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;
 254         int refs;
 255
 256         vp = object->handle;
 257         if (vp == NULL)
 258                 panic("vnode_pager_dealloc: pager already dealloced");
 259
 260         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 261         vm_object_pip_wait(object, "vnpdea");
 262         refs = object->ref_count;
 263
 264         object->handle = NULL;
 265         object->type = OBJT_DEAD;
 266         if (object->flags & OBJ_DISCONNECTWNT) {
 267                 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
 268                 wakeup(object);
 269         }
 270         ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
 271         vp->v_object = NULL;
 272         vp->v_vflag &= ~VV_TEXT;
 273         while (refs-- > 0)
 274                 vunref(vp);
 275 }
 276
 277 static boolean_t
 278 vnode_pager_haspage(object, pindex, before, after)
 279         vm_object_t object;
 280         vm_pindex_t pindex;
 281         int *before;
 282         int *after;
 283 {
 284         struct vnode *vp = object->handle;
 285         daddr_t bn;
 286         int err;
 287         daddr_t reqblock;
 288         int poff;
 289         int bsize;
 290         int pagesperblock, blocksperpage;
 291         int vfslocked;
 292
 293         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 294         /*
 295          * If no vp or vp is doomed or marked transparent to VM, we do not
 296          * have the page.
 297          */
 298         if (vp == NULL || vp->v_iflag & VI_DOOMED)
 299                 return FALSE;
 300         /*
 301          * If the offset is beyond end of file we do
 302          * not have the page.
 303          */
 304         if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
 305                 return FALSE;
 306
 307         bsize = vp->v_mount->mnt_stat.f_iosize;
 308         pagesperblock = bsize / PAGE_SIZE;
 309         blocksperpage = 0;
 310         if (pagesperblock > 0) {
 311                 reqblock = pindex / pagesperblock;
 312         } else {
 313                 blocksperpage = (PAGE_SIZE / bsize);
 314                 reqblock = pindex * blocksperpage;
 315         }
 316         VM_OBJECT_UNLOCK(object);
 317         vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 318         err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
 319         VFS_UNLOCK_GIANT(vfslocked);
 320         VM_OBJECT_LOCK(object);
 321         if (err)
 322                 return TRUE;
 323         if (bn == -1)
 324                 return FALSE;
 325         if (pagesperblock > 0) {
 326                 poff = pindex - (reqblock * pagesperblock);
 327                 if (before) {
 328                         *before *= pagesperblock;
 329                         *before += poff;
 330                 }
 331                 if (after) {
 332                         int numafter;
 333                         *after *= pagesperblock;
 334                         numafter = pagesperblock - (poff + 1);
 335                         if (IDX_TO_OFF(pindex + numafter) >
 336                             object->un_pager.vnp.vnp_size) {
 337                                 numafter =
 338                                     OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
 339                                     pindex;
 340                         }
 341                         *after += numafter;
 342                 }
 343         } else {
 344                 if (before) {
 345                         *before /= blocksperpage;
 346                 }
 347
 348                 if (after) {
 349                         *after /= blocksperpage;
 350                 }
 351         }
 352         return TRUE;
 353 }
 354
 355 /*
 356  * Lets the VM system know about a change in size for a file.
 357  * We adjust our own internal size and flush any cached pages in
 358  * the associated object that are affected by the size change.
 359  *
 360  * Note: this routine may be invoked as a result of a pager put
 361  * operation (possibly at object termination time), so we must be careful.
 362  */
 363 void
 364 vnode_pager_setsize(vp, nsize)
 365         struct vnode *vp;
 366         vm_ooffset_t nsize;
 367 {
 368         vm_object_t object;
 369         vm_page_t m;
 370         vm_pindex_t nobjsize;
 371
 372         if ((object = vp->v_object) == NULL)
 373                 return;
 374 /*      ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
 375         VM_OBJECT_LOCK(object);
 376         if (nsize == object->un_pager.vnp.vnp_size) {
 377                 /*
 378                  * Hasn't changed size
 379                  */
 380                 VM_OBJECT_UNLOCK(object);
 381                 return;
 382         }
 383         nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
 384         if (nsize < object->un_pager.vnp.vnp_size) {
 385                 /*
 386                  * File has shrunk. Toss any cached pages beyond the new EOF.
 387                  */
 388                 if (nobjsize < object->size)
 389                         vm_object_page_remove(object, nobjsize, object->size,
 390                             0);
 391                 /*
 392                  * this gets rid of garbage at the end of a page that is now
 393                  * only partially backed by the vnode.
 394                  *
 395                  * XXX for some reason (I don't know yet), if we take a
 396                  * completely invalid page and mark it partially valid
 397                  * it can screw up NFS reads, so we don't allow the case.
 398                  */
 399                 if ((nsize & PAGE_MASK) &&
 400                     (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
 401                     m->valid != 0) {
 402                         int base = (int)nsize & PAGE_MASK;
 403                         int size = PAGE_SIZE - base;
 404
 405                         /*
 406                          * Clear out partial-page garbage in case
 407                          * the page has been mapped.
 408                          */
 409                         pmap_zero_page_area(m, base, size);
 410
 411                         /*
 412                          * Update the valid bits to reflect the blocks that
 413                          * have been zeroed.  Some of these valid bits may
 414                          * have already been set.
 415                          */
 416                         vm_page_set_valid(m, base, size);
 417
 418                         /*
 419                          * Round "base" to the next block boundary so that the
 420                          * dirty bit for a partially zeroed block is not
 421                          * cleared.
 422                          */
 423                         base = roundup2(base, DEV_BSIZE);
 424
 425                         /*
 426                          * Clear out partial-page dirty bits.
 427                          *
 428                          * note that we do not clear out the valid
 429                          * bits.  This would prevent bogus_page
 430                          * replacement from working properly.
 431                          */
 432                         vm_page_clear_dirty(m, base, PAGE_SIZE - base);
 433                 } else if ((nsize & PAGE_MASK) &&
 434                     __predict_false(object->cache != NULL)) {
 435                         vm_page_cache_free(object, OFF_TO_IDX(nsize),
 436                             nobjsize);
 437                 }
 438         }
 439         object->un_pager.vnp.vnp_size = nsize;
 440         object->size = nobjsize;
 441         VM_OBJECT_UNLOCK(object);
 442 }
 443
 444 /*
 445  * calculate the linear (byte) disk address of specified virtual
 446  * file address
 447  */
 448 static int
 449 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
 450     int *run)
 451 {
 452         int bsize;
 453         int err;
 454         daddr_t vblock;
 455         daddr_t voffset;
 456
 457         if (address < 0)
 458                 return -1;
 459
 460         if (vp->v_iflag & VI_DOOMED)
 461                 return -1;
 462
 463         bsize = vp->v_mount->mnt_stat.f_iosize;
 464         vblock = address / bsize;
 465         voffset = address % bsize;
 466
 467         err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
 468         if (err == 0) {
 469                 if (*rtaddress != -1)
 470                         *rtaddress += voffset / DEV_BSIZE;
 471                 if (run) {
 472                         *run += 1;
 473                         *run *= bsize/PAGE_SIZE;
 474                         *run -= voffset/PAGE_SIZE;
 475                 }
 476         }
 477
 478         return (err);
 479 }
 480
 481 /*
 482  * small block filesystem vnode pager input
 483  */
 484 static int
 485 vnode_pager_input_smlfs(object, m)
 486         vm_object_t object;
 487         vm_page_t m;
 488 {
 489         struct vnode *vp;
 490         struct bufobj *bo;
 491         struct buf *bp;
 492         struct sf_buf *sf;
 493         daddr_t fileaddr;
 494         vm_offset_t bsize;
 495         vm_page_bits_t bits;
 496         int error, i;
 497
 498         error = 0;
 499         vp = object->handle;
 500         if (vp->v_iflag & VI_DOOMED)
 501                 return VM_PAGER_BAD;
 502
 503         bsize = vp->v_mount->mnt_stat.f_iosize;
 504
 505         VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
 506
 507         sf = sf_buf_alloc(m, 0);
 508
 509         for (i = 0; i < PAGE_SIZE / bsize; i++) {
 510                 vm_ooffset_t address;
 511
 512                 bits = vm_page_bits(i * bsize, bsize);
 513                 if (m->valid & bits)
 514                         continue;
 515
 516                 address = IDX_TO_OFF(m->pindex) + i * bsize;
 517                 if (address >= object->un_pager.vnp.vnp_size) {
 518                         fileaddr = -1;
 519                 } else {
 520                         error = vnode_pager_addr(vp, address, &fileaddr, NULL);
 521                         if (error)
 522                                 break;
 523                 }
 524                 if (fileaddr != -1) {
 525                         bp = getpbuf(&vnode_pbuf_freecnt);
 526
 527                         /* build a minimal buffer header */
 528                         bp->b_iocmd = BIO_READ;
 529                         bp->b_iodone = bdone;
 530                         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 531                         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 532                         bp->b_rcred = crhold(curthread->td_ucred);
 533                         bp->b_wcred = crhold(curthread->td_ucred);
 534                         bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
 535                         bp->b_blkno = fileaddr;
 536                         pbgetbo(bo, bp);
 537                         bp->b_bcount = bsize;
 538                         bp->b_bufsize = bsize;
 539                         bp->b_runningbufspace = bp->b_bufsize;
 540                         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 541
 542                         /* do the input */
 543                         bp->b_iooffset = dbtob(bp->b_blkno);
 544                         bstrategy(bp);
 545
 546                         bwait(bp, PVM, "vnsrd");
 547
 548                         if ((bp->b_ioflags & BIO_ERROR) != 0)
 549                                 error = EIO;
 550
 551                         /*
 552                          * free the buffer header back to the swap buffer pool
 553                          */
 554                         pbrelbo(bp);
 555                         relpbuf(bp, &vnode_pbuf_freecnt);
 556                         if (error)
 557                                 break;
 558                 } else
 559                         bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
 560                 KASSERT((m->dirty & bits) == 0,
 561                     ("vnode_pager_input_smlfs: page %p is dirty", m));
 562                 VM_OBJECT_LOCK(object);
 563                 m->valid |= bits;
 564                 VM_OBJECT_UNLOCK(object);
 565         }
 566         sf_buf_free(sf);
 567         if (error) {
 568                 return VM_PAGER_ERROR;
 569         }
 570         return VM_PAGER_OK;
 571 }
 572
 573 /*
 574  * old style vnode pager input routine
 575  */
 576 static int
 577 vnode_pager_input_old(object, m)
 578         vm_object_t object;
 579         vm_page_t m;
 580 {
 581         struct uio auio;
 582         struct iovec aiov;
 583         int error;
 584         int size;
 585         struct sf_buf *sf;
 586         struct vnode *vp;
 587
 588         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 589         error = 0;
 590
 591         /*
 592          * Return failure if beyond current EOF
 593          */
 594         if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
 595                 return VM_PAGER_BAD;
 596         } else {
 597                 size = PAGE_SIZE;
 598                 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
 599                         size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
 600                 vp = object->handle;
 601                 VM_OBJECT_UNLOCK(object);
 602
 603                 /*
 604                  * Allocate a kernel virtual address and initialize so that
 605                  * we can use VOP_READ/WRITE routines.
 606                  */
 607                 sf = sf_buf_alloc(m, 0);
 608
 609                 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
 610                 aiov.iov_len = size;
 611                 auio.uio_iov = &aiov;
 612                 auio.uio_iovcnt = 1;
 613                 auio.uio_offset = IDX_TO_OFF(m->pindex);
 614                 auio.uio_segflg = UIO_SYSSPACE;
 615                 auio.uio_rw = UIO_READ;
 616                 auio.uio_resid = size;
 617                 auio.uio_td = curthread;
 618
 619                 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
 620                 if (!error) {
 621                         int count = size - auio.uio_resid;
 622
 623                         if (count == 0)
 624                                 error = EINVAL;
 625                         else if (count != PAGE_SIZE)
 626                                 bzero((caddr_t)sf_buf_kva(sf) + count,
 627                                     PAGE_SIZE - count);
 628                 }
 629                 sf_buf_free(sf);
 630
 631                 VM_OBJECT_LOCK(object);
 632         }
 633         KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
 634         if (!error)
 635                 m->valid = VM_PAGE_BITS_ALL;
 636         return error ? VM_PAGER_ERROR : VM_PAGER_OK;
 637 }
 638
 639 /*
 640  * generic vnode pager input routine
 641  */
 642
 643 /*
 644  * Local media VFS's that do not implement their own VOP_GETPAGES
 645  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
 646  * to implement the previous behaviour.
 647  *
 648  * All other FS's should use the bypass to get to the local media
 649  * backing vp's VOP_GETPAGES.
 650  */
 651 static int
 652 vnode_pager_getpages(object, m, count, reqpage)
 653         vm_object_t object;
 654         vm_page_t *m;
 655         int count;
 656         int reqpage;
 657 {
 658         int rtval;
 659         struct vnode *vp;
 660         int bytes = count * PAGE_SIZE;
 661         int vfslocked;
 662
 663         vp = object->handle;
 664         VM_OBJECT_UNLOCK(object);
 665         vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 666         rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
 667         KASSERT(rtval != EOPNOTSUPP,
 668             ("vnode_pager: FS getpages not implemented\n"));
 669         VFS_UNLOCK_GIANT(vfslocked);
 670         VM_OBJECT_LOCK(object);
 671         return rtval;
 672 }
 673
 674 /*
 675  * This is now called from local media FS's to operate against their
 676  * own vnodes if they fail to implement VOP_GETPAGES.
 677  */
 678 int
 679 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
 680         struct vnode *vp;
 681         vm_page_t *m;
 682         int bytecount;
 683         int reqpage;
 684 {
 685         vm_object_t object;
 686         vm_offset_t kva;
 687         off_t foff, tfoff, nextoff;
 688         int i, j, size, bsize, first;
 689         daddr_t firstaddr, reqblock;
 690         struct bufobj *bo;
 691         int runpg;
 692         int runend;
 693         struct buf *bp;
 694         int count;
 695         int error;
 696
 697         object = vp->v_object;
 698         count = bytecount / PAGE_SIZE;
 699
 700         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 701             ("vnode_pager_generic_getpages does not support devices"));
 702         if (vp->v_iflag & VI_DOOMED)
 703                 return VM_PAGER_BAD;
 704
 705         bsize = vp->v_mount->mnt_stat.f_iosize;
 706
 707         /* get the UNDERLYING device for the file with VOP_BMAP() */
 708
 709         /*
 710          * originally, we did not check for an error return value -- assuming
 711          * an fs always has a bmap entry point -- that assumption is wrong!!!
 712          */
 713         foff = IDX_TO_OFF(m[reqpage]->pindex);
 714
 715         /*
 716          * if we can't bmap, use old VOP code
 717          */
 718         error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL);
 719         if (error == EOPNOTSUPP) {
 720                 VM_OBJECT_LOCK(object);
 721
 722                 for (i = 0; i < count; i++)
 723                         if (i != reqpage) {
 724                                 vm_page_lock(m[i]);
 725                                 vm_page_free(m[i]);
 726                                 vm_page_unlock(m[i]);
 727                         }
 728                 PCPU_INC(cnt.v_vnodein);
 729                 PCPU_INC(cnt.v_vnodepgsin);
 730                 error = vnode_pager_input_old(object, m[reqpage]);
 731                 VM_OBJECT_UNLOCK(object);
 732                 return (error);
 733         } else if (error != 0) {
 734                 VM_OBJECT_LOCK(object);
 735                 for (i = 0; i < count; i++)
 736                         if (i != reqpage) {
 737                                 vm_page_lock(m[i]);
 738                                 vm_page_free(m[i]);
 739                                 vm_page_unlock(m[i]);
 740                         }
 741                 VM_OBJECT_UNLOCK(object);
 742                 return (VM_PAGER_ERROR);
 743
 744                 /*
 745                  * if the blocksize is smaller than a page size, then use
 746                  * special small filesystem code.  NFS sometimes has a small
 747                  * blocksize, but it can handle large reads itself.
 748                  */
 749         } else if ((PAGE_SIZE / bsize) > 1 &&
 750             (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
 751                 VM_OBJECT_LOCK(object);
 752                 for (i = 0; i < count; i++)
 753                         if (i != reqpage) {
 754                                 vm_page_lock(m[i]);
 755                                 vm_page_free(m[i]);
 756                                 vm_page_unlock(m[i]);
 757                         }
 758                 VM_OBJECT_UNLOCK(object);
 759                 PCPU_INC(cnt.v_vnodein);
 760                 PCPU_INC(cnt.v_vnodepgsin);
 761                 return vnode_pager_input_smlfs(object, m[reqpage]);
 762         }
 763
 764         /*
 765          * If we have a completely valid page available to us, we can
 766          * clean up and return.  Otherwise we have to re-read the
 767          * media.
 768          */
 769         VM_OBJECT_LOCK(object);
 770         if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
 771                 for (i = 0; i < count; i++)
 772                         if (i != reqpage) {
 773                                 vm_page_lock(m[i]);
 774                                 vm_page_free(m[i]);
 775                                 vm_page_unlock(m[i]);
 776                         }
 777                 VM_OBJECT_UNLOCK(object);
 778                 return VM_PAGER_OK;
 779         } else if (reqblock == -1) {
 780                 pmap_zero_page(m[reqpage]);
 781                 KASSERT(m[reqpage]->dirty == 0,
 782                     ("vnode_pager_generic_getpages: page %p is dirty", m));
 783                 m[reqpage]->valid = VM_PAGE_BITS_ALL;
 784                 for (i = 0; i < count; i++)
 785                         if (i != reqpage) {
 786                                 vm_page_lock(m[i]);
 787                                 vm_page_free(m[i]);
 788                                 vm_page_unlock(m[i]);
 789                         }
 790                 VM_OBJECT_UNLOCK(object);
 791                 return (VM_PAGER_OK);
 792         }
 793         m[reqpage]->valid = 0;
 794         VM_OBJECT_UNLOCK(object);
 795
 796         /*
 797          * here on direct device I/O
 798          */
 799         firstaddr = -1;
 800
 801         /*
 802          * calculate the run that includes the required page
 803          */
 804         for (first = 0, i = 0; i < count; i = runend) {
 805                 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr,
 806                     &runpg) != 0) {
 807                         VM_OBJECT_LOCK(object);
 808                         for (; i < count; i++)
 809                                 if (i != reqpage) {
 810                                         vm_page_lock(m[i]);
 811                                         vm_page_free(m[i]);
 812                                         vm_page_unlock(m[i]);
 813                                 }
 814                         VM_OBJECT_UNLOCK(object);
 815                         return (VM_PAGER_ERROR);
 816                 }
 817                 if (firstaddr == -1) {
 818                         VM_OBJECT_LOCK(object);
 819                         if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
 820                                 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
 821                                     (intmax_t)firstaddr, (uintmax_t)(foff >> 32),
 822                                     (uintmax_t)foff,
 823                                     (uintmax_t)
 824                                     (object->un_pager.vnp.vnp_size >> 32),
 825                                     (uintmax_t)object->un_pager.vnp.vnp_size);
 826                         }
 827                         vm_page_lock(m[i]);
 828                         vm_page_free(m[i]);
 829                         vm_page_unlock(m[i]);
 830                         VM_OBJECT_UNLOCK(object);
 831                         runend = i + 1;
 832                         first = runend;
 833                         continue;
 834                 }
 835                 runend = i + runpg;
 836                 if (runend <= reqpage) {
 837                         VM_OBJECT_LOCK(object);
 838                         for (j = i; j < runend; j++) {
 839                                 vm_page_lock(m[j]);
 840                                 vm_page_free(m[j]);
 841                                 vm_page_unlock(m[j]);
 842                         }
 843                         VM_OBJECT_UNLOCK(object);
 844                 } else {
 845                         if (runpg < (count - first)) {
 846                                 VM_OBJECT_LOCK(object);
 847                                 for (i = first + runpg; i < count; i++) {
 848                                         vm_page_lock(m[i]);
 849                                         vm_page_free(m[i]);
 850                                         vm_page_unlock(m[i]);
 851                                 }
 852                                 VM_OBJECT_UNLOCK(object);
 853                                 count = first + runpg;
 854                         }
 855                         break;
 856                 }
 857                 first = runend;
 858         }
 859
 860         /*
 861          * the first and last page have been calculated now, move input pages
 862          * to be zero based...
 863          */
 864         if (first != 0) {
 865                 m += first;
 866                 count -= first;
 867                 reqpage -= first;
 868         }
 869
 870         /*
 871          * calculate the file virtual address for the transfer
 872          */
 873         foff = IDX_TO_OFF(m[0]->pindex);
 874
 875         /*
 876          * calculate the size of the transfer
 877          */
 878         size = count * PAGE_SIZE;
 879         KASSERT(count > 0, ("zero count"));
 880         if ((foff + size) > object->un_pager.vnp.vnp_size)
 881                 size = object->un_pager.vnp.vnp_size - foff;
 882         KASSERT(size > 0, ("zero size"));
 883
 884         /*
 885          * round up physical size for real devices.
 886          */
 887         if (1) {
 888                 int secmask = bo->bo_bsize - 1;
 889                 KASSERT(secmask < PAGE_SIZE && secmask > 0,
 890                     ("vnode_pager_generic_getpages: sector size %d too large",
 891                     secmask + 1));
 892                 size = (size + secmask) & ~secmask;
 893         }
 894
 895         bp = getpbuf(&vnode_pbuf_freecnt);
 896         kva = (vm_offset_t) bp->b_data;
 897
 898         /*
 899          * and map the pages to be read into the kva
 900          */
 901         pmap_qenter(kva, m, count);
 902
 903         /* build a minimal buffer header */
 904         bp->b_iocmd = BIO_READ;
 905         bp->b_iodone = bdone;
 906         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 907         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 908         bp->b_rcred = crhold(curthread->td_ucred);
 909         bp->b_wcred = crhold(curthread->td_ucred);
 910         bp->b_blkno = firstaddr;
 911         pbgetbo(bo, bp);
 912         bp->b_bcount = size;
 913         bp->b_bufsize = size;
 914         bp->b_runningbufspace = bp->b_bufsize;
 915         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 916
 917         PCPU_INC(cnt.v_vnodein);
 918         PCPU_ADD(cnt.v_vnodepgsin, count);
 919
 920         /* do the input */
 921         bp->b_iooffset = dbtob(bp->b_blkno);
 922         bstrategy(bp);
 923
 924         bwait(bp, PVM, "vnread");
 925
 926         if ((bp->b_ioflags & BIO_ERROR) != 0)
 927                 error = EIO;
 928
 929         if (!error) {
 930                 if (size != count * PAGE_SIZE)
 931                         bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
 932         }
 933         pmap_qremove(kva, count);
 934
 935         /*
 936          * free the buffer header back to the swap buffer pool
 937          */
 938         pbrelbo(bp);
 939         relpbuf(bp, &vnode_pbuf_freecnt);
 940
 941         VM_OBJECT_LOCK(object);
 942         for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
 943                 vm_page_t mt;
 944
 945                 nextoff = tfoff + PAGE_SIZE;
 946                 mt = m[i];
 947
 948                 if (nextoff <= object->un_pager.vnp.vnp_size) {
 949                         /*
 950                          * Read filled up entire page.
 951                          */
 952                         mt->valid = VM_PAGE_BITS_ALL;
 953                         KASSERT(mt->dirty == 0,
 954                             ("vnode_pager_generic_getpages: page %p is dirty",
 955                             mt));
 956                         KASSERT(!pmap_page_is_mapped(mt),
 957                             ("vnode_pager_generic_getpages: page %p is mapped",
 958                             mt));
 959                 } else {
 960                         /*
 961                          * Read did not fill up entire page.
 962                          *
 963                          * Currently we do not set the entire page valid,
 964                          * we just try to clear the piece that we couldn't
 965                          * read.
 966                          */
 967                         vm_page_set_valid(mt, 0,
 968                             object->un_pager.vnp.vnp_size - tfoff);
 969                         KASSERT((mt->dirty & vm_page_bits(0,
 970                             object->un_pager.vnp.vnp_size - tfoff)) == 0,
 971                             ("vnode_pager_generic_getpages: page %p is dirty",
 972                             mt));
 973                 }
 974
 975                 if (i != reqpage) {
 976
 977                         /*
 978                          * whether or not to leave the page activated is up in
 979                          * the air, but we should put the page on a page queue
 980                          * somewhere. (it already is in the object). Result:
 981                          * It appears that empirical results show that
 982                          * deactivating pages is best.
 983                          */
 984
 985                         /*
 986                          * just in case someone was asking for this page we
 987                          * now tell them that it is ok to use
 988                          */
 989                         if (!error) {
 990                                 if (mt->oflags & VPO_WANTED) {
 991                                         vm_page_lock(mt);
 992                                         vm_page_activate(mt);
 993                                         vm_page_unlock(mt);
 994                                 } else {
 995                                         vm_page_lock(mt);
 996                                         vm_page_deactivate(mt);
 997                                         vm_page_unlock(mt);
 998                                 }
 999                                 vm_page_wakeup(mt);
1000                         } else {
1001                                 vm_page_lock(mt);
1002                                 vm_page_free(mt);
1003                                 vm_page_unlock(mt);
1004                         }
1005                 }
1006         }
1007         VM_OBJECT_UNLOCK(object);
1008         if (error) {
1009                 printf("vnode_pager_getpages: I/O read error\n");
1010         }
1011         return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
1012 }
1013
1014 /*
1015  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1016  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1017  * vnode_pager_generic_putpages() to implement the previous behaviour.
1018  *
1019  * All other FS's should use the bypass to get to the local media
1020  * backing vp's VOP_PUTPAGES.
1021  */
1022 static void
1023 vnode_pager_putpages(object, m, count, sync, rtvals)
1024         vm_object_t object;
1025         vm_page_t *m;
1026         int count;
1027         boolean_t sync;
1028         int *rtvals;
1029 {
1030         int rtval;
1031         struct vnode *vp;
1032         int bytes = count * PAGE_SIZE;
1033
1034         /*
1035          * Force synchronous operation if we are extremely low on memory
1036          * to prevent a low-memory deadlock.  VOP operations often need to
1037          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1038          * operation ).  The swapper handles the case by limiting the amount
1039          * of asynchronous I/O, but that sort of solution doesn't scale well
1040          * for the vnode pager without a lot of work.
1041          *
1042          * Also, the backing vnode's iodone routine may not wake the pageout
1043          * daemon up.  This should be probably be addressed XXX.
1044          */
1045
1046         if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
1047                 sync |= OBJPC_SYNC;
1048
1049         /*
1050          * Call device-specific putpages function
1051          */
1052         vp = object->handle;
1053         VM_OBJECT_UNLOCK(object);
1054         rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
1055         KASSERT(rtval != EOPNOTSUPP,
1056             ("vnode_pager: stale FS putpages\n"));
1057         VM_OBJECT_LOCK(object);
1058 }
1059
1060
1061 /*
1062  * This is now called from local media FS's to operate against their
1063  * own vnodes if they fail to implement VOP_PUTPAGES.
1064  *
1065  * This is typically called indirectly via the pageout daemon and
1066  * clustering has already typically occured, so in general we ask the
1067  * underlying filesystem to write the data out asynchronously rather
1068  * then delayed.
1069  */
1070 int
1071 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1072     int flags, int *rtvals)
1073 {
1074         int i;
1075         vm_object_t object;
1076         vm_page_t m;
1077         int count;
1078
1079         int maxsize, ncount;
1080         vm_ooffset_t poffset;
1081         struct uio auio;
1082         struct iovec aiov;
1083         int error;
1084         int ioflags;
1085         int ppscheck = 0;
1086         static struct timeval lastfail;
1087         static int curfail;
1088
1089         object = vp->v_object;
1090         count = bytecount / PAGE_SIZE;
1091
1092         for (i = 0; i < count; i++)
1093                 rtvals[i] = VM_PAGER_ERROR;
1094
1095         if ((int64_t)ma[0]->pindex < 0) {
1096                 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1097                     (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1098                 rtvals[0] = VM_PAGER_BAD;
1099                 return VM_PAGER_BAD;
1100         }
1101
1102         maxsize = count * PAGE_SIZE;
1103         ncount = count;
1104
1105         poffset = IDX_TO_OFF(ma[0]->pindex);
1106
1107         /*
1108          * If the page-aligned write is larger then the actual file we
1109          * have to invalidate pages occuring beyond the file EOF.  However,
1110          * there is an edge case where a file may not be page-aligned where
1111          * the last page is partially invalid.  In this case the filesystem
1112          * may not properly clear the dirty bits for the entire page (which
1113          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1114          * With the page locked we are free to fix-up the dirty bits here.
1115          *
1116          * We do not under any circumstances truncate the valid bits, as
1117          * this will screw up bogus page replacement.
1118          */
1119         VM_OBJECT_LOCK(object);
1120         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1121                 if (object->un_pager.vnp.vnp_size > poffset) {
1122                         int pgoff;
1123
1124                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1125                         ncount = btoc(maxsize);
1126                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1127                                 /*
1128                                  * If the object is locked and the following
1129                                  * conditions hold, then the page's dirty
1130                                  * field cannot be concurrently changed by a
1131                                  * pmap operation.
1132                                  */
1133                                 m = ma[ncount - 1];
1134                                 KASSERT(m->busy > 0,
1135                 ("vnode_pager_generic_putpages: page %p is not busy", m));
1136                                 KASSERT((m->aflags & PGA_WRITEABLE) == 0,
1137                 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1138                                 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1139                                     pgoff);
1140                         }
1141                 } else {
1142                         maxsize = 0;
1143                         ncount = 0;
1144                 }
1145                 if (ncount < count) {
1146                         for (i = ncount; i < count; i++) {
1147                                 rtvals[i] = VM_PAGER_BAD;
1148                         }
1149                 }
1150         }
1151         VM_OBJECT_UNLOCK(object);
1152
1153         /*
1154          * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1155          * rather then a bdwrite() to prevent paging I/O from saturating
1156          * the buffer cache.  Dummy-up the sequential heuristic to cause
1157          * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1158          * the system decides how to cluster.
1159          */
1160         ioflags = IO_VMIO;
1161         if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1162                 ioflags |= IO_SYNC;
1163         else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1164                 ioflags |= IO_ASYNC;
1165         ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1166         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1167
1168         aiov.iov_base = (caddr_t) 0;
1169         aiov.iov_len = maxsize;
1170         auio.uio_iov = &aiov;
1171         auio.uio_iovcnt = 1;
1172         auio.uio_offset = poffset;
1173         auio.uio_segflg = UIO_NOCOPY;
1174         auio.uio_rw = UIO_WRITE;
1175         auio.uio_resid = maxsize;
1176         auio.uio_td = (struct thread *) 0;
1177         error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1178         PCPU_INC(cnt.v_vnodeout);
1179         PCPU_ADD(cnt.v_vnodepgsout, ncount);
1180
1181         if (error) {
1182                 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1183                         printf("vnode_pager_putpages: I/O error %d\n", error);
1184         }
1185         if (auio.uio_resid) {
1186                 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1187                         printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1188                             auio.uio_resid, (u_long)ma[0]->pindex);
1189         }
1190         for (i = 0; i < ncount; i++) {
1191                 rtvals[i] = VM_PAGER_OK;
1192         }
1193         return rtvals[0];
1194 }
1195
1196 void
1197 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1198 {
1199         vm_object_t obj;
1200         int i, pos;
1201
1202         if (written == 0)
1203                 return;
1204         obj = ma[0]->object;
1205         VM_OBJECT_LOCK(obj);
1206         for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1207                 if (pos < trunc_page(written)) {
1208                         rtvals[i] = VM_PAGER_OK;
1209                         vm_page_undirty(ma[i]);
1210                 } else {
1211                         /* Partially written page. */
1212                         rtvals[i] = VM_PAGER_AGAIN;
1213                         vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1214                 }
1215         }
1216         VM_OBJECT_UNLOCK(obj);
1217 }