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