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