sys/vm/vnode_pager.c

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