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, vm_ooffset_t);
  87
  88 struct pagerops vnodepagerops = {
  89         .pgo_alloc =    vnode_pager_alloc,
  90         .pgo_dealloc =  vnode_pager_dealloc,
  91         .pgo_getpages = vnode_pager_getpages,
  92         .pgo_putpages = vnode_pager_putpages,
  93         .pgo_haspage =  vnode_pager_haspage,
  94 };
  95
  96 int vnode_pbuf_freecnt;
  97
  98 /* Create the VM system backing object for this vnode */
  99 int
 100 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
 101 {
 102         vm_object_t object;
 103         vm_ooffset_t size = isize;
 104         struct vattr va;
 105
 106         if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
 107                 return (0);
 108
 109         while ((object = vp->v_object) != NULL) {
 110                 VM_OBJECT_LOCK(object);
 111                 if (!(object->flags & OBJ_DEAD)) {
 112                         VM_OBJECT_UNLOCK(object);
 113                         return (0);
 114                 }
 115                 VOP_UNLOCK(vp, 0, td);
 116                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 117                 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0);
 118                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
 119         }
 120
 121         if (size == 0) {
 122                 if (vn_isdisk(vp, NULL)) {
 123                         size = IDX_TO_OFF(INT_MAX);
 124                 } else {
 125                         if (VOP_GETATTR(vp, &va, td->td_ucred, td) != 0)
 126                                 return (0);
 127                         size = va.va_size;
 128                 }
 129         }
 130
 131         object = vnode_pager_alloc(vp, size, 0, 0);
 132         /*
 133          * Dereference the reference we just created.  This assumes
 134          * that the object is associated with the vp.
 135          */
 136         VM_OBJECT_LOCK(object);
 137         object->ref_count--;
 138         VM_OBJECT_UNLOCK(object);
 139         vrele(vp);
 140
 141         KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
 142
 143         return (0);
 144 }
 145
 146 void
 147 vnode_destroy_vobject(struct vnode *vp)
 148 {
 149         struct vm_object *obj;
 150
 151         obj = vp->v_object;
 152         if (obj == NULL)
 153                 return;
 154         ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
 155         VM_OBJECT_LOCK(obj);
 156         if (obj->ref_count == 0) {
 157                 /*
 158                  * vclean() may be called twice. The first time
 159                  * removes the primary reference to the object,
 160                  * the second time goes one further and is a
 161                  * special-case to terminate the object.
 162                  *
 163                  * don't double-terminate the object
 164                  */
 165                 if ((obj->flags & OBJ_DEAD) == 0)
 166                         vm_object_terminate(obj);
 167                 else
 168                         VM_OBJECT_UNLOCK(obj);
 169         } else {
 170                 /*
 171                  * Woe to the process that tries to page now :-).
 172                  */
 173                 vm_pager_deallocate(obj);
 174                 VM_OBJECT_UNLOCK(obj);
 175         }
 176         vp->v_object = NULL;
 177 }
 178
 179
 180 /*
 181  * Allocate (or lookup) pager for a vnode.
 182  * Handle is a vnode pointer.
 183  *
 184  * MPSAFE
 185  */
 186 vm_object_t
 187 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
 188                   vm_ooffset_t offset)
 189 {
 190         vm_object_t object;
 191         struct vnode *vp;
 192
 193         /*
 194          * Pageout to vnode, no can do yet.
 195          */
 196         if (handle == NULL)
 197                 return (NULL);
 198
 199         vp = (struct vnode *) handle;
 200
 201         /*
 202          * If the object is being terminated, wait for it to
 203          * go away.
 204          */
 205 retry:
 206         while ((object = vp->v_object) != NULL) {
 207                 VM_OBJECT_LOCK(object);
 208                 if ((object->flags & OBJ_DEAD) == 0)
 209                         break;
 210                 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
 211                 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0);
 212         }
 213
 214         if (vp->v_usecount == 0)
 215                 panic("vnode_pager_alloc: no vnode reference");
 216
 217         if (object == NULL) {
 218                 /*
 219                  * Add an object of the appropriate size
 220                  */
 221                 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
 222
 223                 object->un_pager.vnp.vnp_size = size;
 224
 225                 object->handle = handle;
 226                 if (VFS_NEEDSGIANT(vp->v_mount))
 227                         vm_object_set_flag(object, OBJ_NEEDGIANT);
 228                 VI_LOCK(vp);
 229                 if (vp->v_object != NULL) {
 230                         /*
 231                          * Object has been created while we were sleeping
 232                          */
 233                         VI_UNLOCK(vp);
 234                         vm_object_destroy(object);
 235                         goto retry;
 236                 }
 237                 vp->v_object = object;
 238                 VI_UNLOCK(vp);
 239         } else {
 240                 object->ref_count++;
 241                 VM_OBJECT_UNLOCK(object);
 242         }
 243         vref(vp);
 244         return (object);
 245 }
 246
 247 /*
 248  *      The object must be locked.
 249  */
 250 static void
 251 vnode_pager_dealloc(object)
 252         vm_object_t object;
 253 {
 254         struct vnode *vp = object->handle;
 255
 256         if (vp == NULL)
 257                 panic("vnode_pager_dealloc: pager already dealloced");
 258
 259         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 260         vm_object_pip_wait(object, "vnpdea");
 261
 262         object->handle = NULL;
 263         object->type = OBJT_DEAD;
 264         if (object->flags & OBJ_DISCONNECTWNT) {
 265                 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
 266                 wakeup(object);
 267         }
 268         ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
 269         vp->v_object = NULL;
 270         vp->v_vflag &= ~VV_TEXT;
 271 }
 272
 273 static boolean_t
 274 vnode_pager_haspage(object, pindex, before, after)
 275         vm_object_t object;
 276         vm_pindex_t pindex;
 277         int *before;
 278         int *after;
 279 {
 280         struct vnode *vp = object->handle;
 281         daddr_t bn;
 282         int err;
 283         daddr_t reqblock;
 284         int poff;
 285         int bsize;
 286         int pagesperblock, blocksperpage;
 287         int vfslocked;
 288
 289         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 290         /*
 291          * If no vp or vp is doomed or marked transparent to VM, we do not
 292          * have the page.
 293          */
 294         if (vp == NULL || vp->v_iflag & VI_DOOMED)
 295                 return FALSE;
 296         /*
 297          * If the offset is beyond end of file we do
 298          * not have the page.
 299          */
 300         if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
 301                 return FALSE;
 302
 303         bsize = vp->v_mount->mnt_stat.f_iosize;
 304         pagesperblock = bsize / PAGE_SIZE;
 305         blocksperpage = 0;
 306         if (pagesperblock > 0) {
 307                 reqblock = pindex / pagesperblock;
 308         } else {
 309                 blocksperpage = (PAGE_SIZE / bsize);
 310                 reqblock = pindex * blocksperpage;
 311         }
 312         VM_OBJECT_UNLOCK(object);
 313         vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 314         err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
 315         VFS_UNLOCK_GIANT(vfslocked);
 316         VM_OBJECT_LOCK(object);
 317         if (err)
 318                 return TRUE;
 319         if (bn == -1)
 320                 return FALSE;
 321         if (pagesperblock > 0) {
 322                 poff = pindex - (reqblock * pagesperblock);
 323                 if (before) {
 324                         *before *= pagesperblock;
 325                         *before += poff;
 326                 }
 327                 if (after) {
 328                         int numafter;
 329                         *after *= pagesperblock;
 330                         numafter = pagesperblock - (poff + 1);
 331                         if (IDX_TO_OFF(pindex + numafter) >
 332                             object->un_pager.vnp.vnp_size) {
 333                                 numafter =
 334                                     OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
 335                                     pindex;
 336                         }
 337                         *after += numafter;
 338                 }
 339         } else {
 340                 if (before) {
 341                         *before /= blocksperpage;
 342                 }
 343
 344                 if (after) {
 345                         *after /= blocksperpage;
 346                 }
 347         }
 348         return TRUE;
 349 }
 350
 351 /*
 352  * Lets the VM system know about a change in size for a file.
 353  * We adjust our own internal size and flush any cached pages in
 354  * the associated object that are affected by the size change.
 355  *
 356  * Note: this routine may be invoked as a result of a pager put
 357  * operation (possibly at object termination time), so we must be careful.
 358  */
 359 void
 360 vnode_pager_setsize(vp, nsize)
 361         struct vnode *vp;
 362         vm_ooffset_t nsize;
 363 {
 364         vm_object_t object;
 365         vm_page_t m;
 366         vm_pindex_t nobjsize;
 367
 368         if ((object = vp->v_object) == NULL)
 369                 return;
 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                          * Clear out partial-page dirty bits.  This
 408                          * has the side effect of setting the valid
 409                          * bits, but that is ok.  There are a bunch
 410                          * of places in the VM system where we expected
 411                          * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
 412                          * case is one of them.  If the page is still
 413                          * partially dirty, make it fully dirty.
 414                          *
 415                          * note that we do not clear out the valid
 416                          * bits.  This would prevent bogus_page
 417                          * replacement from working properly.
 418                          */
 419                         vm_page_lock_queues();
 420                         vm_page_set_validclean(m, base, size);
 421                         if (m->dirty != 0)
 422                                 m->dirty = VM_PAGE_BITS_ALL;
 423                         vm_page_unlock_queues();
 424                 } else if ((nsize & PAGE_MASK) &&
 425                     __predict_false(object->cache != NULL)) {
 426                         vm_page_cache_free(object, OFF_TO_IDX(nsize),
 427                             nobjsize);
 428                 }
 429         }
 430         object->un_pager.vnp.vnp_size = nsize;
 431         object->size = nobjsize;
 432         VM_OBJECT_UNLOCK(object);
 433 }
 434
 435 /*
 436  * calculate the linear (byte) disk address of specified virtual
 437  * file address
 438  */
 439 static int
 440 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
 441     int *run)
 442 {
 443         int bsize;
 444         int err;
 445         daddr_t vblock;
 446         daddr_t voffset;
 447
 448         if (address < 0)
 449                 return -1;
 450
 451         if (vp->v_iflag & VI_DOOMED)
 452                 return -1;
 453
 454         bsize = vp->v_mount->mnt_stat.f_iosize;
 455         vblock = address / bsize;
 456         voffset = address % bsize;
 457
 458         err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
 459         if (err == 0) {
 460                 if (*rtaddress != -1)
 461                         *rtaddress += voffset / DEV_BSIZE;
 462                 if (run) {
 463                         *run += 1;
 464                         *run *= bsize/PAGE_SIZE;
 465                         *run -= voffset/PAGE_SIZE;
 466                 }
 467         }
 468
 469         return (err);
 470 }
 471
 472 /*
 473  * small block filesystem vnode pager input
 474  */
 475 static int
 476 vnode_pager_input_smlfs(object, m)
 477         vm_object_t object;
 478         vm_page_t m;
 479 {
 480         int i;
 481         struct vnode *vp;
 482         struct bufobj *bo;
 483         struct buf *bp;
 484         struct sf_buf *sf;
 485         daddr_t fileaddr;
 486         vm_offset_t bsize;
 487         int error = 0;
 488
 489         vp = object->handle;
 490         if (vp->v_iflag & VI_DOOMED)
 491                 return VM_PAGER_BAD;
 492
 493         bsize = vp->v_mount->mnt_stat.f_iosize;
 494
 495         VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
 496
 497         sf = sf_buf_alloc(m, 0);
 498
 499         for (i = 0; i < PAGE_SIZE / bsize; i++) {
 500                 vm_ooffset_t address;
 501
 502                 if (vm_page_bits(i * bsize, bsize) & m->valid)
 503                         continue;
 504
 505                 address = IDX_TO_OFF(m->pindex) + i * bsize;
 506                 if (address >= object->un_pager.vnp.vnp_size) {
 507                         fileaddr = -1;
 508                 } else {
 509                         error = vnode_pager_addr(vp, address, &fileaddr, NULL);
 510                         if (error)
 511                                 break;
 512                 }
 513                 if (fileaddr != -1) {
 514                         bp = getpbuf(&vnode_pbuf_freecnt);
 515
 516                         /* build a minimal buffer header */
 517                         bp->b_iocmd = BIO_READ;
 518                         bp->b_iodone = bdone;
 519                         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 520                         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 521                         bp->b_rcred = crhold(curthread->td_ucred);
 522                         bp->b_wcred = crhold(curthread->td_ucred);
 523                         bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
 524                         bp->b_blkno = fileaddr;
 525                         pbgetbo(bo, bp);
 526                         bp->b_bcount = bsize;
 527                         bp->b_bufsize = bsize;
 528                         bp->b_runningbufspace = bp->b_bufsize;
 529                         atomic_add_int(&runningbufspace, bp->b_runningbufspace);
 530
 531                         /* do the input */
 532                         bp->b_iooffset = dbtob(bp->b_blkno);
 533                         bstrategy(bp);
 534
 535                         bwait(bp, PVM, "vnsrd");
 536
 537                         if ((bp->b_ioflags & BIO_ERROR) != 0)
 538                                 error = EIO;
 539
 540                         /*
 541                          * free the buffer header back to the swap buffer pool
 542                          */
 543                         pbrelbo(bp);
 544                         relpbuf(bp, &vnode_pbuf_freecnt);
 545                         if (error)
 546                                 break;
 547
 548                         VM_OBJECT_LOCK(object);
 549                         vm_page_lock_queues();
 550                         vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
 551                         vm_page_unlock_queues();
 552                         VM_OBJECT_UNLOCK(object);
 553                 } else {
 554                         VM_OBJECT_LOCK(object);
 555                         vm_page_lock_queues();
 556                         vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
 557                         vm_page_unlock_queues();
 558                         VM_OBJECT_UNLOCK(object);
 559                         bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
 560                 }
 561         }
 562         sf_buf_free(sf);
 563         vm_page_lock_queues();
 564         pmap_clear_modify(m);
 565         vm_page_unlock_queues();
 566         if (error) {
 567                 return VM_PAGER_ERROR;
 568         }
 569         return VM_PAGER_OK;
 570
 571 }
 572
 573
 574 /*
 575  * old style vnode pager input routine
 576  */
 577 static int
 578 vnode_pager_input_old(object, m)
 579         vm_object_t object;
 580         vm_page_t m;
 581 {
 582         struct uio auio;
 583         struct iovec aiov;
 584         int error;
 585         int size;
 586         struct sf_buf *sf;
 587         struct vnode *vp;
 588
 589         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 590         error = 0;
 591
 592         /*
 593          * Return failure if beyond current EOF
 594          */
 595         if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
 596                 return VM_PAGER_BAD;
 597         } else {
 598                 size = PAGE_SIZE;
 599                 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
 600                         size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
 601                 vp = object->handle;
 602                 VM_OBJECT_UNLOCK(object);
 603
 604                 /*
 605                  * Allocate a kernel virtual address and initialize so that
 606                  * we can use VOP_READ/WRITE routines.
 607                  */
 608                 sf = sf_buf_alloc(m, 0);
 609
 610                 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
 611                 aiov.iov_len = size;
 612                 auio.uio_iov = &aiov;
 613                 auio.uio_iovcnt = 1;
 614                 auio.uio_offset = IDX_TO_OFF(m->pindex);
 615                 auio.uio_segflg = UIO_SYSSPACE;
 616                 auio.uio_rw = UIO_READ;
 617                 auio.uio_resid = size;
 618                 auio.uio_td = curthread;
 619
 620                 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
 621                 if (!error) {
 622                         int count = size - auio.uio_resid;
 623
 624                         if (count == 0)
 625                                 error = EINVAL;
 626                         else if (count != PAGE_SIZE)
 627                                 bzero((caddr_t)sf_buf_kva(sf) + count,
 628                                     PAGE_SIZE - count);
 629                 }
 630                 sf_buf_free(sf);
 631
 632                 VM_OBJECT_LOCK(object);
 633         }
 634         vm_page_lock_queues();
 635         pmap_clear_modify(m);
 636         vm_page_undirty(m);
 637         vm_page_unlock_queues();
 638         if (!error)
 639                 m->valid = VM_PAGE_BITS_ALL;
 640         return error ? VM_PAGER_ERROR : VM_PAGER_OK;
 641 }
 642
 643 /*
 644  * generic vnode pager input routine
 645  */
 646
 647 /*
 648  * Local media VFS's that do not implement their own VOP_GETPAGES
 649  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
 650  * to implement the previous behaviour.
 651  *
 652  * All other FS's should use the bypass to get to the local media
 653  * backing vp's VOP_GETPAGES.
 654  */
 655 static int
 656 vnode_pager_getpages(object, m, count, reqpage)
 657         vm_object_t object;
 658         vm_page_t *m;
 659         int count;
 660         int reqpage;
 661 {
 662         int rtval;
 663         struct vnode *vp;
 664         int bytes = count * PAGE_SIZE;
 665         int vfslocked;
 666
 667         vp = object->handle;
 668         VM_OBJECT_UNLOCK(object);
 669         vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 670         rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
 671         KASSERT(rtval != EOPNOTSUPP,
 672             ("vnode_pager: FS getpages not implemented\n"));
 673         VFS_UNLOCK_GIANT(vfslocked);
 674         VM_OBJECT_LOCK(object);
 675         return rtval;
 676 }
 677
 678 /*
 679  * This is now called from local media FS's to operate against their
 680  * own vnodes if they fail to implement VOP_GETPAGES.
 681  */
 682 int
 683 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
 684         struct vnode *vp;
 685         vm_page_t *m;
 686         int bytecount;
 687         int reqpage;
 688 {
 689         vm_object_t object;
 690         vm_offset_t kva;
 691         off_t foff, tfoff, nextoff;
 692         int i, j, size, bsize, first;
 693         daddr_t firstaddr, reqblock;
 694         struct bufobj *bo;
 695         int runpg;
 696         int runend;
 697         struct buf *bp;
 698         int count;
 699         int error;
 700
 701         object = vp->v_object;
 702         count = bytecount / PAGE_SIZE;
 703
 704         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 705             ("vnode_pager_generic_getpages does not support devices"));
 706         if (vp->v_iflag & VI_DOOMED)
 707                 return VM_PAGER_BAD;
 708
 709         bsize = vp->v_mount->mnt_stat.f_iosize;
 710
 711         /* get the UNDERLYING device for the file with VOP_BMAP() */
 712
 713         /*
 714          * originally, we did not check for an error return value -- assuming
 715          * an fs always has a bmap entry point -- that assumption is wrong!!!
 716          */
 717         foff = IDX_TO_OFF(m[reqpage]->pindex);
 718
 719         /*
 720          * if we can't bmap, use old VOP code
 721          */
 722         error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL);
 723         if (error == EOPNOTSUPP) {
 724                 VM_OBJECT_LOCK(object);
 725                 vm_page_lock_queues();
 726                 for (i = 0; i < count; i++)
 727                         if (i != reqpage)
 728                                 vm_page_free(m[i]);
 729                 vm_page_unlock_queues();
 730                 PCPU_INC(cnt.v_vnodein);
 731                 PCPU_INC(cnt.v_vnodepgsin);
 732                 error = vnode_pager_input_old(object, m[reqpage]);
 733                 VM_OBJECT_UNLOCK(object);
 734                 return (error);
 735         } else if (error != 0) {
 736                 VM_OBJECT_LOCK(object);
 737                 vm_page_lock_queues();
 738                 for (i = 0; i < count; i++)
 739                         if (i != reqpage)
 740                                 vm_page_free(m[i]);
 741                 vm_page_unlock_queues();
 742                 VM_OBJECT_UNLOCK(object);
 743                 return (VM_PAGER_ERROR);
 744
 745                 /*
 746                  * if the blocksize is smaller than a page size, then use
 747                  * special small filesystem code.  NFS sometimes has a small
 748                  * blocksize, but it can handle large reads itself.
 749                  */
 750         } else if ((PAGE_SIZE / bsize) > 1 &&
 751             (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
 752                 VM_OBJECT_LOCK(object);
 753                 vm_page_lock_queues();
 754                 for (i = 0; i < count; i++)
 755                         if (i != reqpage)
 756                                 vm_page_free(m[i]);
 757                 vm_page_unlock_queues();
 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                 vm_page_lock_queues();
 772                 for (i = 0; i < count; i++)
 773                         if (i != reqpage)
 774                                 vm_page_free(m[i]);
 775                 vm_page_unlock_queues();
 776                 VM_OBJECT_UNLOCK(object);
 777                 return VM_PAGER_OK;
 778         } else if (reqblock == -1) {
 779                 pmap_zero_page(m[reqpage]);
 780                 vm_page_undirty(m[reqpage]);
 781                 m[reqpage]->valid = VM_PAGE_BITS_ALL;
 782                 vm_page_lock_queues();
 783                 for (i = 0; i < count; i++)
 784                         if (i != reqpage)
 785                                 vm_page_free(m[i]);
 786                 vm_page_unlock_queues();
 787                 VM_OBJECT_UNLOCK(object);
 788                 return (VM_PAGER_OK);
 789         }
 790         m[reqpage]->valid = 0;
 791         VM_OBJECT_UNLOCK(object);
 792
 793         /*
 794          * here on direct device I/O
 795          */
 796         firstaddr = -1;
 797
 798         /*
 799          * calculate the run that includes the required page
 800          */
 801         for (first = 0, i = 0; i < count; i = runend) {
 802                 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr,
 803                     &runpg) != 0) {
 804                         VM_OBJECT_LOCK(object);
 805                         vm_page_lock_queues();
 806                         for (; i < count; i++)
 807                                 if (i != reqpage)
 808                                         vm_page_free(m[i]);
 809                         vm_page_unlock_queues();
 810                         VM_OBJECT_UNLOCK(object);
 811                         return (VM_PAGER_ERROR);
 812                 }
 813                 if (firstaddr == -1) {
 814                         VM_OBJECT_LOCK(object);
 815                         if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
 816                                 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
 817                                     (intmax_t)firstaddr, (uintmax_t)(foff >> 32),
 818                                     (uintmax_t)foff,
 819                                     (uintmax_t)
 820                                     (object->un_pager.vnp.vnp_size >> 32),
 821                                     (uintmax_t)object->un_pager.vnp.vnp_size);
 822                         }
 823                         vm_page_lock_queues();
 824                         vm_page_free(m[i]);
 825                         vm_page_unlock_queues();
 826                         VM_OBJECT_UNLOCK(object);
 827                         runend = i + 1;
 828                         first = runend;
 829                         continue;
 830                 }
 831                 runend = i + runpg;
 832                 if (runend <= reqpage) {
 833                         VM_OBJECT_LOCK(object);
 834                         vm_page_lock_queues();
 835                         for (j = i; j < runend; j++)
 836                                 vm_page_free(m[j]);
 837                         vm_page_unlock_queues();
 838                         VM_OBJECT_UNLOCK(object);
 839                 } else {
 840                         if (runpg < (count - first)) {
 841                                 VM_OBJECT_LOCK(object);
 842                                 vm_page_lock_queues();
 843                                 for (i = first + runpg; i < count; i++)
 844                                         vm_page_free(m[i]);
 845                                 vm_page_unlock_queues();
 846                                 VM_OBJECT_UNLOCK(object);
 847                                 count = first + runpg;
 848                         }
 849                         break;
 850                 }
 851                 first = runend;
 852         }
 853
 854         /*
 855          * the first and last page have been calculated now, move input pages
 856          * to be zero based...
 857          */
 858         if (first != 0) {
 859                 m += first;
 860                 count -= first;
 861                 reqpage -= first;
 862         }
 863
 864         /*
 865          * calculate the file virtual address for the transfer
 866          */
 867         foff = IDX_TO_OFF(m[0]->pindex);
 868
 869         /*
 870          * calculate the size of the transfer
 871          */
 872         size = count * PAGE_SIZE;
 873         KASSERT(count > 0, ("zero count"));
 874         if ((foff + size) > object->un_pager.vnp.vnp_size)
 875                 size = object->un_pager.vnp.vnp_size - foff;
 876         KASSERT(size > 0, ("zero size"));
 877
 878         /*
 879          * round up physical size for real devices.
 880          */
 881         if (1) {
 882                 int secmask = bo->bo_bsize - 1;
 883                 KASSERT(secmask < PAGE_SIZE && secmask > 0,
 884                     ("vnode_pager_generic_getpages: sector size %d too large",
 885                     secmask + 1));
 886                 size = (size + secmask) & ~secmask;
 887         }
 888
 889         bp = getpbuf(&vnode_pbuf_freecnt);
 890         kva = (vm_offset_t) bp->b_data;
 891
 892         /*
 893          * and map the pages to be read into the kva
 894          */
 895         pmap_qenter(kva, m, count);
 896
 897         /* build a minimal buffer header */
 898         bp->b_iocmd = BIO_READ;
 899         bp->b_iodone = bdone;
 900         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 901         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 902         bp->b_rcred = crhold(curthread->td_ucred);
 903         bp->b_wcred = crhold(curthread->td_ucred);
 904         bp->b_blkno = firstaddr;
 905         pbgetbo(bo, bp);
 906         bp->b_bcount = size;
 907         bp->b_bufsize = size;
 908         bp->b_runningbufspace = bp->b_bufsize;
 909         atomic_add_int(&runningbufspace, bp->b_runningbufspace);
 910
 911         PCPU_INC(cnt.v_vnodein);
 912         PCPU_ADD(cnt.v_vnodepgsin, count);
 913
 914         /* do the input */
 915         bp->b_iooffset = dbtob(bp->b_blkno);
 916         bstrategy(bp);
 917
 918         bwait(bp, PVM, "vnread");
 919
 920         if ((bp->b_ioflags & BIO_ERROR) != 0)
 921                 error = EIO;
 922
 923         if (!error) {
 924                 if (size != count * PAGE_SIZE)
 925                         bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
 926         }
 927         pmap_qremove(kva, count);
 928
 929         /*
 930          * free the buffer header back to the swap buffer pool
 931          */
 932         pbrelbo(bp);
 933         relpbuf(bp, &vnode_pbuf_freecnt);
 934
 935         VM_OBJECT_LOCK(object);
 936         vm_page_lock_queues();
 937         for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
 938                 vm_page_t mt;
 939
 940                 nextoff = tfoff + PAGE_SIZE;
 941                 mt = m[i];
 942
 943                 if (nextoff <= object->un_pager.vnp.vnp_size) {
 944                         /*
 945                          * Read filled up entire page.
 946                          */
 947                         mt->valid = VM_PAGE_BITS_ALL;
 948                         vm_page_undirty(mt);    /* should be an assert? XXX */
 949                         pmap_clear_modify(mt);
 950                 } else {
 951                         /*
 952                          * Read did not fill up entire page.  Since this
 953                          * is getpages, the page may be mapped, so we have
 954                          * to zero the invalid portions of the page even
 955                          * though we aren't setting them valid.
 956                          *
 957                          * Currently we do not set the entire page valid,
 958                          * we just try to clear the piece that we couldn't
 959                          * read.
 960                          */
 961                         vm_page_set_validclean(mt, 0,
 962                             object->un_pager.vnp.vnp_size - tfoff);
 963                         /* handled by vm_fault now */
 964                         /* vm_page_zero_invalid(mt, FALSE); */
 965                 }
 966
 967                 if (i != reqpage) {
 968
 969                         /*
 970                          * whether or not to leave the page activated is up in
 971                          * the air, but we should put the page on a page queue
 972                          * somewhere. (it already is in the object). Result:
 973                          * It appears that empirical results show that
 974                          * deactivating pages is best.
 975                          */
 976
 977                         /*
 978                          * just in case someone was asking for this page we
 979                          * now tell them that it is ok to use
 980                          */
 981                         if (!error) {
 982                                 if (mt->oflags & VPO_WANTED)
 983                                         vm_page_activate(mt);
 984                                 else
 985                                         vm_page_deactivate(mt);
 986                                 vm_page_wakeup(mt);
 987                         } else {
 988                                 vm_page_free(mt);
 989                         }
 990                 }
 991         }
 992         vm_page_unlock_queues();
 993         VM_OBJECT_UNLOCK(object);
 994         if (error) {
 995                 printf("vnode_pager_getpages: I/O read error\n");
 996         }
 997         return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
 998 }
 999
1000 /*
1001  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1002  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1003  * vnode_pager_generic_putpages() to implement the previous behaviour.
1004  *
1005  * All other FS's should use the bypass to get to the local media
1006  * backing vp's VOP_PUTPAGES.
1007  */
1008 static void
1009 vnode_pager_putpages(object, m, count, sync, rtvals)
1010         vm_object_t object;
1011         vm_page_t *m;
1012         int count;
1013         boolean_t sync;
1014         int *rtvals;
1015 {
1016         int rtval;
1017         struct vnode *vp;
1018         struct mount *mp;
1019         int bytes = count * PAGE_SIZE;
1020
1021         /*
1022          * Force synchronous operation if we are extremely low on memory
1023          * to prevent a low-memory deadlock.  VOP operations often need to
1024          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1025          * operation ).  The swapper handles the case by limiting the amount
1026          * of asynchronous I/O, but that sort of solution doesn't scale well
1027          * for the vnode pager without a lot of work.
1028          *
1029          * Also, the backing vnode's iodone routine may not wake the pageout
1030          * daemon up.  This should be probably be addressed XXX.
1031          */
1032
1033         if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
1034                 sync |= OBJPC_SYNC;
1035
1036         /*
1037          * Call device-specific putpages function
1038          */
1039         vp = object->handle;
1040         VM_OBJECT_UNLOCK(object);
1041         if (vp->v_type != VREG)
1042                 mp = NULL;
1043         rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
1044         KASSERT(rtval != EOPNOTSUPP,
1045             ("vnode_pager: stale FS putpages\n"));
1046         VM_OBJECT_LOCK(object);
1047 }
1048
1049
1050 /*
1051  * This is now called from local media FS's to operate against their
1052  * own vnodes if they fail to implement VOP_PUTPAGES.
1053  *
1054  * This is typically called indirectly via the pageout daemon and
1055  * clustering has already typically occured, so in general we ask the
1056  * underlying filesystem to write the data out asynchronously rather
1057  * then delayed.
1058  */
1059 int
1060 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
1061         struct vnode *vp;
1062         vm_page_t *m;
1063         int bytecount;
1064         int flags;
1065         int *rtvals;
1066 {
1067         int i;
1068         vm_object_t object;
1069         int count;
1070
1071         int maxsize, ncount;
1072         vm_ooffset_t poffset;
1073         struct uio auio;
1074         struct iovec aiov;
1075         int error;
1076         int ioflags;
1077         int ppscheck = 0;
1078         static struct timeval lastfail;
1079         static int curfail;
1080
1081         object = vp->v_object;
1082         count = bytecount / PAGE_SIZE;
1083
1084         for (i = 0; i < count; i++)
1085                 rtvals[i] = VM_PAGER_AGAIN;
1086
1087         if ((int64_t)m[0]->pindex < 0) {
1088                 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1089                         (long)m[0]->pindex, (u_long)m[0]->dirty);
1090                 rtvals[0] = VM_PAGER_BAD;
1091                 return VM_PAGER_BAD;
1092         }
1093
1094         maxsize = count * PAGE_SIZE;
1095         ncount = count;
1096
1097         poffset = IDX_TO_OFF(m[0]->pindex);
1098
1099         /*
1100          * If the page-aligned write is larger then the actual file we
1101          * have to invalidate pages occuring beyond the file EOF.  However,
1102          * there is an edge case where a file may not be page-aligned where
1103          * the last page is partially invalid.  In this case the filesystem
1104          * may not properly clear the dirty bits for the entire page (which
1105          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1106          * With the page locked we are free to fix-up the dirty bits here.
1107          *
1108          * We do not under any circumstances truncate the valid bits, as
1109          * this will screw up bogus page replacement.
1110          */
1111         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1112                 if (object->un_pager.vnp.vnp_size > poffset) {
1113                         int pgoff;
1114
1115                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1116                         ncount = btoc(maxsize);
1117                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1118                                 vm_page_lock_queues();
1119                                 vm_page_clear_dirty(m[ncount - 1], pgoff,
1120                                         PAGE_SIZE - pgoff);
1121                                 vm_page_unlock_queues();
1122                         }
1123                 } else {
1124                         maxsize = 0;
1125                         ncount = 0;
1126                 }
1127                 if (ncount < count) {
1128                         for (i = ncount; i < count; i++) {
1129                                 rtvals[i] = VM_PAGER_BAD;
1130                         }
1131                 }
1132         }
1133
1134         /*
1135          * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1136          * rather then a bdwrite() to prevent paging I/O from saturating
1137          * the buffer cache.  Dummy-up the sequential heuristic to cause
1138          * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1139          * the system decides how to cluster.
1140          */
1141         ioflags = IO_VMIO;
1142         if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1143                 ioflags |= IO_SYNC;
1144         else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1145                 ioflags |= IO_ASYNC;
1146         ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1147         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1148
1149         aiov.iov_base = (caddr_t) 0;
1150         aiov.iov_len = maxsize;
1151         auio.uio_iov = &aiov;
1152         auio.uio_iovcnt = 1;
1153         auio.uio_offset = poffset;
1154         auio.uio_segflg = UIO_NOCOPY;
1155         auio.uio_rw = UIO_WRITE;
1156         auio.uio_resid = maxsize;
1157         auio.uio_td = (struct thread *) 0;
1158         error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1159         PCPU_INC(cnt.v_vnodeout);
1160         PCPU_ADD(cnt.v_vnodepgsout, ncount);
1161
1162         if (error) {
1163                 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1164                         printf("vnode_pager_putpages: I/O error %d\n", error);
1165         }
1166         if (auio.uio_resid) {
1167                 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1168                         printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1169                             auio.uio_resid, (u_long)m[0]->pindex);
1170         }
1171         for (i = 0; i < ncount; i++) {
1172                 rtvals[i] = VM_PAGER_OK;
1173         }
1174         return rtvals[0];
1175 }
1176
1177 struct vnode *
1178 vnode_pager_lock(vm_object_t first_object)
1179 {
1180         struct vnode *vp;
1181         vm_object_t backing_object, object;
1182
1183         VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
1184         for (object = first_object; object != NULL; object = backing_object) {
1185                 if (object->type != OBJT_VNODE) {
1186                         if ((backing_object = object->backing_object) != NULL)
1187                                 VM_OBJECT_LOCK(backing_object);
1188                         if (object != first_object)
1189                                 VM_OBJECT_UNLOCK(object);
1190                         continue;
1191                 }
1192         retry:
1193                 if (object->flags & OBJ_DEAD) {
1194                         if (object != first_object)
1195                                 VM_OBJECT_UNLOCK(object);
1196                         return NULL;
1197                 }
1198                 vp = object->handle;
1199                 VI_LOCK(vp);
1200                 VM_OBJECT_UNLOCK(object);
1201                 if (first_object != object)
1202                         VM_OBJECT_UNLOCK(first_object);
1203                 VFS_ASSERT_GIANT(vp->v_mount);
1204                 if (vget(vp, LK_CANRECURSE | LK_INTERLOCK |
1205                     LK_RETRY | LK_SHARED, curthread)) {
1206                         VM_OBJECT_LOCK(first_object);
1207                         if (object != first_object)
1208                                 VM_OBJECT_LOCK(object);
1209                         if (object->type != OBJT_VNODE) {
1210                                 if (object != first_object)
1211                                         VM_OBJECT_UNLOCK(object);
1212                                 return NULL;
1213                         }
1214                         printf("vnode_pager_lock: retrying\n");
1215                         goto retry;
1216                 }
1217                 VM_OBJECT_LOCK(first_object);
1218                 return (vp);
1219         }
1220         return NULL;
1221 }