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