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 int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
  79     daddr_t *rtaddress, 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,
  87     vm_ooffset_t, struct ucred *cred);
  88
  89 struct pagerops vnodepagerops = {
  90         .pgo_alloc =    vnode_pager_alloc,
  91         .pgo_dealloc =  vnode_pager_dealloc,
  92         .pgo_getpages = vnode_pager_getpages,
  93         .pgo_putpages = vnode_pager_putpages,
  94         .pgo_haspage =  vnode_pager_haspage,
  95 };
  96
  97 int vnode_pbuf_freecnt;
  98
  99 /* Create the VM system backing object for this vnode */
 100 int
 101 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
 102 {
 103         vm_object_t object;
 104         vm_ooffset_t size = isize;
 105         struct vattr va;
 106
 107         if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
 108                 return (0);
 109
 110         while ((object = vp->v_object) != NULL) {
 111                 VM_OBJECT_LOCK(object);
 112                 if (!(object->flags & OBJ_DEAD)) {
 113                         VM_OBJECT_UNLOCK(object);
 114                         return (0);
 115                 }
 116                 VOP_UNLOCK(vp, 0);
 117                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 118                 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0);
 119                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 120         }
 121
 122         if (size == 0) {
 123                 if (vn_isdisk(vp, NULL)) {
 124                         size = IDX_TO_OFF(INT_MAX);
 125                 } else {
 126                         if (VOP_GETATTR(vp, &va, td->td_ucred))
 127                                 return (0);
 128                         size = va.va_size;
 129                 }
 130         }
 131
 132         object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
 133         /*
 134          * Dereference the reference we just created.  This assumes
 135          * that the object is associated with the vp.
 136          */
 137         VM_OBJECT_LOCK(object);
 138         object->ref_count--;
 139         VM_OBJECT_UNLOCK(object);
 140         vrele(vp);
 141
 142         KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
 143
 144         return (0);
 145 }
 146
 147 void
 148 vnode_destroy_vobject(struct vnode *vp)
 149 {
 150         struct vm_object *obj;
 151
 152         obj = vp->v_object;
 153         if (obj == NULL)
 154                 return;
 155         ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
 156         VM_OBJECT_LOCK(obj);
 157         if (obj->ref_count == 0) {
 158                 /*
 159                  * vclean() may be called twice. The first time
 160                  * removes the primary reference to the object,
 161                  * the second time goes one further and is a
 162                  * special-case to terminate the object.
 163                  *
 164                  * don't double-terminate the object
 165                  */
 166                 if ((obj->flags & OBJ_DEAD) == 0)
 167                         vm_object_terminate(obj);
 168                 else
 169                         VM_OBJECT_UNLOCK(obj);
 170         } else {
 171                 /*
 172                  * Woe to the process that tries to page now :-).
 173                  */
 174                 vm_pager_deallocate(obj);
 175                 VM_OBJECT_UNLOCK(obj);
 176         }
 177         vp->v_object = NULL;
 178 }
 179
 180
 181 /*
 182  * Allocate (or lookup) pager for a vnode.
 183  * Handle is a vnode pointer.
 184  *
 185  * MPSAFE
 186  */
 187 vm_object_t
 188 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
 189     vm_ooffset_t offset, struct ucred *cred)
 190 {
 191         vm_object_t object;
 192         struct vnode *vp;
 193
 194         /*
 195          * Pageout to vnode, no can do yet.
 196          */
 197         if (handle == NULL)
 198                 return (NULL);
 199
 200         vp = (struct vnode *) handle;
 201
 202         /*
 203          * If the object is being terminated, wait for it to
 204          * go away.
 205          */
 206 retry:
 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                  * Add 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                 VI_LOCK(vp);
 228                 if (vp->v_object != NULL) {
 229                         /*
 230                          * Object has been created while we were sleeping
 231                          */
 232                         VI_UNLOCK(vp);
 233                         vm_object_destroy(object);
 234                         goto retry;
 235                 }
 236                 vp->v_object = object;
 237                 VI_UNLOCK(vp);
 238         } else {
 239                 object->ref_count++;
 240                 VM_OBJECT_UNLOCK(object);
 241         }
 242         vref(vp);
 243         return (object);
 244 }
 245
 246 /*
 247  *      The object must be locked.
 248  */
 249 static void
 250 vnode_pager_dealloc(object)
 251         vm_object_t object;
 252 {
 253         struct vnode *vp = object->handle;
 254
 255         if (vp == NULL)
 256                 panic("vnode_pager_dealloc: pager already dealloced");
 257
 258         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 259         vm_object_pip_wait(object, "vnpdea");
 260
 261         object->handle = NULL;
 262         object->type = OBJT_DEAD;
 263         if (object->flags & OBJ_DISCONNECTWNT) {
 264                 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
 265                 wakeup(object);
 266         }
 267         ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
 268         vp->v_object = NULL;
 269         vp->v_vflag &= ~VV_TEXT;
 270 }
 271
 272 static boolean_t
 273 vnode_pager_haspage(object, pindex, before, after)
 274         vm_object_t object;
 275         vm_pindex_t pindex;
 276         int *before;
 277         int *after;
 278 {
 279         struct vnode *vp = object->handle;
 280         daddr_t bn;
 281         int err;
 282         daddr_t reqblock;
 283         int poff;
 284         int bsize;
 285         int pagesperblock, blocksperpage;
 286         int vfslocked;
 287
 288         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 289         /*
 290          * If no vp or vp is doomed or marked transparent to VM, we do not
 291          * have the page.
 292          */
 293         if (vp == NULL || vp->v_iflag & VI_DOOMED)
 294                 return FALSE;
 295         /*
 296          * If the offset is beyond end of file we do
 297          * not have the page.
 298          */
 299         if (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 /*      ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
 370         VM_OBJECT_LOCK(object);
 371         if (nsize == object->un_pager.vnp.vnp_size) {
 372                 /*
 373                  * Hasn't changed size
 374                  */
 375                 VM_OBJECT_UNLOCK(object);
 376                 return;
 377         }
 378         nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
 379         if (nsize < object->un_pager.vnp.vnp_size) {
 380                 /*
 381                  * File has shrunk. Toss any cached pages beyond the new EOF.
 382                  */
 383                 if (nobjsize < object->size)
 384                         vm_object_page_remove(object, nobjsize, object->size,
 385                             FALSE);
 386                 /*
 387                  * this gets rid of garbage at the end of a page that is now
 388                  * only partially backed by the vnode.
 389                  *
 390                  * XXX for some reason (I don't know yet), if we take a
 391                  * completely invalid page and mark it partially valid
 392                  * it can screw up NFS reads, so we don't allow the case.
 393                  */
 394                 if ((nsize & PAGE_MASK) &&
 395                     (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
 396                     m->valid != 0) {
 397                         int base = (int)nsize & PAGE_MASK;
 398                         int size = PAGE_SIZE - base;
 399
 400                         /*
 401                          * Clear out partial-page garbage in case
 402                          * the page has been mapped.
 403                          */
 404                         pmap_zero_page_area(m, base, size);
 405
 406                         /*
 407                          * Update the valid bits to reflect the blocks that
 408                          * have been zeroed.  Some of these valid bits may
 409                          * have already been set.
 410                          */
 411                         vm_page_set_valid(m, base, size);
 412
 413                         /*
 414                          * Round "base" to the next block boundary so that the
 415                          * dirty bit for a partially zeroed block is not
 416                          * cleared.
 417                          */
 418                         base = roundup2(base, DEV_BSIZE);
 419
 420                         /*
 421                          * Clear out partial-page dirty bits.
 422                          *
 423                          * note that we do not clear out the valid
 424                          * bits.  This would prevent bogus_page
 425                          * replacement from working properly.
 426                          */
 427                         vm_page_lock_queues();
 428                         vm_page_clear_dirty(m, base, PAGE_SIZE - base);
 429                         vm_page_unlock_queues();
 430                 } else if ((nsize & PAGE_MASK) &&
 431                     __predict_false(object->cache != NULL)) {
 432                         vm_page_cache_free(object, OFF_TO_IDX(nsize),
 433                             nobjsize);
 434                 }
 435         }
 436         object->un_pager.vnp.vnp_size = nsize;
 437         object->size = nobjsize;
 438         VM_OBJECT_UNLOCK(object);
 439 }
 440
 441 /*
 442  * calculate the linear (byte) disk address of specified virtual
 443  * file address
 444  */
 445 static int
 446 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
 447     int *run)
 448 {
 449         int bsize;
 450         int err;
 451         daddr_t vblock;
 452         daddr_t voffset;
 453
 454         if (address < 0)
 455                 return -1;
 456
 457         if (vp->v_iflag & VI_DOOMED)
 458                 return -1;
 459
 460         bsize = vp->v_mount->mnt_stat.f_iosize;
 461         vblock = address / bsize;
 462         voffset = address % bsize;
 463
 464         err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
 465         if (err == 0) {
 466                 if (*rtaddress != -1)
 467                         *rtaddress += voffset / DEV_BSIZE;
 468                 if (run) {
 469                         *run += 1;
 470                         *run *= bsize/PAGE_SIZE;
 471                         *run -= voffset/PAGE_SIZE;
 472                 }
 473         }
 474
 475         return (err);
 476 }
 477
 478 /*
 479  * small block filesystem vnode pager input
 480  */
 481 static int
 482 vnode_pager_input_smlfs(object, m)
 483         vm_object_t object;
 484         vm_page_t m;
 485 {
 486         int bits, i;
 487         struct vnode *vp;
 488         struct bufobj *bo;
 489         struct buf *bp;
 490         struct sf_buf *sf;
 491         daddr_t fileaddr;
 492         vm_offset_t bsize;
 493         int error = 0;
 494
 495         vp = object->handle;
 496         if (vp->v_iflag & VI_DOOMED)
 497                 return VM_PAGER_BAD;
 498
 499         bsize = vp->v_mount->mnt_stat.f_iosize;
 500
 501         VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
 502
 503         sf = sf_buf_alloc(m, 0);
 504
 505         for (i = 0; i < PAGE_SIZE / bsize; i++) {
 506                 vm_ooffset_t address;
 507
 508                 bits = vm_page_bits(i * bsize, bsize);
 509                 if (m->valid & bits)
 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                         error = vnode_pager_addr(vp, address, &fileaddr, NULL);
 517                         if (error)
 518                                 break;
 519                 }
 520                 if (fileaddr != -1) {
 521                         bp = getpbuf(&vnode_pbuf_freecnt);
 522
 523                         /* build a minimal buffer header */
 524                         bp->b_iocmd = BIO_READ;
 525                         bp->b_iodone = bdone;
 526                         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 527                         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 528                         bp->b_rcred = crhold(curthread->td_ucred);
 529                         bp->b_wcred = crhold(curthread->td_ucred);
 530                         bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
 531                         bp->b_blkno = fileaddr;
 532                         pbgetbo(bo, bp);
 533                         bp->b_bcount = bsize;
 534                         bp->b_bufsize = bsize;
 535                         bp->b_runningbufspace = bp->b_bufsize;
 536                         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 537
 538                         /* do the input */
 539                         bp->b_iooffset = dbtob(bp->b_blkno);
 540                         bstrategy(bp);
 541
 542                         bwait(bp, PVM, "vnsrd");
 543
 544                         if ((bp->b_ioflags & BIO_ERROR) != 0)
 545                                 error = EIO;
 546
 547                         /*
 548                          * free the buffer header back to the swap buffer pool
 549                          */
 550                         pbrelbo(bp);
 551                         relpbuf(bp, &vnode_pbuf_freecnt);
 552                         if (error)
 553                                 break;
 554                 } else
 555                         bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
 556                 KASSERT((m->dirty & bits) == 0,
 557                     ("vnode_pager_input_smlfs: page %p is dirty", m));
 558                 VM_OBJECT_LOCK(object);
 559                 m->valid |= bits;
 560                 VM_OBJECT_UNLOCK(object);
 561         }
 562         sf_buf_free(sf);
 563         if (error) {
 564                 return VM_PAGER_ERROR;
 565         }
 566         return VM_PAGER_OK;
 567 }
 568
 569 /*
 570  * old style vnode pager input routine
 571  */
 572 static int
 573 vnode_pager_input_old(object, m)
 574         vm_object_t object;
 575         vm_page_t m;
 576 {
 577         struct uio auio;
 578         struct iovec aiov;
 579         int error;
 580         int size;
 581         struct sf_buf *sf;
 582         struct vnode *vp;
 583
 584         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 585         error = 0;
 586
 587         /*
 588          * Return failure if beyond current EOF
 589          */
 590         if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
 591                 return VM_PAGER_BAD;
 592         } else {
 593                 size = PAGE_SIZE;
 594                 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
 595                         size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
 596                 vp = object->handle;
 597                 VM_OBJECT_UNLOCK(object);
 598
 599                 /*
 600                  * Allocate a kernel virtual address and initialize so that
 601                  * we can use VOP_READ/WRITE routines.
 602                  */
 603                 sf = sf_buf_alloc(m, 0);
 604
 605                 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
 606                 aiov.iov_len = size;
 607                 auio.uio_iov = &aiov;
 608                 auio.uio_iovcnt = 1;
 609                 auio.uio_offset = IDX_TO_OFF(m->pindex);
 610                 auio.uio_segflg = UIO_SYSSPACE;
 611                 auio.uio_rw = UIO_READ;
 612                 auio.uio_resid = size;
 613                 auio.uio_td = curthread;
 614
 615                 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
 616                 if (!error) {
 617                         int count = size - auio.uio_resid;
 618
 619                         if (count == 0)
 620                                 error = EINVAL;
 621                         else if (count != PAGE_SIZE)
 622                                 bzero((caddr_t)sf_buf_kva(sf) + count,
 623                                     PAGE_SIZE - count);
 624                 }
 625                 sf_buf_free(sf);
 626
 627                 VM_OBJECT_LOCK(object);
 628         }
 629         KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
 630         if (!error)
 631                 m->valid = VM_PAGE_BITS_ALL;
 632         return error ? VM_PAGER_ERROR : VM_PAGER_OK;
 633 }
 634
 635 /*
 636  * generic vnode pager input routine
 637  */
 638
 639 /*
 640  * Local media VFS's that do not implement their own VOP_GETPAGES
 641  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
 642  * to implement the previous behaviour.
 643  *
 644  * All other FS's should use the bypass to get to the local media
 645  * backing vp's VOP_GETPAGES.
 646  */
 647 static int
 648 vnode_pager_getpages(object, m, count, reqpage)
 649         vm_object_t object;
 650         vm_page_t *m;
 651         int count;
 652         int reqpage;
 653 {
 654         int rtval;
 655         struct vnode *vp;
 656         int bytes = count * PAGE_SIZE;
 657         int vfslocked;
 658
 659         vp = object->handle;
 660         VM_OBJECT_UNLOCK(object);
 661         vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 662         rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
 663         KASSERT(rtval != EOPNOTSUPP,
 664             ("vnode_pager: FS getpages not implemented\n"));
 665         VFS_UNLOCK_GIANT(vfslocked);
 666         VM_OBJECT_LOCK(object);
 667         return rtval;
 668 }
 669
 670 /*
 671  * This is now called from local media FS's to operate against their
 672  * own vnodes if they fail to implement VOP_GETPAGES.
 673  */
 674 int
 675 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
 676         struct vnode *vp;
 677         vm_page_t *m;
 678         int bytecount;
 679         int reqpage;
 680 {
 681         vm_object_t object;
 682         vm_offset_t kva;
 683         off_t foff, tfoff, nextoff;
 684         int i, j, size, bsize, first;
 685         daddr_t firstaddr, reqblock;
 686         struct bufobj *bo;
 687         int runpg;
 688         int runend;
 689         struct buf *bp;
 690         int count;
 691         int error;
 692
 693         object = vp->v_object;
 694         count = bytecount / PAGE_SIZE;
 695
 696         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 697             ("vnode_pager_generic_getpages does not support devices"));
 698         if (vp->v_iflag & VI_DOOMED)
 699                 return VM_PAGER_BAD;
 700
 701         bsize = vp->v_mount->mnt_stat.f_iosize;
 702
 703         /* get the UNDERLYING device for the file with VOP_BMAP() */
 704
 705         /*
 706          * originally, we did not check for an error return value -- assuming
 707          * an fs always has a bmap entry point -- that assumption is wrong!!!
 708          */
 709         foff = IDX_TO_OFF(m[reqpage]->pindex);
 710
 711         /*
 712          * if we can't bmap, use old VOP code
 713          */
 714         error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL);
 715         if (error == EOPNOTSUPP) {
 716                 VM_OBJECT_LOCK(object);
 717                 vm_page_lock_queues();
 718                 for (i = 0; i < count; i++)
 719                         if (i != reqpage)
 720                                 vm_page_free(m[i]);
 721                 vm_page_unlock_queues();
 722                 PCPU_INC(cnt.v_vnodein);
 723                 PCPU_INC(cnt.v_vnodepgsin);
 724                 error = vnode_pager_input_old(object, m[reqpage]);
 725                 VM_OBJECT_UNLOCK(object);
 726                 return (error);
 727         } else if (error != 0) {
 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                 VM_OBJECT_UNLOCK(object);
 735                 return (VM_PAGER_ERROR);
 736
 737                 /*
 738                  * if the blocksize is smaller than a page size, then use
 739                  * special small filesystem code.  NFS sometimes has a small
 740                  * blocksize, but it can handle large reads itself.
 741                  */
 742         } else if ((PAGE_SIZE / bsize) > 1 &&
 743             (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
 744                 VM_OBJECT_LOCK(object);
 745                 vm_page_lock_queues();
 746                 for (i = 0; i < count; i++)
 747                         if (i != reqpage)
 748                                 vm_page_free(m[i]);
 749                 vm_page_unlock_queues();
 750                 VM_OBJECT_UNLOCK(object);
 751                 PCPU_INC(cnt.v_vnodein);
 752                 PCPU_INC(cnt.v_vnodepgsin);
 753                 return vnode_pager_input_smlfs(object, m[reqpage]);
 754         }
 755
 756         /*
 757          * If we have a completely valid page available to us, we can
 758          * clean up and return.  Otherwise we have to re-read the
 759          * media.
 760          */
 761         VM_OBJECT_LOCK(object);
 762         if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
 763                 vm_page_lock_queues();
 764                 for (i = 0; i < count; i++)
 765                         if (i != reqpage)
 766                                 vm_page_free(m[i]);
 767                 vm_page_unlock_queues();
 768                 VM_OBJECT_UNLOCK(object);
 769                 return VM_PAGER_OK;
 770         } else if (reqblock == -1) {
 771                 pmap_zero_page(m[reqpage]);
 772                 KASSERT(m[reqpage]->dirty == 0,
 773                     ("vnode_pager_generic_getpages: page %p is dirty", m));
 774                 m[reqpage]->valid = VM_PAGE_BITS_ALL;
 775                 vm_page_lock_queues();
 776                 for (i = 0; i < count; i++)
 777                         if (i != reqpage)
 778                                 vm_page_free(m[i]);
 779                 vm_page_unlock_queues();
 780                 VM_OBJECT_UNLOCK(object);
 781                 return (VM_PAGER_OK);
 782         }
 783         m[reqpage]->valid = 0;
 784         VM_OBJECT_UNLOCK(object);
 785
 786         /*
 787          * here on direct device I/O
 788          */
 789         firstaddr = -1;
 790
 791         /*
 792          * calculate the run that includes the required page
 793          */
 794         for (first = 0, i = 0; i < count; i = runend) {
 795                 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr,
 796                     &runpg) != 0) {
 797                         VM_OBJECT_LOCK(object);
 798                         vm_page_lock_queues();
 799                         for (; i < count; i++)
 800                                 if (i != reqpage)
 801                                         vm_page_free(m[i]);
 802                         vm_page_unlock_queues();
 803                         VM_OBJECT_UNLOCK(object);
 804                         return (VM_PAGER_ERROR);
 805                 }
 806                 if (firstaddr == -1) {
 807                         VM_OBJECT_LOCK(object);
 808                         if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
 809                                 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
 810                                     (intmax_t)firstaddr, (uintmax_t)(foff >> 32),
 811                                     (uintmax_t)foff,
 812                                     (uintmax_t)
 813                                     (object->un_pager.vnp.vnp_size >> 32),
 814                                     (uintmax_t)object->un_pager.vnp.vnp_size);
 815                         }
 816                         vm_page_lock_queues();
 817                         vm_page_free(m[i]);
 818                         vm_page_unlock_queues();
 819                         VM_OBJECT_UNLOCK(object);
 820                         runend = i + 1;
 821                         first = runend;
 822                         continue;
 823                 }
 824                 runend = i + runpg;
 825                 if (runend <= reqpage) {
 826                         VM_OBJECT_LOCK(object);
 827                         vm_page_lock_queues();
 828                         for (j = i; j < runend; j++)
 829                                 vm_page_free(m[j]);
 830                         vm_page_unlock_queues();
 831                         VM_OBJECT_UNLOCK(object);
 832                 } else {
 833                         if (runpg < (count - first)) {
 834                                 VM_OBJECT_LOCK(object);
 835                                 vm_page_lock_queues();
 836                                 for (i = first + runpg; i < count; i++)
 837                                         vm_page_free(m[i]);
 838                                 vm_page_unlock_queues();
 839                                 VM_OBJECT_UNLOCK(object);
 840                                 count = first + runpg;
 841                         }
 842                         break;
 843                 }
 844                 first = runend;
 845         }
 846
 847         /*
 848          * the first and last page have been calculated now, move input pages
 849          * to be zero based...
 850          */
 851         if (first != 0) {
 852                 m += first;
 853                 count -= first;
 854                 reqpage -= first;
 855         }
 856
 857         /*
 858          * calculate the file virtual address for the transfer
 859          */
 860         foff = IDX_TO_OFF(m[0]->pindex);
 861
 862         /*
 863          * calculate the size of the transfer
 864          */
 865         size = count * PAGE_SIZE;
 866         KASSERT(count > 0, ("zero count"));
 867         if ((foff + size) > object->un_pager.vnp.vnp_size)
 868                 size = object->un_pager.vnp.vnp_size - foff;
 869         KASSERT(size > 0, ("zero size"));
 870
 871         /*
 872          * round up physical size for real devices.
 873          */
 874         if (1) {
 875                 int secmask = bo->bo_bsize - 1;
 876                 KASSERT(secmask < PAGE_SIZE && secmask > 0,
 877                     ("vnode_pager_generic_getpages: sector size %d too large",
 878                     secmask + 1));
 879                 size = (size + secmask) & ~secmask;
 880         }
 881
 882         bp = getpbuf(&vnode_pbuf_freecnt);
 883         kva = (vm_offset_t) bp->b_data;
 884
 885         /*
 886          * and map the pages to be read into the kva
 887          */
 888         pmap_qenter(kva, m, count);
 889
 890         /* build a minimal buffer header */
 891         bp->b_iocmd = BIO_READ;
 892         bp->b_iodone = bdone;
 893         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 894         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 895         bp->b_rcred = crhold(curthread->td_ucred);
 896         bp->b_wcred = crhold(curthread->td_ucred);
 897         bp->b_blkno = firstaddr;
 898         pbgetbo(bo, bp);
 899         bp->b_bcount = size;
 900         bp->b_bufsize = size;
 901         bp->b_runningbufspace = bp->b_bufsize;
 902         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 903
 904         PCPU_INC(cnt.v_vnodein);
 905         PCPU_ADD(cnt.v_vnodepgsin, count);
 906
 907         /* do the input */
 908         bp->b_iooffset = dbtob(bp->b_blkno);
 909         bstrategy(bp);
 910
 911         bwait(bp, PVM, "vnread");
 912
 913         if ((bp->b_ioflags & BIO_ERROR) != 0)
 914                 error = EIO;
 915
 916         if (!error) {
 917                 if (size != count * PAGE_SIZE)
 918                         bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
 919         }
 920         pmap_qremove(kva, count);
 921
 922         /*
 923          * free the buffer header back to the swap buffer pool
 924          */
 925         pbrelbo(bp);
 926         relpbuf(bp, &vnode_pbuf_freecnt);
 927
 928         VM_OBJECT_LOCK(object);
 929         vm_page_lock_queues();
 930         for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
 931                 vm_page_t mt;
 932
 933                 nextoff = tfoff + PAGE_SIZE;
 934                 mt = m[i];
 935
 936                 if (nextoff <= object->un_pager.vnp.vnp_size) {
 937                         /*
 938                          * Read filled up entire page.
 939                          */
 940                         mt->valid = VM_PAGE_BITS_ALL;
 941                         KASSERT(mt->dirty == 0,
 942                             ("vnode_pager_generic_getpages: page %p is dirty",
 943                             mt));
 944                         KASSERT(!pmap_page_is_mapped(mt),
 945                             ("vnode_pager_generic_getpages: page %p is mapped",
 946                             mt));
 947                 } else {
 948                         /*
 949                          * Read did not fill up entire page.
 950                          *
 951                          * Currently we do not set the entire page valid,
 952                          * we just try to clear the piece that we couldn't
 953                          * read.
 954                          */
 955                         vm_page_set_valid(mt, 0,
 956                             object->un_pager.vnp.vnp_size - tfoff);
 957                         KASSERT((mt->dirty & vm_page_bits(0,
 958                             object->un_pager.vnp.vnp_size - tfoff)) == 0,
 959                             ("vnode_pager_generic_getpages: page %p is dirty",
 960                             mt));
 961                 }
 962
 963                 if (i != reqpage) {
 964
 965                         /*
 966                          * whether or not to leave the page activated is up in
 967                          * the air, but we should put the page on a page queue
 968                          * somewhere. (it already is in the object). Result:
 969                          * It appears that empirical results show that
 970                          * deactivating pages is best.
 971                          */
 972
 973                         /*
 974                          * just in case someone was asking for this page we
 975                          * now tell them that it is ok to use
 976                          */
 977                         if (!error) {
 978                                 if (mt->oflags & VPO_WANTED)
 979                                         vm_page_activate(mt);
 980                                 else
 981                                         vm_page_deactivate(mt);
 982                                 vm_page_wakeup(mt);
 983                         } else {
 984                                 vm_page_free(mt);
 985                         }
 986                 }
 987         }
 988         vm_page_unlock_queues();
 989         VM_OBJECT_UNLOCK(object);
 990         if (error) {
 991                 printf("vnode_pager_getpages: I/O read error\n");
 992         }
 993         return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
 994 }
 995
 996 /*
 997  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
 998  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
 999  * vnode_pager_generic_putpages() to implement the previous behaviour.
1000  *
1001  * All other FS's should use the bypass to get to the local media
1002  * backing vp's VOP_PUTPAGES.
1003  */
1004 static void
1005 vnode_pager_putpages(object, m, count, sync, rtvals)
1006         vm_object_t object;
1007         vm_page_t *m;
1008         int count;
1009         boolean_t sync;
1010         int *rtvals;
1011 {
1012         int rtval;
1013         struct vnode *vp;
1014         struct mount *mp;
1015         int bytes = count * PAGE_SIZE;
1016
1017         /*
1018          * Force synchronous operation if we are extremely low on memory
1019          * to prevent a low-memory deadlock.  VOP operations often need to
1020          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1021          * operation ).  The swapper handles the case by limiting the amount
1022          * of asynchronous I/O, but that sort of solution doesn't scale well
1023          * for the vnode pager without a lot of work.
1024          *
1025          * Also, the backing vnode's iodone routine may not wake the pageout
1026          * daemon up.  This should be probably be addressed XXX.
1027          */
1028
1029         if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
1030                 sync |= OBJPC_SYNC;
1031
1032         /*
1033          * Call device-specific putpages function
1034          */
1035         vp = object->handle;
1036         VM_OBJECT_UNLOCK(object);
1037         if (vp->v_type != VREG)
1038                 mp = NULL;
1039         rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
1040         KASSERT(rtval != EOPNOTSUPP,
1041             ("vnode_pager: stale FS putpages\n"));
1042         VM_OBJECT_LOCK(object);
1043 }
1044
1045
1046 /*
1047  * This is now called from local media FS's to operate against their
1048  * own vnodes if they fail to implement VOP_PUTPAGES.
1049  *
1050  * This is typically called indirectly via the pageout daemon and
1051  * clustering has already typically occured, so in general we ask the
1052  * underlying filesystem to write the data out asynchronously rather
1053  * then delayed.
1054  */
1055 int
1056 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
1057         struct vnode *vp;
1058         vm_page_t *m;
1059         int bytecount;
1060         int flags;
1061         int *rtvals;
1062 {
1063         int i;
1064         vm_object_t object;
1065         int count;
1066
1067         int maxsize, ncount;
1068         vm_ooffset_t poffset;
1069         struct uio auio;
1070         struct iovec aiov;
1071         int error;
1072         int ioflags;
1073         int ppscheck = 0;
1074         static struct timeval lastfail;
1075         static int curfail;
1076
1077         object = vp->v_object;
1078         count = bytecount / PAGE_SIZE;
1079
1080         for (i = 0; i < count; i++)
1081                 rtvals[i] = VM_PAGER_AGAIN;
1082
1083         if ((int64_t)m[0]->pindex < 0) {
1084                 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1085                         (long)m[0]->pindex, (u_long)m[0]->dirty);
1086                 rtvals[0] = VM_PAGER_BAD;
1087                 return VM_PAGER_BAD;
1088         }
1089
1090         maxsize = count * PAGE_SIZE;
1091         ncount = count;
1092
1093         poffset = IDX_TO_OFF(m[0]->pindex);
1094
1095         /*
1096          * If the page-aligned write is larger then the actual file we
1097          * have to invalidate pages occuring beyond the file EOF.  However,
1098          * there is an edge case where a file may not be page-aligned where
1099          * the last page is partially invalid.  In this case the filesystem
1100          * may not properly clear the dirty bits for the entire page (which
1101          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1102          * With the page locked we are free to fix-up the dirty bits here.
1103          *
1104          * We do not under any circumstances truncate the valid bits, as
1105          * this will screw up bogus page replacement.
1106          */
1107         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1108                 if (object->un_pager.vnp.vnp_size > poffset) {
1109                         int pgoff;
1110
1111                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1112                         ncount = btoc(maxsize);
1113                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1114                                 vm_page_lock_queues();
1115                                 vm_page_clear_dirty(m[ncount - 1], pgoff,
1116                                         PAGE_SIZE - pgoff);
1117                                 vm_page_unlock_queues();
1118                         }
1119                 } else {
1120                         maxsize = 0;
1121                         ncount = 0;
1122                 }
1123                 if (ncount < count) {
1124                         for (i = ncount; i < count; i++) {
1125                                 rtvals[i] = VM_PAGER_BAD;
1126                         }
1127                 }
1128         }
1129
1130         /*
1131          * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1132          * rather then a bdwrite() to prevent paging I/O from saturating
1133          * the buffer cache.  Dummy-up the sequential heuristic to cause
1134          * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1135          * the system decides how to cluster.
1136          */
1137         ioflags = IO_VMIO;
1138         if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1139                 ioflags |= IO_SYNC;
1140         else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1141                 ioflags |= IO_ASYNC;
1142         ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1143         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1144
1145         aiov.iov_base = (caddr_t) 0;
1146         aiov.iov_len = maxsize;
1147         auio.uio_iov = &aiov;
1148         auio.uio_iovcnt = 1;
1149         auio.uio_offset = poffset;
1150         auio.uio_segflg = UIO_NOCOPY;
1151         auio.uio_rw = UIO_WRITE;
1152         auio.uio_resid = maxsize;
1153         auio.uio_td = (struct thread *) 0;
1154         error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1155         PCPU_INC(cnt.v_vnodeout);
1156         PCPU_ADD(cnt.v_vnodepgsout, ncount);
1157
1158         if (error) {
1159                 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1160                         printf("vnode_pager_putpages: I/O error %d\n", error);
1161         }
1162         if (auio.uio_resid) {
1163                 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1164                         printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1165                             auio.uio_resid, (u_long)m[0]->pindex);
1166         }
1167         for (i = 0; i < ncount; i++) {
1168                 rtvals[i] = VM_PAGER_OK;
1169         }
1170         return rtvals[0];
1171 }