehind = min(rbehind, before);
 864         rbehind = min(rbehind, m[0]->pindex);
 865         rahead = min(rahead, after);
 866         rahead = min(rahead, object->size - m[count - 1]->pindex);
 867         KASSERT(rbehind + rahead + count <= sizeof(bp->b_pages),
 868             ("%s: behind %d ahead %d count %d", __func__,
 869             rbehind, rahead, count));
 870
 871         /*
 872          * Fill in the bp->b_pages[] array with requested and optional
 873          * read behind or read ahead pages.  Read behind pages are looked
 874          * up in a backward direction, down to a first cached page.  Same
 875          * for read ahead pages, but there is no need to shift the array
 876          * in case of encountering a cached page.
 877          */
 878         i = bp->b_npages = 0;
 879         if (rbehind) {
 880                 vm_pindex_t startpindex, tpindex;
 881                 vm_page_t p;
 882
 883                 VM_OBJECT_WLOCK(object);
 884                 startpindex = m[0]->pindex - rbehind;
 885                 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
 886                     p->pindex >= startpindex)
 887                         startpindex = p->pindex + 1;
 888
 889                 /* tpindex is unsigned; beware of numeric underflow. */
 890                 for (tpindex = m[0]->pindex - 1;
 891                     tpindex >= startpindex && tpindex < m[0]->pindex;
 892                     tpindex--, i++) {
 893                         p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
 894                         if (p == NULL) {
 895                                 /* Shift the array. */
 896                                 for (int j = 0; j < i; j++)
 897                                         bp->b_pages[j] = bp->b_pages[j +
 898                                             tpindex + 1 - startpindex];
 899                                 break;
 900                         }
 901                         bp->b_pages[tpindex - startpindex] = p;
 902                 }
 903
 904                 bp->b_pgbefore = i;
 905                 bp->b_npages += i;
 906                 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
 907         } else
 908                 bp->b_pgbefore = 0;
 909
 910         /* Requested pages. */
 911         for (int j = 0; j < count; j++, i++)
 912                 bp->b_pages[i] = m[j];
 913         bp->b_npages += count;
 914
 915         if (rahead) {
 916                 vm_pindex_t endpindex, tpindex;
 917                 vm_page_t p;
 918
 919                 if (!VM_OBJECT_WOWNED(object))
 920                         VM_OBJECT_WLOCK(object);
 921                 endpindex = m[count - 1]->pindex + rahead + 1;
 922                 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
 923                     p->pindex < endpindex)
 924                         endpindex = p->pindex;
 925                 if (endpindex > object->size)
 926                         endpindex = object->size;
 927
 928                 for (tpindex = m[count - 1]->pindex + 1;
 929                     tpindex < endpindex; i++, tpindex++) {
 930                         p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
 931                         if (p == NULL)
 932                                 break;
 933                         bp->b_pages[i] = p;
 934                 }
 935
 936                 bp->b_pgafter = i - bp->b_npages;
 937                 bp->b_npages = i;
 938         } else
 939                 bp->b_pgafter = 0;
 940
 941         if (VM_OBJECT_WOWNED(object))
 942                 VM_OBJECT_WUNLOCK(object);
 943
 944         /* Report back actual behind/ahead read. */
 945         if (a_rbehind)
 946                 *a_rbehind = bp->b_pgbefore;
 947         if (a_rahead)
 948                 *a_rahead = bp->b_pgafter;
 949
 950         KASSERT(bp->b_npages <= sizeof(bp->b_pages),
 951             ("%s: buf %p overflowed", __func__, bp));
 952
 953         /*
 954          * Recalculate first offset and bytecount with regards to read behind.
 955          * Truncate bytecount to vnode real size and round up physical size
 956          * for real devices.
 957          */
 958         foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
 959         bytecount = bp->b_npages << PAGE_SHIFT;
 960         if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
 961                 bytecount = object->un_pager.vnp.vnp_size - foff;
 962         secmask = bo->bo_bsize - 1;
 963         KASSERT(secmask < PAGE_SIZE && secmask > 0,
 964             ("%s: sector size %d too large", __func__, secmask + 1));
 965         bytecount = (bytecount + secmask) & ~secmask;
 966
 967         /*
 968          * And map the pages to be read into the kva, if the filesystem
 969          * requires mapped buffers.
 970          */
 971         if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
 972             unmapped_buf_allowed) {
 973                 bp->b_data = unmapped_buf;
 974                 bp->b_offset = 0;
 975         } else {
 976                 bp->b_data = bp->b_kvabase;
 977                 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
 978         }
 979
 980         /* Build a minimal buffer header. */
 981         bp->b_iocmd = BIO_READ;
 982         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 983         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 984         bp->b_rcred = crhold(curthread->td_ucred);
 985         bp->b_wcred = crhold(curthread->td_ucred);
 986         pbgetbo(bo, bp);
 987         bp->b_vp = vp;
 988         bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
 989         bp->b_iooffset = dbtob(bp->b_blkno);
 990
 991         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 992         PCPU_INC(cnt.v_vnodein);
 993         PCPU_ADD(cnt.v_vnodepgsin, bp->b_npages);
 994
 995         if (iodone != NULL) { /* async */
 996                 bp->b_pgiodone = iodone;
 997                 bp->b_caller1 = arg;
 998                 bp->b_iodone = vnode_pager_generic_getpages_done_async;
 999                 bp->b_flags |= B_ASYNC;
1000                 BUF_KERNPROC(bp);
1001                 bstrategy(bp);
1002                 return (VM_PAGER_OK);
1003         } else {
1004                 bp->b_iodone = bdone;
1005                 bstrategy(bp);
1006                 bwait(bp, PVM, "vnread");
1007                 error = vnode_pager_generic_getpages_done(bp);
1008                 for (i = 0; i < bp->b_npages; i++)
1009                         bp->b_pages[i] = NULL;
1010                 bp->b_vp = NULL;
1011                 pbrelbo(bp);
1012                 relpbuf(bp, &vnode_pbuf_freecnt);
1013                 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1014         }
1015 }
1016
1017 static void
1018 vnode_pager_generic_getpages_done_async(struct buf *bp)
1019 {
1020         int error;
1021
1022         error = vnode_pager_generic_getpages_done(bp);
1023         /* Run the iodone upon the requested range. */
1024         bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1025             bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1026         for (int i = 0; i < bp->b_npages; i++)
1027                 bp->b_pages[i] = NULL;
1028         bp->b_vp = NULL;
1029         pbrelbo(bp);
1030         relpbuf(bp, &vnode_async_pbuf_freecnt);
1031 }
1032
1033 static int
1034 vnode_pager_generic_getpages_done(struct buf *bp)
1035 {
1036         vm_object_t object;
1037         off_t tfoff, nextoff;
1038         int i, error;
1039
1040         error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1041         object = bp->b_vp->v_object;
1042
1043         if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1044                 if (!buf_mapped(bp)) {
1045                         bp->b_data = bp->b_kvabase;
1046                         pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1047                             bp->b_npages);
1048                 }
1049                 bzero(bp->b_data + bp->b_bcount,
1050                     PAGE_SIZE * bp->b_npages - bp->b_bcount);
1051         }
1052         if (buf_mapped(bp)) {
1053                 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1054                 bp->b_data = unmapped_buf;
1055         }
1056
1057         VM_OBJECT_WLOCK(object);
1058         for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1059             i < bp->b_npages; i++, tfoff = nextoff) {
1060                 vm_page_t mt;
1061
1062                 nextoff = tfoff + PAGE_SIZE;
1063                 mt = bp->b_pages[i];
1064
1065                 if (nextoff <= object->un_pager.vnp.vnp_size) {
1066                         /*
1067                          * Read filled up entire page.
1068                          */
1069                         mt->valid = VM_PAGE_BITS_ALL;
1070                         KASSERT(mt->dirty == 0,
1071                             ("%s: page %p is dirty", __func__, mt));
1072                         KASSERT(!pmap_page_is_mapped(mt),
1073                             ("%s: page %p is mapped", __func__, mt));
1074                 } else {
1075                         /*
1076                          * Read did not fill up entire page.
1077                          *
1078                          * Currently we do not set the entire page valid,
1079                          * we just try to clear the piece that we couldn't
1080                          * read.
1081                          */
1082                         vm_page_set_valid_range(mt, 0,
1083                             object->un_pager.vnp.vnp_size - tfoff);
1084                         KASSERT((mt->dirty & vm_page_bits(0,
1085                             object->un_pager.vnp.vnp_size - tfoff)) == 0,
1086                             ("%s: page %p is dirty", __func__, mt));
1087                 }
1088
1089                 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1090                         vm_page_readahead_finish(mt);
1091         }
1092         VM_OBJECT_WUNLOCK(object);
1093         if (error != 0)
1094                 printf("%s: I/O read error %d\n", __func__, error);
1095
1096         return (error);
1097 }
1098
1099 /*
1100  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1101  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1102  * vnode_pager_generic_putpages() to implement the previous behaviour.
1103  *
1104  * All other FS's should use the bypass to get to the local media
1105  * backing vp's VOP_PUTPAGES.
1106  */
1107 static void
1108 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1109     int flags, int *rtvals)
1110 {
1111         int rtval;
1112         struct vnode *vp;
1113         int bytes = count * PAGE_SIZE;
1114
1115         /*
1116          * Force synchronous operation if we are extremely low on memory
1117          * to prevent a low-memory deadlock.  VOP operations often need to
1118          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1119          * operation ).  The swapper handles the case by limiting the amount
1120          * of asynchronous I/O, but that sort of solution doesn't scale well
1121          * for the vnode pager without a lot of work.
1122          *
1123          * Also, the backing vnode's iodone routine may not wake the pageout
1124          * daemon up.  This should be probably be addressed XXX.
1125          */
1126
1127         if (vm_cnt.v_free_count < vm_cnt.v_pageout_free_min)
1128                 flags |= VM_PAGER_PUT_SYNC;
1129
1130         /*
1131          * Call device-specific putpages function
1132          */
1133         vp = object->handle;
1134         VM_OBJECT_WUNLOCK(object);
1135         rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1136         KASSERT(rtval != EOPNOTSUPP,
1137             ("vnode_pager: stale FS putpages\n"));
1138         VM_OBJECT_WLOCK(object);
1139 }
1140
1141 static int
1142 vn_off2bidx(vm_ooffset_t offset)
1143 {
1144
1145         return ((offset & PAGE_MASK) / DEV_BSIZE);
1146 }
1147
1148 static bool
1149 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset)
1150 {
1151
1152         KASSERT(IDX_TO_OFF(m->pindex) <= offset &&
1153             offset < IDX_TO_OFF(m->pindex + 1),
1154             ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex,
1155             (uintmax_t)offset));
1156         return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0);
1157 }
1158
1159 /*
1160  * This is now called from local media FS's to operate against their
1161  * own vnodes if they fail to implement VOP_PUTPAGES.
1162  *
1163  * This is typically called indirectly via the pageout daemon and
1164  * clustering has already typically occurred, so in general we ask the
1165  * underlying filesystem to write the data out asynchronously rather
1166  * then delayed.
1167  */
1168 int
1169 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1170     int flags, int *rtvals)
1171 {
1172         vm_object_t object;
1173         vm_page_t m;
1174         vm_ooffset_t maxblksz, next_offset, poffset, prev_offset;
1175         struct uio auio;
1176         struct iovec aiov;
1177         off_t prev_resid, wrsz;
1178         int count, error, i, maxsize, ncount, pgoff, ppscheck;
1179         bool in_hole;
1180         static struct timeval lastfail;
1181         static int curfail;
1182
1183         object = vp->v_object;
1184         count = bytecount / PAGE_SIZE;
1185
1186         for (i = 0; i < count; i++)
1187                 rtvals[i] = VM_PAGER_ERROR;
1188
1189         if ((int64_t)ma[0]->pindex < 0) {
1190                 printf("vnode_pager_generic_putpages: "
1191                     "attempt to write meta-data 0x%jx(%lx)\n",
1192                     (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty);
1193                 rtvals[0] = VM_PAGER_BAD;
1194                 return (VM_PAGER_BAD);
1195         }
1196
1197         maxsize = count * PAGE_SIZE;
1198         ncount = count;
1199
1200         poffset = IDX_TO_OFF(ma[0]->pindex);
1201
1202         /*
1203          * If the page-aligned write is larger then the actual file we
1204          * have to invalidate pages occurring beyond the file EOF.  However,
1205          * there is an edge case where a file may not be page-aligned where
1206          * the last page is partially invalid.  In this case the filesystem
1207          * may not properly clear the dirty bits for the entire page (which
1208          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1209          * With the page locked we are free to fix-up the dirty bits here.
1210          *
1211          * We do not under any circumstances truncate the valid bits, as
1212          * this will screw up bogus page replacement.
1213          */
1214         VM_OBJECT_RLOCK(object);
1215         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1216                 if (!VM_OBJECT_TRYUPGRADE(object)) {
1217                         VM_OBJECT_RUNLOCK(object);
1218                         VM_OBJECT_WLOCK(object);
1219                         if (maxsize + poffset <= object->un_pager.vnp.vnp_size)
1220                                 goto downgrade;
1221                 }
1222                 if (object->un_pager.vnp.vnp_size > poffset) {
1223                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1224                         ncount = btoc(maxsize);
1225                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1226                                 pgoff = roundup2(pgoff, DEV_BSIZE);
1227
1228                                 /*
1229                                  * If the object is locked and the following
1230                                  * conditions hold, then the page's dirty
1231                                  * field cannot be concurrently changed by a
1232                                  * pmap operation.
1233                                  */
1234                                 m = ma[ncount - 1];
1235                                 vm_page_assert_sbusied(m);
1236                                 KASSERT(!pmap_page_is_write_mapped(m),
1237                 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1238                                 MPASS(m->dirty != 0);
1239                                 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1240                                     pgoff);
1241                         }
1242                 } else {
1243                         maxsize = 0;
1244                         ncount = 0;
1245                 }
1246                 for (i = ncount; i < count; i++)
1247                         rtvals[i] = VM_PAGER_BAD;
1248 downgrade:
1249                 VM_OBJECT_LOCK_DOWNGRADE(object);
1250         }
1251
1252         auio.uio_iov = &aiov;
1253         auio.uio_segflg = UIO_NOCOPY;
1254         auio.uio_rw = UIO_WRITE;
1255         auio.uio_td = NULL;
1256         maxblksz = roundup2(poffset + maxsize, DEV_BSIZE);
1257
1258         for (prev_offset = poffset; prev_offset < maxblksz;) {
1259                 /* Skip clean blocks. */
1260                 for (in_hole = true; in_hole && prev_offset < maxblksz;) {
1261                         m = ma[OFF_TO_IDX(prev_offset - poffset)];
1262                         for (i = vn_off2bidx(prev_offset);
1263                             i < sizeof(vm_page_bits_t) * NBBY &&
1264                             prev_offset < maxblksz; i++) {
1265                                 if (vn_dirty_blk(m, prev_offset)) {
1266                                         in_hole = false;
1267                                         break;
1268                                 }
1269                                 prev_offset += DEV_BSIZE;
1270                         }
1271                 }
1272                 if (in_hole)
1273                         goto write_done;
1274
1275                 /* Find longest run of dirty blocks. */
1276                 for (next_offset = prev_offset; next_offset < maxblksz;) {
1277                         m = ma[OFF_TO_IDX(next_offset - poffset)];
1278                         for (i = vn_off2bidx(next_offset);
1279                             i < sizeof(vm_page_bits_t) * NBBY &&
1280                             next_offset < maxblksz; i++) {
1281                                 if (!vn_dirty_blk(m, next_offset))
1282                                         goto start_write;
1283                                 next_offset += DEV_BSIZE;
1284                         }
1285                 }
1286 start_write:
1287                 if (next_offset > poffset + maxsize)
1288                         next_offset = poffset + maxsize;
1289
1290                 /*
1291                  * Getting here requires finding a dirty block in the
1292                  * 'skip clean blocks' loop.
1293                  */
1294                 MPASS(prev_offset < next_offset);
1295
1296                 VM_OBJECT_RUNLOCK(object);
1297                 aiov.iov_base = NULL;
1298                 auio.uio_iovcnt = 1;
1299                 auio.uio_offset = prev_offset;
1300                 prev_resid = auio.uio_resid = aiov.iov_len = next_offset -
1301                     prev_offset;
1302                 error = VOP_WRITE(vp, &auio,
1303                     vnode_pager_putpages_ioflags(flags), curthread->td_ucred);
1304
1305                 wrsz = prev_resid - auio.uio_resid;
1306                 if (wrsz == 0) {
1307                         if (ppsratecheck(&lastfail, &curfail, 1) != 0) {
1308                                 vn_printf(vp, "vnode_pager_putpages: "
1309                                     "zero-length write at %ju resid %zd\n",
1310                                     auio.uio_offset, auio.uio_resid);
1311                         }
1312                         VM_OBJECT_RLOCK(object);
1313                         break;
1314                 }
1315
1316                 /* Adjust the starting offset for next iteration. */
1317                 prev_offset += wrsz;
1318                 MPASS(auio.uio_offset == prev_offset);
1319
1320                 ppscheck = 0;
1321                 if (error != 0 && (ppscheck = ppsratecheck(&lastfail,
1322                     &curfail, 1)) != 0)
1323                         vn_printf(vp, "vnode_pager_putpages: I/O error %d\n",
1324                             error);
1325                 if (auio.uio_resid != 0 && (ppscheck != 0 ||
1326                     ppsratecheck(&lastfail, &curfail, 1) != 0))
1327                         vn_printf(vp, "vnode_pager_putpages: residual I/O %zd "
1328                             "at %ju\n", auio.uio_resid,
1329                             (uintmax_t)ma[0]->pindex);
1330                 VM_OBJECT_RLOCK(object);
1331                 if (error != 0 || auio.uio_resid != 0)
1332                         break;
1333         }
1334 write_done:
1335         /* Mark completely processed pages. */
1336         for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++)
1337                 rtvals[i] = VM_PAGER_OK;
1338         /* Mark partial EOF page. */
1339         if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0)
1340                 rtvals[i++] = VM_PAGER_OK;
1341         /* Unwritten pages in range, free bonus if the page is clean. */
1342         for (; i < ncount; i++)
1343                 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR;
1344         VM_OBJECT_RUNLOCK(object);
1345         PCPU_ADD(cnt.v_vnodepgsout, i);
1346         PCPU_INC(cnt.v_vnodeout);
1347         return (rtvals[0]);
1348 }
1349
1350 int
1351 vnode_pager_putpages_ioflags(int pager_flags)
1352 {
1353         int ioflags;
1354
1355         /*
1356          * Pageouts are already clustered, use IO_ASYNC to force a
1357          * bawrite() rather then a bdwrite() to prevent paging I/O
1358          * from saturating the buffer cache.  Dummy-up the sequential
1359          * heuristic to cause large ranges to cluster.  If neither
1360          * IO_SYNC or IO_ASYNC is set, the system decides how to
1361          * cluster.
1362          */
1363         ioflags = IO_VMIO;
1364         if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0)
1365                 ioflags |= IO_SYNC;
1366         else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0)
1367                 ioflags |= IO_ASYNC;
1368         ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0;
1369         ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0;
1370         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1371         return (ioflags);
1372 }
1373
1374 /*
1375  * vnode_pager_undirty_pages().
1376  *
1377  * A helper to mark pages as clean after pageout that was possibly
1378  * done with a short write.  The lpos argument specifies the page run
1379  * length in bytes, and the written argument specifies how many bytes
1380  * were actually written.  eof is the offset past the last valid byte
1381  * in the vnode using the absolute file position of the first byte in
1382  * the run as the base from which it is computed.
1383  */
1384 void
1385 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof,
1386     int lpos)
1387 {
1388         vm_object_t obj;
1389         int i, pos, pos_devb;
1390
1391         if (written == 0 && eof >= lpos)
1392                 return;
1393         obj = ma[0]->object;
1394         VM_OBJECT_WLOCK(obj);
1395         for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1396                 if (pos < trunc_page(written)) {
1397                         rtvals[i] = VM_PAGER_OK;
1398                         vm_page_undirty(ma[i]);
1399                 } else {
1400                         /* Partially written page. */
1401                         rtvals[i] = VM_PAGER_AGAIN;
1402                         vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1403                 }
1404         }
1405         if (eof >= lpos) /* avoid truncation */
1406                 goto done;
1407         for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) {
1408                 if (pos != trunc_page(pos)) {
1409                         /*
1410                          * The page contains the last valid byte in
1411                          * the vnode, mark the rest of the page as
1412                          * clean, potentially making the whole page
1413                          * clean.
1414                          */
1415                         pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE);
1416                         vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE -
1417                             pos_devb);
1418
1419                         /*
1420                          * If the page was cleaned, report the pageout
1421                          * on it as successful.  msync() no longer
1422                          * needs to write out the page, endlessly
1423                          * creating write requests and dirty buffers.
1424                          */
1425                         if (ma[i]->dirty == 0)
1426                                 rtvals[i] = VM_PAGER_OK;
1427
1428                         pos = round_page(pos);
1429                 } else {
1430                         /* vm_pageout_flush() clears dirty */
1431                         rtvals[i] = VM_PAGER_BAD;
1432                         pos += PAGE_SIZE;
1433                 }
1434         }
1435 done:
1436         VM_OBJECT_WUNLOCK(obj);
1437 }
1438
1439 void
1440 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1441     vm_offset_t end)
1442 {
1443         struct vnode *vp;
1444         vm_ooffset_t old_wm;
1445
1446         VM_OBJECT_WLOCK(object);
1447         if (object->type != OBJT_VNODE) {
1448                 VM_OBJECT_WUNLOCK(object);
1449                 return;
1450         }
1451         old_wm = object->un_pager.vnp.writemappings;
1452         object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1453         vp = object->handle;
1454         if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1455                 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1456                 VOP_ADD_WRITECOUNT(vp, 1);
1457                 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1458                     __func__, vp, vp->v_writecount);
1459         } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1460                 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1461                 VOP_ADD_WRITECOUNT(vp, -1);
1462                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1463                     __func__, vp, vp->v_writecount);
1464         }
1465         VM_OBJECT_WUNLOCK(object);
1466 }
1467
1468 void
1469 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1470     vm_offset_t end)
1471 {
1472         struct vnode *vp;
1473         struct mount *mp;
1474         vm_offset_t inc;
1475
1476         VM_OBJECT_WLOCK(object);
1477
1478         /*
1479          * First, recheck the object type to account for the race when
1480          * the vnode is reclaimed.
1481          */
1482         if (object->type != OBJT_VNODE) {
1483                 VM_OBJECT_WUNLOCK(object);
1484                 return;
1485         }
1486
1487         /*
1488          * Optimize for the case when writemappings is not going to
1489          * zero.
1490          */
1491         inc = end - start;
1492         if (object->un_pager.vnp.writemappings != inc) {
1493                 object->un_pager.vnp.writemappings -= inc;
1494                 VM_OBJECT_WUNLOCK(object);
1495                 return;
1496         }
1497
1498         vp = object->handle;
1499         vhold(vp);
1500         VM_OBJECT_WUNLOCK(object);
1501         mp = NULL;
1502         vn_start_write(vp, &mp, V_WAIT);
1503         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1504
1505         /*
1506          * Decrement the object's writemappings, by swapping the start
1507          * and end arguments for vnode_pager_update_writecount().  If
1508          * there was not a race with vnode reclaimation, then the
1509          * vnode's v_writecount is decremented.
1510          */
1511         vnode_pager_update_writecount(object, end, start);
1512         VOP_UNLOCK(vp, 0);
1513         vdrop(vp);
1514         if (mp != NULL)
1515                 vn_finished_write(mp);
1516 }