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/fs/nfsclient/nfs_clbio.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  * SPDX-License-Identifier: BSD-3-Clause
    3  *
    4  * Copyright (c) 1989, 1993
    5  *      The Regents of the University of California.  All rights reserved.
    6  *
    7  * This code is derived from software contributed to Berkeley by
    8  * Rick Macklem at The University of Guelph.
    9  *
   10  * Redistribution and use in source and binary forms, with or without
   11  * modification, are permitted provided that the following conditions
   12  * are met:
   13  * 1. Redistributions of source code must retain the above copyright
   14  *    notice, this list of conditions and the following disclaimer.
   15  * 2. Redistributions in binary form must reproduce the above copyright
   16  *    notice, this list of conditions and the following disclaimer in the
   17  *    documentation and/or other materials provided with the distribution.
   18  * 3. Neither the name of the University nor the names of its contributors
   19  *    may be used to endorse or promote products derived from this software
   20  *    without specific prior written permission.
   21  *
   22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
   23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
   26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   32  * SUCH DAMAGE.
   33  *
   34  *      @(#)nfs_bio.c   8.9 (Berkeley) 3/30/95
   35  */
   36 
   37 #include <sys/cdefs.h>
   38 __FBSDID("$FreeBSD$");
   39 
   40 #include <sys/param.h>
   41 #include <sys/systm.h>
   42 #include <sys/bio.h>
   43 #include <sys/buf.h>
   44 #include <sys/kernel.h>
   45 #include <sys/mount.h>
   46 #include <sys/rwlock.h>
   47 #include <sys/vmmeter.h>
   48 #include <sys/vnode.h>
   49 
   50 #include <vm/vm.h>
   51 #include <vm/vm_param.h>
   52 #include <vm/vm_extern.h>
   53 #include <vm/vm_page.h>
   54 #include <vm/vm_object.h>
   55 #include <vm/vm_pager.h>
   56 #include <vm/vnode_pager.h>
   57 
   58 #include <fs/nfs/nfsport.h>
   59 #include <fs/nfsclient/nfsmount.h>
   60 #include <fs/nfsclient/nfs.h>
   61 #include <fs/nfsclient/nfsnode.h>
   62 #include <fs/nfsclient/nfs_kdtrace.h>
   63 
   64 extern int newnfs_directio_allow_mmap;
   65 extern struct nfsstatsv1 nfsstatsv1;
   66 extern struct mtx ncl_iod_mutex;
   67 extern int ncl_numasync;
   68 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
   69 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
   70 extern int newnfs_directio_enable;
   71 extern int nfs_keep_dirty_on_error;
   72 
   73 uma_zone_t ncl_pbuf_zone;
   74 
   75 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
   76     struct thread *td);
   77 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
   78     struct ucred *cred, int ioflag);
   79 
   80 /*
   81  * Vnode op for VM getpages.
   82  */
   83 SYSCTL_DECL(_vfs_nfs);
   84 static int use_buf_pager = 1;
   85 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
   86     &use_buf_pager, 0,
   87     "Use buffer pager instead of direct readrpc call");
   88 
   89 static daddr_t
   90 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
   91 {
   92 
   93         return (off / vp->v_bufobj.bo_bsize);
   94 }
   95 
   96 static int
   97 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn, long *sz)
   98 {
   99         struct nfsnode *np;
  100         u_quad_t nsize;
  101         int biosize, bcount;
  102 
  103         np = VTONFS(vp);
  104         NFSLOCKNODE(np);
  105         nsize = np->n_size;
  106         NFSUNLOCKNODE(np);
  107 
  108         biosize = vp->v_bufobj.bo_bsize;
  109         bcount = biosize;
  110         if ((off_t)lbn * biosize >= nsize)
  111                 bcount = 0;
  112         else if ((off_t)(lbn + 1) * biosize > nsize)
  113                 bcount = nsize - (off_t)lbn * biosize;
  114         *sz = bcount;
  115         return (0);
  116 }
  117 
  118 int
  119 ncl_getpages(struct vop_getpages_args *ap)
  120 {
  121         int i, error, nextoff, size, toff, count, npages;
  122         struct uio uio;
  123         struct iovec iov;
  124         vm_offset_t kva;
  125         struct buf *bp;
  126         struct vnode *vp;
  127         struct thread *td;
  128         struct ucred *cred;
  129         struct nfsmount *nmp;
  130         vm_object_t object;
  131         vm_page_t *pages;
  132         struct nfsnode *np;
  133 
  134         vp = ap->a_vp;
  135         np = VTONFS(vp);
  136         td = curthread;
  137         cred = curthread->td_ucred;
  138         nmp = VFSTONFS(vp->v_mount);
  139         pages = ap->a_m;
  140         npages = ap->a_count;
  141 
  142         if ((object = vp->v_object) == NULL) {
  143                 printf("ncl_getpages: called with non-merged cache vnode\n");
  144                 return (VM_PAGER_ERROR);
  145         }
  146 
  147         if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
  148                 NFSLOCKNODE(np);
  149                 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
  150                         NFSUNLOCKNODE(np);
  151                         printf("ncl_getpages: called on non-cacheable vnode\n");
  152                         return (VM_PAGER_ERROR);
  153                 } else
  154                         NFSUNLOCKNODE(np);
  155         }
  156 
  157         mtx_lock(&nmp->nm_mtx);
  158         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  159             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  160                 mtx_unlock(&nmp->nm_mtx);
  161                 /* We'll never get here for v4, because we always have fsinfo */
  162                 (void)ncl_fsinfo(nmp, vp, cred, td);
  163         } else
  164                 mtx_unlock(&nmp->nm_mtx);
  165 
  166         if (use_buf_pager)
  167                 return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
  168                     ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
  169 
  170         /*
  171          * If the requested page is partially valid, just return it and
  172          * allow the pager to zero-out the blanks.  Partially valid pages
  173          * can only occur at the file EOF.
  174          *
  175          * XXXGL: is that true for NFS, where short read can occur???
  176          */
  177         VM_OBJECT_WLOCK(object);
  178         if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
  179                 goto out;
  180         VM_OBJECT_WUNLOCK(object);
  181 
  182         /*
  183          * We use only the kva address for the buffer, but this is extremely
  184          * convenient and fast.
  185          */
  186         bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
  187 
  188         kva = (vm_offset_t) bp->b_data;
  189         pmap_qenter(kva, pages, npages);
  190         VM_CNT_INC(v_vnodein);
  191         VM_CNT_ADD(v_vnodepgsin, npages);
  192 
  193         count = npages << PAGE_SHIFT;
  194         iov.iov_base = (caddr_t) kva;
  195         iov.iov_len = count;
  196         uio.uio_iov = &iov;
  197         uio.uio_iovcnt = 1;
  198         uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
  199         uio.uio_resid = count;
  200         uio.uio_segflg = UIO_SYSSPACE;
  201         uio.uio_rw = UIO_READ;
  202         uio.uio_td = td;
  203 
  204         error = ncl_readrpc(vp, &uio, cred);
  205         pmap_qremove(kva, npages);
  206 
  207         uma_zfree(ncl_pbuf_zone, bp);
  208 
  209         if (error && (uio.uio_resid == count)) {
  210                 printf("ncl_getpages: error %d\n", error);
  211                 return (VM_PAGER_ERROR);
  212         }
  213 
  214         /*
  215          * Calculate the number of bytes read and validate only that number
  216          * of bytes.  Note that due to pending writes, size may be 0.  This
  217          * does not mean that the remaining data is invalid!
  218          */
  219 
  220         size = count - uio.uio_resid;
  221         VM_OBJECT_WLOCK(object);
  222         for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
  223                 vm_page_t m;
  224                 nextoff = toff + PAGE_SIZE;
  225                 m = pages[i];
  226 
  227                 if (nextoff <= size) {
  228                         /*
  229                          * Read operation filled an entire page
  230                          */
  231                         vm_page_valid(m);
  232                         KASSERT(m->dirty == 0,
  233                             ("nfs_getpages: page %p is dirty", m));
  234                 } else if (size > toff) {
  235                         /*
  236                          * Read operation filled a partial page.
  237                          */
  238                         vm_page_invalid(m);
  239                         vm_page_set_valid_range(m, 0, size - toff);
  240                         KASSERT(m->dirty == 0,
  241                             ("nfs_getpages: page %p is dirty", m));
  242                 } else {
  243                         /*
  244                          * Read operation was short.  If no error
  245                          * occurred we may have hit a zero-fill
  246                          * section.  We leave valid set to 0, and page
  247                          * is freed by vm_page_readahead_finish() if
  248                          * its index is not equal to requested, or
  249                          * page is zeroed and set valid by
  250                          * vm_pager_get_pages() for requested page.
  251                          */
  252                         ;
  253                 }
  254         }
  255 out:
  256         VM_OBJECT_WUNLOCK(object);
  257         if (ap->a_rbehind)
  258                 *ap->a_rbehind = 0;
  259         if (ap->a_rahead)
  260                 *ap->a_rahead = 0;
  261         return (VM_PAGER_OK);
  262 }
  263 
  264 /*
  265  * Vnode op for VM putpages.
  266  */
  267 int
  268 ncl_putpages(struct vop_putpages_args *ap)
  269 {
  270         struct uio uio;
  271         struct iovec iov;
  272         int i, error, npages, count;
  273         off_t offset;
  274         int *rtvals;
  275         struct vnode *vp;
  276         struct thread *td;
  277         struct ucred *cred;
  278         struct nfsmount *nmp;
  279         struct nfsnode *np;
  280         vm_page_t *pages;
  281 
  282         vp = ap->a_vp;
  283         np = VTONFS(vp);
  284         td = curthread;                         /* XXX */
  285         /* Set the cred to n_writecred for the write rpcs. */
  286         if (np->n_writecred != NULL)
  287                 cred = crhold(np->n_writecred);
  288         else
  289                 cred = crhold(curthread->td_ucred);     /* XXX */
  290         nmp = VFSTONFS(vp->v_mount);
  291         pages = ap->a_m;
  292         count = ap->a_count;
  293         rtvals = ap->a_rtvals;
  294         npages = btoc(count);
  295         offset = IDX_TO_OFF(pages[0]->pindex);
  296 
  297         mtx_lock(&nmp->nm_mtx);
  298         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  299             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  300                 mtx_unlock(&nmp->nm_mtx);
  301                 (void)ncl_fsinfo(nmp, vp, cred, td);
  302         } else
  303                 mtx_unlock(&nmp->nm_mtx);
  304 
  305         NFSLOCKNODE(np);
  306         if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
  307             (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
  308                 NFSUNLOCKNODE(np);
  309                 printf("ncl_putpages: called on noncache-able vnode\n");
  310                 NFSLOCKNODE(np);
  311         }
  312         /*
  313          * When putting pages, do not extend file past EOF.
  314          */
  315         if (offset + count > np->n_size) {
  316                 count = np->n_size - offset;
  317                 if (count < 0)
  318                         count = 0;
  319         }
  320         NFSUNLOCKNODE(np);
  321 
  322         for (i = 0; i < npages; i++)
  323                 rtvals[i] = VM_PAGER_ERROR;
  324 
  325         VM_CNT_INC(v_vnodeout);
  326         VM_CNT_ADD(v_vnodepgsout, count);
  327 
  328         iov.iov_base = unmapped_buf;
  329         iov.iov_len = count;
  330         uio.uio_iov = &iov;
  331         uio.uio_iovcnt = 1;
  332         uio.uio_offset = offset;
  333         uio.uio_resid = count;
  334         uio.uio_segflg = UIO_NOCOPY;
  335         uio.uio_rw = UIO_WRITE;
  336         uio.uio_td = td;
  337 
  338         error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
  339             cred);
  340         crfree(cred);
  341 
  342         if (error == 0 || !nfs_keep_dirty_on_error) {
  343                 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
  344                     np->n_size - offset, npages * PAGE_SIZE);
  345         }
  346         return (rtvals[0]);
  347 }
  348 
  349 /*
  350  * For nfs, cache consistency can only be maintained approximately.
  351  * Although RFC1094 does not specify the criteria, the following is
  352  * believed to be compatible with the reference port.
  353  * For nfs:
  354  * If the file's modify time on the server has changed since the
  355  * last read rpc or you have written to the file,
  356  * you may have lost data cache consistency with the
  357  * server, so flush all of the file's data out of the cache.
  358  * Then force a getattr rpc to ensure that you have up to date
  359  * attributes.
  360  * NB: This implies that cache data can be read when up to
  361  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
  362  * attributes this could be forced by setting n_attrstamp to 0 before
  363  * the VOP_GETATTR() call.
  364  */
  365 static inline int
  366 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
  367 {
  368         int error = 0;
  369         struct vattr vattr;
  370         struct nfsnode *np = VTONFS(vp);
  371         bool old_lock;
  372 
  373         /*
  374          * Ensure the exclusove access to the node before checking
  375          * whether the cache is consistent.
  376          */
  377         old_lock = ncl_excl_start(vp);
  378         NFSLOCKNODE(np);
  379         if (np->n_flag & NMODIFIED) {
  380                 NFSUNLOCKNODE(np);
  381                 if (vp->v_type != VREG) {
  382                         if (vp->v_type != VDIR)
  383                                 panic("nfs: bioread, not dir");
  384                         ncl_invaldir(vp);
  385                         error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
  386                         if (error != 0)
  387                                 goto out;
  388                 }
  389                 np->n_attrstamp = 0;
  390                 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
  391                 error = VOP_GETATTR(vp, &vattr, cred);
  392                 if (error)
  393                         goto out;
  394                 NFSLOCKNODE(np);
  395                 np->n_mtime = vattr.va_mtime;
  396                 NFSUNLOCKNODE(np);
  397         } else {
  398                 NFSUNLOCKNODE(np);
  399                 error = VOP_GETATTR(vp, &vattr, cred);
  400                 if (error)
  401                         goto out;
  402                 NFSLOCKNODE(np);
  403                 if ((np->n_flag & NSIZECHANGED)
  404                     || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
  405                         NFSUNLOCKNODE(np);
  406                         if (vp->v_type == VDIR)
  407                                 ncl_invaldir(vp);
  408                         error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
  409                         if (error != 0)
  410                                 goto out;
  411                         NFSLOCKNODE(np);
  412                         np->n_mtime = vattr.va_mtime;
  413                         np->n_flag &= ~NSIZECHANGED;
  414                 }
  415                 NFSUNLOCKNODE(np);
  416         }
  417 out:
  418         ncl_excl_finish(vp, old_lock);
  419         return (error);
  420 }
  421 
  422 /*
  423  * Vnode op for read using bio
  424  */
  425 int
  426 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
  427 {
  428         struct nfsnode *np = VTONFS(vp);
  429         struct buf *bp, *rabp;
  430         struct thread *td;
  431         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
  432         daddr_t lbn, rabn;
  433         int biosize, bcount, error, i, n, nra, on, save2, seqcount;
  434         off_t tmp_off;
  435 
  436         KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
  437         if (uio->uio_resid == 0)
  438                 return (0);
  439         if (uio->uio_offset < 0)        /* XXX VDIR cookies can be negative */
  440                 return (EINVAL);
  441         td = uio->uio_td;
  442 
  443         mtx_lock(&nmp->nm_mtx);
  444         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  445             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  446                 mtx_unlock(&nmp->nm_mtx);
  447                 (void)ncl_fsinfo(nmp, vp, cred, td);
  448                 mtx_lock(&nmp->nm_mtx);
  449         }
  450         if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
  451                 (void) newnfs_iosize(nmp);
  452 
  453         tmp_off = uio->uio_offset + uio->uio_resid;
  454         if (vp->v_type != VDIR &&
  455             (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
  456                 mtx_unlock(&nmp->nm_mtx);
  457                 return (EFBIG);
  458         }
  459         mtx_unlock(&nmp->nm_mtx);
  460 
  461         if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
  462                 /* No caching/ no readaheads. Just read data into the user buffer */
  463                 return ncl_readrpc(vp, uio, cred);
  464 
  465         n = 0;
  466         on = 0;
  467         biosize = vp->v_bufobj.bo_bsize;
  468         seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
  469 
  470         error = nfs_bioread_check_cons(vp, td, cred);
  471         if (error)
  472                 return error;
  473 
  474         save2 = curthread_pflags2_set(TDP2_SBPAGES);
  475         do {
  476             u_quad_t nsize;
  477 
  478             NFSLOCKNODE(np);
  479             nsize = np->n_size;
  480             NFSUNLOCKNODE(np);
  481 
  482             switch (vp->v_type) {
  483             case VREG:
  484                 NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
  485                 lbn = uio->uio_offset / biosize;
  486                 on = uio->uio_offset - (lbn * biosize);
  487 
  488                 /*
  489                  * Start the read ahead(s), as required.
  490                  */
  491                 if (nmp->nm_readahead > 0) {
  492                     for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
  493                         (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
  494                         rabn = lbn + 1 + nra;
  495                         if (incore(&vp->v_bufobj, rabn) == NULL) {
  496                             rabp = nfs_getcacheblk(vp, rabn, biosize, td);
  497                             if (!rabp) {
  498                                 error = newnfs_sigintr(nmp, td);
  499                                 if (error == 0)
  500                                         error = EINTR;
  501                                 goto out;
  502                             }
  503                             if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
  504                                 rabp->b_flags |= B_ASYNC;
  505                                 rabp->b_iocmd = BIO_READ;
  506                                 vfs_busy_pages(rabp, 0);
  507                                 if (ncl_asyncio(nmp, rabp, cred, td)) {
  508                                     rabp->b_flags |= B_INVAL;
  509                                     rabp->b_ioflags |= BIO_ERROR;
  510                                     vfs_unbusy_pages(rabp);
  511                                     brelse(rabp);
  512                                     break;
  513                                 }
  514                             } else {
  515                                 brelse(rabp);
  516                             }
  517                         }
  518                     }
  519                 }
  520 
  521                 /* Note that bcount is *not* DEV_BSIZE aligned. */
  522                 bcount = biosize;
  523                 if ((off_t)lbn * biosize >= nsize) {
  524                         bcount = 0;
  525                 } else if ((off_t)(lbn + 1) * biosize > nsize) {
  526                         bcount = nsize - (off_t)lbn * biosize;
  527                 }
  528                 bp = nfs_getcacheblk(vp, lbn, bcount, td);
  529 
  530                 if (!bp) {
  531                         error = newnfs_sigintr(nmp, td);
  532                         if (error == 0)
  533                                 error = EINTR;
  534                         goto out;
  535                 }
  536 
  537                 /*
  538                  * If B_CACHE is not set, we must issue the read.  If this
  539                  * fails, we return an error.
  540                  */
  541 
  542                 if ((bp->b_flags & B_CACHE) == 0) {
  543                     bp->b_iocmd = BIO_READ;
  544                     vfs_busy_pages(bp, 0);
  545                     error = ncl_doio(vp, bp, cred, td, 0);
  546                     if (error) {
  547                         brelse(bp);
  548                         goto out;
  549                     }
  550                 }
  551 
  552                 /*
  553                  * on is the offset into the current bp.  Figure out how many
  554                  * bytes we can copy out of the bp.  Note that bcount is
  555                  * NOT DEV_BSIZE aligned.
  556                  *
  557                  * Then figure out how many bytes we can copy into the uio.
  558                  */
  559 
  560                 n = 0;
  561                 if (on < bcount)
  562                         n = MIN((unsigned)(bcount - on), uio->uio_resid);
  563                 break;
  564             case VLNK:
  565                 NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
  566                 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
  567                 if (!bp) {
  568                         error = newnfs_sigintr(nmp, td);
  569                         if (error == 0)
  570                                 error = EINTR;
  571                         goto out;
  572                 }
  573                 if ((bp->b_flags & B_CACHE) == 0) {
  574                     bp->b_iocmd = BIO_READ;
  575                     vfs_busy_pages(bp, 0);
  576                     error = ncl_doio(vp, bp, cred, td, 0);
  577                     if (error) {
  578                         bp->b_ioflags |= BIO_ERROR;
  579                         brelse(bp);
  580                         goto out;
  581                     }
  582                 }
  583                 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
  584                 on = 0;
  585                 break;
  586             case VDIR:
  587                 NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
  588                 NFSLOCKNODE(np);
  589                 if (np->n_direofoffset
  590                     && uio->uio_offset >= np->n_direofoffset) {
  591                         NFSUNLOCKNODE(np);
  592                         error = 0;
  593                         goto out;
  594                 }
  595                 NFSUNLOCKNODE(np);
  596                 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
  597                 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
  598                 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
  599                 if (!bp) {
  600                         error = newnfs_sigintr(nmp, td);
  601                         if (error == 0)
  602                                 error = EINTR;
  603                         goto out;
  604                 }
  605                 if ((bp->b_flags & B_CACHE) == 0) {
  606                     bp->b_iocmd = BIO_READ;
  607                     vfs_busy_pages(bp, 0);
  608                     error = ncl_doio(vp, bp, cred, td, 0);
  609                     if (error) {
  610                             brelse(bp);
  611                     }
  612                     while (error == NFSERR_BAD_COOKIE) {
  613                         ncl_invaldir(vp);
  614                         error = ncl_vinvalbuf(vp, 0, td, 1);
  615 
  616                         /*
  617                          * Yuck! The directory has been modified on the
  618                          * server. The only way to get the block is by
  619                          * reading from the beginning to get all the
  620                          * offset cookies.
  621                          *
  622                          * Leave the last bp intact unless there is an error.
  623                          * Loop back up to the while if the error is another
  624                          * NFSERR_BAD_COOKIE (double yuch!).
  625                          */
  626                         for (i = 0; i <= lbn && !error; i++) {
  627                             NFSLOCKNODE(np);
  628                             if (np->n_direofoffset
  629                                 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
  630                                     NFSUNLOCKNODE(np);
  631                                     error = 0;
  632                                     goto out;
  633                             }
  634                             NFSUNLOCKNODE(np);
  635                             bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
  636                             if (!bp) {
  637                                 error = newnfs_sigintr(nmp, td);
  638                                 if (error == 0)
  639                                         error = EINTR;
  640                                 goto out;
  641                             }
  642                             if ((bp->b_flags & B_CACHE) == 0) {
  643                                     bp->b_iocmd = BIO_READ;
  644                                     vfs_busy_pages(bp, 0);
  645                                     error = ncl_doio(vp, bp, cred, td, 0);
  646                                     /*
  647                                      * no error + B_INVAL == directory EOF,
  648                                      * use the block.
  649                                      */
  650                                     if (error == 0 && (bp->b_flags & B_INVAL))
  651                                             break;
  652                             }
  653                             /*
  654                              * An error will throw away the block and the
  655                              * for loop will break out.  If no error and this
  656                              * is not the block we want, we throw away the
  657                              * block and go for the next one via the for loop.
  658                              */
  659                             if (error || i < lbn)
  660                                     brelse(bp);
  661                         }
  662                     }
  663                     /*
  664                      * The above while is repeated if we hit another cookie
  665                      * error.  If we hit an error and it wasn't a cookie error,
  666                      * we give up.
  667                      */
  668                     if (error)
  669                             goto out;
  670                 }
  671 
  672                 /*
  673                  * If not eof and read aheads are enabled, start one.
  674                  * (You need the current block first, so that you have the
  675                  *  directory offset cookie of the next block.)
  676                  */
  677                 NFSLOCKNODE(np);
  678                 if (nmp->nm_readahead > 0 &&
  679                     (bp->b_flags & B_INVAL) == 0 &&
  680                     (np->n_direofoffset == 0 ||
  681                     (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
  682                     incore(&vp->v_bufobj, lbn + 1) == NULL) {
  683                         NFSUNLOCKNODE(np);
  684                         rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
  685                         if (rabp) {
  686                             if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
  687                                 rabp->b_flags |= B_ASYNC;
  688                                 rabp->b_iocmd = BIO_READ;
  689                                 vfs_busy_pages(rabp, 0);
  690                                 if (ncl_asyncio(nmp, rabp, cred, td)) {
  691                                     rabp->b_flags |= B_INVAL;
  692                                     rabp->b_ioflags |= BIO_ERROR;
  693                                     vfs_unbusy_pages(rabp);
  694                                     brelse(rabp);
  695                                 }
  696                             } else {
  697                                 brelse(rabp);
  698                             }
  699                         }
  700                         NFSLOCKNODE(np);
  701                 }
  702                 /*
  703                  * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
  704                  * chopped for the EOF condition, we cannot tell how large
  705                  * NFS directories are going to be until we hit EOF.  So
  706                  * an NFS directory buffer is *not* chopped to its EOF.  Now,
  707                  * it just so happens that b_resid will effectively chop it
  708                  * to EOF.  *BUT* this information is lost if the buffer goes
  709                  * away and is reconstituted into a B_CACHE state ( due to
  710                  * being VMIO ) later.  So we keep track of the directory eof
  711                  * in np->n_direofoffset and chop it off as an extra step
  712                  * right here.
  713                  */
  714                 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
  715                 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
  716                         n = np->n_direofoffset - uio->uio_offset;
  717                 NFSUNLOCKNODE(np);
  718                 break;
  719             default:
  720                 printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
  721                 bp = NULL;
  722                 break;
  723             }
  724 
  725             if (n > 0) {
  726                     error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
  727             }
  728             if (vp->v_type == VLNK)
  729                 n = 0;
  730             if (bp != NULL)
  731                 brelse(bp);
  732         } while (error == 0 && uio->uio_resid > 0 && n > 0);
  733 out:
  734         curthread_pflags2_restore(save2);
  735         if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
  736                 NFSLOCKNODE(np);
  737                 ncl_pager_setsize(vp, NULL);
  738         }
  739         return (error);
  740 }
  741 
  742 /*
  743  * The NFS write path cannot handle iovecs with len > 1. So we need to
  744  * break up iovecs accordingly (restricting them to wsize).
  745  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
  746  * For the ASYNC case, 2 copies are needed. The first a copy from the
  747  * user buffer to a staging buffer and then a second copy from the staging
  748  * buffer to mbufs. This can be optimized by copying from the user buffer
  749  * directly into mbufs and passing the chain down, but that requires a
  750  * fair amount of re-working of the relevant codepaths (and can be done
  751  * later).
  752  */
  753 static int
  754 nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred,
  755     int ioflag)
  756 {
  757         int error;
  758         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
  759         struct thread *td = uiop->uio_td;
  760         int size;
  761         int wsize;
  762 
  763         mtx_lock(&nmp->nm_mtx);
  764         wsize = nmp->nm_wsize;
  765         mtx_unlock(&nmp->nm_mtx);
  766         if (ioflag & IO_SYNC) {
  767                 int iomode, must_commit;
  768                 struct uio uio;
  769                 struct iovec iov;
  770 do_sync:
  771                 while (uiop->uio_resid > 0) {
  772                         size = MIN(uiop->uio_resid, wsize);
  773                         size = MIN(uiop->uio_iov->iov_len, size);
  774                         iov.iov_base = uiop->uio_iov->iov_base;
  775                         iov.iov_len = size;
  776                         uio.uio_iov = &iov;
  777                         uio.uio_iovcnt = 1;
  778                         uio.uio_offset = uiop->uio_offset;
  779                         uio.uio_resid = size;
  780                         uio.uio_segflg = uiop->uio_segflg;
  781                         uio.uio_rw = UIO_WRITE;
  782                         uio.uio_td = td;
  783                         iomode = NFSWRITE_FILESYNC;
  784                         /*
  785                          * When doing direct I/O we do not care if the
  786                          * server's write verifier has changed, but we
  787                          * do not want to update the verifier if it has
  788                          * changed, since that hides the change from
  789                          * writes being done through the buffer cache.
  790                          * By passing must_commit in set to two, the code
  791                          * in nfsrpc_writerpc() will not update the
  792                          * verifier on the mount point.
  793                          */
  794                         must_commit = 2;
  795                         error = ncl_writerpc(vp, &uio, cred, &iomode,
  796                             &must_commit, 0, ioflag);
  797                         KASSERT((must_commit == 2),
  798                             ("ncl_directio_write: Updated write verifier"));
  799                         if (error)
  800                                 return (error);
  801                         if (iomode != NFSWRITE_FILESYNC)
  802                                 printf("nfs_directio_write: Broken server "
  803                                     "did not reply FILE_SYNC\n");
  804                         uiop->uio_offset += size;
  805                         uiop->uio_resid -= size;
  806                         if (uiop->uio_iov->iov_len <= size) {
  807                                 uiop->uio_iovcnt--;
  808                                 uiop->uio_iov++;
  809                         } else {
  810                                 uiop->uio_iov->iov_base =
  811                                         (char *)uiop->uio_iov->iov_base + size;
  812                                 uiop->uio_iov->iov_len -= size;
  813                         }
  814                 }
  815         } else {
  816                 struct uio *t_uio;
  817                 struct iovec *t_iov;
  818                 struct buf *bp;
  819 
  820                 /*
  821                  * Break up the write into blocksize chunks and hand these
  822                  * over to nfsiod's for write back.
  823                  * Unfortunately, this incurs a copy of the data. Since
  824                  * the user could modify the buffer before the write is
  825                  * initiated.
  826                  *
  827                  * The obvious optimization here is that one of the 2 copies
  828                  * in the async write path can be eliminated by copying the
  829                  * data here directly into mbufs and passing the mbuf chain
  830                  * down. But that will require a fair amount of re-working
  831                  * of the code and can be done if there's enough interest
  832                  * in NFS directio access.
  833                  */
  834                 while (uiop->uio_resid > 0) {
  835                         size = MIN(uiop->uio_resid, wsize);
  836                         size = MIN(uiop->uio_iov->iov_len, size);
  837                         bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
  838                         t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
  839                         t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
  840                         t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
  841                         t_iov->iov_len = size;
  842                         t_uio->uio_iov = t_iov;
  843                         t_uio->uio_iovcnt = 1;
  844                         t_uio->uio_offset = uiop->uio_offset;
  845                         t_uio->uio_resid = size;
  846                         t_uio->uio_segflg = UIO_SYSSPACE;
  847                         t_uio->uio_rw = UIO_WRITE;
  848                         t_uio->uio_td = td;
  849                         KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
  850                             uiop->uio_segflg == UIO_SYSSPACE,
  851                             ("nfs_directio_write: Bad uio_segflg"));
  852                         if (uiop->uio_segflg == UIO_USERSPACE) {
  853                                 error = copyin(uiop->uio_iov->iov_base,
  854                                     t_iov->iov_base, size);
  855                                 if (error != 0)
  856                                         goto err_free;
  857                         } else
  858                                 /*
  859                                  * UIO_SYSSPACE may never happen, but handle
  860                                  * it just in case it does.
  861                                  */
  862                                 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
  863                                     size);
  864                         bp->b_flags |= B_DIRECT;
  865                         bp->b_iocmd = BIO_WRITE;
  866                         if (cred != NOCRED) {
  867                                 crhold(cred);
  868                                 bp->b_wcred = cred;
  869                         } else
  870                                 bp->b_wcred = NOCRED;
  871                         bp->b_caller1 = (void *)t_uio;
  872                         bp->b_vp = vp;
  873                         error = ncl_asyncio(nmp, bp, NOCRED, td);
  874 err_free:
  875                         if (error) {
  876                                 free(t_iov->iov_base, M_NFSDIRECTIO);
  877                                 free(t_iov, M_NFSDIRECTIO);
  878                                 free(t_uio, M_NFSDIRECTIO);
  879                                 bp->b_vp = NULL;
  880                                 uma_zfree(ncl_pbuf_zone, bp);
  881                                 if (error == EINTR)
  882                                         return (error);
  883                                 goto do_sync;
  884                         }
  885                         uiop->uio_offset += size;
  886                         uiop->uio_resid -= size;
  887                         if (uiop->uio_iov->iov_len <= size) {
  888                                 uiop->uio_iovcnt--;
  889                                 uiop->uio_iov++;
  890                         } else {
  891                                 uiop->uio_iov->iov_base =
  892                                         (char *)uiop->uio_iov->iov_base + size;
  893                                 uiop->uio_iov->iov_len -= size;
  894                         }
  895                 }
  896         }
  897         return (0);
  898 }
  899 
  900 /*
  901  * Vnode op for write using bio
  902  */
  903 int
  904 ncl_write(struct vop_write_args *ap)
  905 {
  906         int biosize;
  907         struct uio *uio = ap->a_uio;
  908         struct thread *td = uio->uio_td;
  909         struct vnode *vp = ap->a_vp;
  910         struct nfsnode *np = VTONFS(vp);
  911         struct ucred *cred = ap->a_cred;
  912         int ioflag = ap->a_ioflag;
  913         struct buf *bp;
  914         struct vattr vattr;
  915         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
  916         daddr_t lbn;
  917         int bcount, noncontig_write, obcount;
  918         int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
  919         size_t orig_resid, local_resid;
  920         off_t orig_size, tmp_off;
  921         struct timespec ts;
  922 
  923         KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
  924         KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
  925             ("ncl_write proc"));
  926         if (vp->v_type != VREG)
  927                 return (EIO);
  928         NFSLOCKNODE(np);
  929         if (np->n_flag & NWRITEERR) {
  930                 np->n_flag &= ~NWRITEERR;
  931                 NFSUNLOCKNODE(np);
  932                 return (np->n_error);
  933         } else
  934                 NFSUNLOCKNODE(np);
  935         mtx_lock(&nmp->nm_mtx);
  936         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  937             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  938                 mtx_unlock(&nmp->nm_mtx);
  939                 (void)ncl_fsinfo(nmp, vp, cred, td);
  940                 mtx_lock(&nmp->nm_mtx);
  941         }
  942         if (nmp->nm_wsize == 0)
  943                 (void) newnfs_iosize(nmp);
  944         mtx_unlock(&nmp->nm_mtx);
  945 
  946         /*
  947          * Synchronously flush pending buffers if we are in synchronous
  948          * mode or if we are appending.
  949          */
  950         if ((ioflag & IO_APPEND) || ((ioflag & IO_SYNC) && (np->n_flag &
  951             NMODIFIED))) {
  952                 /*
  953                  * For the case where IO_APPEND is being done using a
  954                  * direct output (to the NFS server) RPC and
  955                  * newnfs_directio_enable is 0, all buffer cache buffers,
  956                  * including ones not modified, must be invalidated.
  957                  * This ensures that stale data is not read out of the
  958                  * buffer cache.  The call also invalidates all mapped
  959                  * pages and, since the exclusive lock is held on the vnode,
  960                  * new pages cannot be faulted in.
  961                  *
  962                  * For the case where newnfs_directio_enable is set
  963                  * (which is not the default), it is not obvious that
  964                  * stale data should be left in the buffer cache, but
  965                  * the code has been this way for over a decade without
  966                  * complaints.  Note that, unlike doing IO_APPEND via
  967                  * a direct write RPC when newnfs_directio_enable is not set,
  968                  * when newnfs_directio_enable is set, reading is done via
  969                  * direct to NFS server RPCs as well.
  970                  */
  971                 np->n_attrstamp = 0;
  972                 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
  973                 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
  974                     IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
  975                 if (error != 0)
  976                         return (error);
  977         }
  978 
  979         orig_resid = uio->uio_resid;
  980         NFSLOCKNODE(np);
  981         orig_size = np->n_size;
  982         NFSUNLOCKNODE(np);
  983 
  984         /*
  985          * If IO_APPEND then load uio_offset.  We restart here if we cannot
  986          * get the append lock.
  987          */
  988         if (ioflag & IO_APPEND) {
  989                 /*
  990                  * For NFSv4, the AppendWrite will Verify the size against
  991                  * the file's size on the server.  If not the same, the
  992                  * write will then be retried, using the file size returned
  993                  * by the AppendWrite.  However, for NFSv2 and NFSv3, the
  994                  * size must be acquired here via a Getattr RPC.
  995                  * The AppendWrite is not done for a pNFS mount.
  996                  */
  997                 if (!NFSHASNFSV4(nmp) || NFSHASPNFS(nmp)) {
  998                         np->n_attrstamp = 0;
  999                         KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
 1000                         error = VOP_GETATTR(vp, &vattr, cred);
 1001                         if (error)
 1002                                 return (error);
 1003                 }
 1004                 NFSLOCKNODE(np);
 1005                 uio->uio_offset = np->n_size;
 1006                 NFSUNLOCKNODE(np);
 1007         }
 1008 
 1009         if (uio->uio_offset < 0)
 1010                 return (EINVAL);
 1011         tmp_off = uio->uio_offset + uio->uio_resid;
 1012         if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
 1013                 return (EFBIG);
 1014         if (uio->uio_resid == 0)
 1015                 return (0);
 1016 
 1017         /*
 1018          * Do IO_APPEND writing via a synchronous direct write.
 1019          * This can result in a significant performance improvement.
 1020          */
 1021         if ((newnfs_directio_enable && (ioflag & IO_DIRECT)) ||
 1022             (ioflag & IO_APPEND)) {
 1023                 /*
 1024                  * Direct writes to the server must be done NFSWRITE_FILESYNC,
 1025                  * because the write data is not cached and, therefore, the
 1026                  * write cannot be redone after a server reboot.
 1027                  * Set IO_SYNC to make this happen.
 1028                  */
 1029                 ioflag |= IO_SYNC;
 1030                 return (nfs_directio_write(vp, uio, cred, ioflag));
 1031         }
 1032 
 1033         /*
 1034          * Maybe this should be above the vnode op call, but so long as
 1035          * file servers have no limits, i don't think it matters
 1036          */
 1037         error = vn_rlimit_fsize(vp, uio, td);
 1038         if (error != 0)
 1039                 return (error);
 1040 
 1041         save2 = curthread_pflags2_set(TDP2_SBPAGES);
 1042         biosize = vp->v_bufobj.bo_bsize;
 1043         /*
 1044          * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
 1045          * would exceed the local maximum per-file write commit size when
 1046          * combined with those, we must decide whether to flush,
 1047          * go synchronous, or return error.  We don't bother checking
 1048          * IO_UNIT -- we just make all writes atomic anyway, as there's
 1049          * no point optimizing for something that really won't ever happen.
 1050          */
 1051         wouldcommit = 0;
 1052         if (!(ioflag & IO_SYNC)) {
 1053                 int nflag;
 1054 
 1055                 NFSLOCKNODE(np);
 1056                 nflag = np->n_flag;
 1057                 NFSUNLOCKNODE(np);
 1058                 if (nflag & NMODIFIED) {
 1059                         BO_LOCK(&vp->v_bufobj);
 1060                         if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
 1061                                 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
 1062                                     b_bobufs) {
 1063                                         if (bp->b_flags & B_NEEDCOMMIT)
 1064                                                 wouldcommit += bp->b_bcount;
 1065                                 }
 1066                         }
 1067                         BO_UNLOCK(&vp->v_bufobj);
 1068                 }
 1069         }
 1070 
 1071         do {
 1072                 if (!(ioflag & IO_SYNC)) {
 1073                         wouldcommit += biosize;
 1074                         if (wouldcommit > nmp->nm_wcommitsize) {
 1075                                 np->n_attrstamp = 0;
 1076                                 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
 1077                                 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
 1078                                     IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
 1079                                 if (error != 0)
 1080                                         goto out;
 1081                                 wouldcommit = biosize;
 1082                         }
 1083                 }
 1084 
 1085                 NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
 1086                 lbn = uio->uio_offset / biosize;
 1087                 on = uio->uio_offset - (lbn * biosize);
 1088                 n = MIN((unsigned)(biosize - on), uio->uio_resid);
 1089 again:
 1090                 /*
 1091                  * Handle direct append and file extension cases, calculate
 1092                  * unaligned buffer size.
 1093                  */
 1094                 NFSLOCKNODE(np);
 1095                 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
 1096                     (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
 1097                         noncontig_write = 1;
 1098                 else
 1099                         noncontig_write = 0;
 1100                 if ((uio->uio_offset == np->n_size ||
 1101                     (noncontig_write != 0 &&
 1102                     lbn == (np->n_size / biosize) &&
 1103                     uio->uio_offset + n > np->n_size)) && n) {
 1104                         NFSUNLOCKNODE(np);
 1105                         /*
 1106                          * Get the buffer (in its pre-append state to maintain
 1107                          * B_CACHE if it was previously set).  Resize the
 1108                          * nfsnode after we have locked the buffer to prevent
 1109                          * readers from reading garbage.
 1110                          */
 1111                         obcount = np->n_size - (lbn * biosize);
 1112                         bp = nfs_getcacheblk(vp, lbn, obcount, td);
 1113 
 1114                         if (bp != NULL) {
 1115                                 long save;
 1116 
 1117                                 NFSLOCKNODE(np);
 1118                                 np->n_size = uio->uio_offset + n;
 1119                                 np->n_flag |= NMODIFIED;
 1120                                 np->n_flag &= ~NVNSETSZSKIP;
 1121                                 vnode_pager_setsize(vp, np->n_size);
 1122                                 NFSUNLOCKNODE(np);
 1123 
 1124                                 save = bp->b_flags & B_CACHE;
 1125                                 bcount = on + n;
 1126                                 allocbuf(bp, bcount);
 1127                                 bp->b_flags |= save;
 1128                                 if (noncontig_write != 0 && on > obcount)
 1129                                         vfs_bio_bzero_buf(bp, obcount, on -
 1130                                             obcount);
 1131                         }
 1132                 } else {
 1133                         /*
 1134                          * Obtain the locked cache block first, and then
 1135                          * adjust the file's size as appropriate.
 1136                          */
 1137                         bcount = on + n;
 1138                         if ((off_t)lbn * biosize + bcount < np->n_size) {
 1139                                 if ((off_t)(lbn + 1) * biosize < np->n_size)
 1140                                         bcount = biosize;
 1141                                 else
 1142                                         bcount = np->n_size - (off_t)lbn * biosize;
 1143                         }
 1144                         NFSUNLOCKNODE(np);
 1145                         bp = nfs_getcacheblk(vp, lbn, bcount, td);
 1146                         NFSLOCKNODE(np);
 1147                         if (uio->uio_offset + n > np->n_size) {
 1148                                 np->n_size = uio->uio_offset + n;
 1149                                 np->n_flag |= NMODIFIED;
 1150                                 np->n_flag &= ~NVNSETSZSKIP;
 1151                                 vnode_pager_setsize(vp, np->n_size);
 1152                         }
 1153                         NFSUNLOCKNODE(np);
 1154                 }
 1155 
 1156                 if (!bp) {
 1157                         error = newnfs_sigintr(nmp, td);
 1158                         if (!error)
 1159                                 error = EINTR;
 1160                         break;
 1161                 }
 1162 
 1163                 /*
 1164                  * Issue a READ if B_CACHE is not set.  In special-append
 1165                  * mode, B_CACHE is based on the buffer prior to the write
 1166                  * op and is typically set, avoiding the read.  If a read
 1167                  * is required in special append mode, the server will
 1168                  * probably send us a short-read since we extended the file
 1169                  * on our end, resulting in b_resid == 0 and, thusly,
 1170                  * B_CACHE getting set.
 1171                  *
 1172                  * We can also avoid issuing the read if the write covers
 1173                  * the entire buffer.  We have to make sure the buffer state
 1174                  * is reasonable in this case since we will not be initiating
 1175                  * I/O.  See the comments in kern/vfs_bio.c's getblk() for
 1176                  * more information.
 1177                  *
 1178                  * B_CACHE may also be set due to the buffer being cached
 1179                  * normally.
 1180                  */
 1181 
 1182                 bp_cached = 1;
 1183                 if (on == 0 && n == bcount) {
 1184                         if ((bp->b_flags & B_CACHE) == 0)
 1185                                 bp_cached = 0;
 1186                         bp->b_flags |= B_CACHE;
 1187                         bp->b_flags &= ~B_INVAL;
 1188                         bp->b_ioflags &= ~BIO_ERROR;
 1189                 }
 1190 
 1191                 if ((bp->b_flags & B_CACHE) == 0) {
 1192                         bp->b_iocmd = BIO_READ;
 1193                         vfs_busy_pages(bp, 0);
 1194                         error = ncl_doio(vp, bp, cred, td, 0);
 1195                         if (error) {
 1196                                 brelse(bp);
 1197                                 break;
 1198                         }
 1199                 }
 1200                 if (bp->b_wcred == NOCRED)
 1201                         bp->b_wcred = crhold(cred);
 1202                 NFSLOCKNODE(np);
 1203                 np->n_flag |= NMODIFIED;
 1204                 NFSUNLOCKNODE(np);
 1205 
 1206                 /*
 1207                  * If dirtyend exceeds file size, chop it down.  This should
 1208                  * not normally occur but there is an append race where it
 1209                  * might occur XXX, so we log it.
 1210                  *
 1211                  * If the chopping creates a reverse-indexed or degenerate
 1212                  * situation with dirtyoff/end, we 0 both of them.
 1213                  */
 1214 
 1215                 if (bp->b_dirtyend > bcount) {
 1216                         printf("NFS append race @%lx:%d\n",
 1217                             (long)bp->b_blkno * DEV_BSIZE,
 1218                             bp->b_dirtyend - bcount);
 1219                         bp->b_dirtyend = bcount;
 1220                 }
 1221 
 1222                 if (bp->b_dirtyoff >= bp->b_dirtyend)
 1223                         bp->b_dirtyoff = bp->b_dirtyend = 0;
 1224 
 1225                 /*
 1226                  * If the new write will leave a contiguous dirty
 1227                  * area, just update the b_dirtyoff and b_dirtyend,
 1228                  * otherwise force a write rpc of the old dirty area.
 1229                  *
 1230                  * If there has been a file lock applied to this file
 1231                  * or vfs.nfs.old_noncontig_writing is set, do the following:
 1232                  * While it is possible to merge discontiguous writes due to
 1233                  * our having a B_CACHE buffer ( and thus valid read data
 1234                  * for the hole), we don't because it could lead to
 1235                  * significant cache coherency problems with multiple clients,
 1236                  * especially if locking is implemented later on.
 1237                  *
 1238                  * If vfs.nfs.old_noncontig_writing is not set and there has
 1239                  * not been file locking done on this file:
 1240                  * Relax coherency a bit for the sake of performance and
 1241                  * expand the current dirty region to contain the new
 1242                  * write even if it means we mark some non-dirty data as
 1243                  * dirty.
 1244                  */
 1245 
 1246                 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
 1247                     (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
 1248                         if (bwrite(bp) == EINTR) {
 1249                                 error = EINTR;
 1250                                 break;
 1251                         }
 1252                         goto again;
 1253                 }
 1254 
 1255                 local_resid = uio->uio_resid;
 1256                 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
 1257 
 1258                 if (error != 0 && !bp_cached) {
 1259                         /*
 1260                          * This block has no other content then what
 1261                          * possibly was written by the faulty uiomove.
 1262                          * Release it, forgetting the data pages, to
 1263                          * prevent the leak of uninitialized data to
 1264                          * usermode.
 1265                          */
 1266                         bp->b_ioflags |= BIO_ERROR;
 1267                         brelse(bp);
 1268                         uio->uio_offset -= local_resid - uio->uio_resid;
 1269                         uio->uio_resid = local_resid;
 1270                         break;
 1271                 }
 1272 
 1273                 /*
 1274                  * Since this block is being modified, it must be written
 1275                  * again and not just committed.  Since write clustering does
 1276                  * not work for the stage 1 data write, only the stage 2
 1277                  * commit rpc, we have to clear B_CLUSTEROK as well.
 1278                  */
 1279                 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
 1280 
 1281                 /*
 1282                  * Get the partial update on the progress made from
 1283                  * uiomove, if an error occurred.
 1284                  */
 1285                 if (error != 0)
 1286                         n = local_resid - uio->uio_resid;
 1287 
 1288                 /*
 1289                  * Only update dirtyoff/dirtyend if not a degenerate
 1290                  * condition.
 1291                  */
 1292                 if (n > 0) {
 1293                         if (bp->b_dirtyend > 0) {
 1294                                 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
 1295                                 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
 1296                         } else {
 1297                                 bp->b_dirtyoff = on;
 1298                                 bp->b_dirtyend = on + n;
 1299                         }
 1300                         vfs_bio_set_valid(bp, on, n);
 1301                 }
 1302 
 1303                 /*
 1304                  * If IO_SYNC do bwrite().
 1305                  *
 1306                  * IO_INVAL appears to be unused.  The idea appears to be
 1307                  * to turn off caching in this case.  Very odd.  XXX
 1308                  */
 1309                 if ((ioflag & IO_SYNC)) {
 1310                         if (ioflag & IO_INVAL)
 1311                                 bp->b_flags |= B_NOCACHE;
 1312                         error1 = bwrite(bp);
 1313                         if (error1 != 0) {
 1314                                 if (error == 0)
 1315                                         error = error1;
 1316                                 break;
 1317                         }
 1318                 } else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
 1319                         bp->b_flags |= B_ASYNC;
 1320                         (void) ncl_writebp(bp, 0, NULL);
 1321                 } else {
 1322                         bdwrite(bp);
 1323                 }
 1324 
 1325                 if (error != 0)
 1326                         break;
 1327         } while (uio->uio_resid > 0 && n > 0);
 1328 
 1329         if (error == 0) {
 1330                 nanouptime(&ts);
 1331                 NFSLOCKNODE(np);
 1332                 np->n_localmodtime = ts;
 1333                 NFSUNLOCKNODE(np);
 1334         } else {
 1335                 if (ioflag & IO_UNIT) {
 1336                         VATTR_NULL(&vattr);
 1337                         vattr.va_size = orig_size;
 1338                         /* IO_SYNC is handled implicitely */
 1339                         (void)VOP_SETATTR(vp, &vattr, cred);
 1340                         uio->uio_offset -= orig_resid - uio->uio_resid;
 1341                         uio->uio_resid = orig_resid;
 1342                 }
 1343         }
 1344 
 1345 out:
 1346         curthread_pflags2_restore(save2);
 1347         return (error);
 1348 }
 1349 
 1350 /*
 1351  * Get an nfs cache block.
 1352  *
 1353  * Allocate a new one if the block isn't currently in the cache
 1354  * and return the block marked busy. If the calling process is
 1355  * interrupted by a signal for an interruptible mount point, return
 1356  * NULL.
 1357  *
 1358  * The caller must carefully deal with the possible B_INVAL state of
 1359  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
 1360  * indirectly), so synchronous reads can be issued without worrying about
 1361  * the B_INVAL state.  We have to be a little more careful when dealing
 1362  * with writes (see comments in nfs_write()) when extending a file past
 1363  * its EOF.
 1364  */
 1365 static struct buf *
 1366 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
 1367 {
 1368         struct buf *bp;
 1369         struct mount *mp;
 1370         struct nfsmount *nmp;
 1371 
 1372         mp = vp->v_mount;
 1373         nmp = VFSTONFS(mp);
 1374 
 1375         if (nmp->nm_flag & NFSMNT_INT) {
 1376                 sigset_t oldset;
 1377 
 1378                 newnfs_set_sigmask(td, &oldset);
 1379                 bp = getblk(vp, bn, size, PCATCH, 0, 0);
 1380                 newnfs_restore_sigmask(td, &oldset);
 1381                 while (bp == NULL) {
 1382                         if (newnfs_sigintr(nmp, td))
 1383                                 return (NULL);
 1384                         bp = getblk(vp, bn, size, 0, 2 * hz, 0);
 1385                 }
 1386         } else {
 1387                 bp = getblk(vp, bn, size, 0, 0, 0);
 1388         }
 1389 
 1390         if (vp->v_type == VREG)
 1391                 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
 1392         return (bp);
 1393 }
 1394 
 1395 /*
 1396  * Flush and invalidate all dirty buffers. If another process is already
 1397  * doing the flush, just wait for completion.
 1398  */
 1399 int
 1400 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
 1401 {
 1402         struct nfsnode *np = VTONFS(vp);
 1403         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
 1404         int error = 0, slpflag, slptimeo;
 1405         bool old_lock;
 1406         struct timespec ts;
 1407 
 1408         ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
 1409 
 1410         if ((nmp->nm_flag & NFSMNT_INT) == 0)
 1411                 intrflg = 0;
 1412         if (NFSCL_FORCEDISM(nmp->nm_mountp))
 1413                 intrflg = 1;
 1414         if (intrflg) {
 1415                 slpflag = PCATCH;
 1416                 slptimeo = 2 * hz;
 1417         } else {
 1418                 slpflag = 0;
 1419                 slptimeo = 0;
 1420         }
 1421 
 1422         old_lock = ncl_excl_start(vp);
 1423         if (old_lock)
 1424                 flags |= V_ALLOWCLEAN;
 1425 
 1426         /*
 1427          * Now, flush as required.
 1428          */
 1429         if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
 1430              vp->v_bufobj.bo_object != NULL) {
 1431                 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
 1432                 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
 1433                 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
 1434                 /*
 1435                  * If the page clean was interrupted, fail the invalidation.
 1436                  * Not doing so, we run the risk of losing dirty pages in the
 1437                  * vinvalbuf() call below.
 1438                  */
 1439                 if (intrflg && (error = newnfs_sigintr(nmp, td)))
 1440                         goto out;
 1441         }
 1442 
 1443         error = vinvalbuf(vp, flags, slpflag, 0);
 1444         while (error) {
 1445                 if (intrflg && (error = newnfs_sigintr(nmp, td)))
 1446                         goto out;
 1447                 error = vinvalbuf(vp, flags, 0, slptimeo);
 1448         }
 1449         if (NFSHASPNFS(nmp)) {
 1450                 nfscl_layoutcommit(vp, td);
 1451                 nanouptime(&ts);
 1452                 /*
 1453                  * Invalidate the attribute cache, since writes to a DS
 1454                  * won't update the size attribute.
 1455                  */
 1456                 NFSLOCKNODE(np);
 1457                 np->n_attrstamp = 0;
 1458         } else {
 1459                 nanouptime(&ts);
 1460                 NFSLOCKNODE(np);
 1461         }
 1462         if (np->n_directio_asyncwr == 0 && (np->n_flag & NMODIFIED) != 0) {
 1463                 np->n_localmodtime = ts;
 1464                 np->n_flag &= ~NMODIFIED;
 1465         }
 1466         NFSUNLOCKNODE(np);
 1467 out:
 1468         ncl_excl_finish(vp, old_lock);
 1469         return error;
 1470 }
 1471 
 1472 /*
 1473  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
 1474  * This is mainly to avoid queueing async I/O requests when the nfsiods
 1475  * are all hung on a dead server.
 1476  *
 1477  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
 1478  * is eventually dequeued by the async daemon, ncl_doio() *will*.
 1479  */
 1480 int
 1481 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
 1482 {
 1483         int iod;
 1484         int gotiod;
 1485         int slpflag = 0;
 1486         int slptimeo = 0;
 1487         int error, error2;
 1488 
 1489         /*
 1490          * Commits are usually short and sweet so lets save some cpu and
 1491          * leave the async daemons for more important rpc's (such as reads
 1492          * and writes).
 1493          *
 1494          * Readdirplus RPCs do vget()s to acquire the vnodes for entries
 1495          * in the directory in order to update attributes. This can deadlock
 1496          * with another thread that is waiting for async I/O to be done by
 1497          * an nfsiod thread while holding a lock on one of these vnodes.
 1498          * To avoid this deadlock, don't allow the async nfsiod threads to
 1499          * perform Readdirplus RPCs.
 1500          */
 1501         NFSLOCKIOD();
 1502         if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
 1503              (nmp->nm_bufqiods > ncl_numasync / 2)) ||
 1504             (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
 1505                 NFSUNLOCKIOD();
 1506                 return(EIO);
 1507         }
 1508 again:
 1509         if (nmp->nm_flag & NFSMNT_INT)
 1510                 slpflag = PCATCH;
 1511         gotiod = FALSE;
 1512 
 1513         /*
 1514          * Find a free iod to process this request.
 1515          */
 1516         for (iod = 0; iod < ncl_numasync; iod++)
 1517                 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
 1518                         gotiod = TRUE;
 1519                         break;
 1520                 }
 1521 
 1522         /*
 1523          * Try to create one if none are free.
 1524          */
 1525         if (!gotiod)
 1526                 ncl_nfsiodnew();
 1527         else {
 1528                 /*
 1529                  * Found one, so wake it up and tell it which
 1530                  * mount to process.
 1531                  */
 1532                 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
 1533                     iod, nmp));
 1534                 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
 1535                 ncl_iodmount[iod] = nmp;
 1536                 nmp->nm_bufqiods++;
 1537                 wakeup(&ncl_iodwant[iod]);
 1538         }
 1539 
 1540         /*
 1541          * If none are free, we may already have an iod working on this mount
 1542          * point.  If so, it will process our request.
 1543          */
 1544         if (!gotiod) {
 1545                 if (nmp->nm_bufqiods > 0) {
 1546                         NFS_DPF(ASYNCIO,
 1547                                 ("ncl_asyncio: %d iods are already processing mount %p\n",
 1548                                  nmp->nm_bufqiods, nmp));
 1549                         gotiod = TRUE;
 1550                 }
 1551         }
 1552 
 1553         /*
 1554          * If we have an iod which can process the request, then queue
 1555          * the buffer.
 1556          */
 1557         if (gotiod) {
 1558                 /*
 1559                  * Ensure that the queue never grows too large.  We still want
 1560                  * to asynchronize so we block rather then return EIO.
 1561                  */
 1562                 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
 1563                         NFS_DPF(ASYNCIO,
 1564                                 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
 1565                         nmp->nm_bufqwant = TRUE;
 1566                         error = newnfs_msleep(td, &nmp->nm_bufq,
 1567                             &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
 1568                            slptimeo);
 1569                         if (error) {
 1570                                 error2 = newnfs_sigintr(nmp, td);
 1571                                 if (error2) {
 1572                                         NFSUNLOCKIOD();
 1573                                         return (error2);
 1574                                 }
 1575                                 if (slpflag == PCATCH) {
 1576                                         slpflag = 0;
 1577                                         slptimeo = 2 * hz;
 1578                                 }
 1579                         }
 1580                         /*
 1581                          * We might have lost our iod while sleeping,
 1582                          * so check and loop if necessary.
 1583                          */
 1584                         goto again;
 1585                 }
 1586 
 1587                 /* We might have lost our nfsiod */
 1588                 if (nmp->nm_bufqiods == 0) {
 1589                         NFS_DPF(ASYNCIO,
 1590                                 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
 1591                         goto again;
 1592                 }
 1593 
 1594                 if (bp->b_iocmd == BIO_READ) {
 1595                         if (bp->b_rcred == NOCRED && cred != NOCRED)
 1596                                 bp->b_rcred = crhold(cred);
 1597                 } else {
 1598                         if (bp->b_wcred == NOCRED && cred != NOCRED)
 1599                                 bp->b_wcred = crhold(cred);
 1600                 }
 1601 
 1602                 if (bp->b_flags & B_REMFREE)
 1603                         bremfreef(bp);
 1604                 BUF_KERNPROC(bp);
 1605                 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
 1606                 nmp->nm_bufqlen++;
 1607                 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
 1608                         NFSLOCKNODE(VTONFS(bp->b_vp));
 1609                         VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
 1610                         VTONFS(bp->b_vp)->n_directio_asyncwr++;
 1611                         NFSUNLOCKNODE(VTONFS(bp->b_vp));
 1612                 }
 1613                 NFSUNLOCKIOD();
 1614                 return (0);
 1615         }
 1616 
 1617         NFSUNLOCKIOD();
 1618 
 1619         /*
 1620          * All the iods are busy on other mounts, so return EIO to
 1621          * force the caller to process the i/o synchronously.
 1622          */
 1623         NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
 1624         return (EIO);
 1625 }
 1626 
 1627 void
 1628 ncl_doio_directwrite(struct buf *bp)
 1629 {
 1630         int iomode, must_commit;
 1631         struct uio *uiop = (struct uio *)bp->b_caller1;
 1632         char *iov_base = uiop->uio_iov->iov_base;
 1633 
 1634         iomode = NFSWRITE_FILESYNC;
 1635         uiop->uio_td = NULL; /* NULL since we're in nfsiod */
 1636         /*
 1637          * When doing direct I/O we do not care if the
 1638          * server's write verifier has changed, but we
 1639          * do not want to update the verifier if it has
 1640          * changed, since that hides the change from
 1641          * writes being done through the buffer cache.
 1642          * By passing must_commit in set to two, the code
 1643          * in nfsrpc_writerpc() will not update the
 1644          * verifier on the mount point.
 1645          */
 1646         must_commit = 2;
 1647         ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0, 0);
 1648         KASSERT((must_commit == 2), ("ncl_doio_directwrite: Updated write"
 1649             " verifier"));
 1650         if (iomode != NFSWRITE_FILESYNC)
 1651                 printf("ncl_doio_directwrite: Broken server "
 1652                     "did not reply FILE_SYNC\n");
 1653         free(iov_base, M_NFSDIRECTIO);
 1654         free(uiop->uio_iov, M_NFSDIRECTIO);
 1655         free(uiop, M_NFSDIRECTIO);
 1656         if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
 1657                 struct nfsnode *np = VTONFS(bp->b_vp);
 1658                 NFSLOCKNODE(np);
 1659                 if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) {
 1660                         /*
 1661                          * Invalidate the attribute cache, since writes to a DS
 1662                          * won't update the size attribute.
 1663                          */
 1664                         np->n_attrstamp = 0;
 1665                 }
 1666                 np->n_directio_asyncwr--;
 1667                 if (np->n_directio_asyncwr == 0) {
 1668                         np->n_flag &= ~NMODIFIED;
 1669                         if ((np->n_flag & NFSYNCWAIT)) {
 1670                                 np->n_flag &= ~NFSYNCWAIT;
 1671                                 wakeup((caddr_t)&np->n_directio_asyncwr);
 1672                         }
 1673                 }
 1674                 NFSUNLOCKNODE(np);
 1675         }
 1676         bp->b_vp = NULL;
 1677         uma_zfree(ncl_pbuf_zone, bp);
 1678 }
 1679 
 1680 /*
 1681  * Do an I/O operation to/from a cache block. This may be called
 1682  * synchronously or from an nfsiod.
 1683  */
 1684 int
 1685 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
 1686     int called_from_strategy)
 1687 {
 1688         struct uio *uiop;
 1689         struct nfsnode *np;
 1690         struct nfsmount *nmp;
 1691         int error = 0, iomode, must_commit = 0;
 1692         struct uio uio;
 1693         struct iovec io;
 1694         struct proc *p = td ? td->td_proc : NULL;
 1695         uint8_t iocmd;
 1696 
 1697         np = VTONFS(vp);
 1698         nmp = VFSTONFS(vp->v_mount);
 1699         uiop = &uio;
 1700         uiop->uio_iov = &io;
 1701         uiop->uio_iovcnt = 1;
 1702         uiop->uio_segflg = UIO_SYSSPACE;
 1703         uiop->uio_td = td;
 1704 
 1705         /*
 1706          * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
 1707          * do this here so we do not have to do it in all the code that
 1708          * calls us.
 1709          */
 1710         bp->b_flags &= ~B_INVAL;
 1711         bp->b_ioflags &= ~BIO_ERROR;
 1712 
 1713         KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
 1714         iocmd = bp->b_iocmd;
 1715         if (iocmd == BIO_READ) {
 1716             io.iov_len = uiop->uio_resid = bp->b_bcount;
 1717             io.iov_base = bp->b_data;
 1718             uiop->uio_rw = UIO_READ;
 1719 
 1720             switch (vp->v_type) {
 1721             case VREG:
 1722                 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
 1723                 NFSINCRGLOBAL(nfsstatsv1.read_bios);
 1724                 error = ncl_readrpc(vp, uiop, cr);
 1725 
 1726                 if (!error) {
 1727                     if (uiop->uio_resid) {
 1728                         /*
 1729                          * If we had a short read with no error, we must have
 1730                          * hit a file hole.  We should zero-fill the remainder.
 1731                          * This can also occur if the server hits the file EOF.
 1732                          *
 1733                          * Holes used to be able to occur due to pending
 1734                          * writes, but that is not possible any longer.
 1735                          */
 1736                         int nread = bp->b_bcount - uiop->uio_resid;
 1737                         ssize_t left = uiop->uio_resid;
 1738 
 1739                         if (left > 0)
 1740                                 bzero((char *)bp->b_data + nread, left);
 1741                         uiop->uio_resid = 0;
 1742                     }
 1743                 }
 1744                 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
 1745                 if (p && vp->v_writecount <= -1) {
 1746                         NFSLOCKNODE(np);
 1747                         if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
 1748                                 NFSUNLOCKNODE(np);
 1749                                 PROC_LOCK(p);
 1750                                 killproc(p, "text file modification");
 1751                                 PROC_UNLOCK(p);
 1752                         } else
 1753                                 NFSUNLOCKNODE(np);
 1754                 }
 1755                 break;
 1756             case VLNK:
 1757                 uiop->uio_offset = (off_t)0;
 1758                 NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
 1759                 error = ncl_readlinkrpc(vp, uiop, cr);
 1760                 break;
 1761             case VDIR:
 1762                 NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
 1763                 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
 1764                 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
 1765                         error = ncl_readdirplusrpc(vp, uiop, cr, td);
 1766                         if (error == NFSERR_NOTSUPP)
 1767                                 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
 1768                 }
 1769                 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
 1770                         error = ncl_readdirrpc(vp, uiop, cr, td);
 1771                 /*
 1772                  * end-of-directory sets B_INVAL but does not generate an
 1773                  * error.
 1774                  */
 1775                 if (error == 0 && uiop->uio_resid == bp->b_bcount)
 1776                         bp->b_flags |= B_INVAL;
 1777                 break;
 1778             default:
 1779                 printf("ncl_doio:  type %x unexpected\n", vp->v_type);
 1780                 break;
 1781             }
 1782             if (error) {
 1783                 bp->b_ioflags |= BIO_ERROR;
 1784                 bp->b_error = error;
 1785             }
 1786         } else {
 1787             /*
 1788              * If we only need to commit, try to commit
 1789              */
 1790             if (bp->b_flags & B_NEEDCOMMIT) {
 1791                     int retv;
 1792                     off_t off;
 1793 
 1794                     off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
 1795                     retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
 1796                         bp->b_wcred, td);
 1797                     if (NFSCL_FORCEDISM(vp->v_mount) || retv == 0) {
 1798                             bp->b_dirtyoff = bp->b_dirtyend = 0;
 1799                             bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
 1800                             bp->b_resid = 0;
 1801                             bufdone(bp);
 1802                             return (0);
 1803                     }
 1804                     if (retv == NFSERR_STALEWRITEVERF) {
 1805                             ncl_clearcommit(vp->v_mount);
 1806                     }
 1807             }
 1808 
 1809             /*
 1810              * Setup for actual write
 1811              */
 1812             NFSLOCKNODE(np);
 1813             if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
 1814                 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
 1815             NFSUNLOCKNODE(np);
 1816 
 1817             if (bp->b_dirtyend > bp->b_dirtyoff) {
 1818                 io.iov_len = uiop->uio_resid = bp->b_dirtyend
 1819                     - bp->b_dirtyoff;
 1820                 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
 1821                     + bp->b_dirtyoff;
 1822                 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
 1823                 uiop->uio_rw = UIO_WRITE;
 1824                 NFSINCRGLOBAL(nfsstatsv1.write_bios);
 1825 
 1826                 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
 1827                     iomode = NFSWRITE_UNSTABLE;
 1828                 else
 1829                     iomode = NFSWRITE_FILESYNC;
 1830 
 1831                 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
 1832                     called_from_strategy, 0);
 1833 
 1834                 /*
 1835                  * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
 1836                  * to cluster the buffers needing commit.  This will allow
 1837                  * the system to submit a single commit rpc for the whole
 1838                  * cluster.  We can do this even if the buffer is not 100%
 1839                  * dirty (relative to the NFS blocksize), so we optimize the
 1840                  * append-to-file-case.
 1841                  *
 1842                  * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
 1843                  * cleared because write clustering only works for commit
 1844                  * rpc's, not for the data portion of the write).
 1845                  */
 1846 
 1847                 if (!error && iomode == NFSWRITE_UNSTABLE) {
 1848                     bp->b_flags |= B_NEEDCOMMIT;
 1849                     if (bp->b_dirtyoff == 0
 1850                         && bp->b_dirtyend == bp->b_bcount)
 1851                         bp->b_flags |= B_CLUSTEROK;
 1852                 } else {
 1853                     bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
 1854                 }
 1855 
 1856                 /*
 1857                  * For an interrupted write, the buffer is still valid
 1858                  * and the write hasn't been pushed to the server yet,
 1859                  * so we can't set BIO_ERROR and report the interruption
 1860                  * by setting B_EINTR. For the B_ASYNC case, B_EINTR
 1861                  * is not relevant, so the rpc attempt is essentially
 1862                  * a noop.  For the case of a V3 write rpc not being
 1863                  * committed to stable storage, the block is still
 1864                  * dirty and requires either a commit rpc or another
 1865                  * write rpc with iomode == NFSV3WRITE_FILESYNC before
 1866                  * the block is reused. This is indicated by setting
 1867                  * the B_DELWRI and B_NEEDCOMMIT flags.
 1868                  *
 1869                  * EIO is returned by ncl_writerpc() to indicate a recoverable
 1870                  * write error and is handled as above, except that
 1871                  * B_EINTR isn't set. One cause of this is a stale stateid
 1872                  * error for the RPC that indicates recovery is required,
 1873                  * when called with called_from_strategy != 0.
 1874                  *
 1875                  * If the buffer is marked B_PAGING, it does not reside on
 1876                  * the vp's paging queues so we cannot call bdirty().  The
 1877                  * bp in this case is not an NFS cache block so we should
 1878                  * be safe. XXX
 1879                  *
 1880                  * The logic below breaks up errors into recoverable and
 1881                  * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
 1882                  * and keep the buffer around for potential write retries.
 1883                  * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
 1884                  * and save the error in the nfsnode. This is less than ideal
 1885                  * but necessary. Keeping such buffers around could potentially
 1886                  * cause buffer exhaustion eventually (they can never be written
 1887                  * out, so will get constantly be re-dirtied). It also causes
 1888                  * all sorts of vfs panics. For non-recoverable write errors,
 1889                  * also invalidate the attrcache, so we'll be forced to go over
 1890                  * the wire for this object, returning an error to user on next
 1891                  * call (most of the time).
 1892                  */
 1893                 if (error == EINTR || error == EIO || error == ETIMEDOUT
 1894                     || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
 1895                         bp->b_flags &= ~(B_INVAL|B_NOCACHE);
 1896                         if ((bp->b_flags & B_PAGING) == 0) {
 1897                             bdirty(bp);
 1898                             bp->b_flags &= ~B_DONE;
 1899                         }
 1900                         if ((error == EINTR || error == ETIMEDOUT) &&
 1901                             (bp->b_flags & B_ASYNC) == 0)
 1902                             bp->b_flags |= B_EINTR;
 1903                 } else {
 1904                     if (error) {
 1905                         bp->b_ioflags |= BIO_ERROR;
 1906                         bp->b_flags |= B_INVAL;
 1907                         bp->b_error = np->n_error = error;
 1908                         NFSLOCKNODE(np);
 1909                         np->n_flag |= NWRITEERR;
 1910                         np->n_attrstamp = 0;
 1911                         KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
 1912                         NFSUNLOCKNODE(np);
 1913                     }
 1914                     bp->b_dirtyoff = bp->b_dirtyend = 0;
 1915                 }
 1916             } else {
 1917                 bp->b_resid = 0;
 1918                 bufdone(bp);
 1919                 return (0);
 1920             }
 1921         }
 1922         bp->b_resid = uiop->uio_resid;
 1923         if (must_commit == 1)
 1924             ncl_clearcommit(vp->v_mount);
 1925         bufdone(bp);
 1926         return (error);
 1927 }
 1928 
 1929 /*
 1930  * Used to aid in handling ftruncate() operations on the NFS client side.
 1931  * Truncation creates a number of special problems for NFS.  We have to
 1932  * throw away VM pages and buffer cache buffers that are beyond EOF, and
 1933  * we have to properly handle VM pages or (potentially dirty) buffers
 1934  * that straddle the truncation point.
 1935  */
 1936 
 1937 int
 1938 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
 1939 {
 1940         struct nfsnode *np = VTONFS(vp);
 1941         u_quad_t tsize;
 1942         int biosize = vp->v_bufobj.bo_bsize;
 1943         int error = 0;
 1944 
 1945         NFSLOCKNODE(np);
 1946         tsize = np->n_size;
 1947         np->n_size = nsize;
 1948         NFSUNLOCKNODE(np);
 1949 
 1950         if (nsize < tsize) {
 1951                 struct buf *bp;
 1952                 daddr_t lbn;
 1953                 int bufsize;
 1954 
 1955                 /*
 1956                  * vtruncbuf() doesn't get the buffer overlapping the
 1957                  * truncation point.  We may have a B_DELWRI and/or B_CACHE
 1958                  * buffer that now needs to be truncated.
 1959                  */
 1960                 error = vtruncbuf(vp, nsize, biosize);
 1961                 lbn = nsize / biosize;
 1962                 bufsize = nsize - (lbn * biosize);
 1963                 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
 1964                 if (!bp)
 1965                         return EINTR;
 1966                 if (bp->b_dirtyoff > bp->b_bcount)
 1967                         bp->b_dirtyoff = bp->b_bcount;
 1968                 if (bp->b_dirtyend > bp->b_bcount)
 1969                         bp->b_dirtyend = bp->b_bcount;
 1970                 bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
 1971                 brelse(bp);
 1972         } else {
 1973                 vnode_pager_setsize(vp, nsize);
 1974         }
 1975         return(error);
 1976 }

Cache object: 741264fc890f6ebb9caaa4b20c4cc42d


[ 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.