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  * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
  42  * $DragonFly: src/sys/vm/vnode_pager.c,v 1.43 2008/06/19 23:27:39 dillon Exp $
  43  */
  44
  45 /*
  46  * Page to/from files (vnodes).
  47  */
  48
  49 /*
  50  * TODO:
  51  *      Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
  52  *      greatly re-simplify the vnode_pager.
  53  */
  54
  55 #include <sys/param.h>
  56 #include <sys/systm.h>
  57 #include <sys/kernel.h>
  58 #include <sys/proc.h>
  59 #include <sys/vnode.h>
  60 #include <sys/mount.h>
  61 #include <sys/buf.h>
  62 #include <sys/vmmeter.h>
  63 #include <sys/conf.h>
  64 #include <sys/sfbuf.h>
  65
  66 #include <vm/vm.h>
  67 #include <vm/vm_object.h>
  68 #include <vm/vm_page.h>
  69 #include <vm/vm_pager.h>
  70 #include <vm/vm_map.h>
  71 #include <vm/vnode_pager.h>
  72 #include <vm/vm_extern.h>
  73
  74 #include <sys/thread2.h>
  75 #include <vm/vm_page2.h>
  76
  77 static void vnode_pager_dealloc (vm_object_t);
  78 static int vnode_pager_getpages (vm_object_t, vm_page_t *, int, int);
  79 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
  80 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t, int *, int *);
  81
  82 struct pagerops vnodepagerops = {
  83         NULL,
  84         vnode_pager_alloc,
  85         vnode_pager_dealloc,
  86         vnode_pager_getpages,
  87         vnode_pager_putpages,
  88         vnode_pager_haspage,
  89         NULL
  90 };
  91
  92 static struct krate vbadrate = { 1 };
  93 static struct krate vresrate = { 1 };
  94
  95 int vnode_pbuf_freecnt = -1;    /* start out unlimited */
  96
  97 /*
  98  * Allocate (or lookup) pager for a vnode.
  99  * Handle is a vnode pointer.
 100  */
 101 vm_object_t
 102 vnode_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
 103 {
 104         vm_object_t object;
 105         struct vnode *vp;
 106
 107         /*
 108          * Pageout to vnode, no can do yet.
 109          */
 110         if (handle == NULL)
 111                 return (NULL);
 112
 113         /*
 114          * XXX hack - This initialization should be put somewhere else.
 115          */
 116         if (vnode_pbuf_freecnt < 0) {
 117             vnode_pbuf_freecnt = nswbuf / 2 + 1;
 118         }
 119
 120         vp = (struct vnode *) handle;
 121
 122         /*
 123          * Prevent race condition when allocating the object. This
 124          * can happen with NFS vnodes since the nfsnode isn't locked.
 125          */
 126         while (vp->v_flag & VOLOCK) {
 127                 vp->v_flag |= VOWANT;
 128                 tsleep(vp, 0, "vnpobj", 0);
 129         }
 130         vp->v_flag |= VOLOCK;
 131
 132         /*
 133          * If the object is being terminated, wait for it to
 134          * go away.
 135          */
 136         while (((object = vp->v_object) != NULL) &&
 137                 (object->flags & OBJ_DEAD)) {
 138                 vm_object_dead_sleep(object, "vadead");
 139         }
 140
 141         if (vp->v_sysref.refcnt <= 0)
 142                 panic("vnode_pager_alloc: no vnode reference");
 143
 144         if (object == NULL) {
 145                 /*
 146                  * And an object of the appropriate size
 147                  */
 148                 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
 149                 object->flags = 0;
 150                 object->handle = handle;
 151                 vp->v_object = object;
 152                 vp->v_filesize = size;
 153         } else {
 154                 object->ref_count++;
 155                 if (vp->v_filesize != size) {
 156                         kprintf("vnode_pager_alloc: Warning, filesize "
 157                                 "mismatch %lld/%lld\n",
 158                                 (long long)vp->v_filesize,
 159                                 (long long)size);
 160                 }
 161         }
 162         vref(vp);
 163
 164         vp->v_flag &= ~VOLOCK;
 165         if (vp->v_flag & VOWANT) {
 166                 vp->v_flag &= ~VOWANT;
 167                 wakeup(vp);
 168         }
 169         return (object);
 170 }
 171
 172 static void
 173 vnode_pager_dealloc(vm_object_t object)
 174 {
 175         struct vnode *vp = object->handle;
 176
 177         if (vp == NULL)
 178                 panic("vnode_pager_dealloc: pager already dealloced");
 179
 180         vm_object_pip_wait(object, "vnpdea");
 181
 182         object->handle = NULL;
 183         object->type = OBJT_DEAD;
 184         vp->v_object = NULL;
 185         vp->v_filesize = NOOFFSET;
 186         vp->v_flag &= ~(VTEXT | VOBJBUF);
 187 }
 188
 189 /*
 190  * Return whether the vnode pager has the requested page.  Return the
 191  * number of disk-contiguous pages before and after the requested page,
 192  * not including the requested page.
 193  */
 194 static boolean_t
 195 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
 196                     int *after)
 197 {
 198         struct vnode *vp = object->handle;
 199         off_t loffset;
 200         off_t doffset;
 201         int voff;
 202         int bsize;
 203         int error;
 204
 205         /*
 206          * If no vp or vp is doomed or marked transparent to VM, we do not
 207          * have the page.
 208          */
 209         if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
 210                 return FALSE;
 211
 212         /*
 213          * If filesystem no longer mounted or offset beyond end of file we do
 214          * not have the page.
 215          */
 216         loffset = IDX_TO_OFF(pindex);
 217
 218         if (vp->v_mount == NULL || loffset >= vp->v_filesize)
 219                 return FALSE;
 220
 221         bsize = vp->v_mount->mnt_stat.f_iosize;
 222         voff = loffset % bsize;
 223
 224         /*
 225          * BMAP returns byte counts before and after, where after
 226          * is inclusive of the base page.  haspage must return page
 227          * counts before and after where after does not include the
 228          * base page.
 229          *
 230          * BMAP is allowed to return a *after of 0 for backwards
 231          * compatibility.  The base page is still considered valid if
 232          * no error is returned.
 233          */
 234         error = VOP_BMAP(vp, loffset - voff, &doffset, after, before, 0);
 235         if (error) {
 236                 if (before)
 237                         *before = 0;
 238                 if (after)
 239                         *after = 0;
 240                 return TRUE;
 241         }
 242         if (doffset == NOOFFSET)
 243                 return FALSE;
 244
 245         if (before) {
 246                 *before = (*before + voff) >> PAGE_SHIFT;
 247         }
 248         if (after) {
 249                 *after -= voff;
 250                 if (loffset + *after > vp->v_filesize)
 251                         *after = vp->v_filesize - loffset;
 252                 *after >>= PAGE_SHIFT;
 253                 if (*after < 0)
 254                         *after = 0;
 255         }
 256         return TRUE;
 257 }
 258
 259 /*
 260  * Lets the VM system know about a change in size for a file.
 261  * We adjust our own internal size and flush any cached pages in
 262  * the associated object that are affected by the size change.
 263  *
 264  * NOTE: This routine may be invoked as a result of a pager put
 265  * operation (possibly at object termination time), so we must be careful.
 266  *
 267  * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
 268  * we do not blow up on the case.  nsize will always be >= 0, however.
 269  */
 270 void
 271 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
 272 {
 273         vm_pindex_t nobjsize;
 274         vm_pindex_t oobjsize;
 275         vm_object_t object = vp->v_object;
 276
 277         if (object == NULL)
 278                 return;
 279
 280         /*
 281          * Hasn't changed size
 282          */
 283         if (nsize == vp->v_filesize)
 284                 return;
 285
 286         /*
 287          * Has changed size.  Adjust the VM object's size and v_filesize
 288          * before we start scanning pages to prevent new pages from being
 289          * allocated during the scan.
 290          */
 291         nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
 292         oobjsize = object->size;
 293         object->size = nobjsize;
 294
 295         /*
 296          * File has shrunk. Toss any cached pages beyond the new EOF.
 297          */
 298         if (nsize < vp->v_filesize) {
 299                 vp->v_filesize = nsize;
 300                 if (nobjsize < oobjsize) {
 301                         vm_object_page_remove(object, nobjsize, oobjsize,
 302                                               FALSE);
 303                 }
 304                 /*
 305                  * This gets rid of garbage at the end of a page that is now
 306                  * only partially backed by the vnode.  Since we are setting
 307                  * the entire page valid & clean after we are done we have
 308                  * to be sure that the portion of the page within the file
 309                  * bounds is already valid.  If it isn't then making it
 310                  * valid would create a corrupt block.
 311                  */
 312                 if (nsize & PAGE_MASK) {
 313                         vm_offset_t kva;
 314                         vm_page_t m;
 315
 316                         do {
 317                                 m = vm_page_lookup(object, OFF_TO_IDX(nsize));
 318                         } while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
 319
 320                         if (m && m->valid) {
 321                                 int base = (int)nsize & PAGE_MASK;
 322                                 int size = PAGE_SIZE - base;
 323                                 struct sf_buf *sf;
 324
 325                                 /*
 326                                  * Clear out partial-page garbage in case
 327                                  * the page has been mapped.
 328                                  *
 329                                  * This is byte aligned.
 330                                  */
 331                                 vm_page_busy(m);
 332                                 sf = sf_buf_alloc(m, SFB_CPUPRIVATE);
 333                                 kva = sf_buf_kva(sf);
 334                                 bzero((caddr_t)kva + base, size);
 335                                 sf_buf_free(sf);
 336
 337                                 /*
 338                                  * XXX work around SMP data integrity race
 339                                  * by unmapping the page from user processes.
 340                                  * The garbage we just cleared may be mapped
 341                                  * to a user process running on another cpu
 342                                  * and this code is not running through normal
 343                                  * I/O channels which handle SMP issues for
 344                                  * us, so unmap page to synchronize all cpus.
 345                                  *
 346                                  * XXX should vm_pager_unmap_page() have
 347                                  * dealt with this?
 348                                  */
 349                                 vm_page_protect(m, VM_PROT_NONE);
 350
 351                                 /*
 352                                  * Clear out partial-page dirty bits.  This
 353                                  * has the side effect of setting the valid
 354                                  * bits, but that is ok.  There are a bunch
 355                                  * of places in the VM system where we expected
 356                                  * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
 357                                  * case is one of them.  If the page is still
 358                                  * partially dirty, make it fully dirty.
 359                                  *
 360                                  * NOTE: We do not clear out the valid
 361                                  * bits.  This would prevent bogus_page
 362                                  * replacement from working properly.
 363                                  *
 364                                  * NOTE: We do not want to clear the dirty
 365                                  * bit for a partial DEV_BSIZE'd truncation!
 366                                  * This is DEV_BSIZE aligned!
 367                                  */
 368                                 vm_page_clear_dirty_beg_nonincl(m, base, size);
 369                                 if (m->dirty != 0)
 370                                         m->dirty = VM_PAGE_BITS_ALL;
 371                                 vm_page_wakeup(m);
 372                         }
 373                 }
 374         } else {
 375                 vp->v_filesize = nsize;
 376         }
 377 }
 378
 379 /*
 380  * Release a page busied for a getpages operation.  The page may have become
 381  * wired (typically due to being used by the buffer cache) or otherwise been
 382  * soft-busied and cannot be freed in that case.  A held page can still be
 383  * freed.
 384  */
 385 void
 386 vnode_pager_freepage(vm_page_t m)
 387 {
 388         if (m->busy || m->wire_count) {
 389                 vm_page_activate(m);
 390                 vm_page_wakeup(m);
 391         } else {
 392                 vm_page_free(m);
 393         }
 394 }
 395
 396 /*
 397  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
 398  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
 399  * vnode_pager_generic_getpages() to implement the previous behaviour.
 400  *
 401  * All other FS's should use the bypass to get to the local media
 402  * backing vp's VOP_GETPAGES.
 403  */
 404 static int
 405 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
 406 {
 407         int rtval;
 408         struct vnode *vp;
 409         int bytes = count * PAGE_SIZE;
 410
 411         vp = object->handle;
 412         rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
 413         if (rtval == EOPNOTSUPP)
 414                 panic("vnode_pager: vfs's must implement vop_getpages\n");
 415         return rtval;
 416 }
 417
 418 /*
 419  * This is now called from local media FS's to operate against their
 420  * own vnodes if they fail to implement VOP_GETPAGES.
 421  *
 422  * With all the caching local media devices do these days there is really
 423  * very little point to attempting to restrict the I/O size to contiguous
 424  * blocks on-disk, especially if our caller thinks we need all the specified
 425  * pages.  Just construct and issue a READ.
 426  */
 427 int
 428 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount,
 429                              int reqpage)
 430 {
 431         struct iovec aiov;
 432         struct uio auio;
 433         off_t foff;
 434         int error;
 435         int count;
 436         int i;
 437         int ioflags;
 438
 439         /*
 440          * Do not do anything if the vnode is bad.
 441          */
 442         if (vp->v_mount == NULL)
 443                 return VM_PAGER_BAD;
 444
 445         /*
 446          * Calculate the number of pages.  Since we are paging in whole
 447          * pages, adjust bytecount to be an integral multiple of the page
 448          * size.  It will be clipped to the file EOF later on.
 449          */
 450         bytecount = round_page(bytecount);
 451         count = bytecount / PAGE_SIZE;
 452
 453         /*
 454          * If we have a completely valid page available to us, we can
 455          * clean up and return.  Otherwise we have to re-read the
 456          * media.
 457          *
 458          * Note that this does not work with NFS, so NFS has its own
 459          * getpages routine.  The problem is that NFS can have partially
 460          * valid pages associated with the buffer cache due to the piecemeal
 461          * write support.  If we were to fall through and re-read the media
 462          * as we do here, dirty data could be lost.
 463          */
 464         if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
 465                 for (i = 0; i < count; i++) {
 466                         if (i != reqpage)
 467                                 vnode_pager_freepage(m[i]);
 468                 }
 469                 return VM_PAGER_OK;
 470         }
 471
 472         /*
 473          * Discard pages past the file EOF.  If the requested page is past
 474          * the file EOF we just leave its valid bits set to 0, the caller
 475          * expects to maintain ownership of the requested page.  If the
 476          * entire range is past file EOF discard everything and generate
 477          * a pagein error.
 478          */
 479         foff = IDX_TO_OFF(m[0]->pindex);
 480         if (foff >= vp->v_filesize) {
 481                 for (i = 0; i < count; i++) {
 482                         if (i != reqpage)
 483                                 vnode_pager_freepage(m[i]);
 484                 }
 485                 return VM_PAGER_ERROR;
 486         }
 487
 488         if (foff + bytecount > vp->v_filesize) {
 489                 bytecount = vp->v_filesize - foff;
 490                 i = round_page(bytecount) / PAGE_SIZE;
 491                 while (count > i) {
 492                         --count;
 493                         if (count != reqpage)
 494                                 vnode_pager_freepage(m[count]);
 495                 }
 496         }
 497
 498         /*
 499          * The size of the transfer is bytecount.  bytecount will be an
 500          * integral multiple of the page size unless it has been clipped
 501          * to the file EOF.  The transfer cannot exceed the file EOF.
 502          *
 503          * When dealing with real devices we must round-up to the device
 504          * sector size.
 505          */
 506         if (vp->v_type == VBLK || vp->v_type == VCHR) {
 507                 int secmask = vp->v_rdev->si_bsize_phys - 1;
 508                 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
 509                 bytecount = (bytecount + secmask) & ~secmask;
 510         }
 511
 512         /*
 513          * Severe hack to avoid deadlocks with the buffer cache
 514          */
 515         for (i = 0; i < count; ++i) {
 516                 vm_page_t mt = m[i];
 517
 518                 vm_page_io_start(mt);
 519                 vm_page_wakeup(mt);
 520         }
 521
 522         /*
 523          * Issue the I/O without any read-ahead
 524          */
 525         ioflags = IO_VMIO;
 526         /*ioflags |= IO_SEQMAX << IO_SEQSHIFT;*/
 527
 528         aiov.iov_base = (caddr_t) 0;
 529         aiov.iov_len = bytecount;
 530         auio.uio_iov = &aiov;
 531         auio.uio_iovcnt = 1;
 532         auio.uio_offset = foff;
 533         auio.uio_segflg = UIO_NOCOPY;
 534         auio.uio_rw = UIO_READ;
 535         auio.uio_resid = bytecount;
 536         auio.uio_td = NULL;
 537         mycpu->gd_cnt.v_vnodein++;
 538         mycpu->gd_cnt.v_vnodepgsin += count;
 539
 540         error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
 541
 542         /*
 543          * Severe hack to avoid deadlocks with the buffer cache
 544          */
 545         for (i = 0; i < count; ++i) {
 546                 vm_page_t mt = m[i];
 547
 548                 while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
 549                         ;
 550                 vm_page_busy(mt);
 551                 vm_page_io_finish(mt);
 552         }
 553
 554         /*
 555          * Calculate the actual number of bytes read and clean up the
 556          * page list.
 557          */
 558         bytecount -= auio.uio_resid;
 559
 560         for (i = 0; i < count; ++i) {
 561                 vm_page_t mt = m[i];
 562
 563                 if (i != reqpage) {
 564                         if (error == 0 && mt->valid) {
 565                                 if (mt->flags & PG_WANTED)
 566                                         vm_page_activate(mt);
 567                                 else
 568                                         vm_page_deactivate(mt);
 569                                 vm_page_wakeup(mt);
 570                         } else {
 571                                 vnode_pager_freepage(mt);
 572                         }
 573                 } else if (mt->valid == 0) {
 574                         if (error == 0) {
 575                                 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
 576                                 /* whoops, something happened */
 577                                 error = EINVAL;
 578                         }
 579                 } else if (mt->valid != VM_PAGE_BITS_ALL) {
 580                         /*
 581                          * Zero-extend the requested page if necessary (if
 582                          * the filesystem is using a small block size).
 583                          */
 584                         vm_page_zero_invalid(mt, TRUE);
 585                 }
 586         }
 587         if (error) {
 588                 kprintf("vnode_pager_getpages: I/O read error\n");
 589         }
 590         return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
 591 }
 592
 593 /*
 594  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
 595  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
 596  * vnode_pager_generic_putpages() to implement the previous behaviour.
 597  *
 598  * All other FS's should use the bypass to get to the local media
 599  * backing vp's VOP_PUTPAGES.
 600  */
 601 static void
 602 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
 603     boolean_t sync, int *rtvals)
 604 {
 605         int rtval;
 606         struct vnode *vp;
 607         int bytes = count * PAGE_SIZE;
 608
 609         /*
 610          * Force synchronous operation if we are extremely low on memory
 611          * to prevent a low-memory deadlock.  VOP operations often need to
 612          * allocate more memory to initiate the I/O ( i.e. do a BMAP
 613          * operation ).  The swapper handles the case by limiting the amount
 614          * of asynchronous I/O, but that sort of solution doesn't scale well
 615          * for the vnode pager without a lot of work.
 616          *
 617          * Also, the backing vnode's iodone routine may not wake the pageout
 618          * daemon up.  This should be probably be addressed XXX.
 619          */
 620
 621         if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
 622                 sync |= OBJPC_SYNC;
 623
 624         /*
 625          * Call device-specific putpages function
 626          */
 627
 628         vp = object->handle;
 629         rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
 630         if (rtval == EOPNOTSUPP) {
 631             kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
 632             rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
 633         }
 634 }
 635
 636
 637 /*
 638  * This is now called from local media FS's to operate against their
 639  * own vnodes if they fail to implement VOP_PUTPAGES.
 640  *
 641  * This is typically called indirectly via the pageout daemon and
 642  * clustering has already typically occured, so in general we ask the
 643  * underlying filesystem to write the data out asynchronously rather
 644  * then delayed.
 645  */
 646 int
 647 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
 648                              int flags, int *rtvals)
 649 {
 650         int i;
 651         vm_object_t object;
 652         int count;
 653
 654         int maxsize, ncount;
 655         vm_ooffset_t poffset;
 656         struct uio auio;
 657         struct iovec aiov;
 658         int error;
 659         int ioflags;
 660
 661         object = vp->v_object;
 662         count = bytecount / PAGE_SIZE;
 663
 664         for (i = 0; i < count; i++)
 665                 rtvals[i] = VM_PAGER_AGAIN;
 666
 667         if ((int) m[0]->pindex < 0) {
 668                 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
 669                         (long)m[0]->pindex, m[0]->dirty);
 670                 rtvals[0] = VM_PAGER_BAD;
 671                 return VM_PAGER_BAD;
 672         }
 673
 674         maxsize = count * PAGE_SIZE;
 675         ncount = count;
 676
 677         poffset = IDX_TO_OFF(m[0]->pindex);
 678
 679         /*
 680          * If the page-aligned write is larger then the actual file we
 681          * have to invalidate pages occuring beyond the file EOF.  However,
 682          * there is an edge case where a file may not be page-aligned where
 683          * the last page is partially invalid.  In this case the filesystem
 684          * may not properly clear the dirty bits for the entire page (which
 685          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
 686          * With the page locked we are free to fix-up the dirty bits here.
 687          *
 688          * We do not under any circumstances truncate the valid bits, as
 689          * this will screw up bogus page replacement.
 690          *
 691          * The caller has already read-protected the pages.  The VFS must
 692          * use the buffer cache to wrap the pages.  The pages might not
 693          * be immediately flushed by the buffer cache but once under its
 694          * control the pages themselves can wind up being marked clean
 695          * and their covering buffer cache buffer can be marked dirty.
 696          */
 697         if (maxsize + poffset > vp->v_filesize) {
 698                 if (vp->v_filesize > poffset) {
 699                         int pgoff;
 700
 701                         maxsize = vp->v_filesize - poffset;
 702                         ncount = btoc(maxsize);
 703                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
 704                                 vm_page_clear_dirty(m[ncount - 1], pgoff,
 705                                         PAGE_SIZE - pgoff);
 706                         }
 707                 } else {
 708                         maxsize = 0;
 709                         ncount = 0;
 710                 }
 711                 if (ncount < count) {
 712                         for (i = ncount; i < count; i++) {
 713                                 rtvals[i] = VM_PAGER_BAD;
 714                         }
 715                 }
 716         }
 717
 718         /*
 719          * pageouts are already clustered, use IO_ASYNC to force a bawrite()
 720          * rather then a bdwrite() to prevent paging I/O from saturating
 721          * the buffer cache.  Dummy-up the sequential heuristic to cause
 722          * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
 723          * the system decides how to cluster.
 724          */
 725         ioflags = IO_VMIO;
 726         if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
 727                 ioflags |= IO_SYNC;
 728         else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
 729                 ioflags |= IO_ASYNC;
 730         ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
 731         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
 732
 733         aiov.iov_base = (caddr_t) 0;
 734         aiov.iov_len = maxsize;
 735         auio.uio_iov = &aiov;
 736         auio.uio_iovcnt = 1;
 737         auio.uio_offset = poffset;
 738         auio.uio_segflg = UIO_NOCOPY;
 739         auio.uio_rw = UIO_WRITE;
 740         auio.uio_resid = maxsize;
 741         auio.uio_td = NULL;
 742         error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
 743         mycpu->gd_cnt.v_vnodeout++;
 744         mycpu->gd_cnt.v_vnodepgsout += ncount;
 745
 746         if (error) {
 747                 krateprintf(&vbadrate,
 748                             "vnode_pager_putpages: I/O error %d\n", error);
 749         }
 750         if (auio.uio_resid) {
 751                 krateprintf(&vresrate,
 752                             "vnode_pager_putpages: residual I/O %d at %lu\n",
 753                             auio.uio_resid, (u_long)m[0]->pindex);
 754         }
 755         for (i = 0; i < ncount; i++)
 756                 rtvals[i] = VM_PAGER_OK;
 757         return rtvals[0];
 758 }
 759
 760 struct vnode *
 761 vnode_pager_lock(vm_object_t object)
 762 {
 763         struct thread *td = curthread;  /* XXX */
 764         int error;
 765
 766         for (; object != NULL; object = object->backing_object) {
 767                 if (object->type != OBJT_VNODE)
 768                         continue;
 769                 if (object->flags & OBJ_DEAD)
 770                         return NULL;
 771
 772                 for (;;) {
 773                         struct vnode *vp = object->handle;
 774                         error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
 775                         if (error == 0) {
 776                                 if (object->handle != vp) {
 777                                         vput(vp);
 778                                         continue;
 779                                 }
 780                                 return (vp);
 781                         }
 782                         if ((object->flags & OBJ_DEAD) ||
 783                             (object->type != OBJT_VNODE)) {
 784                                 return NULL;
 785                         }
 786                         kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
 787                         tsleep(object->handle, 0, "vnpgrl", hz);
 788                 }
 789         }
 790         return NULL;
 791 }