xref: /csrg-svn/sys/nfs/nfs_bio.c (revision 57783)

/*
 * Copyright (c) 1989 The Regents of the University of California.
 * All rights reserved.
 *
 * This code is derived from software contributed to Berkeley by
 * Rick Macklem at The University of Guelph.
 *
 * %sccs.include.redist.c%
 *
 *	@(#)nfs_bio.c	7.36 (Berkeley) 02/02/93
 */

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/resourcevar.h>
#include <sys/proc.h>
#include <sys/buf.h>
#include <sys/vnode.h>
#include <sys/trace.h>
#include <sys/mount.h>
#include <sys/kernel.h>

#include <vm/vm.h>

#include <nfs/nfsnode.h>
#include <nfs/rpcv2.h>
#include <nfs/nfsv2.h>
#include <nfs/nfs.h>
#include <nfs/nfsmount.h>
#include <nfs/nqnfs.h>

struct buf *nfsincore(), *nfs_getcacheblk(), *nfsgetblk();
extern struct queue_entry nfs_bufq;
extern struct proc *nfs_iodwant[NFS_MAXASYNCDAEMON];
extern int nfs_numasync;

/*
 * Vnode op for read using bio
 * Any similarity to readip() is purely coincidental
 */
nfs_bioread(vp, uio, ioflag, cred)
	register struct vnode *vp;
	register struct uio *uio;
	int ioflag;
	struct ucred *cred;
{
	register struct nfsnode *np = VTONFS(vp);
	register int biosize, diff;
	struct buf *bp, *rabp;
	struct vattr vattr;
	struct proc *p;
	struct nfsmount *nmp;
	daddr_t lbn, bn, rabn;
	caddr_t baddr;
	int got_buf, len, nra, error = 0, n, on, not_readin;

#ifdef lint
	ioflag = ioflag;
#endif /* lint */
#ifdef DIAGNOSTIC
	if (uio->uio_rw != UIO_READ)
		panic("nfs_read mode");
#endif
	if (uio->uio_resid == 0)
		return (0);
	if (uio->uio_offset < 0 && vp->v_type != VDIR)
		return (EINVAL);
	nmp = VFSTONFS(vp->v_mount);
	biosize = nmp->nm_rsize;
	p = uio->uio_procp;
	/*
	 * For nfs, cache consistency can only be maintained approximately.
	 * Although RFC1094 does not specify the criteria, the following is
	 * believed to be compatible with the reference port.
	 * For nqnfs, full cache consistency is maintained within the loop.
	 * For nfs:
	 * If the file's modify time on the server has changed since the
	 * last read rpc or you have written to the file,
	 * you may have lost data cache consistency with the
	 * server, so flush all of the file's data out of the cache.
	 * Then force a getattr rpc to ensure that you have up to date
	 * attributes.
	 * The mount flag NFSMNT_MYWRITE says "Assume that my writes are
	 * the ones changing the modify time.
	 * NB: This implies that cache data can be read when up to
	 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
	 * attributes this could be forced by setting n_attrstamp to 0 before
	 * the VOP_GETATTR() call.
	 */
	if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) {
		if (np->n_flag & NMODIFIED) {
			if ((nmp->nm_flag & NFSMNT_MYWRITE) == 0 ||
			     vp->v_type != VREG) {
				if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
					return (error);
			}
			np->n_attrstamp = 0;
			np->n_direofoffset = 0;
			if (error = VOP_GETATTR(vp, &vattr, cred, p))
				return (error);
			np->n_mtime = vattr.va_mtime.ts_sec;
		} else {
			if (error = VOP_GETATTR(vp, &vattr, cred, p))
				return (error);
			if (np->n_mtime != vattr.va_mtime.ts_sec) {
				np->n_direofoffset = 0;
				if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
					return (error);
				np->n_mtime = vattr.va_mtime.ts_sec;
			}
		}
	}
	do {

	    /*
	     * Get a valid lease. If cached data is stale, flush it.
	     */
	    if (nmp->nm_flag & NFSMNT_NQNFS) {
		if (NQNFS_CKINVALID(vp, np, NQL_READ)) {
		    do {
			error = nqnfs_getlease(vp, NQL_READ, cred, p);
		    } while (error == NQNFS_EXPIRED);
		    if (error)
			return (error);
		    if (np->n_lrev != np->n_brev ||
			((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
			if (vp->v_type == VDIR) {
			    np->n_direofoffset = 0;
			    cache_purge(vp);
			}
			if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
			    return (error);
			np->n_brev = np->n_lrev;
		    }
		} else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
		    np->n_direofoffset = 0;
		    cache_purge(vp);
		    if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
			return (error);
		}
	    }
	    if (np->n_flag & NQNFSNONCACHE) {
		switch (vp->v_type) {
		case VREG:
			error = nfs_readrpc(vp, uio, cred);
			break;
		case VLNK:
			error = nfs_readlinkrpc(vp, uio, cred);
			break;
		case VDIR:
			error = nfs_readdirrpc(vp, uio, cred);
			break;
		};
		return (error);
	    }
	    baddr = (caddr_t)0;
	    switch (vp->v_type) {
	    case VREG:
		nfsstats.biocache_reads++;
		lbn = uio->uio_offset / biosize;
		on = uio->uio_offset & (biosize-1);
		bn = lbn * (biosize / DEV_BSIZE);
		not_readin = 1;

		/*
		 * Start the read ahead(s), as required.
		 */
		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
		    lbn == vp->v_lastr + 1) {
		    for (nra = 0; nra < nmp->nm_readahead &&
			(lbn + 1 + nra) * biosize < np->n_size; nra++) {
			rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE);
			if (!nfsincore(vp, rabn)) {
			    rabp = nfs_getcacheblk(vp, rabn, biosize, p);
			    if (!rabp)
				return (EINTR);
			    if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) {
				rabp->b_flags |= (B_READ | B_ASYNC);
				if (nfs_asyncio(rabp, cred)) {
				    rabp->b_flags |= B_INVAL;
				    brelse(rabp);
				}
			    }
			}
		    }
		}

		/*
		 * If the block is in the cache and has the required data
		 * in a valid region, just copy it out.
		 * Otherwise, get the block and write back/read in,
		 * as required.
		 */
		if ((bp = nfsincore(vp, bn)) &&
		    (bp->b_flags & (B_BUSY | B_WRITEINPROG)) ==
		    (B_BUSY | B_WRITEINPROG))
			got_buf = 0;
		else {
again:
			bp = nfs_getcacheblk(vp, bn, biosize, p);
			if (!bp)
				return (EINTR);
			got_buf = 1;
			if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) {
				bp->b_flags |= B_READ;
				not_readin = 0;
				if (error = nfs_doio(bp, cred, p)) {
				    brelse(bp);
				    return (error);
				}
			}
		}
		n = min((unsigned)(biosize - on), uio->uio_resid);
		diff = np->n_size - uio->uio_offset;
		if (diff < n)
			n = diff;
		if (not_readin && n > 0) {
			if (on < bp->b_validoff || (on + n) > bp->b_validend) {
				if (!got_buf) {
				    bp = nfs_getcacheblk(vp, bn, biosize, p);
				    if (!bp)
					return (EINTR);
				    got_buf = 1;
				}
				bp->b_flags |= B_INVAL;
				if (bp->b_dirtyend > 0) {
				    if ((bp->b_flags & B_DELWRI) == 0)
					panic("nfsbioread");
				    if (VOP_BWRITE(bp) == EINTR)
					return (EINTR);
				} else
				    brelse(bp);
				goto again;
			}
		}
		vp->v_lastr = lbn;
		diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on);
		if (diff < n)
			n = diff;
		break;
	    case VLNK:
		nfsstats.biocache_readlinks++;
		bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
		if (!bp)
			return (EINTR);
		if ((bp->b_flags & B_DONE) == 0) {
			bp->b_flags |= B_READ;
			if (error = nfs_doio(bp, cred, p)) {
				brelse(bp);
				return (error);
			}
		}
		n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
		got_buf = 1;
		on = 0;
		break;
	    case VDIR:
		nfsstats.biocache_readdirs++;
		bn = (daddr_t)uio->uio_offset;
		bp = nfs_getcacheblk(vp, bn, NFS_DIRBLKSIZ, p);
		if (!bp)
			return (EINTR);
		if ((bp->b_flags & B_DONE) == 0) {
			bp->b_flags |= B_READ;
			if (error = nfs_doio(bp, cred, p)) {
				brelse(bp);
				return (error);
			}
		}

		/*
		 * If not eof and read aheads are enabled, start one.
		 * (You need the current block first, so that you have the
		 *  directory offset cookie of the next block.
		 */
		rabn = bp->b_blkno;
		if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
		    rabn != 0 && rabn != np->n_direofoffset &&
		    !nfsincore(vp, rabn)) {
			rabp = nfs_getcacheblk(vp, rabn, NFS_DIRBLKSIZ, p);
			if (rabp) {
			    if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) {
				rabp->b_flags |= (B_READ | B_ASYNC);
				if (nfs_asyncio(rabp, cred)) {
				    rabp->b_flags |= B_INVAL;
				    brelse(rabp);
				}
			    }
			}
		}
		on = 0;
		n = min(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid);
		got_buf = 1;
		break;
	    };

	    if (n > 0) {
		if (!baddr)
			baddr = bp->b_un.b_addr;
		error = uiomove(baddr + on, (int)n, uio);
	    }
	    switch (vp->v_type) {
	    case VREG:
		if (n + on == biosize || uio->uio_offset == np->n_size)
			bp->b_flags |= B_AGE;
		break;
	    case VLNK:
		n = 0;
		break;
	    case VDIR:
		uio->uio_offset = bp->b_blkno;
		break;
	    };
	    if (got_buf)
		brelse(bp);
	} while (error == 0 && uio->uio_resid > 0 && n > 0);
	return (error);
}

/*
 * Vnode op for write using bio
 */
nfs_write(ap)
	struct vop_write_args /* {
		struct vnode *a_vp;
		struct uio *a_uio;
		int  a_ioflag;
		struct ucred *a_cred;
	} */ *ap;
{
	register int biosize;
	register struct uio *uio = ap->a_uio;
	struct proc *p = uio->uio_procp;
	register struct vnode *vp = ap->a_vp;
	struct nfsnode *np = VTONFS(vp);
	register struct ucred *cred = ap->a_cred;
	int ioflag = ap->a_ioflag;
	struct buf *bp;
	struct vattr vattr;
	struct nfsmount *nmp;
	daddr_t lbn, bn;
	int n, on, error = 0;

#ifdef DIAGNOSTIC
	if (uio->uio_rw != UIO_WRITE)
		panic("nfs_write mode");
	if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
		panic("nfs_write proc");
#endif
	if (vp->v_type != VREG)
		return (EIO);
	if (np->n_flag & NWRITEERR) {
		np->n_flag &= ~NWRITEERR;
		return (np->n_error);
	}
	if (ioflag & (IO_APPEND | IO_SYNC)) {
		if (np->n_flag & NMODIFIED) {
			np->n_attrstamp = 0;
			if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
				return (error);
		}
		if (ioflag & IO_APPEND) {
			np->n_attrstamp = 0;
			if (error = VOP_GETATTR(vp, &vattr, cred, p))
				return (error);
			uio->uio_offset = np->n_size;
		}
	}
	nmp = VFSTONFS(vp->v_mount);
	if (uio->uio_offset < 0)
		return (EINVAL);
	if (uio->uio_resid == 0)
		return (0);
	/*
	 * Maybe this should be above the vnode op call, but so long as
	 * file servers have no limits, i don't think it matters
	 */
	if (p && uio->uio_offset + uio->uio_resid >
	      p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
		psignal(p, SIGXFSZ);
		return (EFBIG);
	}
	/*
	 * I use nm_rsize, not nm_wsize so that all buffer cache blocks
	 * will be the same size within a filesystem. nfs_writerpc will
	 * still use nm_wsize when sizing the rpc's.
	 */
	biosize = nmp->nm_rsize;
	do {

		/*
		 * Check for a valid write lease.
		 * If non-cachable, just do the rpc
		 */
		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
		    NQNFS_CKINVALID(vp, np, NQL_WRITE)) {
			do {
				error = nqnfs_getlease(vp, NQL_WRITE, cred, p);
			} while (error == NQNFS_EXPIRED);
			if (error)
				return (error);
			if (np->n_lrev != np->n_brev ||
			    (np->n_flag & NQNFSNONCACHE)) {
				if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
					return (error);
				np->n_brev = np->n_lrev;
			}
		}
		if (np->n_flag & NQNFSNONCACHE)
			return (nfs_writerpc(vp, uio, cred, ioflag));
		nfsstats.biocache_writes++;
		lbn = uio->uio_offset / biosize;
		on = uio->uio_offset & (biosize-1);
		n = min((unsigned)(biosize - on), uio->uio_resid);
		bn = lbn * (biosize / DEV_BSIZE);
again:
		bp = nfs_getcacheblk(vp, bn, biosize, p);
		if (!bp)
			return (EINTR);
		if (bp->b_wcred == NOCRED) {
			crhold(cred);
			bp->b_wcred = cred;
		}
		np->n_flag |= NMODIFIED;
		if (uio->uio_offset + n > np->n_size) {
			np->n_size = uio->uio_offset + n;
			vnode_pager_setsize(vp, (u_long)np->n_size);
		}

		/*
		 * If the new write will leave a contiguous dirty
		 * area, just update the b_dirtyoff and b_dirtyend,
		 * otherwise force a write rpc of the old dirty area.
		 */
		if (bp->b_dirtyend > 0 &&
		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
			bp->b_proc = p;
			if (VOP_BWRITE(bp) == EINTR)
				return (EINTR);
			goto again;
		}

		/*
		 * Check for valid write lease and get one as required.
		 * In case getblk() and/or bwrite() delayed us.
		 */
		if ((nmp->nm_flag & NFSMNT_NQNFS) &&
		    NQNFS_CKINVALID(vp, np, NQL_WRITE)) {
			do {
				error = nqnfs_getlease(vp, NQL_WRITE, cred, p);
			} while (error == NQNFS_EXPIRED);
			if (error) {
				brelse(bp);
				return (error);
			}
			if (np->n_lrev != np->n_brev ||
			    (np->n_flag & NQNFSNONCACHE)) {
				brelse(bp);
				if (error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1))
					return (error);
				np->n_brev = np->n_lrev;
				goto again;
			}
		}
		if (error = uiomove(bp->b_un.b_addr + on, n, uio)) {
			bp->b_flags |= B_ERROR;
			brelse(bp);
			return (error);
		}
		if (bp->b_dirtyend > 0) {
			bp->b_dirtyoff = min(on, bp->b_dirtyoff);
			bp->b_dirtyend = max((on + n), bp->b_dirtyend);
		} else {
			bp->b_dirtyoff = on;
			bp->b_dirtyend = on + n;
		}
#ifndef notdef
		if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff ||
		    bp->b_validoff > bp->b_dirtyend) {
			bp->b_validoff = bp->b_dirtyoff;
			bp->b_validend = bp->b_dirtyend;
		} else {
			bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff);
			bp->b_validend = max(bp->b_validend, bp->b_dirtyend);
		}
#else
		bp->b_validoff = bp->b_dirtyoff;
		bp->b_validend = bp->b_dirtyend;
#endif
		if (ioflag & IO_APPEND)
			bp->b_flags |= B_APPENDWRITE;

		/*
		 * If the lease is non-cachable or IO_SYNC do bwrite().
		 */
		if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
			bp->b_proc = p;
			if (error = VOP_BWRITE(bp))
				return (error);
		} else if ((n + on) == biosize &&
			(nmp->nm_flag & NFSMNT_NQNFS) == 0) {
			bp->b_proc = (struct proc *)0;
			bawrite(bp);
		} else
			bdwrite(bp);
	} while (uio->uio_resid > 0 && n > 0);
	return (0);
}

/*
 * Get an nfs cache block.
 * Allocate a new one if the block isn't currently in the cache
 * and return the block marked busy. If the calling process is
 * interrupted by a signal for an interruptible mount point, return
 * NULL.
 */
struct buf *
nfs_getcacheblk(vp, bn, size, p)
	struct vnode *vp;
	daddr_t bn;
	int size;
	struct proc *p;
{
	register struct buf *bp;
	struct nfsmount *nmp = VFSTONFS(vp->v_mount);

	if (nmp->nm_flag & NFSMNT_INT) {
		bp = nfsgetblk(vp, bn, size, PCATCH, 0);
		while (bp == (struct buf *)0) {
			if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
				return ((struct buf *)0);
			bp = nfsgetblk(vp, bn, size, 0, 2 * hz);
		}
	} else
		bp = nfsgetblk(vp, bn, size, 0, 0);
	return (bp);
}

/*
 * Flush and invalidate all dirty buffers. If another process is already
 * doing the flush, just wait for completion.
 */
nfs_vinvalbuf(vp, flags, cred, p, intrflg)
	struct vnode *vp;
	int flags;
	struct ucred *cred;
	struct proc *p;
	int intrflg;
{
	register struct nfsnode *np = VTONFS(vp);
	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
	int error = 0, slpflag, slptimeo;

	if ((nmp->nm_flag & NFSMNT_INT) == 0)
		intrflg = 0;
	if (intrflg) {
		slpflag = PCATCH;
		slptimeo = 2 * hz;
	} else {
		slpflag = 0;
		slptimeo = 0;
	}
	/*
	 * First wait for any other process doing a flush to complete.
	 */
	while (np->n_flag & NFLUSHINPROG) {
		np->n_flag |= NFLUSHWANT;
		error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
			slptimeo);
		if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
			return (EINTR);
	}

	/*
	 * Now, flush as required.
	 */
	np->n_flag |= NFLUSHINPROG;
	error = nfsvinvalbuf(vp, flags, cred, p, slpflag, 0);
	while (error) {
		if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
			np->n_flag &= ~NFLUSHINPROG;
			if (np->n_flag & NFLUSHWANT) {
				np->n_flag &= ~NFLUSHWANT;
				wakeup((caddr_t)&np->n_flag);
			}
			return (EINTR);
		}
		error = nfsvinvalbuf(vp, flags, cred, p, 0, slptimeo);
	}
	np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
	if (np->n_flag & NFLUSHWANT) {
		np->n_flag &= ~NFLUSHWANT;
		wakeup((caddr_t)&np->n_flag);
	}
	return (0);
}

/*
 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
 * This is mainly to avoid queueing async I/O requests when the nfsiods
 * are all hung on a dead server.
 */
nfs_asyncio(bp, cred)
	register struct buf *bp;
	struct ucred *cred;
{
	register int i;

	if (nfs_numasync == 0)
		return (EIO);
	for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
	    if (nfs_iodwant[i]) {
		if (bp->b_flags & B_READ) {
			if (bp->b_rcred == NOCRED && cred != NOCRED) {
				crhold(cred);
				bp->b_rcred = cred;
			}
		} else {
			if (bp->b_wcred == NOCRED && cred != NOCRED) {
				crhold(cred);
				bp->b_wcred = cred;
			}
		}

		queue_enter_tail(&nfs_bufq, bp, struct buf *, b_freelist);
		nfs_iodwant[i] = (struct proc *)0;
		wakeup((caddr_t)&nfs_iodwant[i]);
		return (0);
	    }
	return (EIO);
}

/*
 * Do an I/O operation to/from a cache block. This may be called
 * synchronously or from an nfsiod.
 */
int
nfs_doio(bp, cr, p)
	register struct buf *bp;
	struct cred *cr;
	struct proc *p;
{
	register struct uio *uiop;
	register struct vnode *vp;
	struct nfsnode *np;
	struct nfsmount *nmp;
	int error, diff, len;
	struct uio uio;
	struct iovec io;

	vp = bp->b_vp;
	np = VTONFS(vp);
	nmp = VFSTONFS(vp->v_mount);
	uiop = &uio;
	uiop->uio_iov = &io;
	uiop->uio_iovcnt = 1;
	uiop->uio_segflg = UIO_SYSSPACE;
	uiop->uio_procp = p;

	/*
	 * Historically, paging was done with physio, but no more.
	 */
	if (bp->b_flags & B_PHYS)
	    panic("doio phys");
	if (bp->b_flags & B_READ) {
	    io.iov_len = uiop->uio_resid = bp->b_bcount;
	    io.iov_base = bp->b_un.b_addr;
	    uiop->uio_rw = UIO_READ;
	    switch (vp->v_type) {
	    case VREG:
		uiop->uio_offset = bp->b_blkno * DEV_BSIZE;
		nfsstats.read_bios++;
		error = nfs_readrpc(vp, uiop, cr);
		if (!error) {
		    bp->b_validoff = 0;
		    if (uiop->uio_resid) {
			/*
			 * If len > 0, there is a hole in the file and
			 * no writes after the hole have been pushed to
			 * the server yet.
			 * Just zero fill the rest of the valid area.
			 */
			diff = bp->b_bcount - uiop->uio_resid;
			len = np->n_size - (bp->b_blkno * DEV_BSIZE
				+ diff);
			if (len > 0) {
			    len = min(len, uiop->uio_resid);
			    bzero(bp->b_un.b_addr + diff, len);
			    bp->b_validend = diff + len;
			} else
			    bp->b_validend = diff;
		    } else
			bp->b_validend = bp->b_bcount;
		}
		if (p && (vp->v_flag & VTEXT) &&
			(((nmp->nm_flag & NFSMNT_NQNFS) &&
			  np->n_lrev != np->n_brev) ||
			 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
			  np->n_mtime != np->n_vattr.va_mtime.ts_sec))) {
			uprintf("Process killed due to text file modification\n");
			psignal(p, SIGKILL);
			p->p_flag |= SKEEP;
		}
		break;
	    case VLNK:
		uiop->uio_offset = 0;
		nfsstats.readlink_bios++;
		error = nfs_readlinkrpc(vp, uiop, cr);
		break;
	    case VDIR:
		uiop->uio_offset = bp->b_lblkno;
		nfsstats.readdir_bios++;
		if (VFSTONFS(vp->v_mount)->nm_flag & NFSMNT_NQNFS)
		    error = nfs_readdirlookrpc(vp, uiop, cr);
		else
		    error = nfs_readdirrpc(vp, uiop, cr);
		/*
		 * Save offset cookie in b_blkno.
		 */
		bp->b_blkno = uiop->uio_offset;
		break;
	    };
	    if (error) {
		bp->b_flags |= B_ERROR;
		bp->b_error = error;
	    }
	} else {
	    io.iov_len = uiop->uio_resid = bp->b_dirtyend
		- bp->b_dirtyoff;
	    uiop->uio_offset = (bp->b_blkno * DEV_BSIZE)
		+ bp->b_dirtyoff;
	    io.iov_base = bp->b_un.b_addr + bp->b_dirtyoff;
	    uiop->uio_rw = UIO_WRITE;
	    nfsstats.write_bios++;
	    if (bp->b_flags & B_APPENDWRITE)
		error = nfs_writerpc(vp, uiop, cr, IO_APPEND);
	    else
		error = nfs_writerpc(vp, uiop, cr, 0);
	    bp->b_flags &= ~(B_WRITEINPROG | B_APPENDWRITE);

	    /*
	     * For an interrupted write, the buffer is still valid and the
	     * write hasn't been pushed to the server yet, so we can't set
	     * B_ERROR and report the interruption by setting B_EINTR. For
	     * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt
	     * is essentially a noop.
	     */
	    if (error == EINTR) {
		bp->b_flags &= ~B_INVAL;
		bp->b_flags |= B_DELWRI;

		/*
		 * Since for the B_ASYNC case, nfs_bwrite() has reassigned the
		 * buffer to the clean list, we have to reassign it back to the
		 * dirty one. Ugh.
		 */
		if (bp->b_flags & B_ASYNC)
		    reassignbuf(bp, vp);
		else
		    bp->b_flags |= B_EINTR;
	    } else {
		if (error) {
		    bp->b_flags |= B_ERROR;
		    bp->b_error = np->n_error = error;
		    np->n_flag |= NWRITEERR;
		}
		bp->b_dirtyoff = bp->b_dirtyend = 0;
	    }
	}
	bp->b_resid = uiop->uio_resid;
	biodone(bp);
	return (error);
}