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-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-2  -  FREEBSD-11-1  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-4  -  FREEBSD-10-3  -  FREEBSD-10-2  -  FREEBSD-10-1  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-3  -  FREEBSD-9-2  -  FREEBSD-9-1  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-4  -  FREEBSD-8-3  -  FREEBSD-8-2  -  FREEBSD-8-1  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-4  -  FREEBSD-7-3  -  FREEBSD-7-2  -  FREEBSD-7-1  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-4  -  FREEBSD-6-3  -  FREEBSD-6-2  -  FREEBSD-6-1  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-5  -  FREEBSD-5-4  -  FREEBSD-5-3  -  FREEBSD-5-2  -  FREEBSD-5-1  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  linux-2.6  -  linux-2.4.22  -  MK83  -  MK84  -  PLAN9  -  DFBSD  -  NETBSD  -  NETBSD5  -  NETBSD4  -  NETBSD3  -  NETBSD20  -  OPENBSD  -  xnu-517  -  xnu-792  -  xnu-792.6.70  -  xnu-1228  -  xnu-1456.1.26  -  xnu-1699.24.8  -  xnu-2050.18.24  -  OPENSOLARIS  -  minix-3-1-1 
SearchContext: -  none  -  3  -  10 

    1 /*-
    2  * Copyright (c) 1989, 1993
    3  *      The Regents of the University of California.  All rights reserved.
    4  *
    5  * This code is derived from software contributed to Berkeley by
    6  * Rick Macklem at The University of Guelph.
    7  *
    8  * Redistribution and use in source and binary forms, with or without
    9  * modification, are permitted provided that the following conditions
   10  * are met:
   11  * 1. Redistributions of source code must retain the above copyright
   12  *    notice, this list of conditions and the following disclaimer.
   13  * 2. Redistributions in binary form must reproduce the above copyright
   14  *    notice, this list of conditions and the following disclaimer in the
   15  *    documentation and/or other materials provided with the distribution.
   16  * 4. Neither the name of the University nor the names of its contributors
   17  *    may be used to endorse or promote products derived from this software
   18  *    without specific prior written permission.
   19  *
   20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
   21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
   24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   30  * SUCH DAMAGE.
   31  *
   32  *      @(#)nfs_bio.c   8.9 (Berkeley) 3/30/95
   33  */
   34 
   35 #include <sys/cdefs.h>
   36 __FBSDID("$FreeBSD: stable/10/sys/fs/nfsclient/nfs_clbio.c 306663 2016-10-03 23:17:57Z rmacklem $");
   37 
   38 #include "opt_kdtrace.h"
   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 nfsstats newnfsstats;
   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 int ncl_pbuf_freecnt = -1;      /* start out unlimited */
   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 int
   84 ncl_getpages(struct vop_getpages_args *ap)
   85 {
   86         int i, error, nextoff, size, toff, count, npages;
   87         struct uio uio;
   88         struct iovec iov;
   89         vm_offset_t kva;
   90         struct buf *bp;
   91         struct vnode *vp;
   92         struct thread *td;
   93         struct ucred *cred;
   94         struct nfsmount *nmp;
   95         vm_object_t object;
   96         vm_page_t *pages;
   97         struct nfsnode *np;
   98 
   99         vp = ap->a_vp;
  100         np = VTONFS(vp);
  101         td = curthread;                         /* XXX */
  102         cred = curthread->td_ucred;             /* XXX */
  103         nmp = VFSTONFS(vp->v_mount);
  104         pages = ap->a_m;
  105         count = ap->a_count;
  106 
  107         if ((object = vp->v_object) == NULL) {
  108                 printf("ncl_getpages: called with non-merged cache vnode\n");
  109                 return (VM_PAGER_ERROR);
  110         }
  111 
  112         if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
  113                 mtx_lock(&np->n_mtx);
  114                 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
  115                         mtx_unlock(&np->n_mtx);
  116                         printf("ncl_getpages: called on non-cacheable vnode\n");
  117                         return (VM_PAGER_ERROR);
  118                 } else
  119                         mtx_unlock(&np->n_mtx);
  120         }
  121 
  122         mtx_lock(&nmp->nm_mtx);
  123         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  124             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  125                 mtx_unlock(&nmp->nm_mtx);
  126                 /* We'll never get here for v4, because we always have fsinfo */
  127                 (void)ncl_fsinfo(nmp, vp, cred, td);
  128         } else
  129                 mtx_unlock(&nmp->nm_mtx);
  130 
  131         npages = btoc(count);
  132 
  133         /*
  134          * If the requested page is partially valid, just return it and
  135          * allow the pager to zero-out the blanks.  Partially valid pages
  136          * can only occur at the file EOF.
  137          */
  138         VM_OBJECT_WLOCK(object);
  139         if (pages[ap->a_reqpage]->valid != 0) {
  140                 for (i = 0; i < npages; ++i) {
  141                         if (i != ap->a_reqpage) {
  142                                 vm_page_lock(pages[i]);
  143                                 vm_page_free(pages[i]);
  144                                 vm_page_unlock(pages[i]);
  145                         }
  146                 }
  147                 VM_OBJECT_WUNLOCK(object);
  148                 return (0);
  149         }
  150         VM_OBJECT_WUNLOCK(object);
  151 
  152         /*
  153          * We use only the kva address for the buffer, but this is extremely
  154          * convienient and fast.
  155          */
  156         bp = getpbuf(&ncl_pbuf_freecnt);
  157 
  158         kva = (vm_offset_t) bp->b_data;
  159         pmap_qenter(kva, pages, npages);
  160         PCPU_INC(cnt.v_vnodein);
  161         PCPU_ADD(cnt.v_vnodepgsin, npages);
  162 
  163         iov.iov_base = (caddr_t) kva;
  164         iov.iov_len = count;
  165         uio.uio_iov = &iov;
  166         uio.uio_iovcnt = 1;
  167         uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
  168         uio.uio_resid = count;
  169         uio.uio_segflg = UIO_SYSSPACE;
  170         uio.uio_rw = UIO_READ;
  171         uio.uio_td = td;
  172 
  173         error = ncl_readrpc(vp, &uio, cred);
  174         pmap_qremove(kva, npages);
  175 
  176         relpbuf(bp, &ncl_pbuf_freecnt);
  177 
  178         if (error && (uio.uio_resid == count)) {
  179                 printf("ncl_getpages: error %d\n", error);
  180                 VM_OBJECT_WLOCK(object);
  181                 for (i = 0; i < npages; ++i) {
  182                         if (i != ap->a_reqpage) {
  183                                 vm_page_lock(pages[i]);
  184                                 vm_page_free(pages[i]);
  185                                 vm_page_unlock(pages[i]);
  186                         }
  187                 }
  188                 VM_OBJECT_WUNLOCK(object);
  189                 return (VM_PAGER_ERROR);
  190         }
  191 
  192         /*
  193          * Calculate the number of bytes read and validate only that number
  194          * of bytes.  Note that due to pending writes, size may be 0.  This
  195          * does not mean that the remaining data is invalid!
  196          */
  197 
  198         size = count - uio.uio_resid;
  199         VM_OBJECT_WLOCK(object);
  200         for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
  201                 vm_page_t m;
  202                 nextoff = toff + PAGE_SIZE;
  203                 m = pages[i];
  204 
  205                 if (nextoff <= size) {
  206                         /*
  207                          * Read operation filled an entire page
  208                          */
  209                         m->valid = VM_PAGE_BITS_ALL;
  210                         KASSERT(m->dirty == 0,
  211                             ("nfs_getpages: page %p is dirty", m));
  212                 } else if (size > toff) {
  213                         /*
  214                          * Read operation filled a partial page.
  215                          */
  216                         m->valid = 0;
  217                         vm_page_set_valid_range(m, 0, size - toff);
  218                         KASSERT(m->dirty == 0,
  219                             ("nfs_getpages: page %p is dirty", m));
  220                 } else {
  221                         /*
  222                          * Read operation was short.  If no error
  223                          * occured we may have hit a zero-fill
  224                          * section.  We leave valid set to 0, and page
  225                          * is freed by vm_page_readahead_finish() if
  226                          * its index is not equal to requested, or
  227                          * page is zeroed and set valid by
  228                          * vm_pager_get_pages() for requested page.
  229                          */
  230                         ;
  231                 }
  232                 if (i != ap->a_reqpage)
  233                         vm_page_readahead_finish(m);
  234         }
  235         VM_OBJECT_WUNLOCK(object);
  236         return (0);
  237 }
  238 
  239 /*
  240  * Vnode op for VM putpages.
  241  */
  242 int
  243 ncl_putpages(struct vop_putpages_args *ap)
  244 {
  245         struct uio uio;
  246         struct iovec iov;
  247         vm_offset_t kva;
  248         struct buf *bp;
  249         int iomode, must_commit, i, error, npages, count;
  250         off_t offset;
  251         int *rtvals;
  252         struct vnode *vp;
  253         struct thread *td;
  254         struct ucred *cred;
  255         struct nfsmount *nmp;
  256         struct nfsnode *np;
  257         vm_page_t *pages;
  258 
  259         vp = ap->a_vp;
  260         np = VTONFS(vp);
  261         td = curthread;                         /* XXX */
  262         /* Set the cred to n_writecred for the write rpcs. */
  263         if (np->n_writecred != NULL)
  264                 cred = crhold(np->n_writecred);
  265         else
  266                 cred = crhold(curthread->td_ucred);     /* XXX */
  267         nmp = VFSTONFS(vp->v_mount);
  268         pages = ap->a_m;
  269         count = ap->a_count;
  270         rtvals = ap->a_rtvals;
  271         npages = btoc(count);
  272         offset = IDX_TO_OFF(pages[0]->pindex);
  273 
  274         mtx_lock(&nmp->nm_mtx);
  275         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  276             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  277                 mtx_unlock(&nmp->nm_mtx);
  278                 (void)ncl_fsinfo(nmp, vp, cred, td);
  279         } else
  280                 mtx_unlock(&nmp->nm_mtx);
  281 
  282         mtx_lock(&np->n_mtx);
  283         if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
  284             (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
  285                 mtx_unlock(&np->n_mtx);
  286                 printf("ncl_putpages: called on noncache-able vnode\n");
  287                 mtx_lock(&np->n_mtx);
  288         }
  289 
  290         for (i = 0; i < npages; i++)
  291                 rtvals[i] = VM_PAGER_ERROR;
  292 
  293         /*
  294          * When putting pages, do not extend file past EOF.
  295          */
  296         if (offset + count > np->n_size) {
  297                 count = np->n_size - offset;
  298                 if (count < 0)
  299                         count = 0;
  300         }
  301         mtx_unlock(&np->n_mtx);
  302 
  303         /*
  304          * We use only the kva address for the buffer, but this is extremely
  305          * convienient and fast.
  306          */
  307         bp = getpbuf(&ncl_pbuf_freecnt);
  308 
  309         kva = (vm_offset_t) bp->b_data;
  310         pmap_qenter(kva, pages, npages);
  311         PCPU_INC(cnt.v_vnodeout);
  312         PCPU_ADD(cnt.v_vnodepgsout, count);
  313 
  314         iov.iov_base = (caddr_t) kva;
  315         iov.iov_len = count;
  316         uio.uio_iov = &iov;
  317         uio.uio_iovcnt = 1;
  318         uio.uio_offset = offset;
  319         uio.uio_resid = count;
  320         uio.uio_segflg = UIO_SYSSPACE;
  321         uio.uio_rw = UIO_WRITE;
  322         uio.uio_td = td;
  323 
  324         if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
  325             iomode = NFSWRITE_UNSTABLE;
  326         else
  327             iomode = NFSWRITE_FILESYNC;
  328 
  329         error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
  330         crfree(cred);
  331 
  332         pmap_qremove(kva, npages);
  333         relpbuf(bp, &ncl_pbuf_freecnt);
  334 
  335         if (error == 0 || !nfs_keep_dirty_on_error) {
  336                 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
  337                 if (must_commit)
  338                         ncl_clearcommit(vp->v_mount);
  339         }
  340         return rtvals[0];
  341 }
  342 
  343 /*
  344  * For nfs, cache consistency can only be maintained approximately.
  345  * Although RFC1094 does not specify the criteria, the following is
  346  * believed to be compatible with the reference port.
  347  * For nfs:
  348  * If the file's modify time on the server has changed since the
  349  * last read rpc or you have written to the file,
  350  * you may have lost data cache consistency with the
  351  * server, so flush all of the file's data out of the cache.
  352  * Then force a getattr rpc to ensure that you have up to date
  353  * attributes.
  354  * NB: This implies that cache data can be read when up to
  355  * NFS_ATTRTIMEO seconds out of date. If you find that you need current
  356  * attributes this could be forced by setting n_attrstamp to 0 before
  357  * the VOP_GETATTR() call.
  358  */
  359 static inline int
  360 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
  361 {
  362         int error = 0;
  363         struct vattr vattr;
  364         struct nfsnode *np = VTONFS(vp);
  365         int old_lock;
  366 
  367         /*
  368          * Grab the exclusive lock before checking whether the cache is
  369          * consistent.
  370          * XXX - We can make this cheaper later (by acquiring cheaper locks).
  371          * But for now, this suffices.
  372          */
  373         old_lock = ncl_upgrade_vnlock(vp);
  374         if (vp->v_iflag & VI_DOOMED) {
  375                 ncl_downgrade_vnlock(vp, old_lock);
  376                 return (EBADF);
  377         }
  378 
  379         mtx_lock(&np->n_mtx);
  380         if (np->n_flag & NMODIFIED) {
  381                 mtx_unlock(&np->n_mtx);
  382                 if (vp->v_type != VREG) {
  383                         if (vp->v_type != VDIR)
  384                                 panic("nfs: bioread, not dir");
  385                         ncl_invaldir(vp);
  386                         error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
  387                         if (error)
  388                                 goto out;
  389                 }
  390                 np->n_attrstamp = 0;
  391                 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
  392                 error = VOP_GETATTR(vp, &vattr, cred);
  393                 if (error)
  394                         goto out;
  395                 mtx_lock(&np->n_mtx);
  396                 np->n_mtime = vattr.va_mtime;
  397                 mtx_unlock(&np->n_mtx);
  398         } else {
  399                 mtx_unlock(&np->n_mtx);
  400                 error = VOP_GETATTR(vp, &vattr, cred);
  401                 if (error)
  402                         return (error);
  403                 mtx_lock(&np->n_mtx);
  404                 if ((np->n_flag & NSIZECHANGED)
  405                     || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
  406                         mtx_unlock(&np->n_mtx);
  407                         if (vp->v_type == VDIR)
  408                                 ncl_invaldir(vp);
  409                         error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
  410                         if (error)
  411                                 goto out;
  412                         mtx_lock(&np->n_mtx);
  413                         np->n_mtime = vattr.va_mtime;
  414                         np->n_flag &= ~NSIZECHANGED;
  415                 }
  416                 mtx_unlock(&np->n_mtx);
  417         }
  418 out:
  419         ncl_downgrade_vnlock(vp, old_lock);
  420         return error;
  421 }
  422 
  423 /*
  424  * Vnode op for read using bio
  425  */
  426 int
  427 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
  428 {
  429         struct nfsnode *np = VTONFS(vp);
  430         int biosize, i;
  431         struct buf *bp, *rabp;
  432         struct thread *td;
  433         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
  434         daddr_t lbn, rabn;
  435         int bcount;
  436         int seqcount;
  437         int nra, error = 0, n = 0, on = 0;
  438         off_t tmp_off;
  439 
  440         KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
  441         if (uio->uio_resid == 0)
  442                 return (0);
  443         if (uio->uio_offset < 0)        /* XXX VDIR cookies can be negative */
  444                 return (EINVAL);
  445         td = uio->uio_td;
  446 
  447         mtx_lock(&nmp->nm_mtx);
  448         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  449             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  450                 mtx_unlock(&nmp->nm_mtx);
  451                 (void)ncl_fsinfo(nmp, vp, cred, td);
  452                 mtx_lock(&nmp->nm_mtx);
  453         }
  454         if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
  455                 (void) newnfs_iosize(nmp);
  456 
  457         tmp_off = uio->uio_offset + uio->uio_resid;
  458         if (vp->v_type != VDIR &&
  459             (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
  460                 mtx_unlock(&nmp->nm_mtx);
  461                 return (EFBIG);
  462         }
  463         mtx_unlock(&nmp->nm_mtx);
  464 
  465         if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
  466                 /* No caching/ no readaheads. Just read data into the user buffer */
  467                 return ncl_readrpc(vp, uio, cred);
  468 
  469         biosize = vp->v_bufobj.bo_bsize;
  470         seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
  471 
  472         error = nfs_bioread_check_cons(vp, td, cred);
  473         if (error)
  474                 return error;
  475 
  476         do {
  477             u_quad_t nsize;
  478 
  479             mtx_lock(&np->n_mtx);
  480             nsize = np->n_size;
  481             mtx_unlock(&np->n_mtx);
  482 
  483             switch (vp->v_type) {
  484             case VREG:
  485                 NFSINCRGLOBAL(newnfsstats.biocache_reads);
  486                 lbn = uio->uio_offset / biosize;
  487                 on = uio->uio_offset - (lbn * biosize);
  488 
  489                 /*
  490                  * Start the read ahead(s), as required.
  491                  */
  492                 if (nmp->nm_readahead > 0) {
  493                     for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
  494                         (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
  495                         rabn = lbn + 1 + nra;
  496                         if (incore(&vp->v_bufobj, rabn) == NULL) {
  497                             rabp = nfs_getcacheblk(vp, rabn, biosize, td);
  498                             if (!rabp) {
  499                                 error = newnfs_sigintr(nmp, td);
  500                                 return (error ? error : EINTR);
  501                             }
  502                             if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
  503                                 rabp->b_flags |= B_ASYNC;
  504                                 rabp->b_iocmd = BIO_READ;
  505                                 vfs_busy_pages(rabp, 0);
  506                                 if (ncl_asyncio(nmp, rabp, cred, td)) {
  507                                     rabp->b_flags |= B_INVAL;
  508                                     rabp->b_ioflags |= BIO_ERROR;
  509                                     vfs_unbusy_pages(rabp);
  510                                     brelse(rabp);
  511                                     break;
  512                                 }
  513                             } else {
  514                                 brelse(rabp);
  515                             }
  516                         }
  517                     }
  518                 }
  519 
  520                 /* Note that bcount is *not* DEV_BSIZE aligned. */
  521                 bcount = biosize;
  522                 if ((off_t)lbn * biosize >= nsize) {
  523                         bcount = 0;
  524                 } else if ((off_t)(lbn + 1) * biosize > nsize) {
  525                         bcount = nsize - (off_t)lbn * biosize;
  526                 }
  527                 bp = nfs_getcacheblk(vp, lbn, bcount, td);
  528 
  529                 if (!bp) {
  530                         error = newnfs_sigintr(nmp, td);
  531                         return (error ? error : EINTR);
  532                 }
  533 
  534                 /*
  535                  * If B_CACHE is not set, we must issue the read.  If this
  536                  * fails, we return an error.
  537                  */
  538 
  539                 if ((bp->b_flags & B_CACHE) == 0) {
  540                     bp->b_iocmd = BIO_READ;
  541                     vfs_busy_pages(bp, 0);
  542                     error = ncl_doio(vp, bp, cred, td, 0);
  543                     if (error) {
  544                         brelse(bp);
  545                         return (error);
  546                     }
  547                 }
  548 
  549                 /*
  550                  * on is the offset into the current bp.  Figure out how many
  551                  * bytes we can copy out of the bp.  Note that bcount is
  552                  * NOT DEV_BSIZE aligned.
  553                  *
  554                  * Then figure out how many bytes we can copy into the uio.
  555                  */
  556 
  557                 n = 0;
  558                 if (on < bcount)
  559                         n = MIN((unsigned)(bcount - on), uio->uio_resid);
  560                 break;
  561             case VLNK:
  562                 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
  563                 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
  564                 if (!bp) {
  565                         error = newnfs_sigintr(nmp, td);
  566                         return (error ? error : EINTR);
  567                 }
  568                 if ((bp->b_flags & B_CACHE) == 0) {
  569                     bp->b_iocmd = BIO_READ;
  570                     vfs_busy_pages(bp, 0);
  571                     error = ncl_doio(vp, bp, cred, td, 0);
  572                     if (error) {
  573                         bp->b_ioflags |= BIO_ERROR;
  574                         brelse(bp);
  575                         return (error);
  576                     }
  577                 }
  578                 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
  579                 on = 0;
  580                 break;
  581             case VDIR:
  582                 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
  583                 if (np->n_direofoffset
  584                     && uio->uio_offset >= np->n_direofoffset) {
  585                     return (0);
  586                 }
  587                 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
  588                 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
  589                 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
  590                 if (!bp) {
  591                     error = newnfs_sigintr(nmp, td);
  592                     return (error ? error : EINTR);
  593                 }
  594                 if ((bp->b_flags & B_CACHE) == 0) {
  595                     bp->b_iocmd = BIO_READ;
  596                     vfs_busy_pages(bp, 0);
  597                     error = ncl_doio(vp, bp, cred, td, 0);
  598                     if (error) {
  599                             brelse(bp);
  600                     }
  601                     while (error == NFSERR_BAD_COOKIE) {
  602                         ncl_invaldir(vp);
  603                         error = ncl_vinvalbuf(vp, 0, td, 1);
  604                         /*
  605                          * Yuck! The directory has been modified on the
  606                          * server. The only way to get the block is by
  607                          * reading from the beginning to get all the
  608                          * offset cookies.
  609                          *
  610                          * Leave the last bp intact unless there is an error.
  611                          * Loop back up to the while if the error is another
  612                          * NFSERR_BAD_COOKIE (double yuch!).
  613                          */
  614                         for (i = 0; i <= lbn && !error; i++) {
  615                             if (np->n_direofoffset
  616                                 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
  617                                     return (0);
  618                             bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
  619                             if (!bp) {
  620                                 error = newnfs_sigintr(nmp, td);
  621                                 return (error ? error : EINTR);
  622                             }
  623                             if ((bp->b_flags & B_CACHE) == 0) {
  624                                     bp->b_iocmd = BIO_READ;
  625                                     vfs_busy_pages(bp, 0);
  626                                     error = ncl_doio(vp, bp, cred, td, 0);
  627                                     /*
  628                                      * no error + B_INVAL == directory EOF,
  629                                      * use the block.
  630                                      */
  631                                     if (error == 0 && (bp->b_flags & B_INVAL))
  632                                             break;
  633                             }
  634                             /*
  635                              * An error will throw away the block and the
  636                              * for loop will break out.  If no error and this
  637                              * is not the block we want, we throw away the
  638                              * block and go for the next one via the for loop.
  639                              */
  640                             if (error || i < lbn)
  641                                     brelse(bp);
  642                         }
  643                     }
  644                     /*
  645                      * The above while is repeated if we hit another cookie
  646                      * error.  If we hit an error and it wasn't a cookie error,
  647                      * we give up.
  648                      */
  649                     if (error)
  650                             return (error);
  651                 }
  652 
  653                 /*
  654                  * If not eof and read aheads are enabled, start one.
  655                  * (You need the current block first, so that you have the
  656                  *  directory offset cookie of the next block.)
  657                  */
  658                 if (nmp->nm_readahead > 0 &&
  659                     (bp->b_flags & B_INVAL) == 0 &&
  660                     (np->n_direofoffset == 0 ||
  661                     (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
  662                     incore(&vp->v_bufobj, lbn + 1) == NULL) {
  663                         rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
  664                         if (rabp) {
  665                             if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
  666                                 rabp->b_flags |= B_ASYNC;
  667                                 rabp->b_iocmd = BIO_READ;
  668                                 vfs_busy_pages(rabp, 0);
  669                                 if (ncl_asyncio(nmp, rabp, cred, td)) {
  670                                     rabp->b_flags |= B_INVAL;
  671                                     rabp->b_ioflags |= BIO_ERROR;
  672                                     vfs_unbusy_pages(rabp);
  673                                     brelse(rabp);
  674                                 }
  675                             } else {
  676                                 brelse(rabp);
  677                             }
  678                         }
  679                 }
  680                 /*
  681                  * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
  682                  * chopped for the EOF condition, we cannot tell how large
  683                  * NFS directories are going to be until we hit EOF.  So
  684                  * an NFS directory buffer is *not* chopped to its EOF.  Now,
  685                  * it just so happens that b_resid will effectively chop it
  686                  * to EOF.  *BUT* this information is lost if the buffer goes
  687                  * away and is reconstituted into a B_CACHE state ( due to
  688                  * being VMIO ) later.  So we keep track of the directory eof
  689                  * in np->n_direofoffset and chop it off as an extra step
  690                  * right here.
  691                  */
  692                 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
  693                 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
  694                         n = np->n_direofoffset - uio->uio_offset;
  695                 break;
  696             default:
  697                 printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
  698                 bp = NULL;
  699                 break;
  700             };
  701 
  702             if (n > 0) {
  703                     error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
  704             }
  705             if (vp->v_type == VLNK)
  706                 n = 0;
  707             if (bp != NULL)
  708                 brelse(bp);
  709         } while (error == 0 && uio->uio_resid > 0 && n > 0);
  710         return (error);
  711 }
  712 
  713 /*
  714  * The NFS write path cannot handle iovecs with len > 1. So we need to
  715  * break up iovecs accordingly (restricting them to wsize).
  716  * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
  717  * For the ASYNC case, 2 copies are needed. The first a copy from the
  718  * user buffer to a staging buffer and then a second copy from the staging
  719  * buffer to mbufs. This can be optimized by copying from the user buffer
  720  * directly into mbufs and passing the chain down, but that requires a
  721  * fair amount of re-working of the relevant codepaths (and can be done
  722  * later).
  723  */
  724 static int
  725 nfs_directio_write(vp, uiop, cred, ioflag)
  726         struct vnode *vp;
  727         struct uio *uiop;
  728         struct ucred *cred;
  729         int ioflag;
  730 {
  731         int error;
  732         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
  733         struct thread *td = uiop->uio_td;
  734         int size;
  735         int wsize;
  736 
  737         mtx_lock(&nmp->nm_mtx);
  738         wsize = nmp->nm_wsize;
  739         mtx_unlock(&nmp->nm_mtx);
  740         if (ioflag & IO_SYNC) {
  741                 int iomode, must_commit;
  742                 struct uio uio;
  743                 struct iovec iov;
  744 do_sync:
  745                 while (uiop->uio_resid > 0) {
  746                         size = MIN(uiop->uio_resid, wsize);
  747                         size = MIN(uiop->uio_iov->iov_len, size);
  748                         iov.iov_base = uiop->uio_iov->iov_base;
  749                         iov.iov_len = size;
  750                         uio.uio_iov = &iov;
  751                         uio.uio_iovcnt = 1;
  752                         uio.uio_offset = uiop->uio_offset;
  753                         uio.uio_resid = size;
  754                         uio.uio_segflg = UIO_USERSPACE;
  755                         uio.uio_rw = UIO_WRITE;
  756                         uio.uio_td = td;
  757                         iomode = NFSWRITE_FILESYNC;
  758                         error = ncl_writerpc(vp, &uio, cred, &iomode,
  759                             &must_commit, 0);
  760                         KASSERT((must_commit == 0),
  761                                 ("ncl_directio_write: Did not commit write"));
  762                         if (error)
  763                                 return (error);
  764                         uiop->uio_offset += size;
  765                         uiop->uio_resid -= size;
  766                         if (uiop->uio_iov->iov_len <= size) {
  767                                 uiop->uio_iovcnt--;
  768                                 uiop->uio_iov++;
  769                         } else {
  770                                 uiop->uio_iov->iov_base =
  771                                         (char *)uiop->uio_iov->iov_base + size;
  772                                 uiop->uio_iov->iov_len -= size;
  773                         }
  774                 }
  775         } else {
  776                 struct uio *t_uio;
  777                 struct iovec *t_iov;
  778                 struct buf *bp;
  779 
  780                 /*
  781                  * Break up the write into blocksize chunks and hand these
  782                  * over to nfsiod's for write back.
  783                  * Unfortunately, this incurs a copy of the data. Since
  784                  * the user could modify the buffer before the write is
  785                  * initiated.
  786                  *
  787                  * The obvious optimization here is that one of the 2 copies
  788                  * in the async write path can be eliminated by copying the
  789                  * data here directly into mbufs and passing the mbuf chain
  790                  * down. But that will require a fair amount of re-working
  791                  * of the code and can be done if there's enough interest
  792                  * in NFS directio access.
  793                  */
  794                 while (uiop->uio_resid > 0) {
  795                         size = MIN(uiop->uio_resid, wsize);
  796                         size = MIN(uiop->uio_iov->iov_len, size);
  797                         bp = getpbuf(&ncl_pbuf_freecnt);
  798                         t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
  799                         t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
  800                         t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
  801                         t_iov->iov_len = size;
  802                         t_uio->uio_iov = t_iov;
  803                         t_uio->uio_iovcnt = 1;
  804                         t_uio->uio_offset = uiop->uio_offset;
  805                         t_uio->uio_resid = size;
  806                         t_uio->uio_segflg = UIO_SYSSPACE;
  807                         t_uio->uio_rw = UIO_WRITE;
  808                         t_uio->uio_td = td;
  809                         KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
  810                             uiop->uio_segflg == UIO_SYSSPACE,
  811                             ("nfs_directio_write: Bad uio_segflg"));
  812                         if (uiop->uio_segflg == UIO_USERSPACE) {
  813                                 error = copyin(uiop->uio_iov->iov_base,
  814                                     t_iov->iov_base, size);
  815                                 if (error != 0)
  816                                         goto err_free;
  817                         } else
  818                                 /*
  819                                  * UIO_SYSSPACE may never happen, but handle
  820                                  * it just in case it does.
  821                                  */
  822                                 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
  823                                     size);
  824                         bp->b_flags |= B_DIRECT;
  825                         bp->b_iocmd = BIO_WRITE;
  826                         if (cred != NOCRED) {
  827                                 crhold(cred);
  828                                 bp->b_wcred = cred;
  829                         } else
  830                                 bp->b_wcred = NOCRED;
  831                         bp->b_caller1 = (void *)t_uio;
  832                         bp->b_vp = vp;
  833                         error = ncl_asyncio(nmp, bp, NOCRED, td);
  834 err_free:
  835                         if (error) {
  836                                 free(t_iov->iov_base, M_NFSDIRECTIO);
  837                                 free(t_iov, M_NFSDIRECTIO);
  838                                 free(t_uio, M_NFSDIRECTIO);
  839                                 bp->b_vp = NULL;
  840                                 relpbuf(bp, &ncl_pbuf_freecnt);
  841                                 if (error == EINTR)
  842                                         return (error);
  843                                 goto do_sync;
  844                         }
  845                         uiop->uio_offset += size;
  846                         uiop->uio_resid -= size;
  847                         if (uiop->uio_iov->iov_len <= size) {
  848                                 uiop->uio_iovcnt--;
  849                                 uiop->uio_iov++;
  850                         } else {
  851                                 uiop->uio_iov->iov_base =
  852                                         (char *)uiop->uio_iov->iov_base + size;
  853                                 uiop->uio_iov->iov_len -= size;
  854                         }
  855                 }
  856         }
  857         return (0);
  858 }
  859 
  860 /*
  861  * Vnode op for write using bio
  862  */
  863 int
  864 ncl_write(struct vop_write_args *ap)
  865 {
  866         int biosize;
  867         struct uio *uio = ap->a_uio;
  868         struct thread *td = uio->uio_td;
  869         struct vnode *vp = ap->a_vp;
  870         struct nfsnode *np = VTONFS(vp);
  871         struct ucred *cred = ap->a_cred;
  872         int ioflag = ap->a_ioflag;
  873         struct buf *bp;
  874         struct vattr vattr;
  875         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
  876         daddr_t lbn;
  877         int bcount, noncontig_write, obcount;
  878         int bp_cached, n, on, error = 0, error1, wouldcommit;
  879         size_t orig_resid, local_resid;
  880         off_t orig_size, tmp_off;
  881 
  882         KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
  883         KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
  884             ("ncl_write proc"));
  885         if (vp->v_type != VREG)
  886                 return (EIO);
  887         mtx_lock(&np->n_mtx);
  888         if (np->n_flag & NWRITEERR) {
  889                 np->n_flag &= ~NWRITEERR;
  890                 mtx_unlock(&np->n_mtx);
  891                 return (np->n_error);
  892         } else
  893                 mtx_unlock(&np->n_mtx);
  894         mtx_lock(&nmp->nm_mtx);
  895         if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
  896             (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
  897                 mtx_unlock(&nmp->nm_mtx);
  898                 (void)ncl_fsinfo(nmp, vp, cred, td);
  899                 mtx_lock(&nmp->nm_mtx);
  900         }
  901         if (nmp->nm_wsize == 0)
  902                 (void) newnfs_iosize(nmp);
  903         mtx_unlock(&nmp->nm_mtx);
  904 
  905         /*
  906          * Synchronously flush pending buffers if we are in synchronous
  907          * mode or if we are appending.
  908          */
  909         if (ioflag & (IO_APPEND | IO_SYNC)) {
  910                 mtx_lock(&np->n_mtx);
  911                 if (np->n_flag & NMODIFIED) {
  912                         mtx_unlock(&np->n_mtx);
  913 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
  914                         /*
  915                          * Require non-blocking, synchronous writes to
  916                          * dirty files to inform the program it needs
  917                          * to fsync(2) explicitly.
  918                          */
  919                         if (ioflag & IO_NDELAY)
  920                                 return (EAGAIN);
  921 #endif
  922                         np->n_attrstamp = 0;
  923                         KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
  924                         error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
  925                         if (error)
  926                                 return (error);
  927                 } else
  928                         mtx_unlock(&np->n_mtx);
  929         }
  930 
  931         orig_resid = uio->uio_resid;
  932         mtx_lock(&np->n_mtx);
  933         orig_size = np->n_size;
  934         mtx_unlock(&np->n_mtx);
  935 
  936         /*
  937          * If IO_APPEND then load uio_offset.  We restart here if we cannot
  938          * get the append lock.
  939          */
  940         if (ioflag & IO_APPEND) {
  941                 np->n_attrstamp = 0;
  942                 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
  943                 error = VOP_GETATTR(vp, &vattr, cred);
  944                 if (error)
  945                         return (error);
  946                 mtx_lock(&np->n_mtx);
  947                 uio->uio_offset = np->n_size;
  948                 mtx_unlock(&np->n_mtx);
  949         }
  950 
  951         if (uio->uio_offset < 0)
  952                 return (EINVAL);
  953         tmp_off = uio->uio_offset + uio->uio_resid;
  954         if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
  955                 return (EFBIG);
  956         if (uio->uio_resid == 0)
  957                 return (0);
  958 
  959         if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
  960                 return nfs_directio_write(vp, uio, cred, ioflag);
  961 
  962         /*
  963          * Maybe this should be above the vnode op call, but so long as
  964          * file servers have no limits, i don't think it matters
  965          */
  966         if (vn_rlimit_fsize(vp, uio, td))
  967                 return (EFBIG);
  968 
  969         biosize = vp->v_bufobj.bo_bsize;
  970         /*
  971          * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
  972          * would exceed the local maximum per-file write commit size when
  973          * combined with those, we must decide whether to flush,
  974          * go synchronous, or return error.  We don't bother checking
  975          * IO_UNIT -- we just make all writes atomic anyway, as there's
  976          * no point optimizing for something that really won't ever happen.
  977          */
  978         wouldcommit = 0;
  979         if (!(ioflag & IO_SYNC)) {
  980                 int nflag;
  981 
  982                 mtx_lock(&np->n_mtx);
  983                 nflag = np->n_flag;
  984                 mtx_unlock(&np->n_mtx);
  985                 if (nflag & NMODIFIED) {
  986                         BO_LOCK(&vp->v_bufobj);
  987                         if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
  988                                 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
  989                                     b_bobufs) {
  990                                         if (bp->b_flags & B_NEEDCOMMIT)
  991                                                 wouldcommit += bp->b_bcount;
  992                                 }
  993                         }
  994                         BO_UNLOCK(&vp->v_bufobj);
  995                 }
  996         }
  997 
  998         do {
  999                 if (!(ioflag & IO_SYNC)) {
 1000                         wouldcommit += biosize;
 1001                         if (wouldcommit > nmp->nm_wcommitsize) {
 1002                                 np->n_attrstamp = 0;
 1003                                 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
 1004                                 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
 1005                                 if (error)
 1006                                         return (error);
 1007                                 wouldcommit = biosize;
 1008                         }
 1009                 }
 1010 
 1011                 NFSINCRGLOBAL(newnfsstats.biocache_writes);
 1012                 lbn = uio->uio_offset / biosize;
 1013                 on = uio->uio_offset - (lbn * biosize);
 1014                 n = MIN((unsigned)(biosize - on), uio->uio_resid);
 1015 again:
 1016                 /*
 1017                  * Handle direct append and file extension cases, calculate
 1018                  * unaligned buffer size.
 1019                  */
 1020                 mtx_lock(&np->n_mtx);
 1021                 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
 1022                     (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
 1023                         noncontig_write = 1;
 1024                 else
 1025                         noncontig_write = 0;
 1026                 if ((uio->uio_offset == np->n_size ||
 1027                     (noncontig_write != 0 &&
 1028                     lbn == (np->n_size / biosize) &&
 1029                     uio->uio_offset + n > np->n_size)) && n) {
 1030                         mtx_unlock(&np->n_mtx);
 1031                         /*
 1032                          * Get the buffer (in its pre-append state to maintain
 1033                          * B_CACHE if it was previously set).  Resize the
 1034                          * nfsnode after we have locked the buffer to prevent
 1035                          * readers from reading garbage.
 1036                          */
 1037                         obcount = np->n_size - (lbn * biosize);
 1038                         bp = nfs_getcacheblk(vp, lbn, obcount, td);
 1039 
 1040                         if (bp != NULL) {
 1041                                 long save;
 1042 
 1043                                 mtx_lock(&np->n_mtx);
 1044                                 np->n_size = uio->uio_offset + n;
 1045                                 np->n_flag |= NMODIFIED;
 1046                                 vnode_pager_setsize(vp, np->n_size);
 1047                                 mtx_unlock(&np->n_mtx);
 1048 
 1049                                 save = bp->b_flags & B_CACHE;
 1050                                 bcount = on + n;
 1051                                 allocbuf(bp, bcount);
 1052                                 bp->b_flags |= save;
 1053                                 if (noncontig_write != 0 && on > obcount)
 1054                                         vfs_bio_bzero_buf(bp, obcount, on -
 1055                                             obcount);
 1056                         }
 1057                 } else {
 1058                         /*
 1059                          * Obtain the locked cache block first, and then
 1060                          * adjust the file's size as appropriate.
 1061                          */
 1062                         bcount = on + n;
 1063                         if ((off_t)lbn * biosize + bcount < np->n_size) {
 1064                                 if ((off_t)(lbn + 1) * biosize < np->n_size)
 1065                                         bcount = biosize;
 1066                                 else
 1067                                         bcount = np->n_size - (off_t)lbn * biosize;
 1068                         }
 1069                         mtx_unlock(&np->n_mtx);
 1070                         bp = nfs_getcacheblk(vp, lbn, bcount, td);
 1071                         mtx_lock(&np->n_mtx);
 1072                         if (uio->uio_offset + n > np->n_size) {
 1073                                 np->n_size = uio->uio_offset + n;
 1074                                 np->n_flag |= NMODIFIED;
 1075                                 vnode_pager_setsize(vp, np->n_size);
 1076                         }
 1077                         mtx_unlock(&np->n_mtx);
 1078                 }
 1079 
 1080                 if (!bp) {
 1081                         error = newnfs_sigintr(nmp, td);
 1082                         if (!error)
 1083                                 error = EINTR;
 1084                         break;
 1085                 }
 1086 
 1087                 /*
 1088                  * Issue a READ if B_CACHE is not set.  In special-append
 1089                  * mode, B_CACHE is based on the buffer prior to the write
 1090                  * op and is typically set, avoiding the read.  If a read
 1091                  * is required in special append mode, the server will
 1092                  * probably send us a short-read since we extended the file
 1093                  * on our end, resulting in b_resid == 0 and, thusly,
 1094                  * B_CACHE getting set.
 1095                  *
 1096                  * We can also avoid issuing the read if the write covers
 1097                  * the entire buffer.  We have to make sure the buffer state
 1098                  * is reasonable in this case since we will not be initiating
 1099                  * I/O.  See the comments in kern/vfs_bio.c's getblk() for
 1100                  * more information.
 1101                  *
 1102                  * B_CACHE may also be set due to the buffer being cached
 1103                  * normally.
 1104                  */
 1105 
 1106                 bp_cached = 1;
 1107                 if (on == 0 && n == bcount) {
 1108                         if ((bp->b_flags & B_CACHE) == 0)
 1109                                 bp_cached = 0;
 1110                         bp->b_flags |= B_CACHE;
 1111                         bp->b_flags &= ~B_INVAL;
 1112                         bp->b_ioflags &= ~BIO_ERROR;
 1113                 }
 1114 
 1115                 if ((bp->b_flags & B_CACHE) == 0) {
 1116                         bp->b_iocmd = BIO_READ;
 1117                         vfs_busy_pages(bp, 0);
 1118                         error = ncl_doio(vp, bp, cred, td, 0);
 1119                         if (error) {
 1120                                 brelse(bp);
 1121                                 break;
 1122                         }
 1123                 }
 1124                 if (bp->b_wcred == NOCRED)
 1125                         bp->b_wcred = crhold(cred);
 1126                 mtx_lock(&np->n_mtx);
 1127                 np->n_flag |= NMODIFIED;
 1128                 mtx_unlock(&np->n_mtx);
 1129 
 1130                 /*
 1131                  * If dirtyend exceeds file size, chop it down.  This should
 1132                  * not normally occur but there is an append race where it
 1133                  * might occur XXX, so we log it.
 1134                  *
 1135                  * If the chopping creates a reverse-indexed or degenerate
 1136                  * situation with dirtyoff/end, we 0 both of them.
 1137                  */
 1138 
 1139                 if (bp->b_dirtyend > bcount) {
 1140                         printf("NFS append race @%lx:%d\n",
 1141                             (long)bp->b_blkno * DEV_BSIZE,
 1142                             bp->b_dirtyend - bcount);
 1143                         bp->b_dirtyend = bcount;
 1144                 }
 1145 
 1146                 if (bp->b_dirtyoff >= bp->b_dirtyend)
 1147                         bp->b_dirtyoff = bp->b_dirtyend = 0;
 1148 
 1149                 /*
 1150                  * If the new write will leave a contiguous dirty
 1151                  * area, just update the b_dirtyoff and b_dirtyend,
 1152                  * otherwise force a write rpc of the old dirty area.
 1153                  *
 1154                  * If there has been a file lock applied to this file
 1155                  * or vfs.nfs.old_noncontig_writing is set, do the following:
 1156                  * While it is possible to merge discontiguous writes due to
 1157                  * our having a B_CACHE buffer ( and thus valid read data
 1158                  * for the hole), we don't because it could lead to
 1159                  * significant cache coherency problems with multiple clients,
 1160                  * especially if locking is implemented later on.
 1161                  *
 1162                  * If vfs.nfs.old_noncontig_writing is not set and there has
 1163                  * not been file locking done on this file:
 1164                  * Relax coherency a bit for the sake of performance and
 1165                  * expand the current dirty region to contain the new
 1166                  * write even if it means we mark some non-dirty data as
 1167                  * dirty.
 1168                  */
 1169 
 1170                 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
 1171                     (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
 1172                         if (bwrite(bp) == EINTR) {
 1173                                 error = EINTR;
 1174                                 break;
 1175                         }
 1176                         goto again;
 1177                 }
 1178 
 1179                 local_resid = uio->uio_resid;
 1180                 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
 1181 
 1182                 if (error != 0 && !bp_cached) {
 1183                         /*
 1184                          * This block has no other content then what
 1185                          * possibly was written by the faulty uiomove.
 1186                          * Release it, forgetting the data pages, to
 1187                          * prevent the leak of uninitialized data to
 1188                          * usermode.
 1189                          */
 1190                         bp->b_ioflags |= BIO_ERROR;
 1191                         brelse(bp);
 1192                         uio->uio_offset -= local_resid - uio->uio_resid;
 1193                         uio->uio_resid = local_resid;
 1194                         break;
 1195                 }
 1196 
 1197                 /*
 1198                  * Since this block is being modified, it must be written
 1199                  * again and not just committed.  Since write clustering does
 1200                  * not work for the stage 1 data write, only the stage 2
 1201                  * commit rpc, we have to clear B_CLUSTEROK as well.
 1202                  */
 1203                 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
 1204 
 1205                 /*
 1206                  * Get the partial update on the progress made from
 1207                  * uiomove, if an error occured.
 1208                  */
 1209                 if (error != 0)
 1210                         n = local_resid - uio->uio_resid;
 1211 
 1212                 /*
 1213                  * Only update dirtyoff/dirtyend if not a degenerate
 1214                  * condition.
 1215                  */
 1216                 if (n > 0) {
 1217                         if (bp->b_dirtyend > 0) {
 1218                                 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
 1219                                 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
 1220                         } else {
 1221                                 bp->b_dirtyoff = on;
 1222                                 bp->b_dirtyend = on + n;
 1223                         }
 1224                         vfs_bio_set_valid(bp, on, n);
 1225                 }
 1226 
 1227                 /*
 1228                  * If IO_SYNC do bwrite().
 1229                  *
 1230                  * IO_INVAL appears to be unused.  The idea appears to be
 1231                  * to turn off caching in this case.  Very odd.  XXX
 1232                  */
 1233                 if ((ioflag & IO_SYNC)) {
 1234                         if (ioflag & IO_INVAL)
 1235                                 bp->b_flags |= B_NOCACHE;
 1236                         error1 = bwrite(bp);
 1237                         if (error1 != 0) {
 1238                                 if (error == 0)
 1239                                         error = error1;
 1240                                 break;
 1241                         }
 1242                 } else if ((n + on) == biosize) {
 1243                         bp->b_flags |= B_ASYNC;
 1244                         (void) ncl_writebp(bp, 0, NULL);
 1245                 } else {
 1246                         bdwrite(bp);
 1247                 }
 1248 
 1249                 if (error != 0)
 1250                         break;
 1251         } while (uio->uio_resid > 0 && n > 0);
 1252 
 1253         if (error != 0) {
 1254                 if (ioflag & IO_UNIT) {
 1255                         VATTR_NULL(&vattr);
 1256                         vattr.va_size = orig_size;
 1257                         /* IO_SYNC is handled implicitely */
 1258                         (void)VOP_SETATTR(vp, &vattr, cred);
 1259                         uio->uio_offset -= orig_resid - uio->uio_resid;
 1260                         uio->uio_resid = orig_resid;
 1261                 }
 1262         }
 1263 
 1264         return (error);
 1265 }
 1266 
 1267 /*
 1268  * Get an nfs cache block.
 1269  *
 1270  * Allocate a new one if the block isn't currently in the cache
 1271  * and return the block marked busy. If the calling process is
 1272  * interrupted by a signal for an interruptible mount point, return
 1273  * NULL.
 1274  *
 1275  * The caller must carefully deal with the possible B_INVAL state of
 1276  * the buffer.  ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
 1277  * indirectly), so synchronous reads can be issued without worrying about
 1278  * the B_INVAL state.  We have to be a little more careful when dealing
 1279  * with writes (see comments in nfs_write()) when extending a file past
 1280  * its EOF.
 1281  */
 1282 static struct buf *
 1283 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
 1284 {
 1285         struct buf *bp;
 1286         struct mount *mp;
 1287         struct nfsmount *nmp;
 1288 
 1289         mp = vp->v_mount;
 1290         nmp = VFSTONFS(mp);
 1291 
 1292         if (nmp->nm_flag & NFSMNT_INT) {
 1293                 sigset_t oldset;
 1294 
 1295                 newnfs_set_sigmask(td, &oldset);
 1296                 bp = getblk(vp, bn, size, PCATCH, 0, 0);
 1297                 newnfs_restore_sigmask(td, &oldset);
 1298                 while (bp == NULL) {
 1299                         if (newnfs_sigintr(nmp, td))
 1300                                 return (NULL);
 1301                         bp = getblk(vp, bn, size, 0, 2 * hz, 0);
 1302                 }
 1303         } else {
 1304                 bp = getblk(vp, bn, size, 0, 0, 0);
 1305         }
 1306 
 1307         if (vp->v_type == VREG)
 1308                 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
 1309         return (bp);
 1310 }
 1311 
 1312 /*
 1313  * Flush and invalidate all dirty buffers. If another process is already
 1314  * doing the flush, just wait for completion.
 1315  */
 1316 int
 1317 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
 1318 {
 1319         struct nfsnode *np = VTONFS(vp);
 1320         struct nfsmount *nmp = VFSTONFS(vp->v_mount);
 1321         int error = 0, slpflag, slptimeo;
 1322         int old_lock = 0;
 1323 
 1324         ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
 1325 
 1326         if ((nmp->nm_flag & NFSMNT_INT) == 0)
 1327                 intrflg = 0;
 1328         if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
 1329                 intrflg = 1;
 1330         if (intrflg) {
 1331                 slpflag = PCATCH;
 1332                 slptimeo = 2 * hz;
 1333         } else {
 1334                 slpflag = 0;
 1335                 slptimeo = 0;
 1336         }
 1337 
 1338         old_lock = ncl_upgrade_vnlock(vp);
 1339         if (vp->v_iflag & VI_DOOMED) {
 1340                 /*
 1341                  * Since vgonel() uses the generic vinvalbuf() to flush
 1342                  * dirty buffers and it does not call this function, it
 1343                  * is safe to just return OK when VI_DOOMED is set.
 1344                  */
 1345                 ncl_downgrade_vnlock(vp, old_lock);
 1346                 return (0);
 1347         }
 1348 
 1349         /*
 1350          * Now, flush as required.
 1351          */
 1352         if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
 1353                 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
 1354                 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
 1355                 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
 1356                 /*
 1357                  * If the page clean was interrupted, fail the invalidation.
 1358                  * Not doing so, we run the risk of losing dirty pages in the
 1359                  * vinvalbuf() call below.
 1360                  */
 1361                 if (intrflg && (error = newnfs_sigintr(nmp, td)))
 1362                         goto out;
 1363         }
 1364 
 1365         error = vinvalbuf(vp, flags, slpflag, 0);
 1366         while (error) {
 1367                 if (intrflg && (error = newnfs_sigintr(nmp, td)))
 1368                         goto out;
 1369                 error = vinvalbuf(vp, flags, 0, slptimeo);
 1370         }
 1371         if (NFSHASPNFS(nmp)) {
 1372                 nfscl_layoutcommit(vp, td);
 1373                 /*
 1374                  * Invalidate the attribute cache, since writes to a DS
 1375                  * won't update the size attribute.
 1376                  */
 1377                 mtx_lock(&np->n_mtx);
 1378                 np->n_attrstamp = 0;
 1379         } else
 1380                 mtx_lock(&np->n_mtx);
 1381         if (np->n_directio_asyncwr == 0)
 1382                 np->n_flag &= ~NMODIFIED;
 1383         mtx_unlock(&np->n_mtx);
 1384 out:
 1385         ncl_downgrade_vnlock(vp, old_lock);
 1386         return error;
 1387 }
 1388 
 1389 /*
 1390  * Initiate asynchronous I/O. Return an error if no nfsiods are available.
 1391  * This is mainly to avoid queueing async I/O requests when the nfsiods
 1392  * are all hung on a dead server.
 1393  *
 1394  * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
 1395  * is eventually dequeued by the async daemon, ncl_doio() *will*.
 1396  */
 1397 int
 1398 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
 1399 {
 1400         int iod;
 1401         int gotiod;
 1402         int slpflag = 0;
 1403         int slptimeo = 0;
 1404         int error, error2;
 1405 
 1406         /*
 1407          * Commits are usually short and sweet so lets save some cpu and
 1408          * leave the async daemons for more important rpc's (such as reads
 1409          * and writes).
 1410          *
 1411          * Readdirplus RPCs do vget()s to acquire the vnodes for entries
 1412          * in the directory in order to update attributes. This can deadlock
 1413          * with another thread that is waiting for async I/O to be done by
 1414          * an nfsiod thread while holding a lock on one of these vnodes.
 1415          * To avoid this deadlock, don't allow the async nfsiod threads to
 1416          * perform Readdirplus RPCs.
 1417          */
 1418         mtx_lock(&ncl_iod_mutex);
 1419         if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
 1420              (nmp->nm_bufqiods > ncl_numasync / 2)) ||
 1421             (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
 1422                 mtx_unlock(&ncl_iod_mutex);
 1423                 return(EIO);
 1424         }
 1425 again:
 1426         if (nmp->nm_flag & NFSMNT_INT)
 1427                 slpflag = PCATCH;
 1428         gotiod = FALSE;
 1429 
 1430         /*
 1431          * Find a free iod to process this request.
 1432          */
 1433         for (iod = 0; iod < ncl_numasync; iod++)
 1434                 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
 1435                         gotiod = TRUE;
 1436                         break;
 1437                 }
 1438 
 1439         /*
 1440          * Try to create one if none are free.
 1441          */
 1442         if (!gotiod)
 1443                 ncl_nfsiodnew();
 1444         else {
 1445                 /*
 1446                  * Found one, so wake it up and tell it which
 1447                  * mount to process.
 1448                  */
 1449                 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
 1450                     iod, nmp));
 1451                 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
 1452                 ncl_iodmount[iod] = nmp;
 1453                 nmp->nm_bufqiods++;
 1454                 wakeup(&ncl_iodwant[iod]);
 1455         }
 1456 
 1457         /*
 1458          * If none are free, we may already have an iod working on this mount
 1459          * point.  If so, it will process our request.
 1460          */
 1461         if (!gotiod) {
 1462                 if (nmp->nm_bufqiods > 0) {
 1463                         NFS_DPF(ASYNCIO,
 1464                                 ("ncl_asyncio: %d iods are already processing mount %p\n",
 1465                                  nmp->nm_bufqiods, nmp));
 1466                         gotiod = TRUE;
 1467                 }
 1468         }
 1469 
 1470         /*
 1471          * If we have an iod which can process the request, then queue
 1472          * the buffer.
 1473          */
 1474         if (gotiod) {
 1475                 /*
 1476                  * Ensure that the queue never grows too large.  We still want
 1477                  * to asynchronize so we block rather then return EIO.
 1478                  */
 1479                 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
 1480                         NFS_DPF(ASYNCIO,
 1481                                 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
 1482                         nmp->nm_bufqwant = TRUE;
 1483                         error = newnfs_msleep(td, &nmp->nm_bufq,
 1484                             &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
 1485                            slptimeo);
 1486                         if (error) {
 1487                                 error2 = newnfs_sigintr(nmp, td);
 1488                                 if (error2) {
 1489                                         mtx_unlock(&ncl_iod_mutex);
 1490                                         return (error2);
 1491                                 }
 1492                                 if (slpflag == PCATCH) {
 1493                                         slpflag = 0;
 1494                                         slptimeo = 2 * hz;
 1495                                 }
 1496                         }
 1497                         /*
 1498                          * We might have lost our iod while sleeping,
 1499                          * so check and loop if nescessary.
 1500                          */
 1501                         goto again;
 1502                 }
 1503 
 1504                 /* We might have lost our nfsiod */
 1505                 if (nmp->nm_bufqiods == 0) {
 1506                         NFS_DPF(ASYNCIO,
 1507                                 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
 1508                         goto again;
 1509                 }
 1510 
 1511                 if (bp->b_iocmd == BIO_READ) {
 1512                         if (bp->b_rcred == NOCRED && cred != NOCRED)
 1513                                 bp->b_rcred = crhold(cred);
 1514                 } else {
 1515                         if (bp->b_wcred == NOCRED && cred != NOCRED)
 1516                                 bp->b_wcred = crhold(cred);
 1517                 }
 1518 
 1519                 if (bp->b_flags & B_REMFREE)
 1520                         bremfreef(bp);
 1521                 BUF_KERNPROC(bp);
 1522                 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
 1523                 nmp->nm_bufqlen++;
 1524                 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
 1525                         mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
 1526                         VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
 1527                         VTONFS(bp->b_vp)->n_directio_asyncwr++;
 1528                         mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
 1529                 }
 1530                 mtx_unlock(&ncl_iod_mutex);
 1531                 return (0);
 1532         }
 1533 
 1534         mtx_unlock(&ncl_iod_mutex);
 1535 
 1536         /*
 1537          * All the iods are busy on other mounts, so return EIO to
 1538          * force the caller to process the i/o synchronously.
 1539          */
 1540         NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
 1541         return (EIO);
 1542 }
 1543 
 1544 void
 1545 ncl_doio_directwrite(struct buf *bp)
 1546 {
 1547         int iomode, must_commit;
 1548         struct uio *uiop = (struct uio *)bp->b_caller1;
 1549         char *iov_base = uiop->uio_iov->iov_base;
 1550 
 1551         iomode = NFSWRITE_FILESYNC;
 1552         uiop->uio_td = NULL; /* NULL since we're in nfsiod */
 1553         ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
 1554         KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
 1555         free(iov_base, M_NFSDIRECTIO);
 1556         free(uiop->uio_iov, M_NFSDIRECTIO);
 1557         free(uiop, M_NFSDIRECTIO);
 1558         if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
 1559                 struct nfsnode *np = VTONFS(bp->b_vp);
 1560                 mtx_lock(&np->n_mtx);
 1561                 if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
 1562                         /*
 1563                          * Invalidate the attribute cache, since writes to a DS
 1564                          * won't update the size attribute.
 1565                          */
 1566                         np->n_attrstamp = 0;
 1567                 }
 1568                 np->n_directio_asyncwr--;
 1569                 if (np->n_directio_asyncwr == 0) {
 1570                         np->n_flag &= ~NMODIFIED;
 1571                         if ((np->n_flag & NFSYNCWAIT)) {
 1572                                 np->n_flag &= ~NFSYNCWAIT;
 1573                                 wakeup((caddr_t)&np->n_directio_asyncwr);
 1574                         }
 1575                 }
 1576                 mtx_unlock(&np->n_mtx);
 1577         }
 1578         bp->b_vp = NULL;
 1579         relpbuf(bp, &ncl_pbuf_freecnt);
 1580 }
 1581 
 1582 /*
 1583  * Do an I/O operation to/from a cache block. This may be called
 1584  * synchronously or from an nfsiod.
 1585  */
 1586 int
 1587 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
 1588     int called_from_strategy)
 1589 {
 1590         struct uio *uiop;
 1591         struct nfsnode *np;
 1592         struct nfsmount *nmp;
 1593         int error = 0, iomode, must_commit = 0;
 1594         struct uio uio;
 1595         struct iovec io;
 1596         struct proc *p = td ? td->td_proc : NULL;
 1597         uint8_t iocmd;
 1598 
 1599         np = VTONFS(vp);
 1600         nmp = VFSTONFS(vp->v_mount);
 1601         uiop = &uio;
 1602         uiop->uio_iov = &io;
 1603         uiop->uio_iovcnt = 1;
 1604         uiop->uio_segflg = UIO_SYSSPACE;
 1605         uiop->uio_td = td;
 1606 
 1607         /*
 1608          * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
 1609          * do this here so we do not have to do it in all the code that
 1610          * calls us.
 1611          */
 1612         bp->b_flags &= ~B_INVAL;
 1613         bp->b_ioflags &= ~BIO_ERROR;
 1614 
 1615         KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
 1616         iocmd = bp->b_iocmd;
 1617         if (iocmd == BIO_READ) {
 1618             io.iov_len = uiop->uio_resid = bp->b_bcount;
 1619             io.iov_base = bp->b_data;
 1620             uiop->uio_rw = UIO_READ;
 1621 
 1622             switch (vp->v_type) {
 1623             case VREG:
 1624                 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
 1625                 NFSINCRGLOBAL(newnfsstats.read_bios);
 1626                 error = ncl_readrpc(vp, uiop, cr);
 1627 
 1628                 if (!error) {
 1629                     if (uiop->uio_resid) {
 1630                         /*
 1631                          * If we had a short read with no error, we must have
 1632                          * hit a file hole.  We should zero-fill the remainder.
 1633                          * This can also occur if the server hits the file EOF.
 1634                          *
 1635                          * Holes used to be able to occur due to pending
 1636                          * writes, but that is not possible any longer.
 1637                          */
 1638                         int nread = bp->b_bcount - uiop->uio_resid;
 1639                         ssize_t left = uiop->uio_resid;
 1640 
 1641                         if (left > 0)
 1642                                 bzero((char *)bp->b_data + nread, left);
 1643                         uiop->uio_resid = 0;
 1644                     }
 1645                 }
 1646                 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
 1647                 if (p && (vp->v_vflag & VV_TEXT)) {
 1648                         mtx_lock(&np->n_mtx);
 1649                         if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
 1650                                 mtx_unlock(&np->n_mtx);
 1651                                 PROC_LOCK(p);
 1652                                 killproc(p, "text file modification");
 1653                                 PROC_UNLOCK(p);
 1654                         } else
 1655                                 mtx_unlock(&np->n_mtx);
 1656                 }
 1657                 break;
 1658             case VLNK:
 1659                 uiop->uio_offset = (off_t)0;
 1660                 NFSINCRGLOBAL(newnfsstats.readlink_bios);
 1661                 error = ncl_readlinkrpc(vp, uiop, cr);
 1662                 break;
 1663             case VDIR:
 1664                 NFSINCRGLOBAL(newnfsstats.readdir_bios);
 1665                 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
 1666                 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
 1667                         error = ncl_readdirplusrpc(vp, uiop, cr, td);
 1668                         if (error == NFSERR_NOTSUPP)
 1669                                 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
 1670                 }
 1671                 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
 1672                         error = ncl_readdirrpc(vp, uiop, cr, td);
 1673                 /*
 1674                  * end-of-directory sets B_INVAL but does not generate an
 1675                  * error.
 1676                  */
 1677                 if (error == 0 && uiop->uio_resid == bp->b_bcount)
 1678                         bp->b_flags |= B_INVAL;
 1679                 break;
 1680             default:
 1681                 printf("ncl_doio:  type %x unexpected\n", vp->v_type);
 1682                 break;
 1683             };
 1684             if (error) {
 1685                 bp->b_ioflags |= BIO_ERROR;
 1686                 bp->b_error = error;
 1687             }
 1688         } else {
 1689             /*
 1690              * If we only need to commit, try to commit
 1691              */
 1692             if (bp->b_flags & B_NEEDCOMMIT) {
 1693                     int retv;
 1694                     off_t off;
 1695 
 1696                     off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
 1697                     retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
 1698                         bp->b_wcred, td);
 1699                     if (retv == 0) {
 1700                             bp->b_dirtyoff = bp->b_dirtyend = 0;
 1701                             bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
 1702                             bp->b_resid = 0;
 1703                             bufdone(bp);
 1704                             return (0);
 1705                     }
 1706                     if (retv == NFSERR_STALEWRITEVERF) {
 1707                             ncl_clearcommit(vp->v_mount);
 1708                     }
 1709             }
 1710 
 1711             /*
 1712              * Setup for actual write
 1713              */
 1714             mtx_lock(&np->n_mtx);
 1715             if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
 1716                 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
 1717             mtx_unlock(&np->n_mtx);
 1718 
 1719             if (bp->b_dirtyend > bp->b_dirtyoff) {
 1720                 io.iov_len = uiop->uio_resid = bp->b_dirtyend
 1721                     - bp->b_dirtyoff;
 1722                 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
 1723                     + bp->b_dirtyoff;
 1724                 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
 1725                 uiop->uio_rw = UIO_WRITE;
 1726                 NFSINCRGLOBAL(newnfsstats.write_bios);
 1727 
 1728                 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
 1729                     iomode = NFSWRITE_UNSTABLE;
 1730                 else
 1731                     iomode = NFSWRITE_FILESYNC;
 1732 
 1733                 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
 1734                     called_from_strategy);
 1735 
 1736                 /*
 1737                  * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
 1738                  * to cluster the buffers needing commit.  This will allow
 1739                  * the system to submit a single commit rpc for the whole
 1740                  * cluster.  We can do this even if the buffer is not 100%
 1741                  * dirty (relative to the NFS blocksize), so we optimize the
 1742                  * append-to-file-case.
 1743                  *
 1744                  * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
 1745                  * cleared because write clustering only works for commit
 1746                  * rpc's, not for the data portion of the write).
 1747                  */
 1748 
 1749                 if (!error && iomode == NFSWRITE_UNSTABLE) {
 1750                     bp->b_flags |= B_NEEDCOMMIT;
 1751                     if (bp->b_dirtyoff == 0
 1752                         && bp->b_dirtyend == bp->b_bcount)
 1753                         bp->b_flags |= B_CLUSTEROK;
 1754                 } else {
 1755                     bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
 1756                 }
 1757 
 1758                 /*
 1759                  * For an interrupted write, the buffer is still valid
 1760                  * and the write hasn't been pushed to the server yet,
 1761                  * so we can't set BIO_ERROR and report the interruption
 1762                  * by setting B_EINTR. For the B_ASYNC case, B_EINTR
 1763                  * is not relevant, so the rpc attempt is essentially
 1764                  * a noop.  For the case of a V3 write rpc not being
 1765                  * committed to stable storage, the block is still
 1766                  * dirty and requires either a commit rpc or another
 1767                  * write rpc with iomode == NFSV3WRITE_FILESYNC before
 1768                  * the block is reused. This is indicated by setting
 1769                  * the B_DELWRI and B_NEEDCOMMIT flags.
 1770                  *
 1771                  * EIO is returned by ncl_writerpc() to indicate a recoverable
 1772                  * write error and is handled as above, except that
 1773                  * B_EINTR isn't set. One cause of this is a stale stateid
 1774                  * error for the RPC that indicates recovery is required,
 1775                  * when called with called_from_strategy != 0.
 1776                  *
 1777                  * If the buffer is marked B_PAGING, it does not reside on
 1778                  * the vp's paging queues so we cannot call bdirty().  The
 1779                  * bp in this case is not an NFS cache block so we should
 1780                  * be safe. XXX
 1781                  *
 1782                  * The logic below breaks up errors into recoverable and
 1783                  * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
 1784                  * and keep the buffer around for potential write retries.
 1785                  * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
 1786                  * and save the error in the nfsnode. This is less than ideal
 1787                  * but necessary. Keeping such buffers around could potentially
 1788                  * cause buffer exhaustion eventually (they can never be written
 1789                  * out, so will get constantly be re-dirtied). It also causes
 1790                  * all sorts of vfs panics. For non-recoverable write errors,
 1791                  * also invalidate the attrcache, so we'll be forced to go over
 1792                  * the wire for this object, returning an error to user on next
 1793                  * call (most of the time).
 1794                  */
 1795                 if (error == EINTR || error == EIO || error == ETIMEDOUT
 1796                     || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
 1797                         int s;
 1798 
 1799                         s = splbio();
 1800                         bp->b_flags &= ~(B_INVAL|B_NOCACHE);
 1801                         if ((bp->b_flags & B_PAGING) == 0) {
 1802                             bdirty(bp);
 1803                             bp->b_flags &= ~B_DONE;
 1804                         }
 1805                         if ((error == EINTR || error == ETIMEDOUT) &&
 1806                             (bp->b_flags & B_ASYNC) == 0)
 1807                             bp->b_flags |= B_EINTR;
 1808                         splx(s);
 1809                 } else {
 1810                     if (error) {
 1811                         bp->b_ioflags |= BIO_ERROR;
 1812                         bp->b_flags |= B_INVAL;
 1813                         bp->b_error = np->n_error = error;
 1814                         mtx_lock(&np->n_mtx);
 1815                         np->n_flag |= NWRITEERR;
 1816                         np->n_attrstamp = 0;
 1817                         KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
 1818                         mtx_unlock(&np->n_mtx);
 1819                     }
 1820                     bp->b_dirtyoff = bp->b_dirtyend = 0;
 1821                 }
 1822             } else {
 1823                 bp->b_resid = 0;
 1824                 bufdone(bp);
 1825                 return (0);
 1826             }
 1827         }
 1828         bp->b_resid = uiop->uio_resid;
 1829         if (must_commit)
 1830             ncl_clearcommit(vp->v_mount);
 1831         bufdone(bp);
 1832         return (error);
 1833 }
 1834 
 1835 /*
 1836  * Used to aid in handling ftruncate() operations on the NFS client side.
 1837  * Truncation creates a number of special problems for NFS.  We have to
 1838  * throw away VM pages and buffer cache buffers that are beyond EOF, and
 1839  * we have to properly handle VM pages or (potentially dirty) buffers
 1840  * that straddle the truncation point.
 1841  */
 1842 
 1843 int
 1844 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
 1845 {
 1846         struct nfsnode *np = VTONFS(vp);
 1847         u_quad_t tsize;
 1848         int biosize = vp->v_bufobj.bo_bsize;
 1849         int error = 0;
 1850 
 1851         mtx_lock(&np->n_mtx);
 1852         tsize = np->n_size;
 1853         np->n_size = nsize;
 1854         mtx_unlock(&np->n_mtx);
 1855 
 1856         if (nsize < tsize) {
 1857                 struct buf *bp;
 1858                 daddr_t lbn;
 1859                 int bufsize;
 1860 
 1861                 /*
 1862                  * vtruncbuf() doesn't get the buffer overlapping the
 1863                  * truncation point.  We may have a B_DELWRI and/or B_CACHE
 1864                  * buffer that now needs to be truncated.
 1865                  */
 1866                 error = vtruncbuf(vp, cred, nsize, biosize);
 1867                 lbn = nsize / biosize;
 1868                 bufsize = nsize - (lbn * biosize);
 1869                 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
 1870                 if (!bp)
 1871                         return EINTR;
 1872                 if (bp->b_dirtyoff > bp->b_bcount)
 1873                         bp->b_dirtyoff = bp->b_bcount;
 1874                 if (bp->b_dirtyend > bp->b_bcount)
 1875                         bp->b_dirtyend = bp->b_bcount;
 1876                 bp->b_flags |= B_RELBUF;  /* don't leave garbage around */
 1877                 brelse(bp);
 1878         } else {
 1879                 vnode_pager_setsize(vp, nsize);
 1880         }
 1881         return(error);
 1882 }
 1883 

Cache object: 02d6ee59c8a8941eba9610965ebf7fd0


[ 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.