The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/nfsclient/nfs_bio.c

Version: -  FREEBSD  -  FREEBSD-13-STABLE  -  FREEBSD-13-0  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  l41  -  OPENBSD  -  linux-2.6  -  MK84  -  PLAN9  -  xnu-8792 
SearchContext: -  none  -  3  -  10 

    1 /*
    2  * Copyright (c) 1989, 1993
    3  *      The Regents of the University of California.  All rights reserved.
    4  *
    5  * This code is derived from software contributed to Berkeley by
    6  * Rick Macklem at The University of Guelph.
    7  *
    8  * Redistribution and use in source and binary forms, with or without
    9  * modification, are permitted provided that the following conditions
   10  * are met:
   11  * 1. Redistributions of source code must retain the above copyright
   12  *    notice, this list of conditions and the following disclaimer.
   13  * 2. Redistributions in binary form must reproduce the above copyright
   14  *    notice, this list of conditions and the following disclaimer in the
   15  *    documentation and/or other materials provided with the distribution.
   16  * 3. All advertising materials mentioning features or use of this software
   17  *    must display the following acknowledgement:
   18  *      This product includes software developed by the University of
   19  *      California, Berkeley and its contributors.
   20  * 4. Neither the name of the University nor the names of its contributors
   21  *    may be used to endorse or promote products derived from this software
   22  *    without specific prior written permission.
   23  *
   24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
   25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
   28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   34  * SUCH DAMAGE.
   35  *
   36  *      @(#)nfs_bio.c   8.9 (Berkeley) 3/30/95
   37  */
   38 
   39 #include <sys/cdefs.h>
   40 __FBSDID("$FreeBSD: releng/5.2/sys/nfsclient/nfs_bio.c 122953 2003-11-22 02:21:49Z alfred $");
   41 
   42 #include <sys/param.h>
   43 #include <sys/systm.h>
   44 #include <sys/bio.h>
   45 #include <sys/buf.h>
   46 #include <sys/kernel.h>
   47 #include <sys/mount.h>
   48 #include <sys/proc.h>
   49 #include <sys/resourcevar.h>
   50 #include <sys/signalvar.h>
   51 #include <sys/vmmeter.h>
   52 #include <sys/vnode.h>
   53 
   54 #include <vm/vm.h>
   55 #include <vm/vm_extern.h>
   56 #include <vm/vm_page.h>
   57 #include <vm/vm_object.h>
   58 #include <vm/vm_pager.h>
   59 #include <vm/vnode_pager.h>
   60 
   61 #include <rpc/rpcclnt.h>
   62 
   63 #include <nfs/rpcv2.h>
   64 #include <nfs/nfsproto.h>
   65 #include <nfsclient/nfs.h>
   66 #include <nfsclient/nfsmount.h>
   67 #include <nfsclient/nfsnode.h>
   68 
   69 #include <nfs4client/nfs4.h>
   70 
   71 /*
   72  * Just call nfs_writebp() with the force argument set to 1.
   73  *
   74  * NOTE: B_DONE may or may not be set in a_bp on call.
   75  */
   76 static int
   77 nfs4_bwrite(struct buf *bp)
   78 {
   79 
   80         return (nfs4_writebp(bp, 1, curthread));
   81 }
   82 
   83 static int
   84 nfs_bwrite(struct buf *bp)
   85 {
   86 
   87         return (nfs_writebp(bp, 1, curthread));
   88 }
   89 
   90 struct buf_ops buf_ops_nfs4 = {
   91         "buf_ops_nfs4",
   92         nfs4_bwrite
   93 };
   94 
   95 struct buf_ops buf_ops_nfs = {
   96         "buf_ops_nfs",
   97         nfs_bwrite
   98 };
   99 
  100 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
  101                     struct thread *td);
  102 
  103 /*
  104  * Vnode op for VM getpages.
  105  */
  106 int
  107 nfs_getpages(struct vop_getpages_args *ap)
  108 {
  109         int i, error, nextoff, size, toff, count, npages;
  110         struct uio uio;
  111         struct iovec iov;
  112         vm_offset_t kva;
  113         struct buf *bp;
  114         struct vnode *vp;
  115         struct thread *td;
  116         struct ucred *cred;
  117         struct nfsmount *nmp;
  118         vm_object_t object;
  119         vm_page_t *pages;
  120 
  121         GIANT_REQUIRED;
  122 
  123         vp = ap->a_vp;
  124         td = curthread;                         /* XXX */
  125         cred = curthread->td_ucred;             /* XXX */
  126         nmp = VFSTONFS(vp->v_mount);
  127         pages = ap->a_m;
  128         count = ap->a_count;
  129 
  130         if ((object = vp->v_object) == NULL) {
  131                 printf("nfs_getpages: called with non-merged cache vnode??\n");
  132                 return VM_PAGER_ERROR;
  133         }
  134 
  135         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  136             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  137                 /* We'll never get here for v4, because we always have fsinfo */
  138                 (void)nfs_fsinfo(nmp, vp, cred, td);
  139         }
  140 
  141         npages = btoc(count);
  142 
  143         /*
  144          * If the requested page is partially valid, just return it and
  145          * allow the pager to zero-out the blanks.  Partially valid pages
  146          * can only occur at the file EOF.
  147          */
  148 
  149         {
  150                 vm_page_t m = pages[ap->a_reqpage];
  151 
  152                 VM_OBJECT_LOCK(object);
  153                 vm_page_lock_queues();
  154                 if (m->valid != 0) {
  155                         /* handled by vm_fault now        */
  156                         /* vm_page_zero_invalid(m, TRUE); */
  157                         for (i = 0; i < npages; ++i) {
  158                                 if (i != ap->a_reqpage)
  159                                         vm_page_free(pages[i]);
  160                         }
  161                         vm_page_unlock_queues();
  162                         VM_OBJECT_UNLOCK(object);
  163                         return(0);
  164                 }
  165                 vm_page_unlock_queues();
  166                 VM_OBJECT_UNLOCK(object);
  167         }
  168 
  169         /*
  170          * We use only the kva address for the buffer, but this is extremely
  171          * convienient and fast.
  172          */
  173         bp = getpbuf(&nfs_pbuf_freecnt);
  174 
  175         kva = (vm_offset_t) bp->b_data;
  176         pmap_qenter(kva, pages, npages);
  177         cnt.v_vnodein++;
  178         cnt.v_vnodepgsin += npages;
  179 
  180         iov.iov_base = (caddr_t) kva;
  181         iov.iov_len = count;
  182         uio.uio_iov = &iov;
  183         uio.uio_iovcnt = 1;
  184         uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
  185         uio.uio_resid = count;
  186         uio.uio_segflg = UIO_SYSSPACE;
  187         uio.uio_rw = UIO_READ;
  188         uio.uio_td = td;
  189 
  190         error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred);
  191         pmap_qremove(kva, npages);
  192 
  193         relpbuf(bp, &nfs_pbuf_freecnt);
  194 
  195         if (error && (uio.uio_resid == count)) {
  196                 printf("nfs_getpages: error %d\n", error);
  197                 VM_OBJECT_LOCK(object);
  198                 vm_page_lock_queues();
  199                 for (i = 0; i < npages; ++i) {
  200                         if (i != ap->a_reqpage)
  201                                 vm_page_free(pages[i]);
  202                 }
  203                 vm_page_unlock_queues();
  204                 VM_OBJECT_UNLOCK(object);
  205                 return VM_PAGER_ERROR;
  206         }
  207 
  208         /*
  209          * Calculate the number of bytes read and validate only that number
  210          * of bytes.  Note that due to pending writes, size may be 0.  This
  211          * does not mean that the remaining data is invalid!
  212          */
  213 
  214         size = count - uio.uio_resid;
  215         VM_OBJECT_LOCK(object);
  216         vm_page_lock_queues();
  217         for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
  218                 vm_page_t m;
  219                 nextoff = toff + PAGE_SIZE;
  220                 m = pages[i];
  221 
  222                 m->flags &= ~PG_ZERO;
  223 
  224                 if (nextoff <= size) {
  225                         /*
  226                          * Read operation filled an entire page
  227                          */
  228                         m->valid = VM_PAGE_BITS_ALL;
  229                         vm_page_undirty(m);
  230                 } else if (size > toff) {
  231                         /*
  232                          * Read operation filled a partial page.
  233                          */
  234                         m->valid = 0;
  235                         vm_page_set_validclean(m, 0, size - toff);
  236                         /* handled by vm_fault now        */
  237                         /* vm_page_zero_invalid(m, TRUE); */
  238                 } else {
  239                         /*
  240                          * Read operation was short.  If no error occured
  241                          * we may have hit a zero-fill section.   We simply
  242                          * leave valid set to 0.
  243                          */
  244                         ;
  245                 }
  246                 if (i != ap->a_reqpage) {
  247                         /*
  248                          * Whether or not to leave the page activated is up in
  249                          * the air, but we should put the page on a page queue
  250                          * somewhere (it already is in the object).  Result:
  251                          * It appears that emperical results show that
  252                          * deactivating pages is best.
  253                          */
  254 
  255                         /*
  256                          * Just in case someone was asking for this page we
  257                          * now tell them that it is ok to use.
  258                          */
  259                         if (!error) {
  260                                 if (m->flags & PG_WANTED)
  261                                         vm_page_activate(m);
  262                                 else
  263                                         vm_page_deactivate(m);
  264                                 vm_page_wakeup(m);
  265                         } else {
  266                                 vm_page_free(m);
  267                         }
  268                 }
  269         }
  270         vm_page_unlock_queues();
  271         VM_OBJECT_UNLOCK(object);
  272         return 0;
  273 }
  274 
  275 /*
  276  * Vnode op for VM putpages.
  277  */
  278 int
  279 nfs_putpages(struct vop_putpages_args *ap)
  280 {
  281         struct uio uio;
  282         struct iovec iov;
  283         vm_offset_t kva;
  284         struct buf *bp;
  285         int iomode, must_commit, i, error, npages, count;
  286         off_t offset;
  287         int *rtvals;
  288         struct vnode *vp;
  289         struct thread *td;
  290         struct ucred *cred;
  291         struct nfsmount *nmp;
  292         struct nfsnode *np;
  293         vm_page_t *pages;
  294 
  295         GIANT_REQUIRED;
  296 
  297         vp = ap->a_vp;
  298         np = VTONFS(vp);
  299         td = curthread;                         /* XXX */
  300         cred = curthread->td_ucred;             /* XXX */
  301         nmp = VFSTONFS(vp->v_mount);
  302         pages = ap->a_m;
  303         count = ap->a_count;
  304         rtvals = ap->a_rtvals;
  305         npages = btoc(count);
  306         offset = IDX_TO_OFF(pages[0]->pindex);
  307 
  308         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  309             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  310                 (void)nfs_fsinfo(nmp, vp, cred, td);
  311         }
  312 
  313         for (i = 0; i < npages; i++)
  314                 rtvals[i] = VM_PAGER_AGAIN;
  315 
  316         /*
  317          * When putting pages, do not extend file past EOF.
  318          */
  319 
  320         if (offset + count > np->n_size) {
  321                 count = np->n_size - offset;
  322                 if (count < 0)
  323                         count = 0;
  324         }
  325 
  326         /*
  327          * We use only the kva address for the buffer, but this is extremely
  328          * convienient and fast.
  329          */
  330         bp = getpbuf(&nfs_pbuf_freecnt);
  331 
  332         kva = (vm_offset_t) bp->b_data;
  333         pmap_qenter(kva, pages, npages);
  334         cnt.v_vnodeout++;
  335         cnt.v_vnodepgsout += count;
  336 
  337         iov.iov_base = (caddr_t) kva;
  338         iov.iov_len = count;
  339         uio.uio_iov = &iov;
  340         uio.uio_iovcnt = 1;
  341         uio.uio_offset = offset;
  342         uio.uio_resid = count;
  343         uio.uio_segflg = UIO_SYSSPACE;
  344         uio.uio_rw = UIO_WRITE;
  345         uio.uio_td = td;
  346 
  347         if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
  348             iomode = NFSV3WRITE_UNSTABLE;
  349         else
  350             iomode = NFSV3WRITE_FILESYNC;
  351 
  352         error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
  353 
  354         pmap_qremove(kva, npages);
  355         relpbuf(bp, &nfs_pbuf_freecnt);
  356 
  357         if (!error) {
  358                 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
  359                 for (i = 0; i < nwritten; i++) {
  360                         rtvals[i] = VM_PAGER_OK;
  361                         vm_page_undirty(pages[i]);
  362                 }
  363                 if (must_commit) {
  364                         nfs_clearcommit(vp->v_mount);
  365                 }
  366         }
  367         return rtvals[0];
  368 }
  369 
  370 /*
  371  * Vnode op for read using bio
  372  */
  373 int
  374 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
  375 {
  376         struct nfsnode *np = VTONFS(vp);
  377         int biosize, i;
  378         struct buf *bp = 0, *rabp;
  379         struct vattr vattr;
  380         struct thread *td;
  381         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
  382         daddr_t lbn, rabn;
  383         int bcount;
  384         int seqcount;
  385         int nra, error = 0, n = 0, on = 0;
  386 
  387 #ifdef DIAGNOSTIC
  388         if (uio->uio_rw != UIO_READ)
  389                 panic("nfs_read mode");
  390 #endif
  391         if (uio->uio_resid == 0)
  392                 return (0);
  393         if (uio->uio_offset < 0)        /* XXX VDIR cookies can be negative */
  394                 return (EINVAL);
  395         td = uio->uio_td;
  396 
  397         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  398             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
  399                 (void)nfs_fsinfo(nmp, vp, cred, td);
  400         if (vp->v_type != VDIR &&
  401             (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
  402                 return (EFBIG);
  403         biosize = vp->v_mount->mnt_stat.f_iosize;
  404         seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
  405         /*
  406          * For nfs, cache consistency can only be maintained approximately.
  407          * Although RFC1094 does not specify the criteria, the following is
  408          * believed to be compatible with the reference port.
  409          * For nfs:
  410          * If the file's modify time on the server has changed since the
  411          * last read rpc or you have written to the file,
  412          * you may have lost data cache consistency with the
  413          * server, so flush all of the file's data out of the cache.
  414          * Then force a getattr rpc to ensure that you have up to date
  415          * attributes.
  416          * NB: This implies that cache data can be read when up to
  417          * NFS_ATTRTIMEO seconds out of date. If you find that you need current
  418          * attributes this could be forced by setting n_attrstamp to 0 before
  419          * the VOP_GETATTR() call.
  420          */
  421         if (np->n_flag & NMODIFIED) {
  422                 if (vp->v_type != VREG) {
  423                         if (vp->v_type != VDIR)
  424                                 panic("nfs: bioread, not dir");
  425                         (nmp->nm_rpcops->nr_invaldir)(vp);
  426                         error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
  427                         if (error)
  428                                 return (error);
  429                 }
  430                 np->n_attrstamp = 0;
  431                 error = VOP_GETATTR(vp, &vattr, cred, td);
  432                 if (error)
  433                         return (error);
  434                 np->n_mtime = vattr.va_mtime.tv_sec;
  435         } else {
  436                 error = VOP_GETATTR(vp, &vattr, cred, td);
  437                 if (error)
  438                         return (error);
  439                 if (np->n_mtime != vattr.va_mtime.tv_sec) {
  440                         if (vp->v_type == VDIR)
  441                                 (nmp->nm_rpcops->nr_invaldir)(vp);
  442                         error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
  443                         if (error)
  444                                 return (error);
  445                         np->n_mtime = vattr.va_mtime.tv_sec;
  446                 }
  447         }
  448         do {
  449             switch (vp->v_type) {
  450             case VREG:
  451                 nfsstats.biocache_reads++;
  452                 lbn = uio->uio_offset / biosize;
  453                 on = uio->uio_offset & (biosize - 1);
  454 
  455                 /*
  456                  * Start the read ahead(s), as required.
  457                  */
  458                 if (nmp->nm_readahead > 0) {
  459                     for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
  460                         (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
  461                         rabn = lbn + 1 + nra;
  462                         if (incore(vp, rabn) == NULL) {
  463                             rabp = nfs_getcacheblk(vp, rabn, biosize, td);
  464                             if (!rabp)
  465                                 return (EINTR);
  466                             if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
  467                                 rabp->b_flags |= B_ASYNC;
  468                                 rabp->b_iocmd = BIO_READ;
  469                                 vfs_busy_pages(rabp, 0);
  470                                 if (nfs_asyncio(rabp, cred, td)) {
  471                                     rabp->b_flags |= B_INVAL;
  472                                     rabp->b_ioflags |= BIO_ERROR;
  473                                     vfs_unbusy_pages(rabp);
  474                                     brelse(rabp);
  475                                     break;
  476                                 }
  477                             } else {
  478                                 brelse(rabp);
  479                             }
  480                         }
  481                     }
  482                 }
  483 
  484                 /*
  485                  * Obtain the buffer cache block.  Figure out the buffer size
  486                  * when we are at EOF.  If we are modifying the size of the
  487                  * buffer based on an EOF condition we need to hold
  488                  * nfs_rslock() through obtaining the buffer to prevent
  489                  * a potential writer-appender from messing with n_size.
  490                  * Otherwise we may accidently truncate the buffer and
  491                  * lose dirty data.
  492                  *
  493                  * Note that bcount is *not* DEV_BSIZE aligned.
  494                  */
  495 
  496 again:
  497                 bcount = biosize;
  498                 if ((off_t)lbn * biosize >= np->n_size) {
  499                         bcount = 0;
  500                 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
  501                         bcount = np->n_size - (off_t)lbn * biosize;
  502                 }
  503                 if (bcount != biosize) {
  504                         switch(nfs_rslock(np, td)) {
  505                         case ENOLCK:
  506                                 goto again;
  507                                 /* not reached */
  508                         case EINTR:
  509                         case ERESTART:
  510                                 return(EINTR);
  511                                 /* not reached */
  512                         default:
  513                                 break;
  514                         }
  515                 }
  516 
  517                 bp = nfs_getcacheblk(vp, lbn, bcount, td);
  518 
  519                 if (bcount != biosize)
  520                         nfs_rsunlock(np, td);
  521                 if (!bp)
  522                         return (EINTR);
  523 
  524                 /*
  525                  * If B_CACHE is not set, we must issue the read.  If this
  526                  * fails, we return an error.
  527                  */
  528 
  529                 if ((bp->b_flags & B_CACHE) == 0) {
  530                     bp->b_iocmd = BIO_READ;
  531                     vfs_busy_pages(bp, 0);
  532                     error = nfs_doio(bp, cred, td);
  533                     if (error) {
  534                         brelse(bp);
  535                         return (error);
  536                     }
  537                 }
  538 
  539                 /*
  540                  * on is the offset into the current bp.  Figure out how many
  541                  * bytes we can copy out of the bp.  Note that bcount is
  542                  * NOT DEV_BSIZE aligned.
  543                  *
  544                  * Then figure out how many bytes we can copy into the uio.
  545                  */
  546 
  547                 n = 0;
  548                 if (on < bcount)
  549                         n = min((unsigned)(bcount - on), uio->uio_resid);
  550                 break;
  551             case VLNK:
  552                 nfsstats.biocache_readlinks++;
  553                 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
  554                 if (!bp)
  555                         return (EINTR);
  556                 if ((bp->b_flags & B_CACHE) == 0) {
  557                     bp->b_iocmd = BIO_READ;
  558                     vfs_busy_pages(bp, 0);
  559                     error = nfs_doio(bp, cred, td);
  560                     if (error) {
  561                         bp->b_ioflags |= BIO_ERROR;
  562                         brelse(bp);
  563                         return (error);
  564                     }
  565                 }
  566                 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
  567                 on = 0;
  568                 break;
  569             case VDIR:
  570                 nfsstats.biocache_readdirs++;
  571                 if (np->n_direofoffset
  572                     && uio->uio_offset >= np->n_direofoffset) {
  573                     return (0);
  574                 }
  575                 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
  576                 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
  577                 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
  578                 if (!bp)
  579                     return (EINTR);
  580                 if ((bp->b_flags & B_CACHE) == 0) {
  581                     bp->b_iocmd = BIO_READ;
  582                     vfs_busy_pages(bp, 0);
  583                     error = nfs_doio(bp, cred, td);
  584                     if (error) {
  585                             brelse(bp);
  586                     }
  587                     while (error == NFSERR_BAD_COOKIE) {
  588                         printf("got bad cookie vp %p bp %p\n", vp, bp);
  589                         (nmp->nm_rpcops->nr_invaldir)(vp);
  590                         error = nfs_vinvalbuf(vp, 0, cred, td, 1);
  591                         /*
  592                          * Yuck! The directory has been modified on the
  593                          * server. The only way to get the block is by
  594                          * reading from the beginning to get all the
  595                          * offset cookies.
  596                          *
  597                          * Leave the last bp intact unless there is an error.
  598                          * Loop back up to the while if the error is another
  599                          * NFSERR_BAD_COOKIE (double yuch!).
  600                          */
  601                         for (i = 0; i <= lbn && !error; i++) {
  602                             if (np->n_direofoffset
  603                                 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
  604                                     return (0);
  605                             bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
  606                             if (!bp)
  607                                 return (EINTR);
  608                             if ((bp->b_flags & B_CACHE) == 0) {
  609                                     bp->b_iocmd = BIO_READ;
  610                                     vfs_busy_pages(bp, 0);
  611                                     error = nfs_doio(bp, cred, td);
  612                                     /*
  613                                      * no error + B_INVAL == directory EOF,
  614                                      * use the block.
  615                                      */
  616                                     if (error == 0 && (bp->b_flags & B_INVAL))
  617                                             break;
  618                             }
  619                             /*
  620                              * An error will throw away the block and the
  621                              * for loop will break out.  If no error and this
  622                              * is not the block we want, we throw away the
  623                              * block and go for the next one via the for loop.
  624                              */
  625                             if (error || i < lbn)
  626                                     brelse(bp);
  627                         }
  628                     }
  629                     /*
  630                      * The above while is repeated if we hit another cookie
  631                      * error.  If we hit an error and it wasn't a cookie error,
  632                      * we give up.
  633                      */
  634                     if (error)
  635                             return (error);
  636                 }
  637 
  638                 /*
  639                  * If not eof and read aheads are enabled, start one.
  640                  * (You need the current block first, so that you have the
  641                  *  directory offset cookie of the next block.)
  642                  */
  643                 if (nmp->nm_readahead > 0 &&
  644                     (bp->b_flags & B_INVAL) == 0 &&
  645                     (np->n_direofoffset == 0 ||
  646                     (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
  647                     incore(vp, lbn + 1) == NULL) {
  648                         rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
  649                         if (rabp) {
  650                             if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
  651                                 rabp->b_flags |= B_ASYNC;
  652                                 rabp->b_iocmd = BIO_READ;
  653                                 vfs_busy_pages(rabp, 0);
  654                                 if (nfs_asyncio(rabp, cred, td)) {
  655                                     rabp->b_flags |= B_INVAL;
  656                                     rabp->b_ioflags |= BIO_ERROR;
  657                                     vfs_unbusy_pages(rabp);
  658                                     brelse(rabp);
  659                                 }
  660                             } else {
  661                                 brelse(rabp);
  662                             }
  663                         }
  664                 }
  665                 /*
  666                  * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
  667                  * chopped for the EOF condition, we cannot tell how large
  668                  * NFS directories are going to be until we hit EOF.  So
  669                  * an NFS directory buffer is *not* chopped to its EOF.  Now,
  670                  * it just so happens that b_resid will effectively chop it
  671                  * to EOF.  *BUT* this information is lost if the buffer goes
  672                  * away and is reconstituted into a B_CACHE state ( due to
  673                  * being VMIO ) later.  So we keep track of the directory eof
  674                  * in np->n_direofoffset and chop it off as an extra step
  675                  * right here.
  676                  */
  677                 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
  678                 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
  679                         n = np->n_direofoffset - uio->uio_offset;
  680                 break;
  681             default:
  682                 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
  683                 break;
  684             };
  685 
  686             if (n > 0) {
  687                     error = uiomove(bp->b_data + on, (int)n, uio);
  688             }
  689             switch (vp->v_type) {
  690             case VREG:
  691                 break;
  692             case VLNK:
  693                 n = 0;
  694                 break;
  695             case VDIR:
  696                 break;
  697             default:
  698                 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
  699             }
  700             brelse(bp);
  701         } while (error == 0 && uio->uio_resid > 0 && n > 0);
  702         return (error);
  703 }
  704 
  705 /*
  706  * Vnode op for write using bio
  707  */
  708 int
  709 nfs_write(struct vop_write_args *ap)
  710 {
  711         int biosize;
  712         struct uio *uio = ap->a_uio;
  713         struct thread *td = uio->uio_td;
  714         struct vnode *vp = ap->a_vp;
  715         struct nfsnode *np = VTONFS(vp);
  716         struct ucred *cred = ap->a_cred;
  717         int ioflag = ap->a_ioflag;
  718         struct buf *bp;
  719         struct vattr vattr;
  720         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
  721         daddr_t lbn;
  722         int bcount;
  723         int n, on, error = 0;
  724         int haverslock = 0;
  725         struct proc *p = td?td->td_proc:NULL;
  726 
  727         GIANT_REQUIRED;
  728 
  729 #ifdef DIAGNOSTIC
  730         if (uio->uio_rw != UIO_WRITE)
  731                 panic("nfs_write mode");
  732         if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
  733                 panic("nfs_write proc");
  734 #endif
  735         if (vp->v_type != VREG)
  736                 return (EIO);
  737         if (np->n_flag & NWRITEERR) {
  738                 np->n_flag &= ~NWRITEERR;
  739                 return (np->n_error);
  740         }
  741         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  742             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
  743                 (void)nfs_fsinfo(nmp, vp, cred, td);
  744 
  745         /*
  746          * Synchronously flush pending buffers if we are in synchronous
  747          * mode or if we are appending.
  748          */
  749         if (ioflag & (IO_APPEND | IO_SYNC)) {
  750                 if (np->n_flag & NMODIFIED) {
  751                         np->n_attrstamp = 0;
  752                         error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
  753                         if (error)
  754                                 return (error);
  755                 }
  756         }
  757 
  758         /*
  759          * If IO_APPEND then load uio_offset.  We restart here if we cannot
  760          * get the append lock.
  761          */
  762 restart:
  763         if (ioflag & IO_APPEND) {
  764                 np->n_attrstamp = 0;
  765                 error = VOP_GETATTR(vp, &vattr, cred, td);
  766                 if (error)
  767                         return (error);
  768                 uio->uio_offset = np->n_size;
  769         }
  770 
  771         if (uio->uio_offset < 0)
  772                 return (EINVAL);
  773         if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
  774                 return (EFBIG);
  775         if (uio->uio_resid == 0)
  776                 return (0);
  777 
  778         /*
  779          * We need to obtain the rslock if we intend to modify np->n_size
  780          * in order to guarentee the append point with multiple contending
  781          * writers, to guarentee that no other appenders modify n_size
  782          * while we are trying to obtain a truncated buffer (i.e. to avoid
  783          * accidently truncating data written by another appender due to
  784          * the race), and to ensure that the buffer is populated prior to
  785          * our extending of the file.  We hold rslock through the entire
  786          * operation.
  787          *
  788          * Note that we do not synchronize the case where someone truncates
  789          * the file while we are appending to it because attempting to lock
  790          * this case may deadlock other parts of the system unexpectedly.
  791          */
  792         if ((ioflag & IO_APPEND) ||
  793             uio->uio_offset + uio->uio_resid > np->n_size) {
  794                 switch(nfs_rslock(np, td)) {
  795                 case ENOLCK:
  796                         goto restart;
  797                         /* not reached */
  798                 case EINTR:
  799                 case ERESTART:
  800                         return(EINTR);
  801                         /* not reached */
  802                 default:
  803                         break;
  804                 }
  805                 haverslock = 1;
  806         }
  807 
  808         /*
  809          * Maybe this should be above the vnode op call, but so long as
  810          * file servers have no limits, i don't think it matters
  811          */
  812         if (p && uio->uio_offset + uio->uio_resid >
  813               p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
  814                 PROC_LOCK(p);
  815                 psignal(p, SIGXFSZ);
  816                 PROC_UNLOCK(p);
  817                 if (haverslock)
  818                         nfs_rsunlock(np, td);
  819                 return (EFBIG);
  820         }
  821 
  822         biosize = vp->v_mount->mnt_stat.f_iosize;
  823 
  824         do {
  825                 nfsstats.biocache_writes++;
  826                 lbn = uio->uio_offset / biosize;
  827                 on = uio->uio_offset & (biosize-1);
  828                 n = min((unsigned)(biosize - on), uio->uio_resid);
  829 again:
  830                 /*
  831                  * Handle direct append and file extension cases, calculate
  832                  * unaligned buffer size.
  833                  */
  834 
  835                 if (uio->uio_offset == np->n_size && n) {
  836                         /*
  837                          * Get the buffer (in its pre-append state to maintain
  838                          * B_CACHE if it was previously set).  Resize the
  839                          * nfsnode after we have locked the buffer to prevent
  840                          * readers from reading garbage.
  841                          */
  842                         bcount = on;
  843                         bp = nfs_getcacheblk(vp, lbn, bcount, td);
  844 
  845                         if (bp != NULL) {
  846                                 long save;
  847 
  848                                 np->n_size = uio->uio_offset + n;
  849                                 np->n_flag |= NMODIFIED;
  850                                 vnode_pager_setsize(vp, np->n_size);
  851 
  852                                 save = bp->b_flags & B_CACHE;
  853                                 bcount += n;
  854                                 allocbuf(bp, bcount);
  855                                 bp->b_flags |= save;
  856                                 bp->b_magic = B_MAGIC_NFS;
  857                                 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0)
  858                                         bp->b_op = &buf_ops_nfs4;
  859                                 else
  860                                         bp->b_op = &buf_ops_nfs;
  861                         }
  862                 } else {
  863                         /*
  864                          * Obtain the locked cache block first, and then
  865                          * adjust the file's size as appropriate.
  866                          */
  867                         bcount = on + n;
  868                         if ((off_t)lbn * biosize + bcount < np->n_size) {
  869                                 if ((off_t)(lbn + 1) * biosize < np->n_size)
  870                                         bcount = biosize;
  871                                 else
  872                                         bcount = np->n_size - (off_t)lbn * biosize;
  873                         }
  874                         bp = nfs_getcacheblk(vp, lbn, bcount, td);
  875                         if (uio->uio_offset + n > np->n_size) {
  876                                 np->n_size = uio->uio_offset + n;
  877                                 np->n_flag |= NMODIFIED;
  878                                 vnode_pager_setsize(vp, np->n_size);
  879                         }
  880                 }
  881 
  882                 if (!bp) {
  883                         error = EINTR;
  884                         break;
  885                 }
  886 
  887                 /*
  888                  * Issue a READ if B_CACHE is not set.  In special-append
  889                  * mode, B_CACHE is based on the buffer prior to the write
  890                  * op and is typically set, avoiding the read.  If a read
  891                  * is required in special append mode, the server will
  892                  * probably send us a short-read since we extended the file
  893                  * on our end, resulting in b_resid == 0 and, thusly,
  894                  * B_CACHE getting set.
  895                  *
  896                  * We can also avoid issuing the read if the write covers
  897                  * the entire buffer.  We have to make sure the buffer state
  898                  * is reasonable in this case since we will not be initiating
  899                  * I/O.  See the comments in kern/vfs_bio.c's getblk() for
  900                  * more information.
  901                  *
  902                  * B_CACHE may also be set due to the buffer being cached
  903                  * normally.
  904                  */
  905 
  906                 if (on == 0 && n == bcount) {
  907                         bp->b_flags |= B_CACHE;
  908                         bp->b_flags &= ~B_INVAL;
  909                         bp->b_ioflags &= ~BIO_ERROR;
  910                 }
  911 
  912                 if ((bp->b_flags & B_CACHE) == 0) {
  913                         bp->b_iocmd = BIO_READ;
  914                         vfs_busy_pages(bp, 0);
  915                         error = nfs_doio(bp, cred, td);
  916                         if (error) {
  917                                 brelse(bp);
  918                                 break;
  919                         }
  920                 }
  921                 if (!bp) {
  922                         error = EINTR;
  923                         break;
  924                 }
  925                 if (bp->b_wcred == NOCRED)
  926                         bp->b_wcred = crhold(cred);
  927                 np->n_flag |= NMODIFIED;
  928 
  929                 /*
  930                  * If dirtyend exceeds file size, chop it down.  This should
  931                  * not normally occur but there is an append race where it
  932                  * might occur XXX, so we log it.
  933                  *
  934                  * If the chopping creates a reverse-indexed or degenerate
  935                  * situation with dirtyoff/end, we 0 both of them.
  936                  */
  937 
  938                 if (bp->b_dirtyend > bcount) {
  939                         printf("NFS append race @%lx:%d\n",
  940                             (long)bp->b_blkno * DEV_BSIZE,
  941                             bp->b_dirtyend - bcount);
  942                         bp->b_dirtyend = bcount;
  943                 }
  944 
  945                 if (bp->b_dirtyoff >= bp->b_dirtyend)
  946                         bp->b_dirtyoff = bp->b_dirtyend = 0;
  947 
  948                 /*
  949                  * If the new write will leave a contiguous dirty
  950                  * area, just update the b_dirtyoff and b_dirtyend,
  951                  * otherwise force a write rpc of the old dirty area.
  952                  *
  953                  * While it is possible to merge discontiguous writes due to
  954                  * our having a B_CACHE buffer ( and thus valid read data
  955                  * for the hole), we don't because it could lead to
  956                  * significant cache coherency problems with multiple clients,
  957                  * especially if locking is implemented later on.
  958                  *
  959                  * as an optimization we could theoretically maintain
  960                  * a linked list of discontinuous areas, but we would still
  961                  * have to commit them separately so there isn't much
  962                  * advantage to it except perhaps a bit of asynchronization.
  963                  */
  964 
  965                 if (bp->b_dirtyend > 0 &&
  966                     (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
  967                         if (BUF_WRITE(bp) == EINTR) {
  968                                 error = EINTR;
  969                                 break;
  970                         }
  971                         goto again;
  972                 }
  973 
  974                 error = uiomove((char *)bp->b_data + on, n, uio);
  975 
  976                 /*
  977                  * Since this block is being modified, it must be written
  978                  * again and not just committed.  Since write clustering does
  979                  * not work for the stage 1 data write, only the stage 2
  980                  * commit rpc, we have to clear B_CLUSTEROK as well.
  981                  */
  982                 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
  983 
  984                 if (error) {
  985                         bp->b_ioflags |= BIO_ERROR;
  986                         brelse(bp);
  987                         break;
  988                 }
  989 
  990                 /*
  991                  * Only update dirtyoff/dirtyend if not a degenerate
  992                  * condition.
  993                  */
  994                 if (n) {
  995                         if (bp->b_dirtyend > 0) {
  996                                 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
  997                                 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
  998                         } else {
  999                                 bp->b_dirtyoff = on;
 1000                                 bp->b_dirtyend = on + n;
 1001                         }
 1002                         vfs_bio_set_validclean(bp, on, n);
 1003                 }
 1004 
 1005                 /*
 1006                  * If IO_SYNC do bwrite().
 1007                  *
 1008                  * IO_INVAL appears to be unused.  The idea appears to be
 1009                  * to turn off caching in this case.  Very odd.  XXX
 1010                  */
 1011                 if ((ioflag & IO_SYNC)) {
 1012                         if (ioflag & IO_INVAL)
 1013                                 bp->b_flags |= B_NOCACHE;
 1014                         error = BUF_WRITE(bp);
 1015                         if (error)
 1016                                 break;
 1017                 } else if ((n + on) == biosize) {
 1018                         bp->b_flags |= B_ASYNC;
 1019                         (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, 0);
 1020                 } else {
 1021                         bdwrite(bp);
 1022                 }
 1023         } while (uio->uio_resid > 0 && n > 0);
 1024 
 1025         if (haverslock)
 1026                 nfs_rsunlock(np, td);
 1027 
 1028         return (error);
 1029 }
 1030 
 1031 /*
 1032  * Get an nfs cache block.
 1033  *
 1034  * Allocate a new one if the block isn't currently in the cache
 1035  * and return the block marked busy. If the calling process is
 1036  * interrupted by a signal for an interruptible mount point, return
 1037  * NULL.
 1038  *
 1039  * The caller must carefully deal with the possible B_INVAL state of
 1040  * the buffer.  nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
 1041  * indirectly), so synchronous reads can be issued without worrying about
 1042  * the B_INVAL state.  We have to be a little more careful when dealing
 1043  * with writes (see comments in nfs_write()) when extending a file past
 1044  * its EOF.
 1045  */
 1046 static struct buf *
 1047 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
 1048 {
 1049         struct buf *bp;
 1050         struct mount *mp;
 1051         struct nfsmount *nmp;
 1052 
 1053         mp = vp->v_mount;
 1054         nmp = VFSTONFS(mp);
 1055 
 1056         if (nmp->nm_flag & NFSMNT_INT) {
 1057                 bp = getblk(vp, bn, size, PCATCH, 0, 0);
 1058                 while (bp == NULL) {
 1059                         if (nfs_sigintr(nmp, NULL, td))
 1060                                 return (NULL);
 1061                         bp = getblk(vp, bn, size, 0, 2 * hz, 0);
 1062                 }
 1063         } else {
 1064                 bp = getblk(vp, bn, size, 0, 0, 0);
 1065         }
 1066 
 1067         if (vp->v_type == VREG) {
 1068                 int biosize;
 1069 
 1070                 biosize = mp->mnt_stat.f_iosize;
 1071                 bp->b_blkno = bn * (biosize / DEV_BSIZE);
 1072         }
 1073         return (bp);
 1074 }
 1075 
 1076 /*
 1077  * Flush and invalidate all dirty buffers. If another process is already
 1078  * doing the flush, just wait for completion.
 1079  */
 1080 int
 1081 nfs_vinvalbuf(struct vnode *vp, int flags, struct ucred *cred,
 1082     struct thread *td, int intrflg)
 1083 {
 1084         struct nfsnode *np = VTONFS(vp);
 1085         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
 1086         int error = 0, slpflag, slptimeo;
 1087 
 1088         ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
 1089 
 1090         /*
 1091          * XXX This check stops us from needlessly doing a vinvalbuf when
 1092          * being called through vclean().  It is not clear that this is
 1093          * unsafe.
 1094          */
 1095         if (vp->v_iflag & VI_XLOCK)
 1096                 return (0);
 1097 
 1098         if ((nmp->nm_flag & NFSMNT_INT) == 0)
 1099                 intrflg = 0;
 1100         if (intrflg) {
 1101                 slpflag = PCATCH;
 1102                 slptimeo = 2 * hz;
 1103         } else {
 1104                 slpflag = 0;
 1105                 slptimeo = 0;
 1106         }
 1107         /*
 1108          * First wait for any other process doing a flush to complete.
 1109          */
 1110         while (np->n_flag & NFLUSHINPROG) {
 1111                 np->n_flag |= NFLUSHWANT;
 1112                 error = tsleep(&np->n_flag, PRIBIO + 2, "nfsvinval",
 1113                         slptimeo);
 1114                 if (error && intrflg &&
 1115                     nfs_sigintr(nmp, NULL, td))
 1116                         return (EINTR);
 1117         }
 1118 
 1119         /*
 1120          * Now, flush as required.
 1121          */
 1122         np->n_flag |= NFLUSHINPROG;
 1123         error = vinvalbuf(vp, flags, cred, td, slpflag, 0);
 1124         while (error) {
 1125                 if (intrflg &&
 1126                     nfs_sigintr(nmp, NULL, td)) {
 1127                         np->n_flag &= ~NFLUSHINPROG;
 1128                         if (np->n_flag & NFLUSHWANT) {
 1129                                 np->n_flag &= ~NFLUSHWANT;
 1130                                 wakeup(&np->n_flag);
 1131                         }
 1132                         return (EINTR);
 1133                 }
 1134                 error = vinvalbuf(vp, flags, cred, td, 0, slptimeo);
 1135         }
 1136         np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
 1137         if (np->n_flag & NFLUSHWANT) {
 1138                 np->n_flag &= ~NFLUSHWANT;
 1139                 wakeup(&np->n_flag);
 1140         }
 1141         return (0);
 1142 }
 1143 
 1144 /*
 1145  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
 1146  * This is mainly to avoid queueing async I/O requests when the nfsiods
 1147  * are all hung on a dead server.
 1148  *
 1149  * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
 1150  * is eventually dequeued by the async daemon, nfs_doio() *will*.
 1151  */
 1152 int
 1153 nfs_asyncio(struct buf *bp, struct ucred *cred, struct thread *td)
 1154 {
 1155         struct nfsmount *nmp;
 1156         int iod;
 1157         int gotiod;
 1158         int slpflag = 0;
 1159         int slptimeo = 0;
 1160         int error;
 1161 
 1162         nmp = VFSTONFS(bp->b_vp->v_mount);
 1163 
 1164         /*
 1165          * Commits are usually short and sweet so lets save some cpu and
 1166          * leave the async daemons for more important rpc's (such as reads
 1167          * and writes).
 1168          */
 1169         if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
 1170             (nmp->nm_bufqiods > nfs_numasync / 2)) {
 1171                 return(EIO);
 1172         }
 1173 
 1174 again:
 1175         if (nmp->nm_flag & NFSMNT_INT)
 1176                 slpflag = PCATCH;
 1177         gotiod = FALSE;
 1178 
 1179         /*
 1180          * Find a free iod to process this request.
 1181          */
 1182         for (iod = 0; iod < nfs_numasync; iod++)
 1183                 if (nfs_iodwant[iod]) {
 1184                         gotiod = TRUE;
 1185                         break;
 1186                 }
 1187 
 1188         /*
 1189          * Try to create one if none are free.
 1190          */
 1191         if (!gotiod) {
 1192                 iod = nfs_nfsiodnew();
 1193                 if (iod != -1)
 1194                         gotiod = TRUE;
 1195         }
 1196 
 1197         if (gotiod) {
 1198                 /*
 1199                  * Found one, so wake it up and tell it which
 1200                  * mount to process.
 1201                  */
 1202                 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
 1203                     iod, nmp));
 1204                 nfs_iodwant[iod] = NULL;
 1205                 nfs_iodmount[iod] = nmp;
 1206                 nmp->nm_bufqiods++;
 1207                 wakeup(&nfs_iodwant[iod]);
 1208         }
 1209 
 1210         /*
 1211          * If none are free, we may already have an iod working on this mount
 1212          * point.  If so, it will process our request.
 1213          */
 1214         if (!gotiod) {
 1215                 if (nmp->nm_bufqiods > 0) {
 1216                         NFS_DPF(ASYNCIO,
 1217                                 ("nfs_asyncio: %d iods are already processing mount %p\n",
 1218                                  nmp->nm_bufqiods, nmp));
 1219                         gotiod = TRUE;
 1220                 }
 1221         }
 1222 
 1223         /*
 1224          * If we have an iod which can process the request, then queue
 1225          * the buffer.
 1226          */
 1227         if (gotiod) {
 1228                 /*
 1229                  * Ensure that the queue never grows too large.  We still want
 1230                  * to asynchronize so we block rather then return EIO.
 1231                  */
 1232                 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
 1233                         NFS_DPF(ASYNCIO,
 1234                                 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
 1235                         nmp->nm_bufqwant = TRUE;
 1236                         error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
 1237                                        "nfsaio", slptimeo);
 1238                         if (error) {
 1239                                 if (nfs_sigintr(nmp, NULL, td))
 1240                                         return (EINTR);
 1241                                 if (slpflag == PCATCH) {
 1242                                         slpflag = 0;
 1243                                         slptimeo = 2 * hz;
 1244                                 }
 1245                         }
 1246                         /*
 1247                          * We might have lost our iod while sleeping,
 1248                          * so check and loop if nescessary.
 1249                          */
 1250                         if (nmp->nm_bufqiods == 0) {
 1251                                 NFS_DPF(ASYNCIO,
 1252                                         ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
 1253                                 goto again;
 1254                         }
 1255                 }
 1256 
 1257                 if (bp->b_iocmd == BIO_READ) {
 1258                         if (bp->b_rcred == NOCRED && cred != NOCRED)
 1259                                 bp->b_rcred = crhold(cred);
 1260                 } else {
 1261                         bp->b_flags |= B_WRITEINPROG;
 1262                         if (bp->b_wcred == NOCRED && cred != NOCRED)
 1263                                 bp->b_wcred = crhold(cred);
 1264                 }
 1265 
 1266                 BUF_KERNPROC(bp);
 1267                 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
 1268                 nmp->nm_bufqlen++;
 1269                 return (0);
 1270         }
 1271 
 1272         /*
 1273          * All the iods are busy on other mounts, so return EIO to
 1274          * force the caller to process the i/o synchronously.
 1275          */
 1276         NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
 1277         return (EIO);
 1278 }
 1279 
 1280 /*
 1281  * Do an I/O operation to/from a cache block. This may be called
 1282  * synchronously or from an nfsiod.
 1283  */
 1284 int
 1285 nfs_doio(struct buf *bp, struct ucred *cr, struct thread *td)
 1286 {
 1287         struct uio *uiop;
 1288         struct vnode *vp;
 1289         struct nfsnode *np;
 1290         struct nfsmount *nmp;
 1291         int error = 0, iomode, must_commit = 0;
 1292         struct uio uio;
 1293         struct iovec io;
 1294         struct proc *p = td ? td->td_proc : NULL;
 1295 
 1296         vp = bp->b_vp;
 1297         np = VTONFS(vp);
 1298         nmp = VFSTONFS(vp->v_mount);
 1299         uiop = &uio;
 1300         uiop->uio_iov = &io;
 1301         uiop->uio_iovcnt = 1;
 1302         uiop->uio_segflg = UIO_SYSSPACE;
 1303         uiop->uio_td = td;
 1304 
 1305         /*
 1306          * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
 1307          * do this here so we do not have to do it in all the code that
 1308          * calls us.
 1309          */
 1310         bp->b_flags &= ~B_INVAL;
 1311         bp->b_ioflags &= ~BIO_ERROR;
 1312 
 1313         KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
 1314 
 1315         if (bp->b_iocmd == BIO_READ) {
 1316             io.iov_len = uiop->uio_resid = bp->b_bcount;
 1317             io.iov_base = bp->b_data;
 1318             uiop->uio_rw = UIO_READ;
 1319 
 1320             switch (vp->v_type) {
 1321             case VREG:
 1322                 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
 1323                 nfsstats.read_bios++;
 1324                 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
 1325 
 1326                 if (!error) {
 1327                     if (uiop->uio_resid) {
 1328                         /*
 1329                          * If we had a short read with no error, we must have
 1330                          * hit a file hole.  We should zero-fill the remainder.
 1331                          * This can also occur if the server hits the file EOF.
 1332                          *
 1333                          * Holes used to be able to occur due to pending
 1334                          * writes, but that is not possible any longer.
 1335                          */
 1336                         int nread = bp->b_bcount - uiop->uio_resid;
 1337                         int left  = uiop->uio_resid;
 1338 
 1339                         if (left > 0)
 1340                                 bzero((char *)bp->b_data + nread, left);
 1341                         uiop->uio_resid = 0;
 1342                     }
 1343                 }
 1344                 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
 1345                 if (p && (vp->v_vflag & VV_TEXT) &&
 1346                         (np->n_mtime != np->n_vattr.va_mtime.tv_sec)) {
 1347                         uprintf("Process killed due to text file modification\n");
 1348                         PROC_LOCK(p);
 1349                         psignal(p, SIGKILL);
 1350                         _PHOLD(p);
 1351                         PROC_UNLOCK(p);
 1352                 }
 1353                 break;
 1354             case VLNK:
 1355                 uiop->uio_offset = (off_t)0;
 1356                 nfsstats.readlink_bios++;
 1357                 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
 1358                 break;
 1359             case VDIR:
 1360                 nfsstats.readdir_bios++;
 1361                 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
 1362                 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0)
 1363                         error = nfs4_readdirrpc(vp, uiop, cr);
 1364                 else {
 1365                         if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
 1366                                 error = nfs_readdirplusrpc(vp, uiop, cr);
 1367                                 if (error == NFSERR_NOTSUPP)
 1368                                         nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
 1369                         }
 1370                         if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
 1371                                 error = nfs_readdirrpc(vp, uiop, cr);
 1372                 }
 1373                 /*
 1374                  * end-of-directory sets B_INVAL but does not generate an
 1375                  * error.
 1376                  */
 1377                 if (error == 0 && uiop->uio_resid == bp->b_bcount)
 1378                         bp->b_flags |= B_INVAL;
 1379                 break;
 1380             default:
 1381                 printf("nfs_doio:  type %x unexpected\n", vp->v_type);
 1382                 break;
 1383             };
 1384             if (error) {
 1385                 bp->b_ioflags |= BIO_ERROR;
 1386                 bp->b_error = error;
 1387             }
 1388         } else {
 1389             /*
 1390              * If we only need to commit, try to commit
 1391              */
 1392             if (bp->b_flags & B_NEEDCOMMIT) {
 1393                     int retv;
 1394                     off_t off;
 1395 
 1396                     off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
 1397                     bp->b_flags |= B_WRITEINPROG;
 1398                     retv = (nmp->nm_rpcops->nr_commit)(
 1399                                 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff,
 1400                                 bp->b_wcred, td);
 1401                     bp->b_flags &= ~B_WRITEINPROG;
 1402                     if (retv == 0) {
 1403                             bp->b_dirtyoff = bp->b_dirtyend = 0;
 1404                             bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
 1405                             bp->b_resid = 0;
 1406                             bufdone(bp);
 1407                             return (0);
 1408                     }
 1409                     if (retv == NFSERR_STALEWRITEVERF) {
 1410                             nfs_clearcommit(bp->b_vp->v_mount);
 1411                     }
 1412             }
 1413 
 1414             /*
 1415              * Setup for actual write
 1416              */
 1417 
 1418             if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
 1419                 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
 1420 
 1421             if (bp->b_dirtyend > bp->b_dirtyoff) {
 1422                 io.iov_len = uiop->uio_resid = bp->b_dirtyend
 1423                     - bp->b_dirtyoff;
 1424                 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
 1425                     + bp->b_dirtyoff;
 1426                 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
 1427                 uiop->uio_rw = UIO_WRITE;
 1428                 nfsstats.write_bios++;
 1429 
 1430                 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
 1431                     iomode = NFSV3WRITE_UNSTABLE;
 1432                 else
 1433                     iomode = NFSV3WRITE_FILESYNC;
 1434 
 1435                 bp->b_flags |= B_WRITEINPROG;
 1436                 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
 1437 
 1438                 /*
 1439                  * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
 1440                  * to cluster the buffers needing commit.  This will allow
 1441                  * the system to submit a single commit rpc for the whole
 1442                  * cluster.  We can do this even if the buffer is not 100%
 1443                  * dirty (relative to the NFS blocksize), so we optimize the
 1444                  * append-to-file-case.
 1445                  *
 1446                  * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
 1447                  * cleared because write clustering only works for commit
 1448                  * rpc's, not for the data portion of the write).
 1449                  */
 1450 
 1451                 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
 1452                     bp->b_flags |= B_NEEDCOMMIT;
 1453                     if (bp->b_dirtyoff == 0
 1454                         && bp->b_dirtyend == bp->b_bcount)
 1455                         bp->b_flags |= B_CLUSTEROK;
 1456                 } else {
 1457                     bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
 1458                 }
 1459                 bp->b_flags &= ~B_WRITEINPROG;
 1460 
 1461                 /*
 1462                  * For an interrupted write, the buffer is still valid
 1463                  * and the write hasn't been pushed to the server yet,
 1464                  * so we can't set BIO_ERROR and report the interruption
 1465                  * by setting B_EINTR. For the B_ASYNC case, B_EINTR
 1466                  * is not relevant, so the rpc attempt is essentially
 1467                  * a noop.  For the case of a V3 write rpc not being
 1468                  * committed to stable storage, the block is still
 1469                  * dirty and requires either a commit rpc or another
 1470                  * write rpc with iomode == NFSV3WRITE_FILESYNC before
 1471                  * the block is reused. This is indicated by setting
 1472                  * the B_DELWRI and B_NEEDCOMMIT flags.
 1473                  *
 1474                  * If the buffer is marked B_PAGING, it does not reside on
 1475                  * the vp's paging queues so we cannot call bdirty().  The
 1476                  * bp in this case is not an NFS cache block so we should
 1477                  * be safe. XXX
 1478                  */
 1479                 if (error == EINTR
 1480                     || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
 1481                         int s;
 1482 
 1483                         s = splbio();
 1484                         bp->b_flags &= ~(B_INVAL|B_NOCACHE);
 1485                         if ((bp->b_flags & B_PAGING) == 0) {
 1486                             bdirty(bp);
 1487                             bp->b_flags &= ~B_DONE;
 1488                         }
 1489                         if (error && (bp->b_flags & B_ASYNC) == 0)
 1490                             bp->b_flags |= B_EINTR;
 1491                         splx(s);
 1492                 } else {
 1493                     if (error) {
 1494                         bp->b_ioflags |= BIO_ERROR;
 1495                         bp->b_error = np->n_error = error;
 1496                         np->n_flag |= NWRITEERR;
 1497                     }
 1498                     bp->b_dirtyoff = bp->b_dirtyend = 0;
 1499                 }
 1500             } else {
 1501                 bp->b_resid = 0;
 1502                 bufdone(bp);
 1503                 return (0);
 1504             }
 1505         }
 1506         bp->b_resid = uiop->uio_resid;
 1507         if (must_commit)
 1508             nfs_clearcommit(vp->v_mount);
 1509         bufdone(bp);
 1510         return (error);
 1511 }
 1512 
 1513 /*
 1514  * Used to aid in handling ftruncate() operations on the NFS client side.
 1515  * Truncation creates a number of special problems for NFS.  We have to
 1516  * throw away VM pages and buffer cache buffers that are beyond EOF, and
 1517  * we have to properly handle VM pages or (potentially dirty) buffers
 1518  * that straddle the truncation point.
 1519  */
 1520 
 1521 int
 1522 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
 1523 {
 1524         struct nfsnode *np = VTONFS(vp);
 1525         u_quad_t tsize = np->n_size;
 1526         int biosize = vp->v_mount->mnt_stat.f_iosize;
 1527         int error = 0;
 1528 
 1529         np->n_size = nsize;
 1530 
 1531         if (np->n_size < tsize) {
 1532                 struct buf *bp;
 1533                 daddr_t lbn;
 1534                 int bufsize;
 1535 
 1536                 /*
 1537                  * vtruncbuf() doesn't get the buffer overlapping the 
 1538                  * truncation point.  We may have a B_DELWRI and/or B_CACHE
 1539                  * buffer that now needs to be truncated.
 1540                  */
 1541                 error = vtruncbuf(vp, cred, td, nsize, biosize);
 1542                 lbn = nsize / biosize;
 1543                 bufsize = nsize & (biosize - 1);
 1544                 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
 1545                 if (bp->b_dirtyoff > bp->b_bcount)
 1546                         bp->b_dirtyoff = bp->b_bcount;
 1547                 if (bp->b_dirtyend > bp->b_bcount)
 1548                         bp->b_dirtyend = bp->b_bcount;
 1549                 bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
 1550                 brelse(bp);
 1551         } else {
 1552                 vnode_pager_setsize(vp, nsize);
 1553         }
 1554         return(error);
 1555 }
 1556 

Cache object: a1ad7dfb0971970cfd21c030853240dd


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.