diff --git a/sys/fs/cd9660/cd9660_lookup.c b/sys/fs/cd9660/cd9660_lookup.c index e23d8f3b030d..590fc1911428 100644 --- a/sys/fs/cd9660/cd9660_lookup.c +++ b/sys/fs/cd9660/cd9660_lookup.c @@ -1,483 +1,485 @@ /*- * Copyright (c) 1989, 1993, 1994 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley * by Pace Willisson (pace@blitz.com). The Rock Ridge Extension * Support code is derived from software contributed to Berkeley * by Atsushi Murai (amurai@spec.co.jp). * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)ufs_lookup.c 7.33 (Berkeley) 5/19/91 * @(#)cd9660_lookup.c 8.2 (Berkeley) 1/23/94 */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include /* * Convert a component of a pathname into a pointer to a locked inode. * This is a very central and rather complicated routine. * If the filesystem is not maintained in a strict tree hierarchy, * this can result in a deadlock situation (see comments in code below). * * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on * whether the name is to be looked up, created, renamed, or deleted. * When CREATE, RENAME, or DELETE is specified, information usable in * creating, renaming, or deleting a directory entry may be calculated. * If flag has LOCKPARENT or'ed into it and the target of the pathname * exists, lookup returns both the target and its parent directory locked. * When creating or renaming and LOCKPARENT is specified, the target may * not be ".". When deleting and LOCKPARENT is specified, the target may * be "."., but the caller must check to ensure it does an vrele and iput * instead of two iputs. * * Overall outline of ufs_lookup: * * search for name in directory, to found or notfound * notfound: * if creating, return locked directory, leaving info on available slots * else return error * found: * if at end of path and deleting, return information to allow delete * if at end of path and rewriting (RENAME and LOCKPARENT), lock target * inode and return info to allow rewrite * if not at end, add name to cache; if at end and neither creating * nor deleting, add name to cache * * NOTE: (LOOKUP | LOCKPARENT) currently returns the parent inode unlocked. */ int cd9660_lookup(ap) struct vop_cachedlookup_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; } */ *ap; { struct vnode *vdp; /* vnode for directory being searched */ struct iso_node *dp; /* inode for directory being searched */ struct iso_mnt *imp; /* filesystem that directory is in */ struct buf *bp; /* a buffer of directory entries */ struct iso_directory_record *ep;/* the current directory entry */ struct iso_directory_record *ep2;/* copy of current directory entry */ int entryoffsetinblock; /* offset of ep in bp's buffer */ int saveoffset = 0; /* offset of last directory entry in dir */ doff_t i_diroff; /* cached i_diroff value. */ doff_t i_offset; /* cached i_offset value. */ int numdirpasses; /* strategy for directory search */ doff_t endsearch; /* offset to end directory search */ struct vnode *pdp; /* saved dp during symlink work */ struct vnode *tdp; /* returned by cd9660_vget_internal */ u_long bmask; /* block offset mask */ int error; ino_t ino, i_ino; int ltype, reclen; u_short namelen; int isoflags; char altname[NAME_MAX]; int res; int assoc, len; char *name; struct mount *mp; struct vnode **vpp = ap->a_vpp; struct componentname *cnp = ap->a_cnp; int flags = cnp->cn_flags; int nameiop = cnp->cn_nameiop; ep2 = ep = NULL; bp = NULL; *vpp = NULL; vdp = ap->a_dvp; dp = VTOI(vdp); imp = dp->i_mnt; /* * We now have a segment name to search for, and a directory to search. */ ino = reclen = 0; i_diroff = dp->i_diroff; len = cnp->cn_namelen; name = cnp->cn_nameptr; /* * A leading `=' means, we are looking for an associated file */ if ((assoc = (imp->iso_ftype != ISO_FTYPE_RRIP && *name == ASSOCCHAR))) { len--; name++; } /* * If there is cached information on a previous search of * this directory, pick up where we last left off. * We cache only lookups as these are the most common * and have the greatest payoff. Caching CREATE has little * benefit as it usually must search the entire directory * to determine that the entry does not exist. Caching the * location of the last DELETE or RENAME has not reduced * profiling time and hence has been removed in the interest * of simplicity. */ bmask = imp->im_bmask; if (nameiop != LOOKUP || i_diroff == 0 || i_diroff > dp->i_size) { entryoffsetinblock = 0; i_offset = 0; numdirpasses = 1; } else { i_offset = i_diroff; if ((entryoffsetinblock = i_offset & bmask) && (error = cd9660_blkatoff(vdp, (off_t)i_offset, NULL, &bp))) return (error); numdirpasses = 2; nchstats.ncs_2passes++; } endsearch = dp->i_size; searchloop: while (i_offset < endsearch) { /* * If offset is on a block boundary, * read the next directory block. * Release previous if it exists. */ if ((i_offset & bmask) == 0) { if (bp != NULL) brelse(bp); if ((error = cd9660_blkatoff(vdp, (off_t)i_offset, NULL, &bp)) != 0) return (error); entryoffsetinblock = 0; } /* * Get pointer to next entry. */ ep = (struct iso_directory_record *) ((char *)bp->b_data + entryoffsetinblock); reclen = isonum_711(ep->length); if (reclen == 0) { /* skip to next block, if any */ i_offset = (i_offset & ~bmask) + imp->logical_block_size; continue; } if (reclen < ISO_DIRECTORY_RECORD_SIZE) /* illegal entry, stop */ break; if (entryoffsetinblock + reclen > imp->logical_block_size) /* entries are not allowed to cross boundaries */ break; namelen = isonum_711(ep->name_len); isoflags = isonum_711(imp->iso_ftype == ISO_FTYPE_HIGH_SIERRA? &ep->date[6]: ep->flags); if (reclen < ISO_DIRECTORY_RECORD_SIZE + namelen) /* illegal entry, stop */ break; /* * Check for a name match. */ switch (imp->iso_ftype) { default: if (!(isoflags & 4) == !assoc) { if ((len == 1 && *name == '.') || (flags & ISDOTDOT)) { if (namelen == 1 && ep->name[0] == ((flags & ISDOTDOT) ? 1 : 0)) { /* * Save directory entry's inode number and * release directory buffer. */ i_ino = isodirino(ep, imp); goto found; } if (namelen != 1 || ep->name[0] != 0) goto notfound; } else if (!(res = isofncmp(name, len, ep->name, namelen, imp->joliet_level, imp->im_flags, imp->im_d2l, imp->im_l2d))) { if (isoflags & 2) ino = isodirino(ep, imp); else ino = dbtob(bp->b_blkno) + entryoffsetinblock; saveoffset = i_offset; } else if (ino) goto foundino; #ifdef NOSORTBUG /* On some CDs directory entries are not sorted correctly */ else if (res < 0) goto notfound; else if (res > 0 && numdirpasses == 2) numdirpasses++; #endif } break; case ISO_FTYPE_RRIP: if (isonum_711(ep->flags)&2) ino = isodirino(ep, imp); else ino = dbtob(bp->b_blkno) + entryoffsetinblock; i_ino = ino; cd9660_rrip_getname(ep, altname, &namelen, &i_ino, imp); if (namelen == cnp->cn_namelen && !bcmp(name,altname,namelen)) goto found; ino = 0; break; } i_offset += reclen; entryoffsetinblock += reclen; } if (ino) { foundino: i_ino = ino; if (saveoffset != i_offset) { if (lblkno(imp, i_offset) != lblkno(imp, saveoffset)) { if (bp != NULL) brelse(bp); if ((error = cd9660_blkatoff(vdp, (off_t)saveoffset, NULL, &bp)) != 0) return (error); } entryoffsetinblock = saveoffset & bmask; ep = (struct iso_directory_record *) ((char *)bp->b_data + entryoffsetinblock); reclen = isonum_711(ep->length); i_offset = saveoffset; } goto found; } notfound: /* * If we started in the middle of the directory and failed * to find our target, we must check the beginning as well. */ if (numdirpasses == 2) { numdirpasses--; i_offset = 0; endsearch = i_diroff; goto searchloop; } if (bp != NULL) brelse(bp); /* * Insert name into cache (as non-existent) if appropriate. */ if (cnp->cn_flags & MAKEENTRY) cache_enter(vdp, *vpp, cnp); if (nameiop == CREATE || nameiop == RENAME) return (EROFS); return (ENOENT); found: if (numdirpasses == 2) nchstats.ncs_pass2++; /* * Found component in pathname. * If the final component of path name, save information * in the cache as to where the entry was found. */ if ((flags & ISLASTCN) && nameiop == LOOKUP) dp->i_diroff = i_offset; /* * Step through the translation in the name. We do not `vput' the * directory because we may need it again if a symbolic link * is relative to the current directory. Instead we save it * unlocked as "pdp". We must get the target inode before unlocking * the directory to insure that the inode will not be removed * before we get it. We prevent deadlock by always fetching * inodes from the root, moving down the directory tree. Thus * when following backward pointers ".." we must unlock the * parent directory before getting the requested directory. * There is a potential race condition here if both the current * and parent directories are removed before the `vget' for the * inode associated with ".." returns. We hope that this occurs * infrequently since we cannot avoid this race condition without * implementing a sophisticated deadlock detection algorithm. * Note also that this simple deadlock detection scheme will not * work if the filesystem has any hard links other than ".." * that point backwards in the directory structure. */ pdp = vdp; /* * Make a copy of the directory entry for non "." lookups so * we can drop the buffer before calling vget() to avoid a * lock order reversal between the vnode lock and the buffer * lock. */ if (dp->i_number != i_ino) { ep2 = malloc(reclen, M_TEMP, M_WAITOK); bcopy(ep, ep2, reclen); ep = ep2; } brelse(bp); /* * If ino is different from i_ino, * it's a relocated directory. */ if (flags & ISDOTDOT) { /* * Expanded copy of vn_vget_ino() so that we can use * cd9660_vget_internal(). */ mp = pdp->v_mount; ltype = VOP_ISLOCKED(pdp); error = vfs_busy(mp, MBF_NOWAIT); if (error != 0) { + vfs_ref(mp); VOP_UNLOCK(pdp, 0); error = vfs_busy(mp, 0); vn_lock(pdp, ltype | LK_RETRY); + vfs_rel(mp); if (error) return (ENOENT); if (pdp->v_iflag & VI_DOOMED) { vfs_unbusy(mp); return (ENOENT); } } VOP_UNLOCK(pdp, 0); error = cd9660_vget_internal(vdp->v_mount, i_ino, cnp->cn_lkflags, &tdp, i_ino != ino, ep); free(ep2, M_TEMP); vfs_unbusy(mp); vn_lock(pdp, ltype | LK_RETRY); if (pdp->v_iflag & VI_DOOMED) { if (error == 0) vput(tdp); error = ENOENT; } if (error) return (error); *vpp = tdp; } else if (dp->i_number == i_ino) { VREF(vdp); /* we want ourself, ie "." */ /* * When we lookup "." we still can be asked to lock it * differently. */ ltype = cnp->cn_lkflags & LK_TYPE_MASK; if (ltype != VOP_ISLOCKED(vdp)) { if (ltype == LK_EXCLUSIVE) vn_lock(vdp, LK_UPGRADE | LK_RETRY); else /* if (ltype == LK_SHARED) */ vn_lock(vdp, LK_DOWNGRADE | LK_RETRY); } *vpp = vdp; } else { error = cd9660_vget_internal(vdp->v_mount, i_ino, cnp->cn_lkflags, &tdp, i_ino != ino, ep); free(ep2, M_TEMP); if (error) return (error); *vpp = tdp; } /* * Insert name into cache if appropriate. */ if (cnp->cn_flags & MAKEENTRY) cache_enter(vdp, *vpp, cnp); return (0); } /* * Return buffer with the contents of block "offset" from the beginning of * directory "ip". If "res" is non-zero, fill it in with a pointer to the * remaining space in the directory. */ int cd9660_blkatoff(vp, offset, res, bpp) struct vnode *vp; off_t offset; char **res; struct buf **bpp; { struct iso_node *ip; struct iso_mnt *imp; struct buf *bp; daddr_t lbn; int bsize, bshift, error; ip = VTOI(vp); imp = ip->i_mnt; lbn = lblkno(imp, offset); bsize = blksize(imp, ip, lbn); bshift = imp->im_bshift; if ((error = bread(vp, lbn, bsize, NOCRED, &bp)) != 0) { brelse(bp); *bpp = NULL; return (error); } /* * We must BMAP the buffer because the directory code may use b_blkno * to calculate the inode for certain types of directory entries. * We could get away with not doing it before we VMIO-backed the * directories because the buffers would get freed atomically with * the invalidation of their data. But with VMIO-backed buffers * the buffers may be freed and then later reconstituted - and the * reconstituted buffer will have no knowledge of b_blkno. */ if (bp->b_blkno == bp->b_lblkno) { bp->b_blkno = (ip->iso_start + bp->b_lblkno) << (bshift - DEV_BSHIFT); } if (res) *res = (char *)bp->b_data + blkoff(imp, offset); *bpp = bp; return (0); } diff --git a/sys/kern/vfs_vnops.c b/sys/kern/vfs_vnops.c index bef2804037b3..f67064613e06 100644 --- a/sys/kern/vfs_vnops.c +++ b/sys/kern/vfs_vnops.c @@ -1,1330 +1,1332 @@ /*- * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94 */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static fo_rdwr_t vn_read; static fo_rdwr_t vn_write; static fo_truncate_t vn_truncate; static fo_ioctl_t vn_ioctl; static fo_poll_t vn_poll; static fo_kqfilter_t vn_kqfilter; static fo_stat_t vn_statfile; static fo_close_t vn_closefile; struct fileops vnops = { .fo_read = vn_read, .fo_write = vn_write, .fo_truncate = vn_truncate, .fo_ioctl = vn_ioctl, .fo_poll = vn_poll, .fo_kqfilter = vn_kqfilter, .fo_stat = vn_statfile, .fo_close = vn_closefile, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE }; int vn_open(ndp, flagp, cmode, fp) struct nameidata *ndp; int *flagp, cmode; struct file *fp; { struct thread *td = ndp->ni_cnd.cn_thread; return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); } /* * Common code for vnode open operations. * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. * * Note that this does NOT free nameidata for the successful case, * due to the NDINIT being done elsewhere. */ int vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, struct ucred *cred, struct file *fp) { struct vnode *vp; struct mount *mp; struct thread *td = ndp->ni_cnd.cn_thread; struct vattr vat; struct vattr *vap = &vat; int fmode, error; accmode_t accmode; int vfslocked, mpsafe; mpsafe = ndp->ni_cnd.cn_flags & MPSAFE; restart: vfslocked = 0; fmode = *flagp; if (fmode & O_CREAT) { ndp->ni_cnd.cn_nameiop = CREATE; ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | MPSAFE; if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) ndp->ni_cnd.cn_flags |= FOLLOW; if (!(vn_open_flags & VN_OPEN_NOAUDIT)) ndp->ni_cnd.cn_flags |= AUDITVNODE1; bwillwrite(); if ((error = namei(ndp)) != 0) return (error); vfslocked = NDHASGIANT(ndp); if (!mpsafe) ndp->ni_cnd.cn_flags &= ~MPSAFE; if (ndp->ni_vp == NULL) { VATTR_NULL(vap); vap->va_type = VREG; vap->va_mode = cmode; if (fmode & O_EXCL) vap->va_vaflags |= VA_EXCLUSIVE; if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { NDFREE(ndp, NDF_ONLY_PNBUF); vput(ndp->ni_dvp); VFS_UNLOCK_GIANT(vfslocked); if ((error = vn_start_write(NULL, &mp, V_XSLEEP | PCATCH)) != 0) return (error); goto restart; } #ifdef MAC error = mac_vnode_check_create(cred, ndp->ni_dvp, &ndp->ni_cnd, vap); if (error == 0) #endif error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, &ndp->ni_cnd, vap); vput(ndp->ni_dvp); vn_finished_write(mp); if (error) { VFS_UNLOCK_GIANT(vfslocked); NDFREE(ndp, NDF_ONLY_PNBUF); return (error); } fmode &= ~O_TRUNC; vp = ndp->ni_vp; } else { if (ndp->ni_dvp == ndp->ni_vp) vrele(ndp->ni_dvp); else vput(ndp->ni_dvp); ndp->ni_dvp = NULL; vp = ndp->ni_vp; if (fmode & O_EXCL) { error = EEXIST; goto bad; } fmode &= ~O_CREAT; } } else { ndp->ni_cnd.cn_nameiop = LOOKUP; ndp->ni_cnd.cn_flags = ISOPEN | ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF | MPSAFE; if (!(fmode & FWRITE)) ndp->ni_cnd.cn_flags |= LOCKSHARED; if (!(vn_open_flags & VN_OPEN_NOAUDIT)) ndp->ni_cnd.cn_flags |= AUDITVNODE1; if ((error = namei(ndp)) != 0) return (error); if (!mpsafe) ndp->ni_cnd.cn_flags &= ~MPSAFE; vfslocked = NDHASGIANT(ndp); vp = ndp->ni_vp; } if (vp->v_type == VLNK) { error = EMLINK; goto bad; } if (vp->v_type == VSOCK) { error = EOPNOTSUPP; goto bad; } accmode = 0; if (fmode & (FWRITE | O_TRUNC)) { if (vp->v_type == VDIR) { error = EISDIR; goto bad; } accmode |= VWRITE; } if (fmode & FREAD) accmode |= VREAD; if (fmode & FEXEC) accmode |= VEXEC; if (fmode & O_APPEND) accmode |= VAPPEND; #ifdef MAC error = mac_vnode_check_open(cred, vp, accmode); if (error) goto bad; #endif if ((fmode & O_CREAT) == 0) { if (accmode & VWRITE) { error = vn_writechk(vp); if (error) goto bad; } if (accmode) { error = VOP_ACCESS(vp, accmode, cred, td); if (error) goto bad; } } if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0) goto bad; if (fmode & FWRITE) vp->v_writecount++; *flagp = fmode; ASSERT_VOP_LOCKED(vp, "vn_open_cred"); if (!mpsafe) VFS_UNLOCK_GIANT(vfslocked); return (0); bad: NDFREE(ndp, NDF_ONLY_PNBUF); vput(vp); VFS_UNLOCK_GIANT(vfslocked); *flagp = fmode; ndp->ni_vp = NULL; return (error); } /* * Check for write permissions on the specified vnode. * Prototype text segments cannot be written. */ int vn_writechk(vp) register struct vnode *vp; { ASSERT_VOP_LOCKED(vp, "vn_writechk"); /* * If there's shared text associated with * the vnode, try to free it up once. If * we fail, we can't allow writing. */ if (vp->v_vflag & VV_TEXT) return (ETXTBSY); return (0); } /* * Vnode close call */ int vn_close(vp, flags, file_cred, td) register struct vnode *vp; int flags; struct ucred *file_cred; struct thread *td; { struct mount *mp; int error, lock_flags; if (!(flags & FWRITE) && vp->v_mount != NULL && vp->v_mount->mnt_kern_flag & MNTK_EXTENDED_SHARED) lock_flags = LK_SHARED; else lock_flags = LK_EXCLUSIVE; VFS_ASSERT_GIANT(vp->v_mount); vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, lock_flags | LK_RETRY); if (flags & FWRITE) { VNASSERT(vp->v_writecount > 0, vp, ("vn_close: negative writecount")); vp->v_writecount--; } error = VOP_CLOSE(vp, flags, file_cred, td); vput(vp); vn_finished_write(mp); return (error); } /* * Heuristic to detect sequential operation. */ static int sequential_heuristic(struct uio *uio, struct file *fp) { /* * Offset 0 is handled specially. open() sets f_seqcount to 1 so * that the first I/O is normally considered to be slightly * sequential. Seeking to offset 0 doesn't change sequentiality * unless previous seeks have reduced f_seqcount to 0, in which * case offset 0 is not special. */ if ((uio->uio_offset == 0 && fp->f_seqcount > 0) || uio->uio_offset == fp->f_nextoff) { /* * f_seqcount is in units of fixed-size blocks so that it * depends mainly on the amount of sequential I/O and not * much on the number of sequential I/O's. The fixed size * of 16384 is hard-coded here since it is (not quite) just * a magic size that works well here. This size is more * closely related to the best I/O size for real disks than * to any block size used by software. */ fp->f_seqcount += howmany(uio->uio_resid, 16384); if (fp->f_seqcount > IO_SEQMAX) fp->f_seqcount = IO_SEQMAX; return (fp->f_seqcount << IO_SEQSHIFT); } /* Not sequential. Quickly draw-down sequentiality. */ if (fp->f_seqcount > 1) fp->f_seqcount = 1; else fp->f_seqcount = 0; return (0); } /* * Package up an I/O request on a vnode into a uio and do it. */ int vn_rdwr(rw, vp, base, len, offset, segflg, ioflg, active_cred, file_cred, aresid, td) enum uio_rw rw; struct vnode *vp; void *base; int len; off_t offset; enum uio_seg segflg; int ioflg; struct ucred *active_cred; struct ucred *file_cred; int *aresid; struct thread *td; { struct uio auio; struct iovec aiov; struct mount *mp; struct ucred *cred; int error, lock_flags; VFS_ASSERT_GIANT(vp->v_mount); if ((ioflg & IO_NODELOCKED) == 0) { mp = NULL; if (rw == UIO_WRITE) { if (vp->v_type != VCHR && (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) return (error); if (MNT_SHARED_WRITES(mp) || ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) { lock_flags = LK_SHARED; } else { lock_flags = LK_EXCLUSIVE; } vn_lock(vp, lock_flags | LK_RETRY); } else vn_lock(vp, LK_SHARED | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); auio.uio_iov = &aiov; auio.uio_iovcnt = 1; aiov.iov_base = base; aiov.iov_len = len; auio.uio_resid = len; auio.uio_offset = offset; auio.uio_segflg = segflg; auio.uio_rw = rw; auio.uio_td = td; error = 0; #ifdef MAC if ((ioflg & IO_NOMACCHECK) == 0) { if (rw == UIO_READ) error = mac_vnode_check_read(active_cred, file_cred, vp); else error = mac_vnode_check_write(active_cred, file_cred, vp); } #endif if (error == 0) { if (file_cred) cred = file_cred; else cred = active_cred; if (rw == UIO_READ) error = VOP_READ(vp, &auio, ioflg, cred); else error = VOP_WRITE(vp, &auio, ioflg, cred); } if (aresid) *aresid = auio.uio_resid; else if (auio.uio_resid && error == 0) error = EIO; if ((ioflg & IO_NODELOCKED) == 0) { if (rw == UIO_WRITE && vp->v_type != VCHR) vn_finished_write(mp); VOP_UNLOCK(vp, 0); } return (error); } /* * Package up an I/O request on a vnode into a uio and do it. The I/O * request is split up into smaller chunks and we try to avoid saturating * the buffer cache while potentially holding a vnode locked, so we * check bwillwrite() before calling vn_rdwr(). We also call uio_yield() * to give other processes a chance to lock the vnode (either other processes * core'ing the same binary, or unrelated processes scanning the directory). */ int vn_rdwr_inchunks(rw, vp, base, len, offset, segflg, ioflg, active_cred, file_cred, aresid, td) enum uio_rw rw; struct vnode *vp; void *base; size_t len; off_t offset; enum uio_seg segflg; int ioflg; struct ucred *active_cred; struct ucred *file_cred; size_t *aresid; struct thread *td; { int error = 0; int iaresid; VFS_ASSERT_GIANT(vp->v_mount); do { int chunk; /* * Force `offset' to a multiple of MAXBSIZE except possibly * for the first chunk, so that filesystems only need to * write full blocks except possibly for the first and last * chunks. */ chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; if (chunk > len) chunk = len; if (rw != UIO_READ && vp->v_type == VREG) bwillwrite(); iaresid = 0; error = vn_rdwr(rw, vp, base, chunk, offset, segflg, ioflg, active_cred, file_cred, &iaresid, td); len -= chunk; /* aresid calc already includes length */ if (error) break; offset += chunk; base = (char *)base + chunk; uio_yield(); } while (len); if (aresid) *aresid = len + iaresid; return (error); } /* * File table vnode read routine. */ static int vn_read(fp, uio, active_cred, flags, td) struct file *fp; struct uio *uio; struct ucred *active_cred; struct thread *td; int flags; { struct vnode *vp; int error, ioflag; struct mtx *mtxp; int vfslocked; KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", uio->uio_td, td)); mtxp = NULL; vp = fp->f_vnode; ioflag = 0; if (fp->f_flag & FNONBLOCK) ioflag |= IO_NDELAY; if (fp->f_flag & O_DIRECT) ioflag |= IO_DIRECT; vfslocked = VFS_LOCK_GIANT(vp->v_mount); /* * According to McKusick the vn lock was protecting f_offset here. * It is now protected by the FOFFSET_LOCKED flag. */ if ((flags & FOF_OFFSET) == 0) { mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); while(fp->f_vnread_flags & FOFFSET_LOCKED) { fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; msleep(&fp->f_vnread_flags, mtxp, PUSER -1, "vnread offlock", 0); } fp->f_vnread_flags |= FOFFSET_LOCKED; mtx_unlock(mtxp); vn_lock(vp, LK_SHARED | LK_RETRY); uio->uio_offset = fp->f_offset; } else vn_lock(vp, LK_SHARED | LK_RETRY); ioflag |= sequential_heuristic(uio, fp); #ifdef MAC error = mac_vnode_check_read(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_READ(vp, uio, ioflag, fp->f_cred); if ((flags & FOF_OFFSET) == 0) { fp->f_offset = uio->uio_offset; mtx_lock(mtxp); if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) wakeup(&fp->f_vnread_flags); fp->f_vnread_flags = 0; mtx_unlock(mtxp); } fp->f_nextoff = uio->uio_offset; VOP_UNLOCK(vp, 0); VFS_UNLOCK_GIANT(vfslocked); return (error); } /* * File table vnode write routine. */ static int vn_write(fp, uio, active_cred, flags, td) struct file *fp; struct uio *uio; struct ucred *active_cred; struct thread *td; int flags; { struct vnode *vp; struct mount *mp; int error, ioflag, lock_flags; int vfslocked; KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", uio->uio_td, td)); vp = fp->f_vnode; vfslocked = VFS_LOCK_GIANT(vp->v_mount); if (vp->v_type == VREG) bwillwrite(); ioflag = IO_UNIT; if (vp->v_type == VREG && (fp->f_flag & O_APPEND)) ioflag |= IO_APPEND; if (fp->f_flag & FNONBLOCK) ioflag |= IO_NDELAY; if (fp->f_flag & O_DIRECT) ioflag |= IO_DIRECT; if ((fp->f_flag & O_FSYNC) || (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS))) ioflag |= IO_SYNC; mp = NULL; if (vp->v_type != VCHR && (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) goto unlock; if ((MNT_SHARED_WRITES(mp) || ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) && (flags & FOF_OFFSET) != 0) { lock_flags = LK_SHARED; } else { lock_flags = LK_EXCLUSIVE; } vn_lock(vp, lock_flags | LK_RETRY); if ((flags & FOF_OFFSET) == 0) uio->uio_offset = fp->f_offset; ioflag |= sequential_heuristic(uio, fp); #ifdef MAC error = mac_vnode_check_write(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); if ((flags & FOF_OFFSET) == 0) fp->f_offset = uio->uio_offset; fp->f_nextoff = uio->uio_offset; VOP_UNLOCK(vp, 0); if (vp->v_type != VCHR) vn_finished_write(mp); unlock: VFS_UNLOCK_GIANT(vfslocked); return (error); } /* * File table truncate routine. */ static int vn_truncate(fp, length, active_cred, td) struct file *fp; off_t length; struct ucred *active_cred; struct thread *td; { struct vattr vattr; struct mount *mp; struct vnode *vp; int vfslocked; int error; vp = fp->f_vnode; vfslocked = VFS_LOCK_GIANT(vp->v_mount); error = vn_start_write(vp, &mp, V_WAIT | PCATCH); if (error) { VFS_UNLOCK_GIANT(vfslocked); return (error); } vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); if (vp->v_type == VDIR) { error = EISDIR; goto out; } #ifdef MAC error = mac_vnode_check_write(active_cred, fp->f_cred, vp); if (error) goto out; #endif error = vn_writechk(vp); if (error == 0) { VATTR_NULL(&vattr); vattr.va_size = length; error = VOP_SETATTR(vp, &vattr, fp->f_cred); } out: VOP_UNLOCK(vp, 0); vn_finished_write(mp); VFS_UNLOCK_GIANT(vfslocked); return (error); } /* * File table vnode stat routine. */ static int vn_statfile(fp, sb, active_cred, td) struct file *fp; struct stat *sb; struct ucred *active_cred; struct thread *td; { struct vnode *vp = fp->f_vnode; int vfslocked; int error; vfslocked = VFS_LOCK_GIANT(vp->v_mount); vn_lock(vp, LK_SHARED | LK_RETRY); error = vn_stat(vp, sb, active_cred, fp->f_cred, td); VOP_UNLOCK(vp, 0); VFS_UNLOCK_GIANT(vfslocked); return (error); } /* * Stat a vnode; implementation for the stat syscall */ int vn_stat(vp, sb, active_cred, file_cred, td) struct vnode *vp; register struct stat *sb; struct ucred *active_cred; struct ucred *file_cred; struct thread *td; { struct vattr vattr; register struct vattr *vap; int error; u_short mode; #ifdef MAC error = mac_vnode_check_stat(active_cred, file_cred, vp); if (error) return (error); #endif vap = &vattr; /* * Initialize defaults for new and unusual fields, so that file * systems which don't support these fields don't need to know * about them. */ vap->va_birthtime.tv_sec = -1; vap->va_birthtime.tv_nsec = 0; vap->va_fsid = VNOVAL; vap->va_rdev = NODEV; error = VOP_GETATTR(vp, vap, active_cred); if (error) return (error); /* * Zero the spare stat fields */ bzero(sb, sizeof *sb); /* * Copy from vattr table */ if (vap->va_fsid != VNOVAL) sb->st_dev = vap->va_fsid; else sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0]; sb->st_ino = vap->va_fileid; mode = vap->va_mode; switch (vap->va_type) { case VREG: mode |= S_IFREG; break; case VDIR: mode |= S_IFDIR; break; case VBLK: mode |= S_IFBLK; break; case VCHR: mode |= S_IFCHR; break; case VLNK: mode |= S_IFLNK; break; case VSOCK: mode |= S_IFSOCK; break; case VFIFO: mode |= S_IFIFO; break; default: return (EBADF); }; sb->st_mode = mode; sb->st_nlink = vap->va_nlink; sb->st_uid = vap->va_uid; sb->st_gid = vap->va_gid; sb->st_rdev = vap->va_rdev; if (vap->va_size > OFF_MAX) return (EOVERFLOW); sb->st_size = vap->va_size; sb->st_atimespec = vap->va_atime; sb->st_mtimespec = vap->va_mtime; sb->st_ctimespec = vap->va_ctime; sb->st_birthtimespec = vap->va_birthtime; /* * According to www.opengroup.org, the meaning of st_blksize is * "a filesystem-specific preferred I/O block size for this * object. In some filesystem types, this may vary from file * to file" * Default to PAGE_SIZE after much discussion. * XXX: min(PAGE_SIZE, vp->v_bufobj.bo_bsize) may be more correct. */ sb->st_blksize = PAGE_SIZE; sb->st_flags = vap->va_flags; if (priv_check(td, PRIV_VFS_GENERATION)) sb->st_gen = 0; else sb->st_gen = vap->va_gen; sb->st_blocks = vap->va_bytes / S_BLKSIZE; return (0); } /* * File table vnode ioctl routine. */ static int vn_ioctl(fp, com, data, active_cred, td) struct file *fp; u_long com; void *data; struct ucred *active_cred; struct thread *td; { struct vnode *vp = fp->f_vnode; struct vattr vattr; int vfslocked; int error; vfslocked = VFS_LOCK_GIANT(vp->v_mount); error = ENOTTY; switch (vp->v_type) { case VREG: case VDIR: if (com == FIONREAD) { vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); error = VOP_GETATTR(vp, &vattr, active_cred); VOP_UNLOCK(vp, 0); if (!error) *(int *)data = vattr.va_size - fp->f_offset; } if (com == FIONBIO || com == FIOASYNC) /* XXX */ error = 0; else error = VOP_IOCTL(vp, com, data, fp->f_flag, active_cred, td); break; default: break; } VFS_UNLOCK_GIANT(vfslocked); return (error); } /* * File table vnode poll routine. */ static int vn_poll(fp, events, active_cred, td) struct file *fp; int events; struct ucred *active_cred; struct thread *td; { struct vnode *vp; int vfslocked; int error; vp = fp->f_vnode; vfslocked = VFS_LOCK_GIANT(vp->v_mount); #ifdef MAC vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); VOP_UNLOCK(vp, 0); if (!error) #endif error = VOP_POLL(vp, events, fp->f_cred, td); VFS_UNLOCK_GIANT(vfslocked); return (error); } /* * Acquire the requested lock and then check for validity. LK_RETRY * permits vn_lock to return doomed vnodes. */ int _vn_lock(struct vnode *vp, int flags, char *file, int line) { int error; VNASSERT((flags & LK_TYPE_MASK) != 0, vp, ("vn_lock called with no locktype.")); do { #ifdef DEBUG_VFS_LOCKS KASSERT(vp->v_holdcnt != 0, ("vn_lock %p: zero hold count", vp)); #endif error = VOP_LOCK1(vp, flags, file, line); flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */ KASSERT((flags & LK_RETRY) == 0 || error == 0, ("LK_RETRY set with incompatible flags (0x%x) or an error occured (%d)", flags, error)); /* * Callers specify LK_RETRY if they wish to get dead vnodes. * If RETRY is not set, we return ENOENT instead. */ if (error == 0 && vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) { VOP_UNLOCK(vp, 0); error = ENOENT; break; } } while (flags & LK_RETRY && error != 0); return (error); } /* * File table vnode close routine. */ static int vn_closefile(fp, td) struct file *fp; struct thread *td; { struct vnode *vp; struct flock lf; int vfslocked; int error; vp = fp->f_vnode; vfslocked = VFS_LOCK_GIANT(vp->v_mount); if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK) { lf.l_whence = SEEK_SET; lf.l_start = 0; lf.l_len = 0; lf.l_type = F_UNLCK; (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); } fp->f_ops = &badfileops; error = vn_close(vp, fp->f_flag, fp->f_cred, td); VFS_UNLOCK_GIANT(vfslocked); return (error); } /* * Preparing to start a filesystem write operation. If the operation is * permitted, then we bump the count of operations in progress and * proceed. If a suspend request is in progress, we wait until the * suspension is over, and then proceed. */ int vn_start_write(vp, mpp, flags) struct vnode *vp; struct mount **mpp; int flags; { struct mount *mp; int error; error = 0; /* * If a vnode is provided, get and return the mount point that * to which it will write. */ if (vp != NULL) { if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { *mpp = NULL; if (error != EOPNOTSUPP) return (error); return (0); } } if ((mp = *mpp) == NULL) return (0); /* * VOP_GETWRITEMOUNT() returns with the mp refcount held through * a vfs_ref(). * As long as a vnode is not provided we need to acquire a * refcount for the provided mountpoint too, in order to * emulate a vfs_ref(). */ MNT_ILOCK(mp); if (vp == NULL) MNT_REF(mp); /* * Check on status of suspension. */ if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || mp->mnt_susp_owner != curthread) { while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { if (flags & V_NOWAIT) { error = EWOULDBLOCK; goto unlock; } error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | (flags & PCATCH), "suspfs", 0); if (error) goto unlock; } } if (flags & V_XSLEEP) goto unlock; mp->mnt_writeopcount++; unlock: MNT_REL(mp); MNT_IUNLOCK(mp); return (error); } /* * Secondary suspension. Used by operations such as vop_inactive * routines that are needed by the higher level functions. These * are allowed to proceed until all the higher level functions have * completed (indicated by mnt_writeopcount dropping to zero). At that * time, these operations are halted until the suspension is over. */ int vn_start_secondary_write(vp, mpp, flags) struct vnode *vp; struct mount **mpp; int flags; { struct mount *mp; int error; retry: if (vp != NULL) { if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { *mpp = NULL; if (error != EOPNOTSUPP) return (error); return (0); } } /* * If we are not suspended or have not yet reached suspended * mode, then let the operation proceed. */ if ((mp = *mpp) == NULL) return (0); /* * VOP_GETWRITEMOUNT() returns with the mp refcount held through * a vfs_ref(). * As long as a vnode is not provided we need to acquire a * refcount for the provided mountpoint too, in order to * emulate a vfs_ref(). */ MNT_ILOCK(mp); if (vp == NULL) MNT_REF(mp); if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { mp->mnt_secondary_writes++; mp->mnt_secondary_accwrites++; MNT_REL(mp); MNT_IUNLOCK(mp); return (0); } if (flags & V_NOWAIT) { MNT_REL(mp); MNT_IUNLOCK(mp); return (EWOULDBLOCK); } /* * Wait for the suspension to finish. */ error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | (flags & PCATCH) | PDROP, "suspfs", 0); vfs_rel(mp); if (error == 0) goto retry; return (error); } /* * Filesystem write operation has completed. If we are suspending and this * operation is the last one, notify the suspender that the suspension is * now in effect. */ void vn_finished_write(mp) struct mount *mp; { if (mp == NULL) return; MNT_ILOCK(mp); mp->mnt_writeopcount--; if (mp->mnt_writeopcount < 0) panic("vn_finished_write: neg cnt"); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && mp->mnt_writeopcount <= 0) wakeup(&mp->mnt_writeopcount); MNT_IUNLOCK(mp); } /* * Filesystem secondary write operation has completed. If we are * suspending and this operation is the last one, notify the suspender * that the suspension is now in effect. */ void vn_finished_secondary_write(mp) struct mount *mp; { if (mp == NULL) return; MNT_ILOCK(mp); mp->mnt_secondary_writes--; if (mp->mnt_secondary_writes < 0) panic("vn_finished_secondary_write: neg cnt"); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && mp->mnt_secondary_writes <= 0) wakeup(&mp->mnt_secondary_writes); MNT_IUNLOCK(mp); } /* * Request a filesystem to suspend write operations. */ int vfs_write_suspend(mp) struct mount *mp; { int error; MNT_ILOCK(mp); if (mp->mnt_susp_owner == curthread) { MNT_IUNLOCK(mp); return (EALREADY); } while (mp->mnt_kern_flag & MNTK_SUSPEND) msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); mp->mnt_kern_flag |= MNTK_SUSPEND; mp->mnt_susp_owner = curthread; if (mp->mnt_writeopcount > 0) (void) msleep(&mp->mnt_writeopcount, MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); else MNT_IUNLOCK(mp); if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) vfs_write_resume(mp); return (error); } /* * Request a filesystem to resume write operations. */ void vfs_write_resume(mp) struct mount *mp; { MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | MNTK_SUSPENDED); mp->mnt_susp_owner = NULL; wakeup(&mp->mnt_writeopcount); wakeup(&mp->mnt_flag); curthread->td_pflags &= ~TDP_IGNSUSP; MNT_IUNLOCK(mp); VFS_SUSP_CLEAN(mp); } else MNT_IUNLOCK(mp); } /* * Implement kqueues for files by translating it to vnode operation. */ static int vn_kqfilter(struct file *fp, struct knote *kn) { int vfslocked; int error; vfslocked = VFS_LOCK_GIANT(fp->f_vnode->v_mount); error = VOP_KQFILTER(fp->f_vnode, kn); VFS_UNLOCK_GIANT(vfslocked); return error; } /* * Simplified in-kernel wrapper calls for extended attribute access. * Both calls pass in a NULL credential, authorizing as "kernel" access. * Set IO_NODELOCKED in ioflg if the vnode is already locked. */ int vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int *buflen, char *buf, struct thread *td) { struct uio auio; struct iovec iov; int error; iov.iov_len = *buflen; iov.iov_base = buf; auio.uio_iov = &iov; auio.uio_iovcnt = 1; auio.uio_rw = UIO_READ; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_offset = 0; auio.uio_resid = *buflen; if ((ioflg & IO_NODELOCKED) == 0) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute retrieval as kernel */ error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) VOP_UNLOCK(vp, 0); if (error == 0) { *buflen = *buflen - auio.uio_resid; } return (error); } /* * XXX failure mode if partially written? */ int vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int buflen, char *buf, struct thread *td) { struct uio auio; struct iovec iov; struct mount *mp; int error; iov.iov_len = buflen; iov.iov_base = buf; auio.uio_iov = &iov; auio.uio_iovcnt = 1; auio.uio_rw = UIO_WRITE; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_offset = 0; auio.uio_resid = buflen; if ((ioflg & IO_NODELOCKED) == 0) { if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute setting as kernel */ error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) { vn_finished_write(mp); VOP_UNLOCK(vp, 0); } return (error); } int vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, struct thread *td) { struct mount *mp; int error; if ((ioflg & IO_NODELOCKED) == 0) { if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute removal as kernel */ error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); if (error == EOPNOTSUPP) error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) { vn_finished_write(mp); VOP_UNLOCK(vp, 0); } return (error); } int vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) { struct mount *mp; int ltype, error; mp = vp->v_mount; ltype = VOP_ISLOCKED(vp); KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, ("vn_vget_ino: vp not locked")); error = vfs_busy(mp, MBF_NOWAIT); if (error != 0) { + vfs_ref(mp); VOP_UNLOCK(vp, 0); error = vfs_busy(mp, 0); vn_lock(vp, ltype | LK_RETRY); + vfs_rel(mp); if (error != 0) return (ENOENT); if (vp->v_iflag & VI_DOOMED) { vfs_unbusy(mp); return (ENOENT); } } VOP_UNLOCK(vp, 0); error = VFS_VGET(mp, ino, lkflags, rvp); vfs_unbusy(mp); vn_lock(vp, ltype | LK_RETRY); if (vp->v_iflag & VI_DOOMED) { if (error == 0) vput(*rvp); error = ENOENT; } return (error); } diff --git a/sys/nfsclient/nfs_vnops.c b/sys/nfsclient/nfs_vnops.c index 060ac54134fd..07a79043681d 100644 --- a/sys/nfsclient/nfs_vnops.c +++ b/sys/nfsclient/nfs_vnops.c @@ -1,3484 +1,3486 @@ /*- * Copyright (c) 1989, 1993 * 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. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)nfs_vnops.c 8.16 (Berkeley) 5/27/95 */ #include __FBSDID("$FreeBSD$"); /* * vnode op calls for Sun NFS version 2 and 3 */ #include "opt_inet.h" #include "opt_kdtrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KDTRACE_HOOKS #include dtrace_nfsclient_accesscache_flush_probe_func_t dtrace_nfsclient_accesscache_flush_done_probe; uint32_t nfsclient_accesscache_flush_done_id; dtrace_nfsclient_accesscache_get_probe_func_t dtrace_nfsclient_accesscache_get_hit_probe, dtrace_nfsclient_accesscache_get_miss_probe; uint32_t nfsclient_accesscache_get_hit_id; uint32_t nfsclient_accesscache_get_miss_id; dtrace_nfsclient_accesscache_load_probe_func_t dtrace_nfsclient_accesscache_load_done_probe; uint32_t nfsclient_accesscache_load_done_id; #endif /* !KDTRACE_HOOKS */ /* Defs */ #define TRUE 1 #define FALSE 0 /* * Ifdef for FreeBSD-current merged buffer cache. It is unfortunate that these * calls are not in getblk() and brelse() so that they would not be necessary * here. */ #ifndef B_VMIO #define vfs_busy_pages(bp, f) #endif static vop_read_t nfsfifo_read; static vop_write_t nfsfifo_write; static vop_close_t nfsfifo_close; static int nfs_flush(struct vnode *, int, int); static int nfs_setattrrpc(struct vnode *, struct vattr *, struct ucred *); static vop_lookup_t nfs_lookup; static vop_create_t nfs_create; static vop_mknod_t nfs_mknod; static vop_open_t nfs_open; static vop_close_t nfs_close; static vop_access_t nfs_access; static vop_getattr_t nfs_getattr; static vop_setattr_t nfs_setattr; static vop_read_t nfs_read; static vop_fsync_t nfs_fsync; static vop_remove_t nfs_remove; static vop_link_t nfs_link; static vop_rename_t nfs_rename; static vop_mkdir_t nfs_mkdir; static vop_rmdir_t nfs_rmdir; static vop_symlink_t nfs_symlink; static vop_readdir_t nfs_readdir; static vop_strategy_t nfs_strategy; static int nfs_lookitup(struct vnode *, const char *, int, struct ucred *, struct thread *, struct nfsnode **); static int nfs_sillyrename(struct vnode *, struct vnode *, struct componentname *); static vop_access_t nfsspec_access; static vop_readlink_t nfs_readlink; static vop_print_t nfs_print; static vop_advlock_t nfs_advlock; static vop_advlockasync_t nfs_advlockasync; /* * Global vfs data structures for nfs */ struct vop_vector nfs_vnodeops = { .vop_default = &default_vnodeops, .vop_access = nfs_access, .vop_advlock = nfs_advlock, .vop_advlockasync = nfs_advlockasync, .vop_close = nfs_close, .vop_create = nfs_create, .vop_fsync = nfs_fsync, .vop_getattr = nfs_getattr, .vop_getpages = nfs_getpages, .vop_putpages = nfs_putpages, .vop_inactive = nfs_inactive, .vop_link = nfs_link, .vop_lookup = nfs_lookup, .vop_mkdir = nfs_mkdir, .vop_mknod = nfs_mknod, .vop_open = nfs_open, .vop_print = nfs_print, .vop_read = nfs_read, .vop_readdir = nfs_readdir, .vop_readlink = nfs_readlink, .vop_reclaim = nfs_reclaim, .vop_remove = nfs_remove, .vop_rename = nfs_rename, .vop_rmdir = nfs_rmdir, .vop_setattr = nfs_setattr, .vop_strategy = nfs_strategy, .vop_symlink = nfs_symlink, .vop_write = nfs_write, }; struct vop_vector nfs_fifoops = { .vop_default = &fifo_specops, .vop_access = nfsspec_access, .vop_close = nfsfifo_close, .vop_fsync = nfs_fsync, .vop_getattr = nfs_getattr, .vop_inactive = nfs_inactive, .vop_print = nfs_print, .vop_read = nfsfifo_read, .vop_reclaim = nfs_reclaim, .vop_setattr = nfs_setattr, .vop_write = nfsfifo_write, }; static int nfs_mknodrpc(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct vattr *vap); static int nfs_removerpc(struct vnode *dvp, const char *name, int namelen, struct ucred *cred, struct thread *td); static int nfs_renamerpc(struct vnode *fdvp, const char *fnameptr, int fnamelen, struct vnode *tdvp, const char *tnameptr, int tnamelen, struct ucred *cred, struct thread *td); static int nfs_renameit(struct vnode *sdvp, struct componentname *scnp, struct sillyrename *sp); /* * Global variables */ struct mtx nfs_iod_mtx; struct proc *nfs_iodwant[NFS_MAXASYNCDAEMON]; struct nfsmount *nfs_iodmount[NFS_MAXASYNCDAEMON]; int nfs_numasync = 0; vop_advlock_t *nfs_advlock_p = nfs_dolock; vop_reclaim_t *nfs_reclaim_p = NULL; #define DIRHDSIZ (sizeof (struct dirent) - (MAXNAMLEN + 1)) SYSCTL_DECL(_vfs_nfs); static int nfsaccess_cache_timeout = NFS_MAXATTRTIMO; SYSCTL_INT(_vfs_nfs, OID_AUTO, access_cache_timeout, CTLFLAG_RW, &nfsaccess_cache_timeout, 0, "NFS ACCESS cache timeout"); static int nfs_prime_access_cache = 0; SYSCTL_INT(_vfs_nfs, OID_AUTO, prime_access_cache, CTLFLAG_RW, &nfs_prime_access_cache, 0, "Prime NFS ACCESS cache when fetching attributes"); static int nfsv3_commit_on_close = 0; SYSCTL_INT(_vfs_nfs, OID_AUTO, nfsv3_commit_on_close, CTLFLAG_RW, &nfsv3_commit_on_close, 0, "write+commit on close, else only write"); static int nfs_clean_pages_on_close = 1; SYSCTL_INT(_vfs_nfs, OID_AUTO, clean_pages_on_close, CTLFLAG_RW, &nfs_clean_pages_on_close, 0, "NFS clean dirty pages on close"); int nfs_directio_enable = 0; SYSCTL_INT(_vfs_nfs, OID_AUTO, nfs_directio_enable, CTLFLAG_RW, &nfs_directio_enable, 0, "Enable NFS directio"); /* * This sysctl allows other processes to mmap a file that has been opened * O_DIRECT by a process. In general, having processes mmap the file while * Direct IO is in progress can lead to Data Inconsistencies. But, we allow * this by default to prevent DoS attacks - to prevent a malicious user from * opening up files O_DIRECT preventing other users from mmap'ing these * files. "Protected" environments where stricter consistency guarantees are * required can disable this knob. The process that opened the file O_DIRECT * cannot mmap() the file, because mmap'ed IO on an O_DIRECT open() is not * meaningful. */ int nfs_directio_allow_mmap = 1; SYSCTL_INT(_vfs_nfs, OID_AUTO, nfs_directio_allow_mmap, CTLFLAG_RW, &nfs_directio_allow_mmap, 0, "Enable mmaped IO on file with O_DIRECT opens"); #if 0 SYSCTL_INT(_vfs_nfs, OID_AUTO, access_cache_hits, CTLFLAG_RD, &nfsstats.accesscache_hits, 0, "NFS ACCESS cache hit count"); SYSCTL_INT(_vfs_nfs, OID_AUTO, access_cache_misses, CTLFLAG_RD, &nfsstats.accesscache_misses, 0, "NFS ACCESS cache miss count"); #endif #define NFSV3ACCESS_ALL (NFSV3ACCESS_READ | NFSV3ACCESS_MODIFY \ | NFSV3ACCESS_EXTEND | NFSV3ACCESS_EXECUTE \ | NFSV3ACCESS_DELETE | NFSV3ACCESS_LOOKUP) /* * SMP Locking Note : * The list of locks after the description of the lock is the ordering * of other locks acquired with the lock held. * np->n_mtx : Protects the fields in the nfsnode. VM Object Lock VI_MTX (acquired indirectly) * nmp->nm_mtx : Protects the fields in the nfsmount. rep->r_mtx * nfs_iod_mtx : Global lock, protects shared nfsiod state. * nfs_reqq_mtx : Global lock, protects the nfs_reqq list. nmp->nm_mtx rep->r_mtx * rep->r_mtx : Protects the fields in an nfsreq. */ static int nfs3_access_otw(struct vnode *vp, int wmode, struct thread *td, struct ucred *cred, uint32_t *retmode) { const int v3 = 1; u_int32_t *tl; int error = 0, attrflag, i, lrupos; struct mbuf *mreq, *mrep, *md, *mb; caddr_t bpos, dpos; u_int32_t rmode; struct nfsnode *np = VTONFS(vp); nfsstats.rpccnt[NFSPROC_ACCESS]++; mreq = nfsm_reqhead(vp, NFSPROC_ACCESS, NFSX_FH(v3) + NFSX_UNSIGNED); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, v3); tl = nfsm_build(u_int32_t *, NFSX_UNSIGNED); *tl = txdr_unsigned(wmode); nfsm_request(vp, NFSPROC_ACCESS, td, cred); nfsm_postop_attr(vp, attrflag); if (!error) { lrupos = 0; tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); rmode = fxdr_unsigned(u_int32_t, *tl); mtx_lock(&np->n_mtx); for (i = 0; i < NFS_ACCESSCACHESIZE; i++) { if (np->n_accesscache[i].uid == cred->cr_uid) { np->n_accesscache[i].mode = rmode; np->n_accesscache[i].stamp = time_second; break; } if (i > 0 && np->n_accesscache[i].stamp < np->n_accesscache[lrupos].stamp) lrupos = i; } if (i == NFS_ACCESSCACHESIZE) { np->n_accesscache[lrupos].uid = cred->cr_uid; np->n_accesscache[lrupos].mode = rmode; np->n_accesscache[lrupos].stamp = time_second; } mtx_unlock(&np->n_mtx); if (retmode != NULL) *retmode = rmode; KDTRACE_NFS_ACCESSCACHE_LOAD_DONE(vp, cred->cr_uid, rmode, 0); } m_freem(mrep); nfsmout: #ifdef KDTRACE_HOOKS if (error) { KDTRACE_NFS_ACCESSCACHE_LOAD_DONE(vp, cred->cr_uid, 0, error); } #endif return (error); } /* * nfs access vnode op. * For nfs version 2, just return ok. File accesses may fail later. * For nfs version 3, use the access rpc to check accessibility. If file modes * are changed on the server, accesses might still fail later. */ static int nfs_access(struct vop_access_args *ap) { struct vnode *vp = ap->a_vp; int error = 0, i, gotahit; u_int32_t mode, rmode, wmode; int v3 = NFS_ISV3(vp); struct nfsnode *np = VTONFS(vp); /* * Disallow write attempts on filesystems mounted read-only; * unless the file is a socket, fifo, or a block or character * device resident on the filesystem. */ if ((ap->a_accmode & VWRITE) && (vp->v_mount->mnt_flag & MNT_RDONLY)) { switch (vp->v_type) { case VREG: case VDIR: case VLNK: return (EROFS); default: break; } } /* * For nfs v3, check to see if we have done this recently, and if * so return our cached result instead of making an ACCESS call. * If not, do an access rpc, otherwise you are stuck emulating * ufs_access() locally using the vattr. This may not be correct, * since the server may apply other access criteria such as * client uid-->server uid mapping that we do not know about. */ if (v3) { if (ap->a_accmode & VREAD) mode = NFSV3ACCESS_READ; else mode = 0; if (vp->v_type != VDIR) { if (ap->a_accmode & VWRITE) mode |= (NFSV3ACCESS_MODIFY | NFSV3ACCESS_EXTEND); if (ap->a_accmode & VEXEC) mode |= NFSV3ACCESS_EXECUTE; } else { if (ap->a_accmode & VWRITE) mode |= (NFSV3ACCESS_MODIFY | NFSV3ACCESS_EXTEND | NFSV3ACCESS_DELETE); if (ap->a_accmode & VEXEC) mode |= NFSV3ACCESS_LOOKUP; } /* XXX safety belt, only make blanket request if caching */ if (nfsaccess_cache_timeout > 0) { wmode = NFSV3ACCESS_READ | NFSV3ACCESS_MODIFY | NFSV3ACCESS_EXTEND | NFSV3ACCESS_EXECUTE | NFSV3ACCESS_DELETE | NFSV3ACCESS_LOOKUP; } else { wmode = mode; } /* * Does our cached result allow us to give a definite yes to * this request? */ gotahit = 0; mtx_lock(&np->n_mtx); for (i = 0; i < NFS_ACCESSCACHESIZE; i++) { if (ap->a_cred->cr_uid == np->n_accesscache[i].uid) { if (time_second < (np->n_accesscache[i].stamp + nfsaccess_cache_timeout) && (np->n_accesscache[i].mode & mode) == mode) { nfsstats.accesscache_hits++; gotahit = 1; } break; } } mtx_unlock(&np->n_mtx); #ifdef KDTRACE_HOOKS if (gotahit) KDTRACE_NFS_ACCESSCACHE_GET_HIT(vp, ap->a_cred->cr_uid, mode); else KDTRACE_NFS_ACCESSCACHE_GET_MISS(vp, ap->a_cred->cr_uid, mode); #endif if (gotahit == 0) { /* * Either a no, or a don't know. Go to the wire. */ nfsstats.accesscache_misses++; error = nfs3_access_otw(vp, wmode, ap->a_td, ap->a_cred, &rmode); if (!error) { if ((rmode & mode) != mode) error = EACCES; } } return (error); } else { if ((error = nfsspec_access(ap)) != 0) { return (error); } /* * Attempt to prevent a mapped root from accessing a file * which it shouldn't. We try to read a byte from the file * if the user is root and the file is not zero length. * After calling nfsspec_access, we should have the correct * file size cached. */ mtx_lock(&np->n_mtx); if (ap->a_cred->cr_uid == 0 && (ap->a_accmode & VREAD) && VTONFS(vp)->n_size > 0) { struct iovec aiov; struct uio auio; char buf[1]; mtx_unlock(&np->n_mtx); aiov.iov_base = buf; aiov.iov_len = 1; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = 0; auio.uio_resid = 1; auio.uio_segflg = UIO_SYSSPACE; auio.uio_rw = UIO_READ; auio.uio_td = ap->a_td; if (vp->v_type == VREG) error = nfs_readrpc(vp, &auio, ap->a_cred); else if (vp->v_type == VDIR) { char* bp; bp = malloc(NFS_DIRBLKSIZ, M_TEMP, M_WAITOK); aiov.iov_base = bp; aiov.iov_len = auio.uio_resid = NFS_DIRBLKSIZ; error = nfs_readdirrpc(vp, &auio, ap->a_cred); free(bp, M_TEMP); } else if (vp->v_type == VLNK) error = nfs_readlinkrpc(vp, &auio, ap->a_cred); else error = EACCES; } else mtx_unlock(&np->n_mtx); return (error); } } int nfs_otw_getattr_avoid = 0; /* * nfs open vnode op * Check to see if the type is ok * and that deletion is not in progress. * For paged in text files, you will need to flush the page cache * if consistency is lost. */ /* ARGSUSED */ static int nfs_open(struct vop_open_args *ap) { struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); struct vattr vattr; int error; int fmode = ap->a_mode; if (vp->v_type != VREG && vp->v_type != VDIR && vp->v_type != VLNK) return (EOPNOTSUPP); /* * Get a valid lease. If cached data is stale, flush it. */ mtx_lock(&np->n_mtx); if (np->n_flag & NMODIFIED) { mtx_unlock(&np->n_mtx); error = nfs_vinvalbuf(vp, V_SAVE, ap->a_td, 1); if (error == EINTR || error == EIO) return (error); np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); if (vp->v_type == VDIR) np->n_direofoffset = 0; error = VOP_GETATTR(vp, &vattr, ap->a_cred); if (error) return (error); mtx_lock(&np->n_mtx); np->n_mtime = vattr.va_mtime; mtx_unlock(&np->n_mtx); } else { struct thread *td = curthread; if (np->n_ac_ts_syscalls != td->td_syscalls || np->n_ac_ts_tid != td->td_tid || td->td_proc == NULL || np->n_ac_ts_pid != td->td_proc->p_pid) { np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); } mtx_unlock(&np->n_mtx); error = VOP_GETATTR(vp, &vattr, ap->a_cred); if (error) return (error); mtx_lock(&np->n_mtx); if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime)) { if (vp->v_type == VDIR) np->n_direofoffset = 0; mtx_unlock(&np->n_mtx); error = nfs_vinvalbuf(vp, V_SAVE, ap->a_td, 1); if (error == EINTR || error == EIO) { return (error); } mtx_lock(&np->n_mtx); np->n_mtime = vattr.va_mtime; } mtx_unlock(&np->n_mtx); } /* * If the object has >= 1 O_DIRECT active opens, we disable caching. */ if (nfs_directio_enable && (fmode & O_DIRECT) && (vp->v_type == VREG)) { if (np->n_directio_opens == 0) { error = nfs_vinvalbuf(vp, V_SAVE, ap->a_td, 1); if (error) return (error); mtx_lock(&np->n_mtx); np->n_flag |= NNONCACHE; mtx_unlock(&np->n_mtx); } np->n_directio_opens++; } vnode_create_vobject(vp, vattr.va_size, ap->a_td); return (0); } /* * nfs close vnode op * What an NFS client should do upon close after writing is a debatable issue. * Most NFS clients push delayed writes to the server upon close, basically for * two reasons: * 1 - So that any write errors may be reported back to the client process * doing the close system call. By far the two most likely errors are * NFSERR_NOSPC and NFSERR_DQUOT to indicate space allocation failure. * 2 - To put a worst case upper bound on cache inconsistency between * multiple clients for the file. * There is also a consistency problem for Version 2 of the protocol w.r.t. * not being able to tell if other clients are writing a file concurrently, * since there is no way of knowing if the changed modify time in the reply * is only due to the write for this client. * (NFS Version 3 provides weak cache consistency data in the reply that * should be sufficient to detect and handle this case.) * * The current code does the following: * for NFS Version 2 - play it safe and flush/invalidate all dirty buffers * for NFS Version 3 - flush dirty buffers to the server but don't invalidate * or commit them (this satisfies 1 and 2 except for the * case where the server crashes after this close but * before the commit RPC, which is felt to be "good * enough". Changing the last argument to nfs_flush() to * a 1 would force a commit operation, if it is felt a * commit is necessary now. */ /* ARGSUSED */ static int nfs_close(struct vop_close_args *ap) { struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); int error = 0; int fmode = ap->a_fflag; if (vp->v_type == VREG) { /* * Examine and clean dirty pages, regardless of NMODIFIED. * This closes a major hole in close-to-open consistency. * We want to push out all dirty pages (and buffers) on * close, regardless of whether they were dirtied by * mmap'ed writes or via write(). */ if (nfs_clean_pages_on_close && vp->v_object) { VM_OBJECT_LOCK(vp->v_object); vm_object_page_clean(vp->v_object, 0, 0, 0); VM_OBJECT_UNLOCK(vp->v_object); } mtx_lock(&np->n_mtx); if (np->n_flag & NMODIFIED) { mtx_unlock(&np->n_mtx); if (NFS_ISV3(vp)) { /* * Under NFSv3 we have dirty buffers to dispose of. We * must flush them to the NFS server. We have the option * of waiting all the way through the commit rpc or just * waiting for the initial write. The default is to only * wait through the initial write so the data is in the * server's cache, which is roughly similar to the state * a standard disk subsystem leaves the file in on close(). * * We cannot clear the NMODIFIED bit in np->n_flag due to * potential races with other processes, and certainly * cannot clear it if we don't commit. */ int cm = nfsv3_commit_on_close ? 1 : 0; error = nfs_flush(vp, MNT_WAIT, cm); /* np->n_flag &= ~NMODIFIED; */ } else error = nfs_vinvalbuf(vp, V_SAVE, ap->a_td, 1); mtx_lock(&np->n_mtx); } if (np->n_flag & NWRITEERR) { np->n_flag &= ~NWRITEERR; error = np->n_error; } mtx_unlock(&np->n_mtx); } if (nfs_directio_enable) KASSERT((np->n_directio_asyncwr == 0), ("nfs_close: dirty unflushed (%d) directio buffers\n", np->n_directio_asyncwr)); if (nfs_directio_enable && (fmode & O_DIRECT) && (vp->v_type == VREG)) { mtx_lock(&np->n_mtx); KASSERT((np->n_directio_opens > 0), ("nfs_close: unexpectedly value (0) of n_directio_opens\n")); np->n_directio_opens--; if (np->n_directio_opens == 0) np->n_flag &= ~NNONCACHE; mtx_unlock(&np->n_mtx); } return (error); } /* * nfs getattr call from vfs. */ static int nfs_getattr(struct vop_getattr_args *ap) { struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); struct thread *td = curthread; struct vattr *vap = ap->a_vap; struct vattr vattr; caddr_t bpos, dpos; int error = 0; struct mbuf *mreq, *mrep, *md, *mb; int v3 = NFS_ISV3(vp); /* * Update local times for special files. */ mtx_lock(&np->n_mtx); if (np->n_flag & (NACC | NUPD)) np->n_flag |= NCHG; mtx_unlock(&np->n_mtx); /* * First look in the cache. */ if (nfs_getattrcache(vp, &vattr) == 0) goto nfsmout; if (v3 && nfs_prime_access_cache && nfsaccess_cache_timeout > 0) { nfsstats.accesscache_misses++; nfs3_access_otw(vp, NFSV3ACCESS_ALL, td, ap->a_cred, NULL); if (nfs_getattrcache(vp, &vattr) == 0) goto nfsmout; } nfsstats.rpccnt[NFSPROC_GETATTR]++; mreq = nfsm_reqhead(vp, NFSPROC_GETATTR, NFSX_FH(v3)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, v3); nfsm_request(vp, NFSPROC_GETATTR, td, ap->a_cred); if (!error) { nfsm_loadattr(vp, &vattr); } m_freem(mrep); nfsmout: vap->va_type = vattr.va_type; vap->va_mode = vattr.va_mode; vap->va_nlink = vattr.va_nlink; vap->va_uid = vattr.va_uid; vap->va_gid = vattr.va_gid; vap->va_fsid = vattr.va_fsid; vap->va_fileid = vattr.va_fileid; vap->va_size = vattr.va_size; vap->va_blocksize = vattr.va_blocksize; vap->va_atime = vattr.va_atime; vap->va_mtime = vattr.va_mtime; vap->va_ctime = vattr.va_ctime; vap->va_gen = vattr.va_gen; vap->va_flags = vattr.va_flags; vap->va_rdev = vattr.va_rdev; vap->va_bytes = vattr.va_bytes; vap->va_filerev = vattr.va_filerev; return (error); } /* * nfs setattr call. */ static int nfs_setattr(struct vop_setattr_args *ap) { struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); struct vattr *vap = ap->a_vap; struct thread *td = curthread; int error = 0; u_quad_t tsize; #ifndef nolint tsize = (u_quad_t)0; #endif /* * Setting of flags is not supported. */ if (vap->va_flags != VNOVAL) return (EOPNOTSUPP); /* * Disallow write attempts if the filesystem is mounted read-only. */ if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL || vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) && (vp->v_mount->mnt_flag & MNT_RDONLY)) { error = EROFS; goto out; } if (vap->va_size != VNOVAL) { switch (vp->v_type) { case VDIR: return (EISDIR); case VCHR: case VBLK: case VSOCK: case VFIFO: if (vap->va_mtime.tv_sec == VNOVAL && vap->va_atime.tv_sec == VNOVAL && vap->va_mode == (mode_t)VNOVAL && vap->va_uid == (uid_t)VNOVAL && vap->va_gid == (gid_t)VNOVAL) return (0); vap->va_size = VNOVAL; break; default: /* * Disallow write attempts if the filesystem is * mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return (EROFS); /* * We run vnode_pager_setsize() early (why?), * we must set np->n_size now to avoid vinvalbuf * V_SAVE races that might setsize a lower * value. */ mtx_lock(&np->n_mtx); tsize = np->n_size; mtx_unlock(&np->n_mtx); error = nfs_meta_setsize(vp, ap->a_cred, td, vap->va_size); mtx_lock(&np->n_mtx); if (np->n_flag & NMODIFIED) { tsize = np->n_size; mtx_unlock(&np->n_mtx); if (vap->va_size == 0) error = nfs_vinvalbuf(vp, 0, td, 1); else error = nfs_vinvalbuf(vp, V_SAVE, td, 1); if (error) { vnode_pager_setsize(vp, tsize); goto out; } } else mtx_unlock(&np->n_mtx); /* * np->n_size has already been set to vap->va_size * in nfs_meta_setsize(). We must set it again since * nfs_loadattrcache() could be called through * nfs_meta_setsize() and could modify np->n_size. */ mtx_lock(&np->n_mtx); np->n_vattr.va_size = np->n_size = vap->va_size; mtx_unlock(&np->n_mtx); }; } else { mtx_lock(&np->n_mtx); if ((vap->va_mtime.tv_sec != VNOVAL || vap->va_atime.tv_sec != VNOVAL) && (np->n_flag & NMODIFIED) && vp->v_type == VREG) { mtx_unlock(&np->n_mtx); if ((error = nfs_vinvalbuf(vp, V_SAVE, td, 1)) != 0 && (error == EINTR || error == EIO)) return error; } else mtx_unlock(&np->n_mtx); } error = nfs_setattrrpc(vp, vap, ap->a_cred); if (error && vap->va_size != VNOVAL) { mtx_lock(&np->n_mtx); np->n_size = np->n_vattr.va_size = tsize; vnode_pager_setsize(vp, tsize); mtx_unlock(&np->n_mtx); } out: return (error); } /* * Do an nfs setattr rpc. */ static int nfs_setattrrpc(struct vnode *vp, struct vattr *vap, struct ucred *cred) { struct nfsv2_sattr *sp; struct nfsnode *np = VTONFS(vp); caddr_t bpos, dpos; u_int32_t *tl; int error = 0, i, wccflag = NFSV3_WCCRATTR; struct mbuf *mreq, *mrep, *md, *mb; int v3 = NFS_ISV3(vp); nfsstats.rpccnt[NFSPROC_SETATTR]++; mreq = nfsm_reqhead(vp, NFSPROC_SETATTR, NFSX_FH(v3) + NFSX_SATTR(v3)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, v3); if (v3) { nfsm_v3attrbuild(vap, TRUE); tl = nfsm_build(u_int32_t *, NFSX_UNSIGNED); *tl = nfs_false; } else { sp = nfsm_build(struct nfsv2_sattr *, NFSX_V2SATTR); if (vap->va_mode == (mode_t)VNOVAL) sp->sa_mode = nfs_xdrneg1; else sp->sa_mode = vtonfsv2_mode(vp->v_type, vap->va_mode); if (vap->va_uid == (uid_t)VNOVAL) sp->sa_uid = nfs_xdrneg1; else sp->sa_uid = txdr_unsigned(vap->va_uid); if (vap->va_gid == (gid_t)VNOVAL) sp->sa_gid = nfs_xdrneg1; else sp->sa_gid = txdr_unsigned(vap->va_gid); sp->sa_size = txdr_unsigned(vap->va_size); txdr_nfsv2time(&vap->va_atime, &sp->sa_atime); txdr_nfsv2time(&vap->va_mtime, &sp->sa_mtime); } nfsm_request(vp, NFSPROC_SETATTR, curthread, cred); if (v3) { mtx_lock(&np->n_mtx); for (i = 0; i < NFS_ACCESSCACHESIZE; i++) np->n_accesscache[i].stamp = 0; mtx_unlock(&np->n_mtx); KDTRACE_NFS_ACCESSCACHE_FLUSH_DONE(vp); nfsm_wcc_data(vp, wccflag); } else nfsm_loadattr(vp, NULL); m_freem(mrep); nfsmout: return (error); } /* * nfs lookup call, one step at a time... * First look in cache * If not found, unlock the directory nfsnode and do the rpc */ static int nfs_lookup(struct vop_lookup_args *ap) { struct componentname *cnp = ap->a_cnp; struct vnode *dvp = ap->a_dvp; struct vnode **vpp = ap->a_vpp; struct mount *mp = dvp->v_mount; struct vattr vattr; int flags = cnp->cn_flags; struct vnode *newvp; struct nfsmount *nmp; caddr_t bpos, dpos; struct mbuf *mreq, *mrep, *md, *mb; long len; nfsfh_t *fhp; struct nfsnode *np; int error = 0, attrflag, fhsize, ltype; int v3 = NFS_ISV3(dvp); struct thread *td = cnp->cn_thread; *vpp = NULLVP; if ((flags & ISLASTCN) && (mp->mnt_flag & MNT_RDONLY) && (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) return (EROFS); if (dvp->v_type != VDIR) return (ENOTDIR); nmp = VFSTONFS(mp); np = VTONFS(dvp); if ((error = VOP_ACCESS(dvp, VEXEC, cnp->cn_cred, td)) != 0) { *vpp = NULLVP; return (error); } error = cache_lookup(dvp, vpp, cnp); if (error > 0 && error != ENOENT) return (error); if (error == -1) { /* * We only accept a positive hit in the cache if the * change time of the file matches our cached copy. * Otherwise, we discard the cache entry and fallback * to doing a lookup RPC. */ newvp = *vpp; if (!VOP_GETATTR(newvp, &vattr, cnp->cn_cred) && vattr.va_ctime.tv_sec == VTONFS(newvp)->n_ctime) { nfsstats.lookupcache_hits++; if (cnp->cn_nameiop != LOOKUP && (flags & ISLASTCN)) cnp->cn_flags |= SAVENAME; return (0); } cache_purge(newvp); if (dvp != newvp) vput(newvp); else vrele(newvp); *vpp = NULLVP; } else if (error == ENOENT) { if (dvp->v_iflag & VI_DOOMED) return (ENOENT); /* * We only accept a negative hit in the cache if the * modification time of the parent directory matches * our cached copy. Otherwise, we discard all of the * negative cache entries for this directory. */ if (VOP_GETATTR(dvp, &vattr, cnp->cn_cred) == 0 && vattr.va_mtime.tv_sec == np->n_dmtime) { nfsstats.lookupcache_hits++; return (ENOENT); } cache_purge_negative(dvp); mtx_lock(&np->n_mtx); np->n_dmtime = 0; mtx_unlock(&np->n_mtx); } error = 0; newvp = NULLVP; nfsstats.lookupcache_misses++; nfsstats.rpccnt[NFSPROC_LOOKUP]++; len = cnp->cn_namelen; mreq = nfsm_reqhead(dvp, NFSPROC_LOOKUP, NFSX_FH(v3) + NFSX_UNSIGNED + nfsm_rndup(len)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(dvp, v3); nfsm_strtom(cnp->cn_nameptr, len, NFS_MAXNAMLEN); nfsm_request(dvp, NFSPROC_LOOKUP, cnp->cn_thread, cnp->cn_cred); if (error) { if (v3) { nfsm_postop_attr(dvp, attrflag); m_freem(mrep); } goto nfsmout; } nfsm_getfh(fhp, fhsize, v3); /* * Handle RENAME case... */ if (cnp->cn_nameiop == RENAME && (flags & ISLASTCN)) { if (NFS_CMPFH(np, fhp, fhsize)) { m_freem(mrep); return (EISDIR); } error = nfs_nget(mp, fhp, fhsize, &np, LK_EXCLUSIVE); if (error) { m_freem(mrep); return (error); } newvp = NFSTOV(np); if (v3) { nfsm_postop_attr(newvp, attrflag); nfsm_postop_attr(dvp, attrflag); } else nfsm_loadattr(newvp, NULL); *vpp = newvp; m_freem(mrep); cnp->cn_flags |= SAVENAME; return (0); } if (flags & ISDOTDOT) { ltype = VOP_ISLOCKED(dvp); error = vfs_busy(mp, MBF_NOWAIT); if (error != 0) { + vfs_ref(mp); VOP_UNLOCK(dvp, 0); error = vfs_busy(mp, 0); vn_lock(dvp, ltype | LK_RETRY); + vfs_rel(mp); if (error == 0 && (dvp->v_iflag & VI_DOOMED)) { vfs_unbusy(mp); error = ENOENT; } if (error != 0) { m_freem(mrep); return (error); } } VOP_UNLOCK(dvp, 0); error = nfs_nget(mp, fhp, fhsize, &np, cnp->cn_lkflags); if (error == 0) newvp = NFSTOV(np); vfs_unbusy(mp); vn_lock(dvp, ltype | LK_RETRY); if (dvp->v_iflag & VI_DOOMED) { if (error == 0) { if (newvp == dvp) vrele(newvp); else vput(newvp); } error = ENOENT; } if (error) { m_freem(mrep); return (error); } } else if (NFS_CMPFH(np, fhp, fhsize)) { VREF(dvp); newvp = dvp; } else { error = nfs_nget(mp, fhp, fhsize, &np, cnp->cn_lkflags); if (error) { m_freem(mrep); return (error); } newvp = NFSTOV(np); } if (v3) { nfsm_postop_attr(newvp, attrflag); nfsm_postop_attr(dvp, attrflag); } else nfsm_loadattr(newvp, NULL); if (cnp->cn_nameiop != LOOKUP && (flags & ISLASTCN)) cnp->cn_flags |= SAVENAME; if ((cnp->cn_flags & MAKEENTRY) && (cnp->cn_nameiop != DELETE || !(flags & ISLASTCN))) { np->n_ctime = np->n_vattr.va_ctime.tv_sec; cache_enter(dvp, newvp, cnp); } *vpp = newvp; m_freem(mrep); nfsmout: if (error) { if (newvp != NULLVP) { vput(newvp); *vpp = NULLVP; } if (error != ENOENT) goto done; /* The requested file was not found. */ if ((cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME) && (flags & ISLASTCN)) { /* * XXX: UFS does a full VOP_ACCESS(dvp, * VWRITE) here instead of just checking * MNT_RDONLY. */ if (mp->mnt_flag & MNT_RDONLY) return (EROFS); cnp->cn_flags |= SAVENAME; return (EJUSTRETURN); } if ((cnp->cn_flags & MAKEENTRY) && cnp->cn_nameiop != CREATE) { /* * Maintain n_dmtime as the modification time * of the parent directory when the oldest -ve * name cache entry for this directory was * added. */ mtx_lock(&np->n_mtx); if (np->n_dmtime == 0) np->n_dmtime = np->n_vattr.va_mtime.tv_sec; mtx_unlock(&np->n_mtx); cache_enter(dvp, NULL, cnp); } return (ENOENT); } done: return (error); } /* * nfs read call. * Just call nfs_bioread() to do the work. */ static int nfs_read(struct vop_read_args *ap) { struct vnode *vp = ap->a_vp; switch (vp->v_type) { case VREG: return (nfs_bioread(vp, ap->a_uio, ap->a_ioflag, ap->a_cred)); case VDIR: return (EISDIR); default: return (EOPNOTSUPP); } } /* * nfs readlink call */ static int nfs_readlink(struct vop_readlink_args *ap) { struct vnode *vp = ap->a_vp; if (vp->v_type != VLNK) return (EINVAL); return (nfs_bioread(vp, ap->a_uio, 0, ap->a_cred)); } /* * Do a readlink rpc. * Called by nfs_doio() from below the buffer cache. */ int nfs_readlinkrpc(struct vnode *vp, struct uio *uiop, struct ucred *cred) { caddr_t bpos, dpos; int error = 0, len, attrflag; struct mbuf *mreq, *mrep, *md, *mb; int v3 = NFS_ISV3(vp); nfsstats.rpccnt[NFSPROC_READLINK]++; mreq = nfsm_reqhead(vp, NFSPROC_READLINK, NFSX_FH(v3)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, v3); nfsm_request(vp, NFSPROC_READLINK, uiop->uio_td, cred); if (v3) nfsm_postop_attr(vp, attrflag); if (!error) { nfsm_strsiz(len, NFS_MAXPATHLEN); if (len == NFS_MAXPATHLEN) { struct nfsnode *np = VTONFS(vp); mtx_lock(&np->n_mtx); if (np->n_size && np->n_size < NFS_MAXPATHLEN) len = np->n_size; mtx_unlock(&np->n_mtx); } nfsm_mtouio(uiop, len); } m_freem(mrep); nfsmout: return (error); } /* * nfs read rpc call * Ditto above */ int nfs_readrpc(struct vnode *vp, struct uio *uiop, struct ucred *cred) { u_int32_t *tl; caddr_t bpos, dpos; struct mbuf *mreq, *mrep, *md, *mb; struct nfsmount *nmp; int error = 0, len, retlen, tsiz, eof, attrflag; int v3 = NFS_ISV3(vp); int rsize; #ifndef nolint eof = 0; #endif nmp = VFSTONFS(vp->v_mount); tsiz = uiop->uio_resid; mtx_lock(&nmp->nm_mtx); if (uiop->uio_offset + tsiz > nmp->nm_maxfilesize) { mtx_unlock(&nmp->nm_mtx); return (EFBIG); } rsize = nmp->nm_rsize; mtx_unlock(&nmp->nm_mtx); while (tsiz > 0) { nfsstats.rpccnt[NFSPROC_READ]++; len = (tsiz > rsize) ? rsize : tsiz; mreq = nfsm_reqhead(vp, NFSPROC_READ, NFSX_FH(v3) + NFSX_UNSIGNED * 3); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, v3); tl = nfsm_build(u_int32_t *, NFSX_UNSIGNED * 3); if (v3) { txdr_hyper(uiop->uio_offset, tl); *(tl + 2) = txdr_unsigned(len); } else { *tl++ = txdr_unsigned(uiop->uio_offset); *tl++ = txdr_unsigned(len); *tl = 0; } nfsm_request(vp, NFSPROC_READ, uiop->uio_td, cred); if (v3) { nfsm_postop_attr(vp, attrflag); if (error) { m_freem(mrep); goto nfsmout; } tl = nfsm_dissect(u_int32_t *, 2 * NFSX_UNSIGNED); eof = fxdr_unsigned(int, *(tl + 1)); } else { nfsm_loadattr(vp, NULL); } nfsm_strsiz(retlen, rsize); nfsm_mtouio(uiop, retlen); m_freem(mrep); tsiz -= retlen; if (v3) { if (eof || retlen == 0) { tsiz = 0; } } else if (retlen < len) { tsiz = 0; } } nfsmout: return (error); } /* * nfs write call */ int nfs_writerpc(struct vnode *vp, struct uio *uiop, struct ucred *cred, int *iomode, int *must_commit) { u_int32_t *tl; int32_t backup; caddr_t bpos, dpos; struct mbuf *mreq, *mrep, *md, *mb; struct nfsmount *nmp = VFSTONFS(vp->v_mount); int error = 0, len, tsiz, wccflag = NFSV3_WCCRATTR, rlen, commit; int v3 = NFS_ISV3(vp), committed = NFSV3WRITE_FILESYNC; int wsize; #ifndef DIAGNOSTIC if (uiop->uio_iovcnt != 1) panic("nfs: writerpc iovcnt > 1"); #endif *must_commit = 0; tsiz = uiop->uio_resid; mtx_lock(&nmp->nm_mtx); if (uiop->uio_offset + tsiz > nmp->nm_maxfilesize) { mtx_unlock(&nmp->nm_mtx); return (EFBIG); } wsize = nmp->nm_wsize; mtx_unlock(&nmp->nm_mtx); while (tsiz > 0) { nfsstats.rpccnt[NFSPROC_WRITE]++; len = (tsiz > wsize) ? wsize : tsiz; mreq = nfsm_reqhead(vp, NFSPROC_WRITE, NFSX_FH(v3) + 5 * NFSX_UNSIGNED + nfsm_rndup(len)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, v3); if (v3) { tl = nfsm_build(u_int32_t *, 5 * NFSX_UNSIGNED); txdr_hyper(uiop->uio_offset, tl); tl += 2; *tl++ = txdr_unsigned(len); *tl++ = txdr_unsigned(*iomode); *tl = txdr_unsigned(len); } else { u_int32_t x; tl = nfsm_build(u_int32_t *, 4 * NFSX_UNSIGNED); /* Set both "begin" and "current" to non-garbage. */ x = txdr_unsigned((u_int32_t)uiop->uio_offset); *tl++ = x; /* "begin offset" */ *tl++ = x; /* "current offset" */ x = txdr_unsigned(len); *tl++ = x; /* total to this offset */ *tl = x; /* size of this write */ } nfsm_uiotom(uiop, len); nfsm_request(vp, NFSPROC_WRITE, uiop->uio_td, cred); if (v3) { wccflag = NFSV3_WCCCHK; nfsm_wcc_data(vp, wccflag); if (!error) { tl = nfsm_dissect(u_int32_t *, 2 * NFSX_UNSIGNED + NFSX_V3WRITEVERF); rlen = fxdr_unsigned(int, *tl++); if (rlen == 0) { error = NFSERR_IO; m_freem(mrep); break; } else if (rlen < len) { backup = len - rlen; uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base - backup; uiop->uio_iov->iov_len += backup; uiop->uio_offset -= backup; uiop->uio_resid += backup; len = rlen; } commit = fxdr_unsigned(int, *tl++); /* * Return the lowest committment level * obtained by any of the RPCs. */ if (committed == NFSV3WRITE_FILESYNC) committed = commit; else if (committed == NFSV3WRITE_DATASYNC && commit == NFSV3WRITE_UNSTABLE) committed = commit; mtx_lock(&nmp->nm_mtx); if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0){ bcopy((caddr_t)tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF); nmp->nm_state |= NFSSTA_HASWRITEVERF; } else if (bcmp((caddr_t)tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF)) { *must_commit = 1; bcopy((caddr_t)tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF); } mtx_unlock(&nmp->nm_mtx); } } else { nfsm_loadattr(vp, NULL); } if (wccflag) { mtx_lock(&(VTONFS(vp))->n_mtx); VTONFS(vp)->n_mtime = VTONFS(vp)->n_vattr.va_mtime; mtx_unlock(&(VTONFS(vp))->n_mtx); } m_freem(mrep); if (error) break; tsiz -= len; } nfsmout: if (vp->v_mount->mnt_kern_flag & MNTK_ASYNC) committed = NFSV3WRITE_FILESYNC; *iomode = committed; if (error) uiop->uio_resid = tsiz; return (error); } /* * nfs mknod rpc * For NFS v2 this is a kludge. Use a create rpc but with the IFMT bits of the * mode set to specify the file type and the size field for rdev. */ static int nfs_mknodrpc(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct vattr *vap) { struct nfsv2_sattr *sp; u_int32_t *tl; struct vnode *newvp = NULL; struct nfsnode *np = NULL; struct vattr vattr; caddr_t bpos, dpos; int error = 0, wccflag = NFSV3_WCCRATTR, gotvp = 0; struct mbuf *mreq, *mrep, *md, *mb; u_int32_t rdev; int v3 = NFS_ISV3(dvp); if (vap->va_type == VCHR || vap->va_type == VBLK) rdev = txdr_unsigned(vap->va_rdev); else if (vap->va_type == VFIFO || vap->va_type == VSOCK) rdev = nfs_xdrneg1; else { return (EOPNOTSUPP); } if ((error = VOP_GETATTR(dvp, &vattr, cnp->cn_cred)) != 0) return (error); nfsstats.rpccnt[NFSPROC_MKNOD]++; mreq = nfsm_reqhead(dvp, NFSPROC_MKNOD, NFSX_FH(v3) + 4 * NFSX_UNSIGNED + + nfsm_rndup(cnp->cn_namelen) + NFSX_SATTR(v3)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(dvp, v3); nfsm_strtom(cnp->cn_nameptr, cnp->cn_namelen, NFS_MAXNAMLEN); if (v3) { tl = nfsm_build(u_int32_t *, NFSX_UNSIGNED); *tl++ = vtonfsv3_type(vap->va_type); nfsm_v3attrbuild(vap, FALSE); if (vap->va_type == VCHR || vap->va_type == VBLK) { tl = nfsm_build(u_int32_t *, 2 * NFSX_UNSIGNED); *tl++ = txdr_unsigned(major(vap->va_rdev)); *tl = txdr_unsigned(minor(vap->va_rdev)); } } else { sp = nfsm_build(struct nfsv2_sattr *, NFSX_V2SATTR); sp->sa_mode = vtonfsv2_mode(vap->va_type, vap->va_mode); sp->sa_uid = nfs_xdrneg1; sp->sa_gid = nfs_xdrneg1; sp->sa_size = rdev; txdr_nfsv2time(&vap->va_atime, &sp->sa_atime); txdr_nfsv2time(&vap->va_mtime, &sp->sa_mtime); } nfsm_request(dvp, NFSPROC_MKNOD, cnp->cn_thread, cnp->cn_cred); if (!error) { nfsm_mtofh(dvp, newvp, v3, gotvp); if (!gotvp) { if (newvp) { vput(newvp); newvp = NULL; } error = nfs_lookitup(dvp, cnp->cn_nameptr, cnp->cn_namelen, cnp->cn_cred, cnp->cn_thread, &np); if (!error) newvp = NFSTOV(np); } } if (v3) nfsm_wcc_data(dvp, wccflag); m_freem(mrep); nfsmout: if (error) { if (newvp) vput(newvp); } else { if (cnp->cn_flags & MAKEENTRY) cache_enter(dvp, newvp, cnp); *vpp = newvp; } mtx_lock(&(VTONFS(dvp))->n_mtx); VTONFS(dvp)->n_flag |= NMODIFIED; if (!wccflag) { VTONFS(dvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(dvp); } mtx_unlock(&(VTONFS(dvp))->n_mtx); return (error); } /* * nfs mknod vop * just call nfs_mknodrpc() to do the work. */ /* ARGSUSED */ static int nfs_mknod(struct vop_mknod_args *ap) { return (nfs_mknodrpc(ap->a_dvp, ap->a_vpp, ap->a_cnp, ap->a_vap)); } static u_long create_verf; /* * nfs file create call */ static int nfs_create(struct vop_create_args *ap) { struct vnode *dvp = ap->a_dvp; struct vattr *vap = ap->a_vap; struct componentname *cnp = ap->a_cnp; struct nfsv2_sattr *sp; u_int32_t *tl; struct nfsnode *np = NULL; struct vnode *newvp = NULL; caddr_t bpos, dpos; int error = 0, wccflag = NFSV3_WCCRATTR, gotvp = 0, fmode = 0; struct mbuf *mreq, *mrep, *md, *mb; struct vattr vattr; int v3 = NFS_ISV3(dvp); /* * Oops, not for me.. */ if (vap->va_type == VSOCK) return (nfs_mknodrpc(dvp, ap->a_vpp, cnp, vap)); if ((error = VOP_GETATTR(dvp, &vattr, cnp->cn_cred)) != 0) return (error); if (vap->va_vaflags & VA_EXCLUSIVE) fmode |= O_EXCL; again: nfsstats.rpccnt[NFSPROC_CREATE]++; mreq = nfsm_reqhead(dvp, NFSPROC_CREATE, NFSX_FH(v3) + 2 * NFSX_UNSIGNED + nfsm_rndup(cnp->cn_namelen) + NFSX_SATTR(v3)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(dvp, v3); nfsm_strtom(cnp->cn_nameptr, cnp->cn_namelen, NFS_MAXNAMLEN); if (v3) { tl = nfsm_build(u_int32_t *, NFSX_UNSIGNED); if (fmode & O_EXCL) { *tl = txdr_unsigned(NFSV3CREATE_EXCLUSIVE); tl = nfsm_build(u_int32_t *, NFSX_V3CREATEVERF); #ifdef INET INIT_VNET_INET(curvnet); IN_IFADDR_RLOCK(); if (!TAILQ_EMPTY(&V_in_ifaddrhead)) *tl++ = IA_SIN(TAILQ_FIRST(&V_in_ifaddrhead))->sin_addr.s_addr; else #endif *tl++ = create_verf; #ifdef INET IN_IFADDR_RUNLOCK(); #endif *tl = ++create_verf; } else { *tl = txdr_unsigned(NFSV3CREATE_UNCHECKED); nfsm_v3attrbuild(vap, FALSE); } } else { sp = nfsm_build(struct nfsv2_sattr *, NFSX_V2SATTR); sp->sa_mode = vtonfsv2_mode(vap->va_type, vap->va_mode); sp->sa_uid = nfs_xdrneg1; sp->sa_gid = nfs_xdrneg1; sp->sa_size = 0; txdr_nfsv2time(&vap->va_atime, &sp->sa_atime); txdr_nfsv2time(&vap->va_mtime, &sp->sa_mtime); } nfsm_request(dvp, NFSPROC_CREATE, cnp->cn_thread, cnp->cn_cred); if (!error) { nfsm_mtofh(dvp, newvp, v3, gotvp); if (!gotvp) { if (newvp) { vput(newvp); newvp = NULL; } error = nfs_lookitup(dvp, cnp->cn_nameptr, cnp->cn_namelen, cnp->cn_cred, cnp->cn_thread, &np); if (!error) newvp = NFSTOV(np); } } if (v3) nfsm_wcc_data(dvp, wccflag); m_freem(mrep); nfsmout: if (error) { if (v3 && (fmode & O_EXCL) && error == NFSERR_NOTSUPP) { fmode &= ~O_EXCL; goto again; } if (newvp) vput(newvp); } else if (v3 && (fmode & O_EXCL)) { /* * We are normally called with only a partially initialized * VAP. Since the NFSv3 spec says that server may use the * file attributes to store the verifier, the spec requires * us to do a SETATTR RPC. FreeBSD servers store the verifier * in atime, but we can't really assume that all servers will * so we ensure that our SETATTR sets both atime and mtime. */ if (vap->va_mtime.tv_sec == VNOVAL) vfs_timestamp(&vap->va_mtime); if (vap->va_atime.tv_sec == VNOVAL) vap->va_atime = vap->va_mtime; error = nfs_setattrrpc(newvp, vap, cnp->cn_cred); if (error) vput(newvp); } if (!error) { if (cnp->cn_flags & MAKEENTRY) cache_enter(dvp, newvp, cnp); *ap->a_vpp = newvp; } mtx_lock(&(VTONFS(dvp))->n_mtx); VTONFS(dvp)->n_flag |= NMODIFIED; if (!wccflag) { VTONFS(dvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(dvp); } mtx_unlock(&(VTONFS(dvp))->n_mtx); return (error); } /* * nfs file remove call * To try and make nfs semantics closer to ufs semantics, a file that has * other processes using the vnode is renamed instead of removed and then * removed later on the last close. * - If v_usecount > 1 * If a rename is not already in the works * call nfs_sillyrename() to set it up * else * do the remove rpc */ static int nfs_remove(struct vop_remove_args *ap) { struct vnode *vp = ap->a_vp; struct vnode *dvp = ap->a_dvp; struct componentname *cnp = ap->a_cnp; struct nfsnode *np = VTONFS(vp); int error = 0; struct vattr vattr; #ifndef DIAGNOSTIC if ((cnp->cn_flags & HASBUF) == 0) panic("nfs_remove: no name"); if (vrefcnt(vp) < 1) panic("nfs_remove: bad v_usecount"); #endif if (vp->v_type == VDIR) error = EPERM; else if (vrefcnt(vp) == 1 || (np->n_sillyrename && !VOP_GETATTR(vp, &vattr, cnp->cn_cred) && vattr.va_nlink > 1)) { /* * Purge the name cache so that the chance of a lookup for * the name succeeding while the remove is in progress is * minimized. Without node locking it can still happen, such * that an I/O op returns ESTALE, but since you get this if * another host removes the file.. */ cache_purge(vp); /* * throw away biocache buffers, mainly to avoid * unnecessary delayed writes later. */ error = nfs_vinvalbuf(vp, 0, cnp->cn_thread, 1); /* Do the rpc */ if (error != EINTR && error != EIO) error = nfs_removerpc(dvp, cnp->cn_nameptr, cnp->cn_namelen, cnp->cn_cred, cnp->cn_thread); /* * Kludge City: If the first reply to the remove rpc is lost.. * the reply to the retransmitted request will be ENOENT * since the file was in fact removed * Therefore, we cheat and return success. */ if (error == ENOENT) error = 0; } else if (!np->n_sillyrename) error = nfs_sillyrename(dvp, vp, cnp); np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); return (error); } /* * nfs file remove rpc called from nfs_inactive */ int nfs_removeit(struct sillyrename *sp) { /* * Make sure that the directory vnode is still valid. * XXX we should lock sp->s_dvp here. */ if (sp->s_dvp->v_type == VBAD) return (0); return (nfs_removerpc(sp->s_dvp, sp->s_name, sp->s_namlen, sp->s_cred, NULL)); } /* * Nfs remove rpc, called from nfs_remove() and nfs_removeit(). */ static int nfs_removerpc(struct vnode *dvp, const char *name, int namelen, struct ucred *cred, struct thread *td) { caddr_t bpos, dpos; int error = 0, wccflag = NFSV3_WCCRATTR; struct mbuf *mreq, *mrep, *md, *mb; int v3 = NFS_ISV3(dvp); nfsstats.rpccnt[NFSPROC_REMOVE]++; mreq = nfsm_reqhead(dvp, NFSPROC_REMOVE, NFSX_FH(v3) + NFSX_UNSIGNED + nfsm_rndup(namelen)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(dvp, v3); nfsm_strtom(name, namelen, NFS_MAXNAMLEN); nfsm_request(dvp, NFSPROC_REMOVE, td, cred); if (v3) nfsm_wcc_data(dvp, wccflag); m_freem(mrep); nfsmout: mtx_lock(&(VTONFS(dvp))->n_mtx); VTONFS(dvp)->n_flag |= NMODIFIED; if (!wccflag) { VTONFS(dvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(dvp); } mtx_unlock(&(VTONFS(dvp))->n_mtx); return (error); } /* * nfs file rename call */ static int nfs_rename(struct vop_rename_args *ap) { struct vnode *fvp = ap->a_fvp; struct vnode *tvp = ap->a_tvp; struct vnode *fdvp = ap->a_fdvp; struct vnode *tdvp = ap->a_tdvp; struct componentname *tcnp = ap->a_tcnp; struct componentname *fcnp = ap->a_fcnp; int error; #ifndef DIAGNOSTIC if ((tcnp->cn_flags & HASBUF) == 0 || (fcnp->cn_flags & HASBUF) == 0) panic("nfs_rename: no name"); #endif /* Check for cross-device rename */ if ((fvp->v_mount != tdvp->v_mount) || (tvp && (fvp->v_mount != tvp->v_mount))) { error = EXDEV; goto out; } if (fvp == tvp) { nfs_printf("nfs_rename: fvp == tvp (can't happen)\n"); error = 0; goto out; } if ((error = vn_lock(fvp, LK_EXCLUSIVE)) != 0) goto out; /* * We have to flush B_DELWRI data prior to renaming * the file. If we don't, the delayed-write buffers * can be flushed out later after the file has gone stale * under NFSV3. NFSV2 does not have this problem because * ( as far as I can tell ) it flushes dirty buffers more * often. * * Skip the rename operation if the fsync fails, this can happen * due to the server's volume being full, when we pushed out data * that was written back to our cache earlier. Not checking for * this condition can result in potential (silent) data loss. */ error = VOP_FSYNC(fvp, MNT_WAIT, fcnp->cn_thread); VOP_UNLOCK(fvp, 0); if (!error && tvp) error = VOP_FSYNC(tvp, MNT_WAIT, tcnp->cn_thread); if (error) goto out; /* * If the tvp exists and is in use, sillyrename it before doing the * rename of the new file over it. * XXX Can't sillyrename a directory. */ if (tvp && vrefcnt(tvp) > 1 && !VTONFS(tvp)->n_sillyrename && tvp->v_type != VDIR && !nfs_sillyrename(tdvp, tvp, tcnp)) { vput(tvp); tvp = NULL; } error = nfs_renamerpc(fdvp, fcnp->cn_nameptr, fcnp->cn_namelen, tdvp, tcnp->cn_nameptr, tcnp->cn_namelen, tcnp->cn_cred, tcnp->cn_thread); if (fvp->v_type == VDIR) { if (tvp != NULL && tvp->v_type == VDIR) cache_purge(tdvp); cache_purge(fdvp); } out: if (tdvp == tvp) vrele(tdvp); else vput(tdvp); if (tvp) vput(tvp); vrele(fdvp); vrele(fvp); /* * Kludge: Map ENOENT => 0 assuming that it is a reply to a retry. */ if (error == ENOENT) error = 0; return (error); } /* * nfs file rename rpc called from nfs_remove() above */ static int nfs_renameit(struct vnode *sdvp, struct componentname *scnp, struct sillyrename *sp) { return (nfs_renamerpc(sdvp, scnp->cn_nameptr, scnp->cn_namelen, sdvp, sp->s_name, sp->s_namlen, scnp->cn_cred, scnp->cn_thread)); } /* * Do an nfs rename rpc. Called from nfs_rename() and nfs_renameit(). */ static int nfs_renamerpc(struct vnode *fdvp, const char *fnameptr, int fnamelen, struct vnode *tdvp, const char *tnameptr, int tnamelen, struct ucred *cred, struct thread *td) { caddr_t bpos, dpos; int error = 0, fwccflag = NFSV3_WCCRATTR, twccflag = NFSV3_WCCRATTR; struct mbuf *mreq, *mrep, *md, *mb; int v3 = NFS_ISV3(fdvp); nfsstats.rpccnt[NFSPROC_RENAME]++; mreq = nfsm_reqhead(fdvp, NFSPROC_RENAME, (NFSX_FH(v3) + NFSX_UNSIGNED)*2 + nfsm_rndup(fnamelen) + nfsm_rndup(tnamelen)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(fdvp, v3); nfsm_strtom(fnameptr, fnamelen, NFS_MAXNAMLEN); nfsm_fhtom(tdvp, v3); nfsm_strtom(tnameptr, tnamelen, NFS_MAXNAMLEN); nfsm_request(fdvp, NFSPROC_RENAME, td, cred); if (v3) { nfsm_wcc_data(fdvp, fwccflag); nfsm_wcc_data(tdvp, twccflag); } m_freem(mrep); nfsmout: mtx_lock(&(VTONFS(fdvp))->n_mtx); VTONFS(fdvp)->n_flag |= NMODIFIED; mtx_unlock(&(VTONFS(fdvp))->n_mtx); mtx_lock(&(VTONFS(tdvp))->n_mtx); VTONFS(tdvp)->n_flag |= NMODIFIED; mtx_unlock(&(VTONFS(tdvp))->n_mtx); if (!fwccflag) { VTONFS(fdvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(fdvp); } if (!twccflag) { VTONFS(tdvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(tdvp); } return (error); } /* * nfs hard link create call */ static int nfs_link(struct vop_link_args *ap) { struct vnode *vp = ap->a_vp; struct vnode *tdvp = ap->a_tdvp; struct componentname *cnp = ap->a_cnp; caddr_t bpos, dpos; int error = 0, wccflag = NFSV3_WCCRATTR, attrflag = 0; struct mbuf *mreq, *mrep, *md, *mb; int v3; if (vp->v_mount != tdvp->v_mount) { return (EXDEV); } /* * Push all writes to the server, so that the attribute cache * doesn't get "out of sync" with the server. * XXX There should be a better way! */ VOP_FSYNC(vp, MNT_WAIT, cnp->cn_thread); v3 = NFS_ISV3(vp); nfsstats.rpccnt[NFSPROC_LINK]++; mreq = nfsm_reqhead(vp, NFSPROC_LINK, NFSX_FH(v3)*2 + NFSX_UNSIGNED + nfsm_rndup(cnp->cn_namelen)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, v3); nfsm_fhtom(tdvp, v3); nfsm_strtom(cnp->cn_nameptr, cnp->cn_namelen, NFS_MAXNAMLEN); nfsm_request(vp, NFSPROC_LINK, cnp->cn_thread, cnp->cn_cred); if (v3) { nfsm_postop_attr(vp, attrflag); nfsm_wcc_data(tdvp, wccflag); } m_freem(mrep); nfsmout: mtx_lock(&(VTONFS(tdvp))->n_mtx); VTONFS(tdvp)->n_flag |= NMODIFIED; mtx_unlock(&(VTONFS(tdvp))->n_mtx); if (!attrflag) { VTONFS(vp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); } if (!wccflag) { VTONFS(tdvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(tdvp); } return (error); } /* * nfs symbolic link create call */ static int nfs_symlink(struct vop_symlink_args *ap) { struct vnode *dvp = ap->a_dvp; struct vattr *vap = ap->a_vap; struct componentname *cnp = ap->a_cnp; struct nfsv2_sattr *sp; caddr_t bpos, dpos; int slen, error = 0, wccflag = NFSV3_WCCRATTR, gotvp; struct mbuf *mreq, *mrep, *md, *mb; struct vnode *newvp = NULL; int v3 = NFS_ISV3(dvp); nfsstats.rpccnt[NFSPROC_SYMLINK]++; slen = strlen(ap->a_target); mreq = nfsm_reqhead(dvp, NFSPROC_SYMLINK, NFSX_FH(v3) + 2*NFSX_UNSIGNED + nfsm_rndup(cnp->cn_namelen) + nfsm_rndup(slen) + NFSX_SATTR(v3)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(dvp, v3); nfsm_strtom(cnp->cn_nameptr, cnp->cn_namelen, NFS_MAXNAMLEN); if (v3) { nfsm_v3attrbuild(vap, FALSE); } nfsm_strtom(ap->a_target, slen, NFS_MAXPATHLEN); if (!v3) { sp = nfsm_build(struct nfsv2_sattr *, NFSX_V2SATTR); sp->sa_mode = vtonfsv2_mode(VLNK, vap->va_mode); sp->sa_uid = nfs_xdrneg1; sp->sa_gid = nfs_xdrneg1; sp->sa_size = nfs_xdrneg1; txdr_nfsv2time(&vap->va_atime, &sp->sa_atime); txdr_nfsv2time(&vap->va_mtime, &sp->sa_mtime); } /* * Issue the NFS request and get the rpc response. * * Only NFSv3 responses returning an error of 0 actually return * a file handle that can be converted into newvp without having * to do an extra lookup rpc. */ nfsm_request(dvp, NFSPROC_SYMLINK, cnp->cn_thread, cnp->cn_cred); if (v3) { if (error == 0) nfsm_mtofh(dvp, newvp, v3, gotvp); nfsm_wcc_data(dvp, wccflag); } /* * out code jumps -> here, mrep is also freed. */ m_freem(mrep); nfsmout: /* * If we do not have an error and we could not extract the newvp from * the response due to the request being NFSv2, we have to do a * lookup in order to obtain a newvp to return. */ if (error == 0 && newvp == NULL) { struct nfsnode *np = NULL; error = nfs_lookitup(dvp, cnp->cn_nameptr, cnp->cn_namelen, cnp->cn_cred, cnp->cn_thread, &np); if (!error) newvp = NFSTOV(np); } if (error) { if (newvp) vput(newvp); } else { *ap->a_vpp = newvp; } mtx_lock(&(VTONFS(dvp))->n_mtx); VTONFS(dvp)->n_flag |= NMODIFIED; mtx_unlock(&(VTONFS(dvp))->n_mtx); if (!wccflag) { VTONFS(dvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(dvp); } return (error); } /* * nfs make dir call */ static int nfs_mkdir(struct vop_mkdir_args *ap) { struct vnode *dvp = ap->a_dvp; struct vattr *vap = ap->a_vap; struct componentname *cnp = ap->a_cnp; struct nfsv2_sattr *sp; int len; struct nfsnode *np = NULL; struct vnode *newvp = NULL; caddr_t bpos, dpos; int error = 0, wccflag = NFSV3_WCCRATTR; int gotvp = 0; struct mbuf *mreq, *mrep, *md, *mb; struct vattr vattr; int v3 = NFS_ISV3(dvp); if ((error = VOP_GETATTR(dvp, &vattr, cnp->cn_cred)) != 0) return (error); len = cnp->cn_namelen; nfsstats.rpccnt[NFSPROC_MKDIR]++; mreq = nfsm_reqhead(dvp, NFSPROC_MKDIR, NFSX_FH(v3) + NFSX_UNSIGNED + nfsm_rndup(len) + NFSX_SATTR(v3)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(dvp, v3); nfsm_strtom(cnp->cn_nameptr, len, NFS_MAXNAMLEN); if (v3) { nfsm_v3attrbuild(vap, FALSE); } else { sp = nfsm_build(struct nfsv2_sattr *, NFSX_V2SATTR); sp->sa_mode = vtonfsv2_mode(VDIR, vap->va_mode); sp->sa_uid = nfs_xdrneg1; sp->sa_gid = nfs_xdrneg1; sp->sa_size = nfs_xdrneg1; txdr_nfsv2time(&vap->va_atime, &sp->sa_atime); txdr_nfsv2time(&vap->va_mtime, &sp->sa_mtime); } nfsm_request(dvp, NFSPROC_MKDIR, cnp->cn_thread, cnp->cn_cred); if (!error) nfsm_mtofh(dvp, newvp, v3, gotvp); if (v3) nfsm_wcc_data(dvp, wccflag); m_freem(mrep); nfsmout: mtx_lock(&(VTONFS(dvp))->n_mtx); VTONFS(dvp)->n_flag |= NMODIFIED; mtx_unlock(&(VTONFS(dvp))->n_mtx); if (!wccflag) { VTONFS(dvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(dvp); } if (error == 0 && newvp == NULL) { error = nfs_lookitup(dvp, cnp->cn_nameptr, len, cnp->cn_cred, cnp->cn_thread, &np); if (!error) { newvp = NFSTOV(np); if (newvp->v_type != VDIR) error = EEXIST; } } if (error) { if (newvp) vput(newvp); } else *ap->a_vpp = newvp; return (error); } /* * nfs remove directory call */ static int nfs_rmdir(struct vop_rmdir_args *ap) { struct vnode *vp = ap->a_vp; struct vnode *dvp = ap->a_dvp; struct componentname *cnp = ap->a_cnp; caddr_t bpos, dpos; int error = 0, wccflag = NFSV3_WCCRATTR; struct mbuf *mreq, *mrep, *md, *mb; int v3 = NFS_ISV3(dvp); if (dvp == vp) return (EINVAL); nfsstats.rpccnt[NFSPROC_RMDIR]++; mreq = nfsm_reqhead(dvp, NFSPROC_RMDIR, NFSX_FH(v3) + NFSX_UNSIGNED + nfsm_rndup(cnp->cn_namelen)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(dvp, v3); nfsm_strtom(cnp->cn_nameptr, cnp->cn_namelen, NFS_MAXNAMLEN); nfsm_request(dvp, NFSPROC_RMDIR, cnp->cn_thread, cnp->cn_cred); if (v3) nfsm_wcc_data(dvp, wccflag); m_freem(mrep); nfsmout: mtx_lock(&(VTONFS(dvp))->n_mtx); VTONFS(dvp)->n_flag |= NMODIFIED; mtx_unlock(&(VTONFS(dvp))->n_mtx); if (!wccflag) { VTONFS(dvp)->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(dvp); } cache_purge(dvp); cache_purge(vp); /* * Kludge: Map ENOENT => 0 assuming that you have a reply to a retry. */ if (error == ENOENT) error = 0; return (error); } /* * nfs readdir call */ static int nfs_readdir(struct vop_readdir_args *ap) { struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); struct uio *uio = ap->a_uio; int tresid, error = 0; struct vattr vattr; if (vp->v_type != VDIR) return(EPERM); /* * First, check for hit on the EOF offset cache */ if (np->n_direofoffset > 0 && uio->uio_offset >= np->n_direofoffset && (np->n_flag & NMODIFIED) == 0) { if (VOP_GETATTR(vp, &vattr, ap->a_cred) == 0) { mtx_lock(&np->n_mtx); if (!NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime)) { mtx_unlock(&np->n_mtx); nfsstats.direofcache_hits++; goto out; } else mtx_unlock(&np->n_mtx); } } /* * Call nfs_bioread() to do the real work. */ tresid = uio->uio_resid; error = nfs_bioread(vp, uio, 0, ap->a_cred); if (!error && uio->uio_resid == tresid) { nfsstats.direofcache_misses++; } out: return (error); } /* * Readdir rpc call. * Called from below the buffer cache by nfs_doio(). */ int nfs_readdirrpc(struct vnode *vp, struct uio *uiop, struct ucred *cred) { int len, left; struct dirent *dp = NULL; u_int32_t *tl; caddr_t cp; nfsuint64 *cookiep; caddr_t bpos, dpos; struct mbuf *mreq, *mrep, *md, *mb; nfsuint64 cookie; struct nfsmount *nmp = VFSTONFS(vp->v_mount); struct nfsnode *dnp = VTONFS(vp); u_quad_t fileno; int error = 0, tlen, more_dirs = 1, blksiz = 0, bigenough = 1; int attrflag; int v3 = NFS_ISV3(vp); #ifndef DIAGNOSTIC if (uiop->uio_iovcnt != 1 || (uiop->uio_offset & (DIRBLKSIZ - 1)) || (uiop->uio_resid & (DIRBLKSIZ - 1))) panic("nfs readdirrpc bad uio"); #endif /* * If there is no cookie, assume directory was stale. */ nfs_dircookie_lock(dnp); cookiep = nfs_getcookie(dnp, uiop->uio_offset, 0); if (cookiep) { cookie = *cookiep; nfs_dircookie_unlock(dnp); } else { nfs_dircookie_unlock(dnp); return (NFSERR_BAD_COOKIE); } /* * Loop around doing readdir rpc's of size nm_readdirsize * truncated to a multiple of DIRBLKSIZ. * The stopping criteria is EOF or buffer full. */ while (more_dirs && bigenough) { nfsstats.rpccnt[NFSPROC_READDIR]++; mreq = nfsm_reqhead(vp, NFSPROC_READDIR, NFSX_FH(v3) + NFSX_READDIR(v3)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, v3); if (v3) { tl = nfsm_build(u_int32_t *, 5 * NFSX_UNSIGNED); *tl++ = cookie.nfsuquad[0]; *tl++ = cookie.nfsuquad[1]; mtx_lock(&dnp->n_mtx); *tl++ = dnp->n_cookieverf.nfsuquad[0]; *tl++ = dnp->n_cookieverf.nfsuquad[1]; mtx_unlock(&dnp->n_mtx); } else { tl = nfsm_build(u_int32_t *, 2 * NFSX_UNSIGNED); *tl++ = cookie.nfsuquad[0]; } *tl = txdr_unsigned(nmp->nm_readdirsize); nfsm_request(vp, NFSPROC_READDIR, uiop->uio_td, cred); if (v3) { nfsm_postop_attr(vp, attrflag); if (!error) { tl = nfsm_dissect(u_int32_t *, 2 * NFSX_UNSIGNED); mtx_lock(&dnp->n_mtx); dnp->n_cookieverf.nfsuquad[0] = *tl++; dnp->n_cookieverf.nfsuquad[1] = *tl; mtx_unlock(&dnp->n_mtx); } else { m_freem(mrep); goto nfsmout; } } tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); more_dirs = fxdr_unsigned(int, *tl); /* loop thru the dir entries, doctoring them to 4bsd form */ while (more_dirs && bigenough) { if (v3) { tl = nfsm_dissect(u_int32_t *, 3 * NFSX_UNSIGNED); fileno = fxdr_hyper(tl); len = fxdr_unsigned(int, *(tl + 2)); } else { tl = nfsm_dissect(u_int32_t *, 2 * NFSX_UNSIGNED); fileno = fxdr_unsigned(u_quad_t, *tl++); len = fxdr_unsigned(int, *tl); } if (len <= 0 || len > NFS_MAXNAMLEN) { error = EBADRPC; m_freem(mrep); goto nfsmout; } tlen = nfsm_rndup(len); if (tlen == len) tlen += 4; /* To ensure null termination */ left = DIRBLKSIZ - blksiz; if ((tlen + DIRHDSIZ) > left) { dp->d_reclen += left; uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + left; uiop->uio_iov->iov_len -= left; uiop->uio_offset += left; uiop->uio_resid -= left; blksiz = 0; } if ((tlen + DIRHDSIZ) > uiop->uio_resid) bigenough = 0; if (bigenough) { dp = (struct dirent *)uiop->uio_iov->iov_base; dp->d_fileno = (int)fileno; dp->d_namlen = len; dp->d_reclen = tlen + DIRHDSIZ; dp->d_type = DT_UNKNOWN; blksiz += dp->d_reclen; if (blksiz == DIRBLKSIZ) blksiz = 0; uiop->uio_offset += DIRHDSIZ; uiop->uio_resid -= DIRHDSIZ; uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + DIRHDSIZ; uiop->uio_iov->iov_len -= DIRHDSIZ; nfsm_mtouio(uiop, len); cp = uiop->uio_iov->iov_base; tlen -= len; *cp = '\0'; /* null terminate */ uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + tlen; uiop->uio_iov->iov_len -= tlen; uiop->uio_offset += tlen; uiop->uio_resid -= tlen; } else nfsm_adv(nfsm_rndup(len)); if (v3) { tl = nfsm_dissect(u_int32_t *, 3 * NFSX_UNSIGNED); } else { tl = nfsm_dissect(u_int32_t *, 2 * NFSX_UNSIGNED); } if (bigenough) { cookie.nfsuquad[0] = *tl++; if (v3) cookie.nfsuquad[1] = *tl++; } else if (v3) tl += 2; else tl++; more_dirs = fxdr_unsigned(int, *tl); } /* * If at end of rpc data, get the eof boolean */ if (!more_dirs) { tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); more_dirs = (fxdr_unsigned(int, *tl) == 0); } m_freem(mrep); } /* * Fill last record, iff any, out to a multiple of DIRBLKSIZ * by increasing d_reclen for the last record. */ if (blksiz > 0) { left = DIRBLKSIZ - blksiz; dp->d_reclen += left; uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + left; uiop->uio_iov->iov_len -= left; uiop->uio_offset += left; uiop->uio_resid -= left; } /* * We are now either at the end of the directory or have filled the * block. */ if (bigenough) dnp->n_direofoffset = uiop->uio_offset; else { if (uiop->uio_resid > 0) nfs_printf("EEK! readdirrpc resid > 0\n"); nfs_dircookie_lock(dnp); cookiep = nfs_getcookie(dnp, uiop->uio_offset, 1); *cookiep = cookie; nfs_dircookie_unlock(dnp); } nfsmout: return (error); } /* * NFS V3 readdir plus RPC. Used in place of nfs_readdirrpc(). */ int nfs_readdirplusrpc(struct vnode *vp, struct uio *uiop, struct ucred *cred) { int len, left; struct dirent *dp; u_int32_t *tl; caddr_t cp; struct vnode *newvp; nfsuint64 *cookiep; caddr_t bpos, dpos, dpossav1, dpossav2; struct mbuf *mreq, *mrep, *md, *mb, *mdsav1, *mdsav2; struct nameidata nami, *ndp = &nami; struct componentname *cnp = &ndp->ni_cnd; nfsuint64 cookie; struct nfsmount *nmp = VFSTONFS(vp->v_mount); struct nfsnode *dnp = VTONFS(vp), *np; nfsfh_t *fhp; u_quad_t fileno; int error = 0, tlen, more_dirs = 1, blksiz = 0, doit, bigenough = 1, i; int attrflag, fhsize; #ifndef nolint dp = NULL; #endif #ifndef DIAGNOSTIC if (uiop->uio_iovcnt != 1 || (uiop->uio_offset & (DIRBLKSIZ - 1)) || (uiop->uio_resid & (DIRBLKSIZ - 1))) panic("nfs readdirplusrpc bad uio"); #endif ndp->ni_dvp = vp; newvp = NULLVP; /* * If there is no cookie, assume directory was stale. */ nfs_dircookie_lock(dnp); cookiep = nfs_getcookie(dnp, uiop->uio_offset, 0); if (cookiep) { cookie = *cookiep; nfs_dircookie_unlock(dnp); } else { nfs_dircookie_unlock(dnp); return (NFSERR_BAD_COOKIE); } /* * Loop around doing readdir rpc's of size nm_readdirsize * truncated to a multiple of DIRBLKSIZ. * The stopping criteria is EOF or buffer full. */ while (more_dirs && bigenough) { nfsstats.rpccnt[NFSPROC_READDIRPLUS]++; mreq = nfsm_reqhead(vp, NFSPROC_READDIRPLUS, NFSX_FH(1) + 6 * NFSX_UNSIGNED); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, 1); tl = nfsm_build(u_int32_t *, 6 * NFSX_UNSIGNED); *tl++ = cookie.nfsuquad[0]; *tl++ = cookie.nfsuquad[1]; mtx_lock(&dnp->n_mtx); *tl++ = dnp->n_cookieverf.nfsuquad[0]; *tl++ = dnp->n_cookieverf.nfsuquad[1]; mtx_unlock(&dnp->n_mtx); *tl++ = txdr_unsigned(nmp->nm_readdirsize); *tl = txdr_unsigned(nmp->nm_rsize); nfsm_request(vp, NFSPROC_READDIRPLUS, uiop->uio_td, cred); nfsm_postop_attr(vp, attrflag); if (error) { m_freem(mrep); goto nfsmout; } tl = nfsm_dissect(u_int32_t *, 3 * NFSX_UNSIGNED); mtx_lock(&dnp->n_mtx); dnp->n_cookieverf.nfsuquad[0] = *tl++; dnp->n_cookieverf.nfsuquad[1] = *tl++; mtx_unlock(&dnp->n_mtx); more_dirs = fxdr_unsigned(int, *tl); /* loop thru the dir entries, doctoring them to 4bsd form */ while (more_dirs && bigenough) { tl = nfsm_dissect(u_int32_t *, 3 * NFSX_UNSIGNED); fileno = fxdr_hyper(tl); len = fxdr_unsigned(int, *(tl + 2)); if (len <= 0 || len > NFS_MAXNAMLEN) { error = EBADRPC; m_freem(mrep); goto nfsmout; } tlen = nfsm_rndup(len); if (tlen == len) tlen += 4; /* To ensure null termination*/ left = DIRBLKSIZ - blksiz; if ((tlen + DIRHDSIZ) > left) { dp->d_reclen += left; uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + left; uiop->uio_iov->iov_len -= left; uiop->uio_offset += left; uiop->uio_resid -= left; blksiz = 0; } if ((tlen + DIRHDSIZ) > uiop->uio_resid) bigenough = 0; if (bigenough) { dp = (struct dirent *)uiop->uio_iov->iov_base; dp->d_fileno = (int)fileno; dp->d_namlen = len; dp->d_reclen = tlen + DIRHDSIZ; dp->d_type = DT_UNKNOWN; blksiz += dp->d_reclen; if (blksiz == DIRBLKSIZ) blksiz = 0; uiop->uio_offset += DIRHDSIZ; uiop->uio_resid -= DIRHDSIZ; uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + DIRHDSIZ; uiop->uio_iov->iov_len -= DIRHDSIZ; cnp->cn_nameptr = uiop->uio_iov->iov_base; cnp->cn_namelen = len; nfsm_mtouio(uiop, len); cp = uiop->uio_iov->iov_base; tlen -= len; *cp = '\0'; uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + tlen; uiop->uio_iov->iov_len -= tlen; uiop->uio_offset += tlen; uiop->uio_resid -= tlen; } else nfsm_adv(nfsm_rndup(len)); tl = nfsm_dissect(u_int32_t *, 3 * NFSX_UNSIGNED); if (bigenough) { cookie.nfsuquad[0] = *tl++; cookie.nfsuquad[1] = *tl++; } else tl += 2; /* * Since the attributes are before the file handle * (sigh), we must skip over the attributes and then * come back and get them. */ attrflag = fxdr_unsigned(int, *tl); if (attrflag) { dpossav1 = dpos; mdsav1 = md; nfsm_adv(NFSX_V3FATTR); tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); doit = fxdr_unsigned(int, *tl); /* * Skip loading the attrs for "..". There's a * race between loading the attrs here and * lookups that look for the directory currently * being read (in the parent). We try to acquire * the exclusive lock on ".." here, owning the * lock on the directory being read. Lookup will * hold the lock on ".." and try to acquire the * lock on the directory being read. * * There are other ways of fixing this, one would * be to do a trylock on the ".." vnode and skip * loading the attrs on ".." if it happens to be * locked by another process. But skipping the * attrload on ".." seems the easiest option. */ if (strcmp(dp->d_name, "..") == 0) { doit = 0; /* * We've already skipped over the attrs, * skip over the filehandle. And store d_type * as VDIR. */ tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); i = fxdr_unsigned(int, *tl); nfsm_adv(nfsm_rndup(i)); dp->d_type = IFTODT(VTTOIF(VDIR)); } if (doit) { nfsm_getfh(fhp, fhsize, 1); if (NFS_CMPFH(dnp, fhp, fhsize)) { VREF(vp); newvp = vp; np = dnp; } else { error = nfs_nget(vp->v_mount, fhp, fhsize, &np, LK_EXCLUSIVE); if (error) doit = 0; else newvp = NFSTOV(np); } } if (doit && bigenough) { dpossav2 = dpos; dpos = dpossav1; mdsav2 = md; md = mdsav1; nfsm_loadattr(newvp, NULL); dpos = dpossav2; md = mdsav2; dp->d_type = IFTODT(VTTOIF(np->n_vattr.va_type)); ndp->ni_vp = newvp; /* Update n_ctime, so subsequent lookup doesn't purge entry */ np->n_ctime = np->n_vattr.va_ctime.tv_sec; cache_enter(ndp->ni_dvp, ndp->ni_vp, cnp); } } else { /* Just skip over the file handle */ tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); i = fxdr_unsigned(int, *tl); if (i) { tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); fhsize = fxdr_unsigned(int, *tl); nfsm_adv(nfsm_rndup(fhsize)); } } if (newvp != NULLVP) { if (newvp == vp) vrele(newvp); else vput(newvp); newvp = NULLVP; } tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); more_dirs = fxdr_unsigned(int, *tl); } /* * If at end of rpc data, get the eof boolean */ if (!more_dirs) { tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED); more_dirs = (fxdr_unsigned(int, *tl) == 0); } m_freem(mrep); } /* * Fill last record, iff any, out to a multiple of DIRBLKSIZ * by increasing d_reclen for the last record. */ if (blksiz > 0) { left = DIRBLKSIZ - blksiz; dp->d_reclen += left; uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + left; uiop->uio_iov->iov_len -= left; uiop->uio_offset += left; uiop->uio_resid -= left; } /* * We are now either at the end of the directory or have filled the * block. */ if (bigenough) dnp->n_direofoffset = uiop->uio_offset; else { if (uiop->uio_resid > 0) nfs_printf("EEK! readdirplusrpc resid > 0\n"); nfs_dircookie_lock(dnp); cookiep = nfs_getcookie(dnp, uiop->uio_offset, 1); *cookiep = cookie; nfs_dircookie_unlock(dnp); } nfsmout: if (newvp != NULLVP) { if (newvp == vp) vrele(newvp); else vput(newvp); newvp = NULLVP; } return (error); } /* * Silly rename. To make the NFS filesystem that is stateless look a little * more like the "ufs" a remove of an active vnode is translated to a rename * to a funny looking filename that is removed by nfs_inactive on the * nfsnode. There is the potential for another process on a different client * to create the same funny name between the nfs_lookitup() fails and the * nfs_rename() completes, but... */ static int nfs_sillyrename(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) { struct sillyrename *sp; struct nfsnode *np; int error; short pid; unsigned int lticks; cache_purge(dvp); np = VTONFS(vp); #ifndef DIAGNOSTIC if (vp->v_type == VDIR) panic("nfs: sillyrename dir"); #endif sp = malloc(sizeof (struct sillyrename), M_NFSREQ, M_WAITOK); sp->s_cred = crhold(cnp->cn_cred); sp->s_dvp = dvp; sp->s_removeit = nfs_removeit; VREF(dvp); /* * Fudge together a funny name. * Changing the format of the funny name to accomodate more * sillynames per directory. * The name is now changed to .nfs...4, where ticks is * CPU ticks since boot. */ pid = cnp->cn_thread->td_proc->p_pid; lticks = (unsigned int)ticks; for ( ; ; ) { sp->s_namlen = sprintf(sp->s_name, ".nfs.%08x.%04x4.4", lticks, pid); if (nfs_lookitup(dvp, sp->s_name, sp->s_namlen, sp->s_cred, cnp->cn_thread, NULL)) break; lticks++; } error = nfs_renameit(dvp, cnp, sp); if (error) goto bad; error = nfs_lookitup(dvp, sp->s_name, sp->s_namlen, sp->s_cred, cnp->cn_thread, &np); np->n_sillyrename = sp; return (0); bad: vrele(sp->s_dvp); crfree(sp->s_cred); free((caddr_t)sp, M_NFSREQ); return (error); } /* * Look up a file name and optionally either update the file handle or * allocate an nfsnode, depending on the value of npp. * npp == NULL --> just do the lookup * *npp == NULL --> allocate a new nfsnode and make sure attributes are * handled too * *npp != NULL --> update the file handle in the vnode */ static int nfs_lookitup(struct vnode *dvp, const char *name, int len, struct ucred *cred, struct thread *td, struct nfsnode **npp) { struct vnode *newvp = NULL; struct nfsnode *np, *dnp = VTONFS(dvp); caddr_t bpos, dpos; int error = 0, fhlen, attrflag; struct mbuf *mreq, *mrep, *md, *mb; nfsfh_t *nfhp; int v3 = NFS_ISV3(dvp); nfsstats.rpccnt[NFSPROC_LOOKUP]++; mreq = nfsm_reqhead(dvp, NFSPROC_LOOKUP, NFSX_FH(v3) + NFSX_UNSIGNED + nfsm_rndup(len)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(dvp, v3); nfsm_strtom(name, len, NFS_MAXNAMLEN); nfsm_request(dvp, NFSPROC_LOOKUP, td, cred); if (npp && !error) { nfsm_getfh(nfhp, fhlen, v3); if (*npp) { np = *npp; if (np->n_fhsize > NFS_SMALLFH && fhlen <= NFS_SMALLFH) { free((caddr_t)np->n_fhp, M_NFSBIGFH); np->n_fhp = &np->n_fh; } else if (np->n_fhsize <= NFS_SMALLFH && fhlen>NFS_SMALLFH) np->n_fhp =(nfsfh_t *)malloc(fhlen, M_NFSBIGFH, M_WAITOK); bcopy((caddr_t)nfhp, (caddr_t)np->n_fhp, fhlen); np->n_fhsize = fhlen; newvp = NFSTOV(np); } else if (NFS_CMPFH(dnp, nfhp, fhlen)) { VREF(dvp); newvp = dvp; } else { error = nfs_nget(dvp->v_mount, nfhp, fhlen, &np, LK_EXCLUSIVE); if (error) { m_freem(mrep); return (error); } newvp = NFSTOV(np); } if (v3) { nfsm_postop_attr(newvp, attrflag); if (!attrflag && *npp == NULL) { m_freem(mrep); if (newvp == dvp) vrele(newvp); else vput(newvp); return (ENOENT); } } else nfsm_loadattr(newvp, NULL); } m_freem(mrep); nfsmout: if (npp && *npp == NULL) { if (error) { if (newvp) { if (newvp == dvp) vrele(newvp); else vput(newvp); } } else *npp = np; } return (error); } /* * Nfs Version 3 commit rpc */ int nfs_commit(struct vnode *vp, u_quad_t offset, int cnt, struct ucred *cred, struct thread *td) { u_int32_t *tl; struct nfsmount *nmp = VFSTONFS(vp->v_mount); caddr_t bpos, dpos; int error = 0, wccflag = NFSV3_WCCRATTR; struct mbuf *mreq, *mrep, *md, *mb; mtx_lock(&nmp->nm_mtx); if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0) { mtx_unlock(&nmp->nm_mtx); return (0); } mtx_unlock(&nmp->nm_mtx); nfsstats.rpccnt[NFSPROC_COMMIT]++; mreq = nfsm_reqhead(vp, NFSPROC_COMMIT, NFSX_FH(1)); mb = mreq; bpos = mtod(mb, caddr_t); nfsm_fhtom(vp, 1); tl = nfsm_build(u_int32_t *, 3 * NFSX_UNSIGNED); txdr_hyper(offset, tl); tl += 2; *tl = txdr_unsigned(cnt); nfsm_request(vp, NFSPROC_COMMIT, td, cred); nfsm_wcc_data(vp, wccflag); if (!error) { tl = nfsm_dissect(u_int32_t *, NFSX_V3WRITEVERF); if (bcmp((caddr_t)nmp->nm_verf, (caddr_t)tl, NFSX_V3WRITEVERF)) { bcopy((caddr_t)tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF); error = NFSERR_STALEWRITEVERF; } } m_freem(mrep); nfsmout: return (error); } /* * Strategy routine. * For async requests when nfsiod(s) are running, queue the request by * calling nfs_asyncio(), otherwise just all nfs_doio() to do the * request. */ static int nfs_strategy(struct vop_strategy_args *ap) { struct buf *bp = ap->a_bp; struct ucred *cr; KASSERT(!(bp->b_flags & B_DONE), ("nfs_strategy: buffer %p unexpectedly marked B_DONE", bp)); BUF_ASSERT_HELD(bp); if (bp->b_iocmd == BIO_READ) cr = bp->b_rcred; else cr = bp->b_wcred; /* * If the op is asynchronous and an i/o daemon is waiting * queue the request, wake it up and wait for completion * otherwise just do it ourselves. */ if ((bp->b_flags & B_ASYNC) == 0 || nfs_asyncio(VFSTONFS(ap->a_vp->v_mount), bp, NOCRED, curthread)) (void)nfs_doio(ap->a_vp, bp, cr, curthread); return (0); } /* * fsync vnode op. Just call nfs_flush() with commit == 1. */ /* ARGSUSED */ static int nfs_fsync(struct vop_fsync_args *ap) { return (nfs_flush(ap->a_vp, ap->a_waitfor, 1)); } /* * Flush all the blocks associated with a vnode. * Walk through the buffer pool and push any dirty pages * associated with the vnode. */ static int nfs_flush(struct vnode *vp, int waitfor, int commit) { struct nfsnode *np = VTONFS(vp); struct buf *bp; int i; struct buf *nbp; struct nfsmount *nmp = VFSTONFS(vp->v_mount); int error = 0, slptimeo = 0, slpflag = 0, retv, bvecpos; int passone = 1; u_quad_t off, endoff, toff; struct ucred* wcred = NULL; struct buf **bvec = NULL; struct bufobj *bo; struct thread *td = curthread; #ifndef NFS_COMMITBVECSIZ #define NFS_COMMITBVECSIZ 20 #endif struct buf *bvec_on_stack[NFS_COMMITBVECSIZ]; int bvecsize = 0, bveccount; if (nmp->nm_flag & NFSMNT_INT) slpflag = PCATCH; if (!commit) passone = 0; bo = &vp->v_bufobj; /* * A b_flags == (B_DELWRI | B_NEEDCOMMIT) block has been written to the * server, but has not been committed to stable storage on the server * yet. On the first pass, the byte range is worked out and the commit * rpc is done. On the second pass, nfs_writebp() is called to do the * job. */ again: off = (u_quad_t)-1; endoff = 0; bvecpos = 0; if (NFS_ISV3(vp) && commit) { if (bvec != NULL && bvec != bvec_on_stack) free(bvec, M_TEMP); /* * Count up how many buffers waiting for a commit. */ bveccount = 0; BO_LOCK(bo); TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (!BUF_ISLOCKED(bp) && (bp->b_flags & (B_DELWRI | B_NEEDCOMMIT)) == (B_DELWRI | B_NEEDCOMMIT)) bveccount++; } /* * Allocate space to remember the list of bufs to commit. It is * important to use M_NOWAIT here to avoid a race with nfs_write. * If we can't get memory (for whatever reason), we will end up * committing the buffers one-by-one in the loop below. */ if (bveccount > NFS_COMMITBVECSIZ) { /* * Release the vnode interlock to avoid a lock * order reversal. */ BO_UNLOCK(bo); bvec = (struct buf **) malloc(bveccount * sizeof(struct buf *), M_TEMP, M_NOWAIT); BO_LOCK(bo); if (bvec == NULL) { bvec = bvec_on_stack; bvecsize = NFS_COMMITBVECSIZ; } else bvecsize = bveccount; } else { bvec = bvec_on_stack; bvecsize = NFS_COMMITBVECSIZ; } TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (bvecpos >= bvecsize) break; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { nbp = TAILQ_NEXT(bp, b_bobufs); continue; } if ((bp->b_flags & (B_DELWRI | B_NEEDCOMMIT)) != (B_DELWRI | B_NEEDCOMMIT)) { BUF_UNLOCK(bp); nbp = TAILQ_NEXT(bp, b_bobufs); continue; } BO_UNLOCK(bo); bremfree(bp); /* * Work out if all buffers are using the same cred * so we can deal with them all with one commit. * * NOTE: we are not clearing B_DONE here, so we have * to do it later on in this routine if we intend to * initiate I/O on the bp. * * Note: to avoid loopback deadlocks, we do not * assign b_runningbufspace. */ if (wcred == NULL) wcred = bp->b_wcred; else if (wcred != bp->b_wcred) wcred = NOCRED; vfs_busy_pages(bp, 1); BO_LOCK(bo); /* * bp is protected by being locked, but nbp is not * and vfs_busy_pages() may sleep. We have to * recalculate nbp. */ nbp = TAILQ_NEXT(bp, b_bobufs); /* * A list of these buffers is kept so that the * second loop knows which buffers have actually * been committed. This is necessary, since there * may be a race between the commit rpc and new * uncommitted writes on the file. */ bvec[bvecpos++] = bp; toff = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; if (toff < off) off = toff; toff += (u_quad_t)(bp->b_dirtyend - bp->b_dirtyoff); if (toff > endoff) endoff = toff; } BO_UNLOCK(bo); } if (bvecpos > 0) { /* * Commit data on the server, as required. * If all bufs are using the same wcred, then use that with * one call for all of them, otherwise commit each one * separately. */ if (wcred != NOCRED) retv = nfs_commit(vp, off, (int)(endoff - off), wcred, td); else { retv = 0; for (i = 0; i < bvecpos; i++) { off_t off, size; bp = bvec[i]; off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; size = (u_quad_t)(bp->b_dirtyend - bp->b_dirtyoff); retv = nfs_commit(vp, off, (int)size, bp->b_wcred, td); if (retv) break; } } if (retv == NFSERR_STALEWRITEVERF) nfs_clearcommit(vp->v_mount); /* * Now, either mark the blocks I/O done or mark the * blocks dirty, depending on whether the commit * succeeded. */ for (i = 0; i < bvecpos; i++) { bp = bvec[i]; bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); if (retv) { /* * Error, leave B_DELWRI intact */ vfs_unbusy_pages(bp); brelse(bp); } else { /* * Success, remove B_DELWRI ( bundirty() ). * * b_dirtyoff/b_dirtyend seem to be NFS * specific. We should probably move that * into bundirty(). XXX */ bufobj_wref(bo); bp->b_flags |= B_ASYNC; bundirty(bp); bp->b_flags &= ~B_DONE; bp->b_ioflags &= ~BIO_ERROR; bp->b_dirtyoff = bp->b_dirtyend = 0; bufdone(bp); } } } /* * Start/do any write(s) that are required. */ loop: BO_LOCK(bo); TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { if (waitfor != MNT_WAIT || passone) continue; error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo), "nfsfsync", slpflag, slptimeo); if (error == 0) { BUF_UNLOCK(bp); goto loop; } if (error == ENOLCK) { error = 0; goto loop; } if (nfs_sigintr(nmp, td)) { error = EINTR; goto done; } if (slpflag == PCATCH) { slpflag = 0; slptimeo = 2 * hz; } goto loop; } if ((bp->b_flags & B_DELWRI) == 0) panic("nfs_fsync: not dirty"); if ((passone || !commit) && (bp->b_flags & B_NEEDCOMMIT)) { BUF_UNLOCK(bp); continue; } BO_UNLOCK(bo); bremfree(bp); if (passone || !commit) bp->b_flags |= B_ASYNC; else bp->b_flags |= B_ASYNC; bwrite(bp); if (nfs_sigintr(nmp, td)) { error = EINTR; goto done; } goto loop; } if (passone) { passone = 0; BO_UNLOCK(bo); goto again; } if (waitfor == MNT_WAIT) { while (bo->bo_numoutput) { error = bufobj_wwait(bo, slpflag, slptimeo); if (error) { BO_UNLOCK(bo); error = nfs_sigintr(nmp, td); if (error) goto done; if (slpflag == PCATCH) { slpflag = 0; slptimeo = 2 * hz; } BO_LOCK(bo); } } if (bo->bo_dirty.bv_cnt != 0 && commit) { BO_UNLOCK(bo); goto loop; } /* * Wait for all the async IO requests to drain */ BO_UNLOCK(bo); mtx_lock(&np->n_mtx); while (np->n_directio_asyncwr > 0) { np->n_flag |= NFSYNCWAIT; error = nfs_msleep(td, (caddr_t)&np->n_directio_asyncwr, &np->n_mtx, slpflag | (PRIBIO + 1), "nfsfsync", 0); if (error) { if (nfs_sigintr(nmp, td)) { mtx_unlock(&np->n_mtx); error = EINTR; goto done; } } } mtx_unlock(&np->n_mtx); } else BO_UNLOCK(bo); mtx_lock(&np->n_mtx); if (np->n_flag & NWRITEERR) { error = np->n_error; np->n_flag &= ~NWRITEERR; } if (commit && bo->bo_dirty.bv_cnt == 0 && bo->bo_numoutput == 0 && np->n_directio_asyncwr == 0) np->n_flag &= ~NMODIFIED; mtx_unlock(&np->n_mtx); done: if (bvec != NULL && bvec != bvec_on_stack) free(bvec, M_TEMP); return (error); } /* * NFS advisory byte-level locks. */ static int nfs_advlock(struct vop_advlock_args *ap) { struct vnode *vp = ap->a_vp; u_quad_t size; int error; error = vn_lock(vp, LK_SHARED); if (error) return (error); if ((VFSTONFS(vp->v_mount)->nm_flag & NFSMNT_NOLOCKD) != 0) { size = VTONFS(vp)->n_size; VOP_UNLOCK(vp, 0); error = lf_advlock(ap, &(vp->v_lockf), size); } else { if (nfs_advlock_p) error = nfs_advlock_p(ap); else error = ENOLCK; } return (error); } /* * NFS advisory byte-level locks. */ static int nfs_advlockasync(struct vop_advlockasync_args *ap) { struct vnode *vp = ap->a_vp; u_quad_t size; int error; error = vn_lock(vp, LK_SHARED); if (error) return (error); if ((VFSTONFS(vp->v_mount)->nm_flag & NFSMNT_NOLOCKD) != 0) { size = VTONFS(vp)->n_size; VOP_UNLOCK(vp, 0); error = lf_advlockasync(ap, &(vp->v_lockf), size); } else { VOP_UNLOCK(vp, 0); error = EOPNOTSUPP; } return (error); } /* * Print out the contents of an nfsnode. */ static int nfs_print(struct vop_print_args *ap) { struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); nfs_printf("\tfileid %ld fsid 0x%x", np->n_vattr.va_fileid, np->n_vattr.va_fsid); if (vp->v_type == VFIFO) fifo_printinfo(vp); printf("\n"); return (0); } /* * This is the "real" nfs::bwrite(struct buf*). * We set B_CACHE if this is a VMIO buffer. */ int nfs_writebp(struct buf *bp, int force __unused, struct thread *td) { int s; int oldflags = bp->b_flags; #if 0 int retv = 1; off_t off; #endif BUF_ASSERT_HELD(bp); if (bp->b_flags & B_INVAL) { brelse(bp); return(0); } bp->b_flags |= B_CACHE; /* * Undirty the bp. We will redirty it later if the I/O fails. */ s = splbio(); bundirty(bp); bp->b_flags &= ~B_DONE; bp->b_ioflags &= ~BIO_ERROR; bp->b_iocmd = BIO_WRITE; bufobj_wref(bp->b_bufobj); curthread->td_ru.ru_oublock++; splx(s); /* * Note: to avoid loopback deadlocks, we do not * assign b_runningbufspace. */ vfs_busy_pages(bp, 1); BUF_KERNPROC(bp); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); if( (oldflags & B_ASYNC) == 0) { int rtval = bufwait(bp); if (oldflags & B_DELWRI) { s = splbio(); reassignbuf(bp); splx(s); } brelse(bp); return (rtval); } return (0); } /* * nfs special file access vnode op. * Essentially just get vattr and then imitate iaccess() since the device is * local to the client. */ static int nfsspec_access(struct vop_access_args *ap) { struct vattr *vap; struct ucred *cred = ap->a_cred; struct vnode *vp = ap->a_vp; accmode_t accmode = ap->a_accmode; struct vattr vattr; int error; /* * Disallow write attempts on filesystems mounted read-only; * unless the file is a socket, fifo, or a block or character * device resident on the filesystem. */ if ((accmode & VWRITE) && (vp->v_mount->mnt_flag & MNT_RDONLY)) { switch (vp->v_type) { case VREG: case VDIR: case VLNK: return (EROFS); default: break; } } vap = &vattr; error = VOP_GETATTR(vp, vap, cred); if (error) goto out; error = vaccess(vp->v_type, vap->va_mode, vap->va_uid, vap->va_gid, accmode, cred, NULL); out: return error; } /* * Read wrapper for fifos. */ static int nfsfifo_read(struct vop_read_args *ap) { struct nfsnode *np = VTONFS(ap->a_vp); int error; /* * Set access flag. */ mtx_lock(&np->n_mtx); np->n_flag |= NACC; getnanotime(&np->n_atim); mtx_unlock(&np->n_mtx); error = fifo_specops.vop_read(ap); return error; } /* * Write wrapper for fifos. */ static int nfsfifo_write(struct vop_write_args *ap) { struct nfsnode *np = VTONFS(ap->a_vp); /* * Set update flag. */ mtx_lock(&np->n_mtx); np->n_flag |= NUPD; getnanotime(&np->n_mtim); mtx_unlock(&np->n_mtx); return(fifo_specops.vop_write(ap)); } /* * Close wrapper for fifos. * * Update the times on the nfsnode then do fifo close. */ static int nfsfifo_close(struct vop_close_args *ap) { struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); struct vattr vattr; struct timespec ts; mtx_lock(&np->n_mtx); if (np->n_flag & (NACC | NUPD)) { getnanotime(&ts); if (np->n_flag & NACC) np->n_atim = ts; if (np->n_flag & NUPD) np->n_mtim = ts; np->n_flag |= NCHG; if (vrefcnt(vp) == 1 && (vp->v_mount->mnt_flag & MNT_RDONLY) == 0) { VATTR_NULL(&vattr); if (np->n_flag & NACC) vattr.va_atime = np->n_atim; if (np->n_flag & NUPD) vattr.va_mtime = np->n_mtim; mtx_unlock(&np->n_mtx); (void)VOP_SETATTR(vp, &vattr, ap->a_cred); goto out; } } mtx_unlock(&np->n_mtx); out: return (fifo_specops.vop_close(ap)); } /* * Just call nfs_writebp() with the force argument set to 1. * * NOTE: B_DONE may or may not be set in a_bp on call. */ static int nfs_bwrite(struct buf *bp) { return (nfs_writebp(bp, 1, curthread)); } struct buf_ops buf_ops_nfs = { .bop_name = "buf_ops_nfs", .bop_write = nfs_bwrite, .bop_strategy = bufstrategy, .bop_sync = bufsync, .bop_bdflush = bufbdflush, }; diff --git a/sys/ufs/ffs/ffs_softdep.c b/sys/ufs/ffs/ffs_softdep.c index d7b8818721fc..72522b2594e5 100644 --- a/sys/ufs/ffs/ffs_softdep.c +++ b/sys/ufs/ffs/ffs_softdep.c @@ -1,6384 +1,6386 @@ /*- * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. * * The soft updates code is derived from the appendix of a University * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, * "Soft Updates: A Solution to the Metadata Update Problem in File * Systems", CSE-TR-254-95, August 1995). * * Further information about soft updates can be obtained from: * * Marshall Kirk McKusick http://www.mckusick.com/softdep/ * 1614 Oxford Street mckusick@mckusick.com * Berkeley, CA 94709-1608 +1-510-843-9542 * USA * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 */ #include __FBSDID("$FreeBSD$"); #include "opt_ffs.h" #include "opt_ddb.h" /* * For now we want the safety net that the DEBUG flag provides. */ #ifndef DEBUG #define DEBUG #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifndef SOFTUPDATES int softdep_flushfiles(oldmnt, flags, td) struct mount *oldmnt; int flags; struct thread *td; { panic("softdep_flushfiles called"); } int softdep_mount(devvp, mp, fs, cred) struct vnode *devvp; struct mount *mp; struct fs *fs; struct ucred *cred; { return (0); } void softdep_initialize() { return; } void softdep_uninitialize() { return; } void softdep_setup_inomapdep(bp, ip, newinum) struct buf *bp; struct inode *ip; ino_t newinum; { panic("softdep_setup_inomapdep called"); } void softdep_setup_blkmapdep(bp, mp, newblkno) struct buf *bp; struct mount *mp; ufs2_daddr_t newblkno; { panic("softdep_setup_blkmapdep called"); } void softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) struct inode *ip; ufs_lbn_t lbn; ufs2_daddr_t newblkno; ufs2_daddr_t oldblkno; long newsize; long oldsize; struct buf *bp; { panic("softdep_setup_allocdirect called"); } void softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) struct inode *ip; ufs_lbn_t lbn; ufs2_daddr_t newblkno; ufs2_daddr_t oldblkno; long newsize; long oldsize; struct buf *bp; { panic("softdep_setup_allocext called"); } void softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) struct inode *ip; ufs_lbn_t lbn; struct buf *bp; int ptrno; ufs2_daddr_t newblkno; ufs2_daddr_t oldblkno; struct buf *nbp; { panic("softdep_setup_allocindir_page called"); } void softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) struct buf *nbp; struct inode *ip; struct buf *bp; int ptrno; ufs2_daddr_t newblkno; { panic("softdep_setup_allocindir_meta called"); } void softdep_setup_freeblocks(ip, length, flags) struct inode *ip; off_t length; int flags; { panic("softdep_setup_freeblocks called"); } void softdep_freefile(pvp, ino, mode) struct vnode *pvp; ino_t ino; int mode; { panic("softdep_freefile called"); } int softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) struct buf *bp; struct inode *dp; off_t diroffset; ino_t newinum; struct buf *newdirbp; int isnewblk; { panic("softdep_setup_directory_add called"); } void softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) struct inode *dp; caddr_t base; caddr_t oldloc; caddr_t newloc; int entrysize; { panic("softdep_change_directoryentry_offset called"); } void softdep_setup_remove(bp, dp, ip, isrmdir) struct buf *bp; struct inode *dp; struct inode *ip; int isrmdir; { panic("softdep_setup_remove called"); } void softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) struct buf *bp; struct inode *dp; struct inode *ip; ino_t newinum; int isrmdir; { panic("softdep_setup_directory_change called"); } void softdep_change_linkcnt(ip) struct inode *ip; { panic("softdep_change_linkcnt called"); } void softdep_load_inodeblock(ip) struct inode *ip; { panic("softdep_load_inodeblock called"); } void softdep_update_inodeblock(ip, bp, waitfor) struct inode *ip; struct buf *bp; int waitfor; { panic("softdep_update_inodeblock called"); } int softdep_fsync(vp) struct vnode *vp; /* the "in_core" copy of the inode */ { return (0); } void softdep_fsync_mountdev(vp) struct vnode *vp; { return; } int softdep_flushworklist(oldmnt, countp, td) struct mount *oldmnt; int *countp; struct thread *td; { *countp = 0; return (0); } int softdep_sync_metadata(struct vnode *vp) { return (0); } int softdep_slowdown(vp) struct vnode *vp; { panic("softdep_slowdown called"); } void softdep_releasefile(ip) struct inode *ip; /* inode with the zero effective link count */ { panic("softdep_releasefile called"); } int softdep_request_cleanup(fs, vp) struct fs *fs; struct vnode *vp; { return (0); } int softdep_check_suspend(struct mount *mp, struct vnode *devvp, int softdep_deps, int softdep_accdeps, int secondary_writes, int secondary_accwrites) { struct bufobj *bo; int error; (void) softdep_deps, (void) softdep_accdeps; bo = &devvp->v_bufobj; ASSERT_BO_LOCKED(bo); MNT_ILOCK(mp); while (mp->mnt_secondary_writes != 0) { BO_UNLOCK(bo); msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), (PUSER - 1) | PDROP, "secwr", 0); BO_LOCK(bo); MNT_ILOCK(mp); } /* * Reasons for needing more work before suspend: * - Dirty buffers on devvp. * - Secondary writes occurred after start of vnode sync loop */ error = 0; if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0 || secondary_writes != 0 || mp->mnt_secondary_writes != 0 || secondary_accwrites != mp->mnt_secondary_accwrites) error = EAGAIN; BO_UNLOCK(bo); return (error); } void softdep_get_depcounts(struct mount *mp, int *softdepactivep, int *softdepactiveaccp) { (void) mp; *softdepactivep = 0; *softdepactiveaccp = 0; } #else /* * These definitions need to be adapted to the system to which * this file is being ported. */ /* * malloc types defined for the softdep system. */ static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block"); static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes"); #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) #define D_PAGEDEP 0 #define D_INODEDEP 1 #define D_NEWBLK 2 #define D_BMSAFEMAP 3 #define D_ALLOCDIRECT 4 #define D_INDIRDEP 5 #define D_ALLOCINDIR 6 #define D_FREEFRAG 7 #define D_FREEBLKS 8 #define D_FREEFILE 9 #define D_DIRADD 10 #define D_MKDIR 11 #define D_DIRREM 12 #define D_NEWDIRBLK 13 #define D_LAST D_NEWDIRBLK /* * translate from workitem type to memory type * MUST match the defines above, such that memtype[D_XXX] == M_XXX */ static struct malloc_type *memtype[] = { M_PAGEDEP, M_INODEDEP, M_NEWBLK, M_BMSAFEMAP, M_ALLOCDIRECT, M_INDIRDEP, M_ALLOCINDIR, M_FREEFRAG, M_FREEBLKS, M_FREEFILE, M_DIRADD, M_MKDIR, M_DIRREM, M_NEWDIRBLK }; #define DtoM(type) (memtype[type]) /* * Names of malloc types. */ #define TYPENAME(type) \ ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") /* * End system adaptation definitions. */ /* * Forward declarations. */ struct inodedep_hashhead; struct newblk_hashhead; struct pagedep_hashhead; /* * Internal function prototypes. */ static void softdep_error(char *, int); static void drain_output(struct vnode *); static struct buf *getdirtybuf(struct buf *, struct mtx *, int); static void clear_remove(struct thread *); static void clear_inodedeps(struct thread *); static int flush_pagedep_deps(struct vnode *, struct mount *, struct diraddhd *); static int flush_inodedep_deps(struct mount *, ino_t); static int flush_deplist(struct allocdirectlst *, int, int *); static int handle_written_filepage(struct pagedep *, struct buf *); static void diradd_inode_written(struct diradd *, struct inodedep *); static int handle_written_inodeblock(struct inodedep *, struct buf *); static void handle_allocdirect_partdone(struct allocdirect *); static void handle_allocindir_partdone(struct allocindir *); static void initiate_write_filepage(struct pagedep *, struct buf *); static void handle_written_mkdir(struct mkdir *, int); static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); static void handle_workitem_freefile(struct freefile *); static void handle_workitem_remove(struct dirrem *, struct vnode *); static struct dirrem *newdirrem(struct buf *, struct inode *, struct inode *, int, struct dirrem **); static void free_diradd(struct diradd *); static void free_allocindir(struct allocindir *, struct inodedep *); static void free_newdirblk(struct newdirblk *); static int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t, ufs2_daddr_t *); static void deallocate_dependencies(struct buf *, struct inodedep *); static void free_allocdirect(struct allocdirectlst *, struct allocdirect *, int); static int check_inode_unwritten(struct inodedep *); static int free_inodedep(struct inodedep *); static void handle_workitem_freeblocks(struct freeblks *, int); static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); static void setup_allocindir_phase2(struct buf *, struct inode *, struct allocindir *); static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, ufs2_daddr_t); static void handle_workitem_freefrag(struct freefrag *); static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long); static void allocdirect_merge(struct allocdirectlst *, struct allocdirect *, struct allocdirect *); static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *); static int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t, struct newblk **); static int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **); static int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t, struct inodedep **); static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); static int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **); static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, struct mount *mp, int, struct pagedep **); static void pause_timer(void *); static int request_cleanup(struct mount *, int); static int process_worklist_item(struct mount *, int); static void add_to_worklist(struct worklist *); static void softdep_flush(void); static int softdep_speedup(void); /* * Exported softdep operations. */ static void softdep_disk_io_initiation(struct buf *); static void softdep_disk_write_complete(struct buf *); static void softdep_deallocate_dependencies(struct buf *); static int softdep_count_dependencies(struct buf *bp, int); static struct mtx lk; MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF); #define TRY_ACQUIRE_LOCK(lk) mtx_trylock(lk) #define ACQUIRE_LOCK(lk) mtx_lock(lk) #define FREE_LOCK(lk) mtx_unlock(lk) #define BUF_AREC(bp) ((bp)->b_lock.lock_object.lo_flags |= LO_RECURSABLE) #define BUF_NOREC(bp) ((bp)->b_lock.lock_object.lo_flags &= ~LO_RECURSABLE) /* * Worklist queue management. * These routines require that the lock be held. */ #ifndef /* NOT */ DEBUG #define WORKLIST_INSERT(head, item) do { \ (item)->wk_state |= ONWORKLIST; \ LIST_INSERT_HEAD(head, item, wk_list); \ } while (0) #define WORKLIST_REMOVE(item) do { \ (item)->wk_state &= ~ONWORKLIST; \ LIST_REMOVE(item, wk_list); \ } while (0) #else /* DEBUG */ static void worklist_insert(struct workhead *, struct worklist *); static void worklist_remove(struct worklist *); #define WORKLIST_INSERT(head, item) worklist_insert(head, item) #define WORKLIST_REMOVE(item) worklist_remove(item) static void worklist_insert(head, item) struct workhead *head; struct worklist *item; { mtx_assert(&lk, MA_OWNED); if (item->wk_state & ONWORKLIST) panic("worklist_insert: already on list"); item->wk_state |= ONWORKLIST; LIST_INSERT_HEAD(head, item, wk_list); } static void worklist_remove(item) struct worklist *item; { mtx_assert(&lk, MA_OWNED); if ((item->wk_state & ONWORKLIST) == 0) panic("worklist_remove: not on list"); item->wk_state &= ~ONWORKLIST; LIST_REMOVE(item, wk_list); } #endif /* DEBUG */ /* * Routines for tracking and managing workitems. */ static void workitem_free(struct worklist *, int); static void workitem_alloc(struct worklist *, int, struct mount *); #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type)) static void workitem_free(item, type) struct worklist *item; int type; { struct ufsmount *ump; mtx_assert(&lk, MA_OWNED); #ifdef DEBUG if (item->wk_state & ONWORKLIST) panic("workitem_free: still on list"); if (item->wk_type != type) panic("workitem_free: type mismatch"); #endif ump = VFSTOUFS(item->wk_mp); if (--ump->softdep_deps == 0 && ump->softdep_req) wakeup(&ump->softdep_deps); free(item, DtoM(type)); } static void workitem_alloc(item, type, mp) struct worklist *item; int type; struct mount *mp; { item->wk_type = type; item->wk_mp = mp; item->wk_state = 0; ACQUIRE_LOCK(&lk); VFSTOUFS(mp)->softdep_deps++; VFSTOUFS(mp)->softdep_accdeps++; FREE_LOCK(&lk); } /* * Workitem queue management */ static int max_softdeps; /* maximum number of structs before slowdown */ static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ static int tickdelay = 2; /* number of ticks to pause during slowdown */ static int proc_waiting; /* tracks whether we have a timeout posted */ static int *stat_countp; /* statistic to count in proc_waiting timeout */ static struct callout softdep_callout; static int req_pending; static int req_clear_inodedeps; /* syncer process flush some inodedeps */ #define FLUSH_INODES 1 static int req_clear_remove; /* syncer process flush some freeblks */ #define FLUSH_REMOVE 2 #define FLUSH_REMOVE_WAIT 3 /* * runtime statistics */ static int stat_worklist_push; /* number of worklist cleanups */ static int stat_blk_limit_push; /* number of times block limit neared */ static int stat_ino_limit_push; /* number of times inode limit neared */ static int stat_blk_limit_hit; /* number of times block slowdown imposed */ static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, ""); SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); /* SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &softdep_on_worklist, 0, ""); */ SYSCTL_DECL(_vfs_ffs); static int compute_summary_at_mount = 0; /* Whether to recompute the summary at mount time */ SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, &compute_summary_at_mount, 0, "Recompute summary at mount"); static struct proc *softdepproc; static struct kproc_desc softdep_kp = { "softdepflush", softdep_flush, &softdepproc }; SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, &softdep_kp); static void softdep_flush(void) { struct mount *nmp; struct mount *mp; struct ufsmount *ump; struct thread *td; int remaining; int vfslocked; td = curthread; td->td_pflags |= TDP_NORUNNINGBUF; for (;;) { kproc_suspend_check(softdepproc); vfslocked = VFS_LOCK_GIANT((struct mount *)NULL); ACQUIRE_LOCK(&lk); /* * If requested, try removing inode or removal dependencies. */ if (req_clear_inodedeps) { clear_inodedeps(td); req_clear_inodedeps -= 1; wakeup_one(&proc_waiting); } if (req_clear_remove) { clear_remove(td); req_clear_remove -= 1; wakeup_one(&proc_waiting); } FREE_LOCK(&lk); VFS_UNLOCK_GIANT(vfslocked); remaining = 0; mtx_lock(&mountlist_mtx); for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { nmp = TAILQ_NEXT(mp, mnt_list); if ((mp->mnt_flag & MNT_SOFTDEP) == 0) continue; if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) continue; vfslocked = VFS_LOCK_GIANT(mp); softdep_process_worklist(mp, 0); ump = VFSTOUFS(mp); remaining += ump->softdep_on_worklist - ump->softdep_on_worklist_inprogress; VFS_UNLOCK_GIANT(vfslocked); mtx_lock(&mountlist_mtx); nmp = TAILQ_NEXT(mp, mnt_list); vfs_unbusy(mp); } mtx_unlock(&mountlist_mtx); if (remaining) continue; ACQUIRE_LOCK(&lk); if (!req_pending) msleep(&req_pending, &lk, PVM, "sdflush", hz); req_pending = 0; FREE_LOCK(&lk); } } static int softdep_speedup(void) { mtx_assert(&lk, MA_OWNED); if (req_pending == 0) { req_pending = 1; wakeup(&req_pending); } return speedup_syncer(); } /* * Add an item to the end of the work queue. * This routine requires that the lock be held. * This is the only routine that adds items to the list. * The following routine is the only one that removes items * and does so in order from first to last. */ static void add_to_worklist(wk) struct worklist *wk; { struct ufsmount *ump; mtx_assert(&lk, MA_OWNED); ump = VFSTOUFS(wk->wk_mp); if (wk->wk_state & ONWORKLIST) panic("add_to_worklist: already on list"); wk->wk_state |= ONWORKLIST; if (LIST_EMPTY(&ump->softdep_workitem_pending)) LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); else LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); ump->softdep_worklist_tail = wk; ump->softdep_on_worklist += 1; } /* * Process that runs once per second to handle items in the background queue. * * Note that we ensure that everything is done in the order in which they * appear in the queue. The code below depends on this property to ensure * that blocks of a file are freed before the inode itself is freed. This * ordering ensures that no new triples will be generated * until all the old ones have been purged from the dependency lists. */ int softdep_process_worklist(mp, full) struct mount *mp; int full; { struct thread *td = curthread; int cnt, matchcnt, loopcount; struct ufsmount *ump; long starttime; KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); /* * Record the process identifier of our caller so that we can give * this process preferential treatment in request_cleanup below. */ matchcnt = 0; ump = VFSTOUFS(mp); ACQUIRE_LOCK(&lk); loopcount = 1; starttime = time_second; while (ump->softdep_on_worklist > 0) { if ((cnt = process_worklist_item(mp, 0)) == -1) break; else matchcnt += cnt; /* * If requested, try removing inode or removal dependencies. */ if (req_clear_inodedeps) { clear_inodedeps(td); req_clear_inodedeps -= 1; wakeup_one(&proc_waiting); } if (req_clear_remove) { clear_remove(td); req_clear_remove -= 1; wakeup_one(&proc_waiting); } /* * We do not generally want to stop for buffer space, but if * we are really being a buffer hog, we will stop and wait. */ if (loopcount++ % 128 == 0) { FREE_LOCK(&lk); uio_yield(); bwillwrite(); ACQUIRE_LOCK(&lk); } /* * Never allow processing to run for more than one * second. Otherwise the other mountpoints may get * excessively backlogged. */ if (!full && starttime != time_second) { matchcnt = -1; break; } } FREE_LOCK(&lk); return (matchcnt); } /* * Process one item on the worklist. */ static int process_worklist_item(mp, flags) struct mount *mp; int flags; { struct worklist *wk, *wkend; struct ufsmount *ump; struct vnode *vp; int matchcnt = 0; mtx_assert(&lk, MA_OWNED); KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); /* * If we are being called because of a process doing a * copy-on-write, then it is not safe to write as we may * recurse into the copy-on-write routine. */ if (curthread->td_pflags & TDP_COWINPROGRESS) return (-1); /* * Normally we just process each item on the worklist in order. * However, if we are in a situation where we cannot lock any * inodes, we have to skip over any dirrem requests whose * vnodes are resident and locked. */ ump = VFSTOUFS(mp); vp = NULL; LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) { if (wk->wk_state & INPROGRESS) continue; if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) break; wk->wk_state |= INPROGRESS; ump->softdep_on_worklist_inprogress++; FREE_LOCK(&lk); ffs_vgetf(mp, WK_DIRREM(wk)->dm_oldinum, LK_NOWAIT | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ); ACQUIRE_LOCK(&lk); wk->wk_state &= ~INPROGRESS; ump->softdep_on_worklist_inprogress--; if (vp != NULL) break; } if (wk == 0) return (-1); /* * Remove the item to be processed. If we are removing the last * item on the list, we need to recalculate the tail pointer. * As this happens rarely and usually when the list is short, * we just run down the list to find it rather than tracking it * in the above loop. */ WORKLIST_REMOVE(wk); if (wk == ump->softdep_worklist_tail) { LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list) if (LIST_NEXT(wkend, wk_list) == NULL) break; ump->softdep_worklist_tail = wkend; } ump->softdep_on_worklist -= 1; FREE_LOCK(&lk); if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) panic("process_worklist_item: suspended filesystem"); matchcnt++; switch (wk->wk_type) { case D_DIRREM: /* removal of a directory entry */ handle_workitem_remove(WK_DIRREM(wk), vp); break; case D_FREEBLKS: /* releasing blocks and/or fragments from a file */ handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT); break; case D_FREEFRAG: /* releasing a fragment when replaced as a file grows */ handle_workitem_freefrag(WK_FREEFRAG(wk)); break; case D_FREEFILE: /* releasing an inode when its link count drops to 0 */ handle_workitem_freefile(WK_FREEFILE(wk)); break; default: panic("%s_process_worklist: Unknown type %s", "softdep", TYPENAME(wk->wk_type)); /* NOTREACHED */ } vn_finished_secondary_write(mp); ACQUIRE_LOCK(&lk); return (matchcnt); } /* * Move dependencies from one buffer to another. */ void softdep_move_dependencies(oldbp, newbp) struct buf *oldbp; struct buf *newbp; { struct worklist *wk, *wktail; if (!LIST_EMPTY(&newbp->b_dep)) panic("softdep_move_dependencies: need merge code"); wktail = 0; ACQUIRE_LOCK(&lk); while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { LIST_REMOVE(wk, wk_list); if (wktail == 0) LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); else LIST_INSERT_AFTER(wktail, wk, wk_list); wktail = wk; } FREE_LOCK(&lk); } /* * Purge the work list of all items associated with a particular mount point. */ int softdep_flushworklist(oldmnt, countp, td) struct mount *oldmnt; int *countp; struct thread *td; { struct vnode *devvp; int count, error = 0; struct ufsmount *ump; /* * Alternately flush the block device associated with the mount * point and process any dependencies that the flushing * creates. We continue until no more worklist dependencies * are found. */ *countp = 0; ump = VFSTOUFS(oldmnt); devvp = ump->um_devvp; while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { *countp += count; vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); error = VOP_FSYNC(devvp, MNT_WAIT, td); VOP_UNLOCK(devvp, 0); if (error) break; } return (error); } int softdep_waitidle(struct mount *mp) { struct ufsmount *ump; int error; int i; ump = VFSTOUFS(mp); ACQUIRE_LOCK(&lk); for (i = 0; i < 10 && ump->softdep_deps; i++) { ump->softdep_req = 1; if (ump->softdep_on_worklist) panic("softdep_waitidle: work added after flush."); msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1); } ump->softdep_req = 0; FREE_LOCK(&lk); error = 0; if (i == 10) { error = EBUSY; printf("softdep_waitidle: Failed to flush worklist for %p\n", mp); } return (error); } /* * Flush all vnodes and worklist items associated with a specified mount point. */ int softdep_flushfiles(oldmnt, flags, td) struct mount *oldmnt; int flags; struct thread *td; { int error, depcount, loopcnt, retry_flush_count, retry; loopcnt = 10; retry_flush_count = 3; retry_flush: error = 0; /* * Alternately flush the vnodes associated with the mount * point and process any dependencies that the flushing * creates. In theory, this loop can happen at most twice, * but we give it a few extra just to be sure. */ for (; loopcnt > 0; loopcnt--) { /* * Do another flush in case any vnodes were brought in * as part of the cleanup operations. */ if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) break; if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || depcount == 0) break; } /* * If we are unmounting then it is an error to fail. If we * are simply trying to downgrade to read-only, then filesystem * activity can keep us busy forever, so we just fail with EBUSY. */ if (loopcnt == 0) { if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) panic("softdep_flushfiles: looping"); error = EBUSY; } if (!error) error = softdep_waitidle(oldmnt); if (!error) { if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { retry = 0; MNT_ILOCK(oldmnt); KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, ("softdep_flushfiles: !MNTK_NOINSMNTQ")); if (oldmnt->mnt_nvnodelistsize > 0) { if (--retry_flush_count > 0) { retry = 1; loopcnt = 3; } else error = EBUSY; } MNT_IUNLOCK(oldmnt); if (retry) goto retry_flush; } } return (error); } /* * Structure hashing. * * There are three types of structures that can be looked up: * 1) pagedep structures identified by mount point, inode number, * and logical block. * 2) inodedep structures identified by mount point and inode number. * 3) newblk structures identified by mount point and * physical block number. * * The "pagedep" and "inodedep" dependency structures are hashed * separately from the file blocks and inodes to which they correspond. * This separation helps when the in-memory copy of an inode or * file block must be replaced. It also obviates the need to access * an inode or file page when simply updating (or de-allocating) * dependency structures. Lookup of newblk structures is needed to * find newly allocated blocks when trying to associate them with * their allocdirect or allocindir structure. * * The lookup routines optionally create and hash a new instance when * an existing entry is not found. */ #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ #define NODELAY 0x0002 /* cannot do background work */ /* * Structures and routines associated with pagedep caching. */ LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; u_long pagedep_hash; /* size of hash table - 1 */ #define PAGEDEP_HASH(mp, inum, lbn) \ (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ pagedep_hash]) static int pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp) struct pagedep_hashhead *pagedephd; ino_t ino; ufs_lbn_t lbn; struct mount *mp; int flags; struct pagedep **pagedeppp; { struct pagedep *pagedep; LIST_FOREACH(pagedep, pagedephd, pd_hash) if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn && mp == pagedep->pd_list.wk_mp) break; if (pagedep) { *pagedeppp = pagedep; if ((flags & DEPALLOC) != 0 && (pagedep->pd_state & ONWORKLIST) == 0) return (0); return (1); } *pagedeppp = NULL; return (0); } /* * Look up a pagedep. Return 1 if found, 0 if not found or found * when asked to allocate but not associated with any buffer. * If not found, allocate if DEPALLOC flag is passed. * Found or allocated entry is returned in pagedeppp. * This routine must be called with splbio interrupts blocked. */ static int pagedep_lookup(ip, lbn, flags, pagedeppp) struct inode *ip; ufs_lbn_t lbn; int flags; struct pagedep **pagedeppp; { struct pagedep *pagedep; struct pagedep_hashhead *pagedephd; struct mount *mp; int ret; int i; mtx_assert(&lk, MA_OWNED); mp = ITOV(ip)->v_mount; pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); if (*pagedeppp || (flags & DEPALLOC) == 0) return (ret); FREE_LOCK(&lk); pagedep = malloc(sizeof(struct pagedep), M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); ACQUIRE_LOCK(&lk); ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); if (*pagedeppp) { WORKITEM_FREE(pagedep, D_PAGEDEP); return (ret); } pagedep->pd_ino = ip->i_number; pagedep->pd_lbn = lbn; LIST_INIT(&pagedep->pd_dirremhd); LIST_INIT(&pagedep->pd_pendinghd); for (i = 0; i < DAHASHSZ; i++) LIST_INIT(&pagedep->pd_diraddhd[i]); LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); *pagedeppp = pagedep; return (0); } /* * Structures and routines associated with inodedep caching. */ LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; static u_long inodedep_hash; /* size of hash table - 1 */ static long num_inodedep; /* number of inodedep allocated */ #define INODEDEP_HASH(fs, inum) \ (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) static int inodedep_find(inodedephd, fs, inum, inodedeppp) struct inodedep_hashhead *inodedephd; struct fs *fs; ino_t inum; struct inodedep **inodedeppp; { struct inodedep *inodedep; LIST_FOREACH(inodedep, inodedephd, id_hash) if (inum == inodedep->id_ino && fs == inodedep->id_fs) break; if (inodedep) { *inodedeppp = inodedep; return (1); } *inodedeppp = NULL; return (0); } /* * Look up an inodedep. Return 1 if found, 0 if not found. * If not found, allocate if DEPALLOC flag is passed. * Found or allocated entry is returned in inodedeppp. * This routine must be called with splbio interrupts blocked. */ static int inodedep_lookup(mp, inum, flags, inodedeppp) struct mount *mp; ino_t inum; int flags; struct inodedep **inodedeppp; { struct inodedep *inodedep; struct inodedep_hashhead *inodedephd; struct fs *fs; mtx_assert(&lk, MA_OWNED); fs = VFSTOUFS(mp)->um_fs; inodedephd = INODEDEP_HASH(fs, inum); if (inodedep_find(inodedephd, fs, inum, inodedeppp)) return (1); if ((flags & DEPALLOC) == 0) return (0); /* * If we are over our limit, try to improve the situation. */ if (num_inodedep > max_softdeps && (flags & NODELAY) == 0) request_cleanup(mp, FLUSH_INODES); FREE_LOCK(&lk); inodedep = malloc(sizeof(struct inodedep), M_INODEDEP, M_SOFTDEP_FLAGS); workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); ACQUIRE_LOCK(&lk); if (inodedep_find(inodedephd, fs, inum, inodedeppp)) { WORKITEM_FREE(inodedep, D_INODEDEP); return (1); } num_inodedep += 1; inodedep->id_fs = fs; inodedep->id_ino = inum; inodedep->id_state = ALLCOMPLETE; inodedep->id_nlinkdelta = 0; inodedep->id_savedino1 = NULL; inodedep->id_savedsize = -1; inodedep->id_savedextsize = -1; inodedep->id_buf = NULL; LIST_INIT(&inodedep->id_pendinghd); LIST_INIT(&inodedep->id_inowait); LIST_INIT(&inodedep->id_bufwait); TAILQ_INIT(&inodedep->id_inoupdt); TAILQ_INIT(&inodedep->id_newinoupdt); TAILQ_INIT(&inodedep->id_extupdt); TAILQ_INIT(&inodedep->id_newextupdt); LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); *inodedeppp = inodedep; return (0); } /* * Structures and routines associated with newblk caching. */ LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; u_long newblk_hash; /* size of hash table - 1 */ #define NEWBLK_HASH(fs, inum) \ (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) static int newblk_find(newblkhd, fs, newblkno, newblkpp) struct newblk_hashhead *newblkhd; struct fs *fs; ufs2_daddr_t newblkno; struct newblk **newblkpp; { struct newblk *newblk; LIST_FOREACH(newblk, newblkhd, nb_hash) if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) break; if (newblk) { *newblkpp = newblk; return (1); } *newblkpp = NULL; return (0); } /* * Look up a newblk. Return 1 if found, 0 if not found. * If not found, allocate if DEPALLOC flag is passed. * Found or allocated entry is returned in newblkpp. */ static int newblk_lookup(fs, newblkno, flags, newblkpp) struct fs *fs; ufs2_daddr_t newblkno; int flags; struct newblk **newblkpp; { struct newblk *newblk; struct newblk_hashhead *newblkhd; newblkhd = NEWBLK_HASH(fs, newblkno); if (newblk_find(newblkhd, fs, newblkno, newblkpp)) return (1); if ((flags & DEPALLOC) == 0) return (0); FREE_LOCK(&lk); newblk = malloc(sizeof(struct newblk), M_NEWBLK, M_SOFTDEP_FLAGS); ACQUIRE_LOCK(&lk); if (newblk_find(newblkhd, fs, newblkno, newblkpp)) { free(newblk, M_NEWBLK); return (1); } newblk->nb_state = 0; newblk->nb_fs = fs; newblk->nb_newblkno = newblkno; LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); *newblkpp = newblk; return (0); } /* * Executed during filesystem system initialization before * mounting any filesystems. */ void softdep_initialize() { LIST_INIT(&mkdirlisthd); max_softdeps = desiredvnodes * 4; pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, &pagedep_hash); inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); /* initialise bioops hack */ bioops.io_start = softdep_disk_io_initiation; bioops.io_complete = softdep_disk_write_complete; bioops.io_deallocate = softdep_deallocate_dependencies; bioops.io_countdeps = softdep_count_dependencies; /* Initialize the callout with an mtx. */ callout_init_mtx(&softdep_callout, &lk, 0); } /* * Executed after all filesystems have been unmounted during * filesystem module unload. */ void softdep_uninitialize() { callout_drain(&softdep_callout); hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); } /* * Called at mount time to notify the dependency code that a * filesystem wishes to use it. */ int softdep_mount(devvp, mp, fs, cred) struct vnode *devvp; struct mount *mp; struct fs *fs; struct ucred *cred; { struct csum_total cstotal; struct ufsmount *ump; struct cg *cgp; struct buf *bp; int error, cyl; MNT_ILOCK(mp); mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | MNTK_SOFTDEP; mp->mnt_noasync++; } MNT_IUNLOCK(mp); ump = VFSTOUFS(mp); LIST_INIT(&ump->softdep_workitem_pending); ump->softdep_worklist_tail = NULL; ump->softdep_on_worklist = 0; ump->softdep_deps = 0; /* * When doing soft updates, the counters in the * superblock may have gotten out of sync. Recomputation * can take a long time and can be deferred for background * fsck. However, the old behavior of scanning the cylinder * groups and recalculating them at mount time is available * by setting vfs.ffs.compute_summary_at_mount to one. */ if (compute_summary_at_mount == 0 || fs->fs_clean != 0) return (0); bzero(&cstotal, sizeof cstotal); for (cyl = 0; cyl < fs->fs_ncg; cyl++) { if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), fs->fs_cgsize, cred, &bp)) != 0) { brelse(bp); return (error); } cgp = (struct cg *)bp->b_data; cstotal.cs_nffree += cgp->cg_cs.cs_nffree; cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; cstotal.cs_nifree += cgp->cg_cs.cs_nifree; cstotal.cs_ndir += cgp->cg_cs.cs_ndir; fs->fs_cs(fs, cyl) = cgp->cg_cs; brelse(bp); } #ifdef DEBUG if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); #endif bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); return (0); } /* * Protecting the freemaps (or bitmaps). * * To eliminate the need to execute fsck before mounting a filesystem * after a power failure, one must (conservatively) guarantee that the * on-disk copy of the bitmaps never indicate that a live inode or block is * free. So, when a block or inode is allocated, the bitmap should be * updated (on disk) before any new pointers. When a block or inode is * freed, the bitmap should not be updated until all pointers have been * reset. The latter dependency is handled by the delayed de-allocation * approach described below for block and inode de-allocation. The former * dependency is handled by calling the following procedure when a block or * inode is allocated. When an inode is allocated an "inodedep" is created * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. * Each "inodedep" is also inserted into the hash indexing structure so * that any additional link additions can be made dependent on the inode * allocation. * * The ufs filesystem maintains a number of free block counts (e.g., per * cylinder group, per cylinder and per pair) * in addition to the bitmaps. These counts are used to improve efficiency * during allocation and therefore must be consistent with the bitmaps. * There is no convenient way to guarantee post-crash consistency of these * counts with simple update ordering, for two main reasons: (1) The counts * and bitmaps for a single cylinder group block are not in the same disk * sector. If a disk write is interrupted (e.g., by power failure), one may * be written and the other not. (2) Some of the counts are located in the * superblock rather than the cylinder group block. So, we focus our soft * updates implementation on protecting the bitmaps. When mounting a * filesystem, we recompute the auxiliary counts from the bitmaps. */ /* * Called just after updating the cylinder group block to allocate an inode. */ void softdep_setup_inomapdep(bp, ip, newinum) struct buf *bp; /* buffer for cylgroup block with inode map */ struct inode *ip; /* inode related to allocation */ ino_t newinum; /* new inode number being allocated */ { struct inodedep *inodedep; struct bmsafemap *bmsafemap; /* * Create a dependency for the newly allocated inode. * Panic if it already exists as something is seriously wrong. * Otherwise add it to the dependency list for the buffer holding * the cylinder group map from which it was allocated. */ ACQUIRE_LOCK(&lk); if ((inodedep_lookup(UFSTOVFS(ip->i_ump), newinum, DEPALLOC|NODELAY, &inodedep))) panic("softdep_setup_inomapdep: dependency for new inode " "already exists"); inodedep->id_buf = bp; inodedep->id_state &= ~DEPCOMPLETE; bmsafemap = bmsafemap_lookup(inodedep->id_list.wk_mp, bp); LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); FREE_LOCK(&lk); } /* * Called just after updating the cylinder group block to * allocate block or fragment. */ void softdep_setup_blkmapdep(bp, mp, newblkno) struct buf *bp; /* buffer for cylgroup block with block map */ struct mount *mp; /* filesystem doing allocation */ ufs2_daddr_t newblkno; /* number of newly allocated block */ { struct newblk *newblk; struct bmsafemap *bmsafemap; struct fs *fs; fs = VFSTOUFS(mp)->um_fs; /* * Create a dependency for the newly allocated block. * Add it to the dependency list for the buffer holding * the cylinder group map from which it was allocated. */ ACQUIRE_LOCK(&lk); if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) panic("softdep_setup_blkmapdep: found block"); newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp); LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); FREE_LOCK(&lk); } /* * Find the bmsafemap associated with a cylinder group buffer. * If none exists, create one. The buffer must be locked when * this routine is called and this routine must be called with * splbio interrupts blocked. */ static struct bmsafemap * bmsafemap_lookup(mp, bp) struct mount *mp; struct buf *bp; { struct bmsafemap *bmsafemap; struct worklist *wk; mtx_assert(&lk, MA_OWNED); LIST_FOREACH(wk, &bp->b_dep, wk_list) if (wk->wk_type == D_BMSAFEMAP) return (WK_BMSAFEMAP(wk)); FREE_LOCK(&lk); bmsafemap = malloc(sizeof(struct bmsafemap), M_BMSAFEMAP, M_SOFTDEP_FLAGS); workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); bmsafemap->sm_buf = bp; LIST_INIT(&bmsafemap->sm_allocdirecthd); LIST_INIT(&bmsafemap->sm_allocindirhd); LIST_INIT(&bmsafemap->sm_inodedephd); LIST_INIT(&bmsafemap->sm_newblkhd); ACQUIRE_LOCK(&lk); WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); return (bmsafemap); } /* * Direct block allocation dependencies. * * When a new block is allocated, the corresponding disk locations must be * initialized (with zeros or new data) before the on-disk inode points to * them. Also, the freemap from which the block was allocated must be * updated (on disk) before the inode's pointer. These two dependencies are * independent of each other and are needed for all file blocks and indirect * blocks that are pointed to directly by the inode. Just before the * "in-core" version of the inode is updated with a newly allocated block * number, a procedure (below) is called to setup allocation dependency * structures. These structures are removed when the corresponding * dependencies are satisfied or when the block allocation becomes obsolete * (i.e., the file is deleted, the block is de-allocated, or the block is a * fragment that gets upgraded). All of these cases are handled in * procedures described later. * * When a file extension causes a fragment to be upgraded, either to a larger * fragment or to a full block, the on-disk location may change (if the * previous fragment could not simply be extended). In this case, the old * fragment must be de-allocated, but not until after the inode's pointer has * been updated. In most cases, this is handled by later procedures, which * will construct a "freefrag" structure to be added to the workitem queue * when the inode update is complete (or obsolete). The main exception to * this is when an allocation occurs while a pending allocation dependency * (for the same block pointer) remains. This case is handled in the main * allocation dependency setup procedure by immediately freeing the * unreferenced fragments. */ void softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) struct inode *ip; /* inode to which block is being added */ ufs_lbn_t lbn; /* block pointer within inode */ ufs2_daddr_t newblkno; /* disk block number being added */ ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ long newsize; /* size of new block */ long oldsize; /* size of new block */ struct buf *bp; /* bp for allocated block */ { struct allocdirect *adp, *oldadp; struct allocdirectlst *adphead; struct bmsafemap *bmsafemap; struct inodedep *inodedep; struct pagedep *pagedep; struct newblk *newblk; struct mount *mp; mp = UFSTOVFS(ip->i_ump); adp = malloc(sizeof(struct allocdirect), M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp); adp->ad_lbn = lbn; adp->ad_newblkno = newblkno; adp->ad_oldblkno = oldblkno; adp->ad_newsize = newsize; adp->ad_oldsize = oldsize; adp->ad_state = ATTACHED; LIST_INIT(&adp->ad_newdirblk); if (newblkno == oldblkno) adp->ad_freefrag = NULL; else adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); ACQUIRE_LOCK(&lk); if (lbn >= NDADDR) { /* allocating an indirect block */ if (oldblkno != 0) panic("softdep_setup_allocdirect: non-zero indir"); } else { /* * Allocating a direct block. * * If we are allocating a directory block, then we must * allocate an associated pagedep to track additions and * deletions. */ if ((ip->i_mode & IFMT) == IFDIR && pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); } if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) panic("softdep_setup_allocdirect: lost block"); if (newblk->nb_state == DEPCOMPLETE) { adp->ad_state |= DEPCOMPLETE; adp->ad_buf = NULL; } else { bmsafemap = newblk->nb_bmsafemap; adp->ad_buf = bmsafemap->sm_buf; LIST_REMOVE(newblk, nb_deps); LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); } LIST_REMOVE(newblk, nb_hash); free(newblk, M_NEWBLK); inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); adp->ad_inodedep = inodedep; WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); /* * The list of allocdirects must be kept in sorted and ascending * order so that the rollback routines can quickly determine the * first uncommitted block (the size of the file stored on disk * ends at the end of the lowest committed fragment, or if there * are no fragments, at the end of the highest committed block). * Since files generally grow, the typical case is that the new * block is to be added at the end of the list. We speed this * special case by checking against the last allocdirect in the * list before laboriously traversing the list looking for the * insertion point. */ adphead = &inodedep->id_newinoupdt; oldadp = TAILQ_LAST(adphead, allocdirectlst); if (oldadp == NULL || oldadp->ad_lbn <= lbn) { /* insert at end of list */ TAILQ_INSERT_TAIL(adphead, adp, ad_next); if (oldadp != NULL && oldadp->ad_lbn == lbn) allocdirect_merge(adphead, adp, oldadp); FREE_LOCK(&lk); return; } TAILQ_FOREACH(oldadp, adphead, ad_next) { if (oldadp->ad_lbn >= lbn) break; } if (oldadp == NULL) panic("softdep_setup_allocdirect: lost entry"); /* insert in middle of list */ TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); if (oldadp->ad_lbn == lbn) allocdirect_merge(adphead, adp, oldadp); FREE_LOCK(&lk); } /* * Replace an old allocdirect dependency with a newer one. * This routine must be called with splbio interrupts blocked. */ static void allocdirect_merge(adphead, newadp, oldadp) struct allocdirectlst *adphead; /* head of list holding allocdirects */ struct allocdirect *newadp; /* allocdirect being added */ struct allocdirect *oldadp; /* existing allocdirect being checked */ { struct worklist *wk; struct freefrag *freefrag; struct newdirblk *newdirblk; mtx_assert(&lk, MA_OWNED); if (newadp->ad_oldblkno != oldadp->ad_newblkno || newadp->ad_oldsize != oldadp->ad_newsize || newadp->ad_lbn >= NDADDR) panic("%s %jd != new %jd || old size %ld != new %ld", "allocdirect_merge: old blkno", (intmax_t)newadp->ad_oldblkno, (intmax_t)oldadp->ad_newblkno, newadp->ad_oldsize, oldadp->ad_newsize); newadp->ad_oldblkno = oldadp->ad_oldblkno; newadp->ad_oldsize = oldadp->ad_oldsize; /* * If the old dependency had a fragment to free or had never * previously had a block allocated, then the new dependency * can immediately post its freefrag and adopt the old freefrag. * This action is done by swapping the freefrag dependencies. * The new dependency gains the old one's freefrag, and the * old one gets the new one and then immediately puts it on * the worklist when it is freed by free_allocdirect. It is * not possible to do this swap when the old dependency had a * non-zero size but no previous fragment to free. This condition * arises when the new block is an extension of the old block. * Here, the first part of the fragment allocated to the new * dependency is part of the block currently claimed on disk by * the old dependency, so cannot legitimately be freed until the * conditions for the new dependency are fulfilled. */ if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { freefrag = newadp->ad_freefrag; newadp->ad_freefrag = oldadp->ad_freefrag; oldadp->ad_freefrag = freefrag; } /* * If we are tracking a new directory-block allocation, * move it from the old allocdirect to the new allocdirect. */ if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { newdirblk = WK_NEWDIRBLK(wk); WORKLIST_REMOVE(&newdirblk->db_list); if (!LIST_EMPTY(&oldadp->ad_newdirblk)) panic("allocdirect_merge: extra newdirblk"); WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list); } free_allocdirect(adphead, oldadp, 0); } /* * Allocate a new freefrag structure if needed. */ static struct freefrag * newfreefrag(ip, blkno, size) struct inode *ip; ufs2_daddr_t blkno; long size; { struct freefrag *freefrag; struct fs *fs; if (blkno == 0) return (NULL); fs = ip->i_fs; if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) panic("newfreefrag: frag size"); freefrag = malloc(sizeof(struct freefrag), M_FREEFRAG, M_SOFTDEP_FLAGS); workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); freefrag->ff_inum = ip->i_number; freefrag->ff_blkno = blkno; freefrag->ff_fragsize = size; return (freefrag); } /* * This workitem de-allocates fragments that were replaced during * file block allocation. */ static void handle_workitem_freefrag(freefrag) struct freefrag *freefrag; { struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, freefrag->ff_fragsize, freefrag->ff_inum); ACQUIRE_LOCK(&lk); WORKITEM_FREE(freefrag, D_FREEFRAG); FREE_LOCK(&lk); } /* * Set up a dependency structure for an external attributes data block. * This routine follows much of the structure of softdep_setup_allocdirect. * See the description of softdep_setup_allocdirect above for details. */ void softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) struct inode *ip; ufs_lbn_t lbn; ufs2_daddr_t newblkno; ufs2_daddr_t oldblkno; long newsize; long oldsize; struct buf *bp; { struct allocdirect *adp, *oldadp; struct allocdirectlst *adphead; struct bmsafemap *bmsafemap; struct inodedep *inodedep; struct newblk *newblk; struct mount *mp; mp = UFSTOVFS(ip->i_ump); adp = malloc(sizeof(struct allocdirect), M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp); adp->ad_lbn = lbn; adp->ad_newblkno = newblkno; adp->ad_oldblkno = oldblkno; adp->ad_newsize = newsize; adp->ad_oldsize = oldsize; adp->ad_state = ATTACHED | EXTDATA; LIST_INIT(&adp->ad_newdirblk); if (newblkno == oldblkno) adp->ad_freefrag = NULL; else adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); ACQUIRE_LOCK(&lk); if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) panic("softdep_setup_allocext: lost block"); inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); adp->ad_inodedep = inodedep; if (newblk->nb_state == DEPCOMPLETE) { adp->ad_state |= DEPCOMPLETE; adp->ad_buf = NULL; } else { bmsafemap = newblk->nb_bmsafemap; adp->ad_buf = bmsafemap->sm_buf; LIST_REMOVE(newblk, nb_deps); LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); } LIST_REMOVE(newblk, nb_hash); free(newblk, M_NEWBLK); WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); if (lbn >= NXADDR) panic("softdep_setup_allocext: lbn %lld > NXADDR", (long long)lbn); /* * The list of allocdirects must be kept in sorted and ascending * order so that the rollback routines can quickly determine the * first uncommitted block (the size of the file stored on disk * ends at the end of the lowest committed fragment, or if there * are no fragments, at the end of the highest committed block). * Since files generally grow, the typical case is that the new * block is to be added at the end of the list. We speed this * special case by checking against the last allocdirect in the * list before laboriously traversing the list looking for the * insertion point. */ adphead = &inodedep->id_newextupdt; oldadp = TAILQ_LAST(adphead, allocdirectlst); if (oldadp == NULL || oldadp->ad_lbn <= lbn) { /* insert at end of list */ TAILQ_INSERT_TAIL(adphead, adp, ad_next); if (oldadp != NULL && oldadp->ad_lbn == lbn) allocdirect_merge(adphead, adp, oldadp); FREE_LOCK(&lk); return; } TAILQ_FOREACH(oldadp, adphead, ad_next) { if (oldadp->ad_lbn >= lbn) break; } if (oldadp == NULL) panic("softdep_setup_allocext: lost entry"); /* insert in middle of list */ TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); if (oldadp->ad_lbn == lbn) allocdirect_merge(adphead, adp, oldadp); FREE_LOCK(&lk); } /* * Indirect block allocation dependencies. * * The same dependencies that exist for a direct block also exist when * a new block is allocated and pointed to by an entry in a block of * indirect pointers. The undo/redo states described above are also * used here. Because an indirect block contains many pointers that * may have dependencies, a second copy of the entire in-memory indirect * block is kept. The buffer cache copy is always completely up-to-date. * The second copy, which is used only as a source for disk writes, * contains only the safe pointers (i.e., those that have no remaining * update dependencies). The second copy is freed when all pointers * are safe. The cache is not allowed to replace indirect blocks with * pending update dependencies. If a buffer containing an indirect * block with dependencies is written, these routines will mark it * dirty again. It can only be successfully written once all the * dependencies are removed. The ffs_fsync routine in conjunction with * softdep_sync_metadata work together to get all the dependencies * removed so that a file can be successfully written to disk. Three * procedures are used when setting up indirect block pointer * dependencies. The division is necessary because of the organization * of the "balloc" routine and because of the distinction between file * pages and file metadata blocks. */ /* * Allocate a new allocindir structure. */ static struct allocindir * newallocindir(ip, ptrno, newblkno, oldblkno) struct inode *ip; /* inode for file being extended */ int ptrno; /* offset of pointer in indirect block */ ufs2_daddr_t newblkno; /* disk block number being added */ ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ { struct allocindir *aip; aip = malloc(sizeof(struct allocindir), M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO); workitem_alloc(&aip->ai_list, D_ALLOCINDIR, UFSTOVFS(ip->i_ump)); aip->ai_state = ATTACHED; aip->ai_offset = ptrno; aip->ai_newblkno = newblkno; aip->ai_oldblkno = oldblkno; aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); return (aip); } /* * Called just before setting an indirect block pointer * to a newly allocated file page. */ void softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) struct inode *ip; /* inode for file being extended */ ufs_lbn_t lbn; /* allocated block number within file */ struct buf *bp; /* buffer with indirect blk referencing page */ int ptrno; /* offset of pointer in indirect block */ ufs2_daddr_t newblkno; /* disk block number being added */ ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ struct buf *nbp; /* buffer holding allocated page */ { struct allocindir *aip; struct pagedep *pagedep; ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); aip = newallocindir(ip, ptrno, newblkno, oldblkno); ACQUIRE_LOCK(&lk); /* * If we are allocating a directory page, then we must * allocate an associated pagedep to track additions and * deletions. */ if ((ip->i_mode & IFMT) == IFDIR && pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); setup_allocindir_phase2(bp, ip, aip); FREE_LOCK(&lk); } /* * Called just before setting an indirect block pointer to a * newly allocated indirect block. */ void softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) struct buf *nbp; /* newly allocated indirect block */ struct inode *ip; /* inode for file being extended */ struct buf *bp; /* indirect block referencing allocated block */ int ptrno; /* offset of pointer in indirect block */ ufs2_daddr_t newblkno; /* disk block number being added */ { struct allocindir *aip; ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); aip = newallocindir(ip, ptrno, newblkno, 0); ACQUIRE_LOCK(&lk); WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); setup_allocindir_phase2(bp, ip, aip); FREE_LOCK(&lk); } /* * Called to finish the allocation of the "aip" allocated * by one of the two routines above. */ static void setup_allocindir_phase2(bp, ip, aip) struct buf *bp; /* in-memory copy of the indirect block */ struct inode *ip; /* inode for file being extended */ struct allocindir *aip; /* allocindir allocated by the above routines */ { struct worklist *wk; struct indirdep *indirdep, *newindirdep; struct bmsafemap *bmsafemap; struct allocindir *oldaip; struct freefrag *freefrag; struct newblk *newblk; ufs2_daddr_t blkno; mtx_assert(&lk, MA_OWNED); if (bp->b_lblkno >= 0) panic("setup_allocindir_phase2: not indir blk"); for (indirdep = NULL, newindirdep = NULL; ; ) { LIST_FOREACH(wk, &bp->b_dep, wk_list) { if (wk->wk_type != D_INDIRDEP) continue; indirdep = WK_INDIRDEP(wk); break; } if (indirdep == NULL && newindirdep) { indirdep = newindirdep; WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); newindirdep = NULL; } if (indirdep) { if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, &newblk) == 0) panic("setup_allocindir: lost block"); if (newblk->nb_state == DEPCOMPLETE) { aip->ai_state |= DEPCOMPLETE; aip->ai_buf = NULL; } else { bmsafemap = newblk->nb_bmsafemap; aip->ai_buf = bmsafemap->sm_buf; LIST_REMOVE(newblk, nb_deps); LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, aip, ai_deps); } LIST_REMOVE(newblk, nb_hash); free(newblk, M_NEWBLK); aip->ai_indirdep = indirdep; /* * Check to see if there is an existing dependency * for this block. If there is, merge the old * dependency into the new one. */ if (aip->ai_oldblkno == 0) oldaip = NULL; else LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) if (oldaip->ai_offset == aip->ai_offset) break; freefrag = NULL; if (oldaip != NULL) { if (oldaip->ai_newblkno != aip->ai_oldblkno) panic("setup_allocindir_phase2: blkno"); aip->ai_oldblkno = oldaip->ai_oldblkno; freefrag = aip->ai_freefrag; aip->ai_freefrag = oldaip->ai_freefrag; oldaip->ai_freefrag = NULL; free_allocindir(oldaip, NULL); } LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); if (ip->i_ump->um_fstype == UFS1) ((ufs1_daddr_t *)indirdep->ir_savebp->b_data) [aip->ai_offset] = aip->ai_oldblkno; else ((ufs2_daddr_t *)indirdep->ir_savebp->b_data) [aip->ai_offset] = aip->ai_oldblkno; FREE_LOCK(&lk); if (freefrag != NULL) handle_workitem_freefrag(freefrag); } else FREE_LOCK(&lk); if (newindirdep) { newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; brelse(newindirdep->ir_savebp); ACQUIRE_LOCK(&lk); WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); if (indirdep) break; FREE_LOCK(&lk); } if (indirdep) { ACQUIRE_LOCK(&lk); break; } newindirdep = malloc(sizeof(struct indirdep), M_INDIRDEP, M_SOFTDEP_FLAGS); workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, UFSTOVFS(ip->i_ump)); newindirdep->ir_state = ATTACHED; if (ip->i_ump->um_fstype == UFS1) newindirdep->ir_state |= UFS1FMT; LIST_INIT(&newindirdep->ir_deplisthd); LIST_INIT(&newindirdep->ir_donehd); if (bp->b_blkno == bp->b_lblkno) { ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, NULL, NULL); bp->b_blkno = blkno; } newindirdep->ir_savebp = getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); BUF_KERNPROC(newindirdep->ir_savebp); bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); ACQUIRE_LOCK(&lk); } } /* * Block de-allocation dependencies. * * When blocks are de-allocated, the on-disk pointers must be nullified before * the blocks are made available for use by other files. (The true * requirement is that old pointers must be nullified before new on-disk * pointers are set. We chose this slightly more stringent requirement to * reduce complexity.) Our implementation handles this dependency by updating * the inode (or indirect block) appropriately but delaying the actual block * de-allocation (i.e., freemap and free space count manipulation) until * after the updated versions reach stable storage. After the disk is * updated, the blocks can be safely de-allocated whenever it is convenient. * This implementation handles only the common case of reducing a file's * length to zero. Other cases are handled by the conventional synchronous * write approach. * * The ffs implementation with which we worked double-checks * the state of the block pointers and file size as it reduces * a file's length. Some of this code is replicated here in our * soft updates implementation. The freeblks->fb_chkcnt field is * used to transfer a part of this information to the procedure * that eventually de-allocates the blocks. * * This routine should be called from the routine that shortens * a file's length, before the inode's size or block pointers * are modified. It will save the block pointer information for * later release and zero the inode so that the calling routine * can release it. */ void softdep_setup_freeblocks(ip, length, flags) struct inode *ip; /* The inode whose length is to be reduced */ off_t length; /* The new length for the file */ int flags; /* IO_EXT and/or IO_NORMAL */ { struct freeblks *freeblks; struct inodedep *inodedep; struct allocdirect *adp; struct bufobj *bo; struct vnode *vp; struct buf *bp; struct fs *fs; ufs2_daddr_t extblocks, datablocks; struct mount *mp; int i, delay, error; fs = ip->i_fs; mp = UFSTOVFS(ip->i_ump); if (length != 0) panic("softdep_setup_freeblocks: non-zero length"); freeblks = malloc(sizeof(struct freeblks), M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); freeblks->fb_state = ATTACHED; freeblks->fb_uid = ip->i_uid; freeblks->fb_previousinum = ip->i_number; freeblks->fb_devvp = ip->i_devvp; extblocks = 0; if (fs->fs_magic == FS_UFS2_MAGIC) extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); datablocks = DIP(ip, i_blocks) - extblocks; if ((flags & IO_NORMAL) == 0) { freeblks->fb_oldsize = 0; freeblks->fb_chkcnt = 0; } else { freeblks->fb_oldsize = ip->i_size; ip->i_size = 0; DIP_SET(ip, i_size, 0); freeblks->fb_chkcnt = datablocks; for (i = 0; i < NDADDR; i++) { freeblks->fb_dblks[i] = DIP(ip, i_db[i]); DIP_SET(ip, i_db[i], 0); } for (i = 0; i < NIADDR; i++) { freeblks->fb_iblks[i] = DIP(ip, i_ib[i]); DIP_SET(ip, i_ib[i], 0); } /* * If the file was removed, then the space being freed was * accounted for then (see softdep_releasefile()). If the * file is merely being truncated, then we account for it now. */ if ((ip->i_flag & IN_SPACECOUNTED) == 0) { UFS_LOCK(ip->i_ump); fs->fs_pendingblocks += datablocks; UFS_UNLOCK(ip->i_ump); } } if ((flags & IO_EXT) == 0) { freeblks->fb_oldextsize = 0; } else { freeblks->fb_oldextsize = ip->i_din2->di_extsize; ip->i_din2->di_extsize = 0; freeblks->fb_chkcnt += extblocks; for (i = 0; i < NXADDR; i++) { freeblks->fb_eblks[i] = ip->i_din2->di_extb[i]; ip->i_din2->di_extb[i] = 0; } } DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt); /* * Push the zero'ed inode to to its disk buffer so that we are free * to delete its dependencies below. Once the dependencies are gone * the buffer can be safely released. */ if ((error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), (int)fs->fs_bsize, NOCRED, &bp)) != 0) { brelse(bp); softdep_error("softdep_setup_freeblocks", error); } if (ip->i_ump->um_fstype == UFS1) *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; else *((struct ufs2_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; /* * Find and eliminate any inode dependencies. */ ACQUIRE_LOCK(&lk); (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); if ((inodedep->id_state & IOSTARTED) != 0) panic("softdep_setup_freeblocks: inode busy"); /* * Add the freeblks structure to the list of operations that * must await the zero'ed inode being written to disk. If we * still have a bitmap dependency (delay == 0), then the inode * has never been written to disk, so we can process the * freeblks below once we have deleted the dependencies. */ delay = (inodedep->id_state & DEPCOMPLETE); if (delay) WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); /* * Because the file length has been truncated to zero, any * pending block allocation dependency structures associated * with this inode are obsolete and can simply be de-allocated. * We must first merge the two dependency lists to get rid of * any duplicate freefrag structures, then purge the merged list. * If we still have a bitmap dependency, then the inode has never * been written to disk, so we can free any fragments without delay. */ if (flags & IO_NORMAL) { merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) free_allocdirect(&inodedep->id_inoupdt, adp, delay); } if (flags & IO_EXT) { merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) free_allocdirect(&inodedep->id_extupdt, adp, delay); } FREE_LOCK(&lk); bdwrite(bp); /* * We must wait for any I/O in progress to finish so that * all potential buffers on the dirty list will be visible. * Once they are all there, walk the list and get rid of * any dependencies. */ vp = ITOV(ip); bo = &vp->v_bufobj; BO_LOCK(bo); drain_output(vp); restart: TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) continue; if ((bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT)) == NULL) goto restart; BO_UNLOCK(bo); ACQUIRE_LOCK(&lk); (void) inodedep_lookup(mp, ip->i_number, 0, &inodedep); deallocate_dependencies(bp, inodedep); FREE_LOCK(&lk); bp->b_flags |= B_INVAL | B_NOCACHE; brelse(bp); BO_LOCK(bo); goto restart; } BO_UNLOCK(bo); ACQUIRE_LOCK(&lk); if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) (void) free_inodedep(inodedep); if(delay) { freeblks->fb_state |= DEPCOMPLETE; /* * If the inode with zeroed block pointers is now on disk * we can start freeing blocks. Add freeblks to the worklist * instead of calling handle_workitem_freeblocks directly as * it is more likely that additional IO is needed to complete * the request here than in the !delay case. */ if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) add_to_worklist(&freeblks->fb_list); } FREE_LOCK(&lk); /* * If the inode has never been written to disk (delay == 0), * then we can process the freeblks now that we have deleted * the dependencies. */ if (!delay) handle_workitem_freeblocks(freeblks, 0); } /* * Reclaim any dependency structures from a buffer that is about to * be reallocated to a new vnode. The buffer must be locked, thus, * no I/O completion operations can occur while we are manipulating * its associated dependencies. The mutex is held so that other I/O's * associated with related dependencies do not occur. */ static void deallocate_dependencies(bp, inodedep) struct buf *bp; struct inodedep *inodedep; { struct worklist *wk; struct indirdep *indirdep; struct allocindir *aip; struct pagedep *pagedep; struct dirrem *dirrem; struct diradd *dap; int i; mtx_assert(&lk, MA_OWNED); while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { switch (wk->wk_type) { case D_INDIRDEP: indirdep = WK_INDIRDEP(wk); /* * None of the indirect pointers will ever be visible, * so they can simply be tossed. GOINGAWAY ensures * that allocated pointers will be saved in the buffer * cache until they are freed. Note that they will * only be able to be found by their physical address * since the inode mapping the logical address will * be gone. The save buffer used for the safe copy * was allocated in setup_allocindir_phase2 using * the physical address so it could be used for this * purpose. Hence we swap the safe copy with the real * copy, allowing the safe copy to be freed and holding * on to the real copy for later use in indir_trunc. */ if (indirdep->ir_state & GOINGAWAY) panic("deallocate_dependencies: already gone"); indirdep->ir_state |= GOINGAWAY; VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1; while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) free_allocindir(aip, inodedep); if (bp->b_lblkno >= 0 || bp->b_blkno != indirdep->ir_savebp->b_lblkno) panic("deallocate_dependencies: not indir"); bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); WORKLIST_REMOVE(wk); WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); continue; case D_PAGEDEP: pagedep = WK_PAGEDEP(wk); /* * None of the directory additions will ever be * visible, so they can simply be tossed. */ for (i = 0; i < DAHASHSZ; i++) while ((dap = LIST_FIRST(&pagedep->pd_diraddhd[i]))) free_diradd(dap); while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) free_diradd(dap); /* * Copy any directory remove dependencies to the list * to be processed after the zero'ed inode is written. * If the inode has already been written, then they * can be dumped directly onto the work list. */ LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { LIST_REMOVE(dirrem, dm_next); dirrem->dm_dirinum = pagedep->pd_ino; if (inodedep == NULL || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) add_to_worklist(&dirrem->dm_list); else WORKLIST_INSERT(&inodedep->id_bufwait, &dirrem->dm_list); } if ((pagedep->pd_state & NEWBLOCK) != 0) { LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list) if (wk->wk_type == D_NEWDIRBLK && WK_NEWDIRBLK(wk)->db_pagedep == pagedep) break; if (wk != NULL) { WORKLIST_REMOVE(wk); free_newdirblk(WK_NEWDIRBLK(wk)); } else panic("deallocate_dependencies: " "lost pagedep"); } WORKLIST_REMOVE(&pagedep->pd_list); LIST_REMOVE(pagedep, pd_hash); WORKITEM_FREE(pagedep, D_PAGEDEP); continue; case D_ALLOCINDIR: free_allocindir(WK_ALLOCINDIR(wk), inodedep); continue; case D_ALLOCDIRECT: case D_INODEDEP: panic("deallocate_dependencies: Unexpected type %s", TYPENAME(wk->wk_type)); /* NOTREACHED */ default: panic("deallocate_dependencies: Unknown type %s", TYPENAME(wk->wk_type)); /* NOTREACHED */ } } } /* * Free an allocdirect. Generate a new freefrag work request if appropriate. * This routine must be called with splbio interrupts blocked. */ static void free_allocdirect(adphead, adp, delay) struct allocdirectlst *adphead; struct allocdirect *adp; int delay; { struct newdirblk *newdirblk; struct worklist *wk; mtx_assert(&lk, MA_OWNED); if ((adp->ad_state & DEPCOMPLETE) == 0) LIST_REMOVE(adp, ad_deps); TAILQ_REMOVE(adphead, adp, ad_next); if ((adp->ad_state & COMPLETE) == 0) WORKLIST_REMOVE(&adp->ad_list); if (adp->ad_freefrag != NULL) { if (delay) WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, &adp->ad_freefrag->ff_list); else add_to_worklist(&adp->ad_freefrag->ff_list); } if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) { newdirblk = WK_NEWDIRBLK(wk); WORKLIST_REMOVE(&newdirblk->db_list); if (!LIST_EMPTY(&adp->ad_newdirblk)) panic("free_allocdirect: extra newdirblk"); if (delay) WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, &newdirblk->db_list); else free_newdirblk(newdirblk); } WORKITEM_FREE(adp, D_ALLOCDIRECT); } /* * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. * This routine must be called with splbio interrupts blocked. */ static void free_newdirblk(newdirblk) struct newdirblk *newdirblk; { struct pagedep *pagedep; struct diradd *dap; int i; mtx_assert(&lk, MA_OWNED); /* * If the pagedep is still linked onto the directory buffer * dependency chain, then some of the entries on the * pd_pendinghd list may not be committed to disk yet. In * this case, we will simply clear the NEWBLOCK flag and * let the pd_pendinghd list be processed when the pagedep * is next written. If the pagedep is no longer on the buffer * dependency chain, then all the entries on the pd_pending * list are committed to disk and we can free them here. */ pagedep = newdirblk->db_pagedep; pagedep->pd_state &= ~NEWBLOCK; if ((pagedep->pd_state & ONWORKLIST) == 0) while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) free_diradd(dap); /* * If no dependencies remain, the pagedep will be freed. */ for (i = 0; i < DAHASHSZ; i++) if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) break; if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) { LIST_REMOVE(pagedep, pd_hash); WORKITEM_FREE(pagedep, D_PAGEDEP); } WORKITEM_FREE(newdirblk, D_NEWDIRBLK); } /* * Prepare an inode to be freed. The actual free operation is not * done until the zero'ed inode has been written to disk. */ void softdep_freefile(pvp, ino, mode) struct vnode *pvp; ino_t ino; int mode; { struct inode *ip = VTOI(pvp); struct inodedep *inodedep; struct freefile *freefile; /* * This sets up the inode de-allocation dependency. */ freefile = malloc(sizeof(struct freefile), M_FREEFILE, M_SOFTDEP_FLAGS); workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); freefile->fx_mode = mode; freefile->fx_oldinum = ino; freefile->fx_devvp = ip->i_devvp; if ((ip->i_flag & IN_SPACECOUNTED) == 0) { UFS_LOCK(ip->i_ump); ip->i_fs->fs_pendinginodes += 1; UFS_UNLOCK(ip->i_ump); } /* * If the inodedep does not exist, then the zero'ed inode has * been written to disk. If the allocated inode has never been * written to disk, then the on-disk inode is zero'ed. In either * case we can free the file immediately. */ ACQUIRE_LOCK(&lk); if (inodedep_lookup(pvp->v_mount, ino, 0, &inodedep) == 0 || check_inode_unwritten(inodedep)) { FREE_LOCK(&lk); handle_workitem_freefile(freefile); return; } WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); FREE_LOCK(&lk); if (ip->i_number == ino) ip->i_flag |= IN_MODIFIED; } /* * Check to see if an inode has never been written to disk. If * so free the inodedep and return success, otherwise return failure. * This routine must be called with splbio interrupts blocked. * * If we still have a bitmap dependency, then the inode has never * been written to disk. Drop the dependency as it is no longer * necessary since the inode is being deallocated. We set the * ALLCOMPLETE flags since the bitmap now properly shows that the * inode is not allocated. Even if the inode is actively being * written, it has been rolled back to its zero'ed state, so we * are ensured that a zero inode is what is on the disk. For short * lived files, this change will usually result in removing all the * dependencies from the inode so that it can be freed immediately. */ static int check_inode_unwritten(inodedep) struct inodedep *inodedep; { mtx_assert(&lk, MA_OWNED); if ((inodedep->id_state & DEPCOMPLETE) != 0 || !LIST_EMPTY(&inodedep->id_pendinghd) || !LIST_EMPTY(&inodedep->id_bufwait) || !LIST_EMPTY(&inodedep->id_inowait) || !TAILQ_EMPTY(&inodedep->id_inoupdt) || !TAILQ_EMPTY(&inodedep->id_newinoupdt) || !TAILQ_EMPTY(&inodedep->id_extupdt) || !TAILQ_EMPTY(&inodedep->id_newextupdt) || inodedep->id_nlinkdelta != 0) return (0); /* * Another process might be in initiate_write_inodeblock_ufs[12] * trying to allocate memory without holding "Softdep Lock". */ if ((inodedep->id_state & IOSTARTED) != 0 && inodedep->id_savedino1 == NULL) return (0); inodedep->id_state |= ALLCOMPLETE; LIST_REMOVE(inodedep, id_deps); inodedep->id_buf = NULL; if (inodedep->id_state & ONWORKLIST) WORKLIST_REMOVE(&inodedep->id_list); if (inodedep->id_savedino1 != NULL) { free(inodedep->id_savedino1, M_SAVEDINO); inodedep->id_savedino1 = NULL; } if (free_inodedep(inodedep) == 0) panic("check_inode_unwritten: busy inode"); return (1); } /* * Try to free an inodedep structure. Return 1 if it could be freed. */ static int free_inodedep(inodedep) struct inodedep *inodedep; { mtx_assert(&lk, MA_OWNED); if ((inodedep->id_state & ONWORKLIST) != 0 || (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || !LIST_EMPTY(&inodedep->id_pendinghd) || !LIST_EMPTY(&inodedep->id_bufwait) || !LIST_EMPTY(&inodedep->id_inowait) || !TAILQ_EMPTY(&inodedep->id_inoupdt) || !TAILQ_EMPTY(&inodedep->id_newinoupdt) || !TAILQ_EMPTY(&inodedep->id_extupdt) || !TAILQ_EMPTY(&inodedep->id_newextupdt) || inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL) return (0); LIST_REMOVE(inodedep, id_hash); WORKITEM_FREE(inodedep, D_INODEDEP); num_inodedep -= 1; return (1); } /* * This workitem routine performs the block de-allocation. * The workitem is added to the pending list after the updated * inode block has been written to disk. As mentioned above, * checks regarding the number of blocks de-allocated (compared * to the number of blocks allocated for the file) are also * performed in this function. */ static void handle_workitem_freeblocks(freeblks, flags) struct freeblks *freeblks; int flags; { struct inode *ip; struct vnode *vp; struct fs *fs; struct ufsmount *ump; int i, nblocks, level, bsize; ufs2_daddr_t bn, blocksreleased = 0; int error, allerror = 0; ufs_lbn_t baselbns[NIADDR], tmpval; int fs_pendingblocks; ump = VFSTOUFS(freeblks->fb_list.wk_mp); fs = ump->um_fs; fs_pendingblocks = 0; tmpval = 1; baselbns[0] = NDADDR; for (i = 1; i < NIADDR; i++) { tmpval *= NINDIR(fs); baselbns[i] = baselbns[i - 1] + tmpval; } nblocks = btodb(fs->fs_bsize); blocksreleased = 0; /* * Release all extended attribute blocks or frags. */ if (freeblks->fb_oldextsize > 0) { for (i = (NXADDR - 1); i >= 0; i--) { if ((bn = freeblks->fb_eblks[i]) == 0) continue; bsize = sblksize(fs, freeblks->fb_oldextsize, i); ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, freeblks->fb_previousinum); blocksreleased += btodb(bsize); } } /* * Release all data blocks or frags. */ if (freeblks->fb_oldsize > 0) { /* * Indirect blocks first. */ for (level = (NIADDR - 1); level >= 0; level--) { if ((bn = freeblks->fb_iblks[level]) == 0) continue; if ((error = indir_trunc(freeblks, fsbtodb(fs, bn), level, baselbns[level], &blocksreleased)) != 0) allerror = error; ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, fs->fs_bsize, freeblks->fb_previousinum); fs_pendingblocks += nblocks; blocksreleased += nblocks; } /* * All direct blocks or frags. */ for (i = (NDADDR - 1); i >= 0; i--) { if ((bn = freeblks->fb_dblks[i]) == 0) continue; bsize = sblksize(fs, freeblks->fb_oldsize, i); ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, freeblks->fb_previousinum); fs_pendingblocks += btodb(bsize); blocksreleased += btodb(bsize); } } UFS_LOCK(ump); fs->fs_pendingblocks -= fs_pendingblocks; UFS_UNLOCK(ump); /* * If we still have not finished background cleanup, then check * to see if the block count needs to be adjusted. */ if (freeblks->fb_chkcnt != blocksreleased && (fs->fs_flags & FS_UNCLEAN) != 0 && ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_previousinum, (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ) == 0) { ip = VTOI(vp); DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \ freeblks->fb_chkcnt - blocksreleased); ip->i_flag |= IN_CHANGE; vput(vp); } #ifdef INVARIANTS if (freeblks->fb_chkcnt != blocksreleased && ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0)) printf("handle_workitem_freeblocks: block count\n"); if (allerror) softdep_error("handle_workitem_freeblks", allerror); #endif /* INVARIANTS */ ACQUIRE_LOCK(&lk); WORKITEM_FREE(freeblks, D_FREEBLKS); FREE_LOCK(&lk); } /* * Release blocks associated with the inode ip and stored in the indirect * block dbn. If level is greater than SINGLE, the block is an indirect block * and recursive calls to indirtrunc must be used to cleanse other indirect * blocks. */ static int indir_trunc(freeblks, dbn, level, lbn, countp) struct freeblks *freeblks; ufs2_daddr_t dbn; int level; ufs_lbn_t lbn; ufs2_daddr_t *countp; { struct buf *bp; struct fs *fs; struct worklist *wk; struct indirdep *indirdep; struct ufsmount *ump; ufs1_daddr_t *bap1 = 0; ufs2_daddr_t nb, *bap2 = 0; ufs_lbn_t lbnadd; int i, nblocks, ufs1fmt; int error, allerror = 0; int fs_pendingblocks; ump = VFSTOUFS(freeblks->fb_list.wk_mp); fs = ump->um_fs; fs_pendingblocks = 0; lbnadd = 1; for (i = level; i > 0; i--) lbnadd *= NINDIR(fs); /* * Get buffer of block pointers to be freed. This routine is not * called until the zero'ed inode has been written, so it is safe * to free blocks as they are encountered. Because the inode has * been zero'ed, calls to bmap on these blocks will fail. So, we * have to use the on-disk address and the block device for the * filesystem to look them up. If the file was deleted before its * indirect blocks were all written to disk, the routine that set * us up (deallocate_dependencies) will have arranged to leave * a complete copy of the indirect block in memory for our use. * Otherwise we have to read the blocks in from the disk. */ #ifdef notyet bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0, GB_NOCREAT); #else bp = incore(&freeblks->fb_devvp->v_bufobj, dbn); #endif ACQUIRE_LOCK(&lk); if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) { if (wk->wk_type != D_INDIRDEP || (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || (indirdep->ir_state & GOINGAWAY) == 0) panic("indir_trunc: lost indirdep"); WORKLIST_REMOVE(wk); WORKITEM_FREE(indirdep, D_INDIRDEP); if (!LIST_EMPTY(&bp->b_dep)) panic("indir_trunc: dangling dep"); ump->um_numindirdeps -= 1; FREE_LOCK(&lk); } else { #ifdef notyet if (bp) brelse(bp); #endif FREE_LOCK(&lk); error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp); if (error) { brelse(bp); return (error); } } /* * Recursively free indirect blocks. */ if (ump->um_fstype == UFS1) { ufs1fmt = 1; bap1 = (ufs1_daddr_t *)bp->b_data; } else { ufs1fmt = 0; bap2 = (ufs2_daddr_t *)bp->b_data; } nblocks = btodb(fs->fs_bsize); for (i = NINDIR(fs) - 1; i >= 0; i--) { if (ufs1fmt) nb = bap1[i]; else nb = bap2[i]; if (nb == 0) continue; if (level != 0) { if ((error = indir_trunc(freeblks, fsbtodb(fs, nb), level - 1, lbn + (i * lbnadd), countp)) != 0) allerror = error; } ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize, freeblks->fb_previousinum); fs_pendingblocks += nblocks; *countp += nblocks; } UFS_LOCK(ump); fs->fs_pendingblocks -= fs_pendingblocks; UFS_UNLOCK(ump); bp->b_flags |= B_INVAL | B_NOCACHE; brelse(bp); return (allerror); } /* * Free an allocindir. * This routine must be called with splbio interrupts blocked. */ static void free_allocindir(aip, inodedep) struct allocindir *aip; struct inodedep *inodedep; { struct freefrag *freefrag; mtx_assert(&lk, MA_OWNED); if ((aip->ai_state & DEPCOMPLETE) == 0) LIST_REMOVE(aip, ai_deps); if (aip->ai_state & ONWORKLIST) WORKLIST_REMOVE(&aip->ai_list); LIST_REMOVE(aip, ai_next); if ((freefrag = aip->ai_freefrag) != NULL) { if (inodedep == NULL) add_to_worklist(&freefrag->ff_list); else WORKLIST_INSERT(&inodedep->id_bufwait, &freefrag->ff_list); } WORKITEM_FREE(aip, D_ALLOCINDIR); } /* * Directory entry addition dependencies. * * When adding a new directory entry, the inode (with its incremented link * count) must be written to disk before the directory entry's pointer to it. * Also, if the inode is newly allocated, the corresponding freemap must be * updated (on disk) before the directory entry's pointer. These requirements * are met via undo/redo on the directory entry's pointer, which consists * simply of the inode number. * * As directory entries are added and deleted, the free space within a * directory block can become fragmented. The ufs filesystem will compact * a fragmented directory block to make space for a new entry. When this * occurs, the offsets of previously added entries change. Any "diradd" * dependency structures corresponding to these entries must be updated with * the new offsets. */ /* * This routine is called after the in-memory inode's link * count has been incremented, but before the directory entry's * pointer to the inode has been set. */ int softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) struct buf *bp; /* buffer containing directory block */ struct inode *dp; /* inode for directory */ off_t diroffset; /* offset of new entry in directory */ ino_t newinum; /* inode referenced by new directory entry */ struct buf *newdirbp; /* non-NULL => contents of new mkdir */ int isnewblk; /* entry is in a newly allocated block */ { int offset; /* offset of new entry within directory block */ ufs_lbn_t lbn; /* block in directory containing new entry */ struct fs *fs; struct diradd *dap; struct allocdirect *adp; struct pagedep *pagedep; struct inodedep *inodedep; struct newdirblk *newdirblk = 0; struct mkdir *mkdir1, *mkdir2; struct mount *mp; /* * Whiteouts have no dependencies. */ if (newinum == WINO) { if (newdirbp != NULL) bdwrite(newdirbp); return (0); } mp = UFSTOVFS(dp->i_ump); fs = dp->i_fs; lbn = lblkno(fs, diroffset); offset = blkoff(fs, diroffset); dap = malloc(sizeof(struct diradd), M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); workitem_alloc(&dap->da_list, D_DIRADD, mp); dap->da_offset = offset; dap->da_newinum = newinum; dap->da_state = ATTACHED; if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) { newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, M_SOFTDEP_FLAGS); workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); } if (newdirbp == NULL) { dap->da_state |= DEPCOMPLETE; ACQUIRE_LOCK(&lk); } else { dap->da_state |= MKDIR_BODY | MKDIR_PARENT; mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); mkdir1->md_state = MKDIR_BODY; mkdir1->md_diradd = dap; mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); mkdir2->md_state = MKDIR_PARENT; mkdir2->md_diradd = dap; /* * Dependency on "." and ".." being written to disk. */ mkdir1->md_buf = newdirbp; ACQUIRE_LOCK(&lk); LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); FREE_LOCK(&lk); bdwrite(newdirbp); /* * Dependency on link count increase for parent directory */ ACQUIRE_LOCK(&lk); if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { dap->da_state &= ~MKDIR_PARENT; WORKITEM_FREE(mkdir2, D_MKDIR); } else { LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); } } /* * Link into parent directory pagedep to await its being written. */ if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); dap->da_pagedep = pagedep; LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, da_pdlist); /* * Link into its inodedep. Put it on the id_bufwait list if the inode * is not yet written. If it is written, do the post-inode write * processing to put it on the id_pendinghd list. */ (void) inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) diradd_inode_written(dap, inodedep); else WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); if (isnewblk) { /* * Directories growing into indirect blocks are rare * enough and the frequency of new block allocation * in those cases even more rare, that we choose not * to bother tracking them. Rather we simply force the * new directory entry to disk. */ if (lbn >= NDADDR) { FREE_LOCK(&lk); /* * We only have a new allocation when at the * beginning of a new block, not when we are * expanding into an existing block. */ if (blkoff(fs, diroffset) == 0) return (1); return (0); } /* * We only have a new allocation when at the beginning * of a new fragment, not when we are expanding into an * existing fragment. Also, there is nothing to do if we * are already tracking this block. */ if (fragoff(fs, diroffset) != 0) { FREE_LOCK(&lk); return (0); } if ((pagedep->pd_state & NEWBLOCK) != 0) { WORKITEM_FREE(newdirblk, D_NEWDIRBLK); FREE_LOCK(&lk); return (0); } /* * Find our associated allocdirect and have it track us. */ if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0) panic("softdep_setup_directory_add: lost inodedep"); adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst); if (adp == NULL || adp->ad_lbn != lbn) panic("softdep_setup_directory_add: lost entry"); pagedep->pd_state |= NEWBLOCK; newdirblk->db_pagedep = pagedep; WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list); } FREE_LOCK(&lk); return (0); } /* * This procedure is called to change the offset of a directory * entry when compacting a directory block which must be owned * exclusively by the caller. Note that the actual entry movement * must be done in this procedure to ensure that no I/O completions * occur while the move is in progress. */ void softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) struct inode *dp; /* inode for directory */ caddr_t base; /* address of dp->i_offset */ caddr_t oldloc; /* address of old directory location */ caddr_t newloc; /* address of new directory location */ int entrysize; /* size of directory entry */ { int offset, oldoffset, newoffset; struct pagedep *pagedep; struct diradd *dap; ufs_lbn_t lbn; ACQUIRE_LOCK(&lk); lbn = lblkno(dp->i_fs, dp->i_offset); offset = blkoff(dp->i_fs, dp->i_offset); if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) goto done; oldoffset = offset + (oldloc - base); newoffset = offset + (newloc - base); LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { if (dap->da_offset != oldoffset) continue; dap->da_offset = newoffset; if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) break; LIST_REMOVE(dap, da_pdlist); LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], dap, da_pdlist); break; } if (dap == NULL) { LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { if (dap->da_offset == oldoffset) { dap->da_offset = newoffset; break; } } } done: bcopy(oldloc, newloc, entrysize); FREE_LOCK(&lk); } /* * Free a diradd dependency structure. This routine must be called * with splbio interrupts blocked. */ static void free_diradd(dap) struct diradd *dap; { struct dirrem *dirrem; struct pagedep *pagedep; struct inodedep *inodedep; struct mkdir *mkdir, *nextmd; mtx_assert(&lk, MA_OWNED); WORKLIST_REMOVE(&dap->da_list); LIST_REMOVE(dap, da_pdlist); if ((dap->da_state & DIRCHG) == 0) { pagedep = dap->da_pagedep; } else { dirrem = dap->da_previous; pagedep = dirrem->dm_pagedep; dirrem->dm_dirinum = pagedep->pd_ino; add_to_worklist(&dirrem->dm_list); } if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 0, &inodedep) != 0) (void) free_inodedep(inodedep); if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { nextmd = LIST_NEXT(mkdir, md_mkdirs); if (mkdir->md_diradd != dap) continue; dap->da_state &= ~mkdir->md_state; WORKLIST_REMOVE(&mkdir->md_list); LIST_REMOVE(mkdir, md_mkdirs); WORKITEM_FREE(mkdir, D_MKDIR); } if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) panic("free_diradd: unfound ref"); } WORKITEM_FREE(dap, D_DIRADD); } /* * Directory entry removal dependencies. * * When removing a directory entry, the entry's inode pointer must be * zero'ed on disk before the corresponding inode's link count is decremented * (possibly freeing the inode for re-use). This dependency is handled by * updating the directory entry but delaying the inode count reduction until * after the directory block has been written to disk. After this point, the * inode count can be decremented whenever it is convenient. */ /* * This routine should be called immediately after removing * a directory entry. The inode's link count should not be * decremented by the calling procedure -- the soft updates * code will do this task when it is safe. */ void softdep_setup_remove(bp, dp, ip, isrmdir) struct buf *bp; /* buffer containing directory block */ struct inode *dp; /* inode for the directory being modified */ struct inode *ip; /* inode for directory entry being removed */ int isrmdir; /* indicates if doing RMDIR */ { struct dirrem *dirrem, *prevdirrem; /* * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. */ dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); /* * If the COMPLETE flag is clear, then there were no active * entries and we want to roll back to a zeroed entry until * the new inode is committed to disk. If the COMPLETE flag is * set then we have deleted an entry that never made it to * disk. If the entry we deleted resulted from a name change, * then the old name still resides on disk. We cannot delete * its inode (returned to us in prevdirrem) until the zeroed * directory entry gets to disk. The new inode has never been * referenced on the disk, so can be deleted immediately. */ if ((dirrem->dm_state & COMPLETE) == 0) { LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, dm_next); FREE_LOCK(&lk); } else { if (prevdirrem != NULL) LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, prevdirrem, dm_next); dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; FREE_LOCK(&lk); handle_workitem_remove(dirrem, NULL); } } /* * Allocate a new dirrem if appropriate and return it along with * its associated pagedep. Called without a lock, returns with lock. */ static long num_dirrem; /* number of dirrem allocated */ static struct dirrem * newdirrem(bp, dp, ip, isrmdir, prevdirremp) struct buf *bp; /* buffer containing directory block */ struct inode *dp; /* inode for the directory being modified */ struct inode *ip; /* inode for directory entry being removed */ int isrmdir; /* indicates if doing RMDIR */ struct dirrem **prevdirremp; /* previously referenced inode, if any */ { int offset; ufs_lbn_t lbn; struct diradd *dap; struct dirrem *dirrem; struct pagedep *pagedep; /* * Whiteouts have no deletion dependencies. */ if (ip == NULL) panic("newdirrem: whiteout"); /* * If we are over our limit, try to improve the situation. * Limiting the number of dirrem structures will also limit * the number of freefile and freeblks structures. */ ACQUIRE_LOCK(&lk); if (!(ip->i_flags & SF_SNAPSHOT) && num_dirrem > max_softdeps / 2) (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE); num_dirrem += 1; FREE_LOCK(&lk); dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); workitem_alloc(&dirrem->dm_list, D_DIRREM, ITOV(dp)->v_mount); dirrem->dm_state = isrmdir ? RMDIR : 0; dirrem->dm_oldinum = ip->i_number; *prevdirremp = NULL; ACQUIRE_LOCK(&lk); lbn = lblkno(dp->i_fs, dp->i_offset); offset = blkoff(dp->i_fs, dp->i_offset); if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); dirrem->dm_pagedep = pagedep; /* * Check for a diradd dependency for the same directory entry. * If present, then both dependencies become obsolete and can * be de-allocated. Check for an entry on both the pd_dirraddhd * list and the pd_pendinghd list. */ LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) if (dap->da_offset == offset) break; if (dap == NULL) { LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) if (dap->da_offset == offset) break; if (dap == NULL) return (dirrem); } /* * Must be ATTACHED at this point. */ if ((dap->da_state & ATTACHED) == 0) panic("newdirrem: not ATTACHED"); if (dap->da_newinum != ip->i_number) panic("newdirrem: inum %d should be %d", ip->i_number, dap->da_newinum); /* * If we are deleting a changed name that never made it to disk, * then return the dirrem describing the previous inode (which * represents the inode currently referenced from this entry on disk). */ if ((dap->da_state & DIRCHG) != 0) { *prevdirremp = dap->da_previous; dap->da_state &= ~DIRCHG; dap->da_pagedep = pagedep; } /* * We are deleting an entry that never made it to disk. * Mark it COMPLETE so we can delete its inode immediately. */ dirrem->dm_state |= COMPLETE; free_diradd(dap); return (dirrem); } /* * Directory entry change dependencies. * * Changing an existing directory entry requires that an add operation * be completed first followed by a deletion. The semantics for the addition * are identical to the description of adding a new entry above except * that the rollback is to the old inode number rather than zero. Once * the addition dependency is completed, the removal is done as described * in the removal routine above. */ /* * This routine should be called immediately after changing * a directory entry. The inode's link count should not be * decremented by the calling procedure -- the soft updates * code will perform this task when it is safe. */ void softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) struct buf *bp; /* buffer containing directory block */ struct inode *dp; /* inode for the directory being modified */ struct inode *ip; /* inode for directory entry being removed */ ino_t newinum; /* new inode number for changed entry */ int isrmdir; /* indicates if doing RMDIR */ { int offset; struct diradd *dap = NULL; struct dirrem *dirrem, *prevdirrem; struct pagedep *pagedep; struct inodedep *inodedep; struct mount *mp; offset = blkoff(dp->i_fs, dp->i_offset); mp = UFSTOVFS(dp->i_ump); /* * Whiteouts do not need diradd dependencies. */ if (newinum != WINO) { dap = malloc(sizeof(struct diradd), M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); workitem_alloc(&dap->da_list, D_DIRADD, mp); dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; dap->da_offset = offset; dap->da_newinum = newinum; } /* * Allocate a new dirrem and ACQUIRE_LOCK. */ dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); pagedep = dirrem->dm_pagedep; /* * The possible values for isrmdir: * 0 - non-directory file rename * 1 - directory rename within same directory * inum - directory rename to new directory of given inode number * When renaming to a new directory, we are both deleting and * creating a new directory entry, so the link count on the new * directory should not change. Thus we do not need the followup * dirrem which is usually done in handle_workitem_remove. We set * the DIRCHG flag to tell handle_workitem_remove to skip the * followup dirrem. */ if (isrmdir > 1) dirrem->dm_state |= DIRCHG; /* * Whiteouts have no additional dependencies, * so just put the dirrem on the correct list. */ if (newinum == WINO) { if ((dirrem->dm_state & COMPLETE) == 0) { LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, dm_next); } else { dirrem->dm_dirinum = pagedep->pd_ino; add_to_worklist(&dirrem->dm_list); } FREE_LOCK(&lk); return; } /* * If the COMPLETE flag is clear, then there were no active * entries and we want to roll back to the previous inode until * the new inode is committed to disk. If the COMPLETE flag is * set, then we have deleted an entry that never made it to disk. * If the entry we deleted resulted from a name change, then the old * inode reference still resides on disk. Any rollback that we do * needs to be to that old inode (returned to us in prevdirrem). If * the entry we deleted resulted from a create, then there is * no entry on the disk, so we want to roll back to zero rather * than the uncommitted inode. In either of the COMPLETE cases we * want to immediately free the unwritten and unreferenced inode. */ if ((dirrem->dm_state & COMPLETE) == 0) { dap->da_previous = dirrem; } else { if (prevdirrem != NULL) { dap->da_previous = prevdirrem; } else { dap->da_state &= ~DIRCHG; dap->da_pagedep = pagedep; } dirrem->dm_dirinum = pagedep->pd_ino; add_to_worklist(&dirrem->dm_list); } /* * Link into its inodedep. Put it on the id_bufwait list if the inode * is not yet written. If it is written, do the post-inode write * processing to put it on the id_pendinghd list. */ if (inodedep_lookup(mp, newinum, DEPALLOC, &inodedep) == 0 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { dap->da_state |= COMPLETE; LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); } else { LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, da_pdlist); WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); } FREE_LOCK(&lk); } /* * Called whenever the link count on an inode is changed. * It creates an inode dependency so that the new reference(s) * to the inode cannot be committed to disk until the updated * inode has been written. */ void softdep_change_linkcnt(ip) struct inode *ip; /* the inode with the increased link count */ { struct inodedep *inodedep; ACQUIRE_LOCK(&lk); (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, &inodedep); if (ip->i_nlink < ip->i_effnlink) panic("softdep_change_linkcnt: bad delta"); inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; FREE_LOCK(&lk); } /* * Called when the effective link count and the reference count * on an inode drops to zero. At this point there are no names * referencing the file in the filesystem and no active file * references. The space associated with the file will be freed * as soon as the necessary soft dependencies are cleared. */ void softdep_releasefile(ip) struct inode *ip; /* inode with the zero effective link count */ { struct inodedep *inodedep; struct fs *fs; int extblocks; if (ip->i_effnlink > 0) panic("softdep_releasefile: file still referenced"); /* * We may be called several times as the on-disk link count * drops to zero. We only want to account for the space once. */ if (ip->i_flag & IN_SPACECOUNTED) return; /* * We have to deactivate a snapshot otherwise copyonwrites may * add blocks and the cleanup may remove blocks after we have * tried to account for them. */ if ((ip->i_flags & SF_SNAPSHOT) != 0) ffs_snapremove(ITOV(ip)); /* * If we are tracking an nlinkdelta, we have to also remember * whether we accounted for the freed space yet. */ ACQUIRE_LOCK(&lk); if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep))) inodedep->id_state |= SPACECOUNTED; FREE_LOCK(&lk); fs = ip->i_fs; extblocks = 0; if (fs->fs_magic == FS_UFS2_MAGIC) extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); UFS_LOCK(ip->i_ump); ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks; ip->i_fs->fs_pendinginodes += 1; UFS_UNLOCK(ip->i_ump); ip->i_flag |= IN_SPACECOUNTED; } /* * This workitem decrements the inode's link count. * If the link count reaches zero, the file is removed. */ static void handle_workitem_remove(dirrem, xp) struct dirrem *dirrem; struct vnode *xp; { struct thread *td = curthread; struct inodedep *inodedep; struct vnode *vp; struct inode *ip; ino_t oldinum; int error; if ((vp = xp) == NULL && (error = ffs_vgetf(dirrem->dm_list.wk_mp, dirrem->dm_oldinum, LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ)) != 0) { softdep_error("handle_workitem_remove: vget", error); return; } ip = VTOI(vp); ACQUIRE_LOCK(&lk); if ((inodedep_lookup(dirrem->dm_list.wk_mp, dirrem->dm_oldinum, 0, &inodedep)) == 0) panic("handle_workitem_remove: lost inodedep"); /* * Normal file deletion. */ if ((dirrem->dm_state & RMDIR) == 0) { ip->i_nlink--; DIP_SET(ip, i_nlink, ip->i_nlink); ip->i_flag |= IN_CHANGE; if (ip->i_nlink < ip->i_effnlink) panic("handle_workitem_remove: bad file delta"); inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; num_dirrem -= 1; WORKITEM_FREE(dirrem, D_DIRREM); FREE_LOCK(&lk); vput(vp); return; } /* * Directory deletion. Decrement reference count for both the * just deleted parent directory entry and the reference for ".". * Next truncate the directory to length zero. When the * truncation completes, arrange to have the reference count on * the parent decremented to account for the loss of "..". */ ip->i_nlink -= 2; DIP_SET(ip, i_nlink, ip->i_nlink); ip->i_flag |= IN_CHANGE; if (ip->i_nlink < ip->i_effnlink) panic("handle_workitem_remove: bad dir delta"); inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; FREE_LOCK(&lk); if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0) softdep_error("handle_workitem_remove: truncate", error); ACQUIRE_LOCK(&lk); /* * Rename a directory to a new parent. Since, we are both deleting * and creating a new directory entry, the link count on the new * directory should not change. Thus we skip the followup dirrem. */ if (dirrem->dm_state & DIRCHG) { num_dirrem -= 1; WORKITEM_FREE(dirrem, D_DIRREM); FREE_LOCK(&lk); vput(vp); return; } /* * If the inodedep does not exist, then the zero'ed inode has * been written to disk. If the allocated inode has never been * written to disk, then the on-disk inode is zero'ed. In either * case we can remove the file immediately. */ dirrem->dm_state = 0; oldinum = dirrem->dm_oldinum; dirrem->dm_oldinum = dirrem->dm_dirinum; if (inodedep_lookup(dirrem->dm_list.wk_mp, oldinum, 0, &inodedep) == 0 || check_inode_unwritten(inodedep)) { if (xp != NULL) add_to_worklist(&dirrem->dm_list); FREE_LOCK(&lk); vput(vp); if (xp == NULL) handle_workitem_remove(dirrem, NULL); return; } WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); FREE_LOCK(&lk); ip->i_flag |= IN_CHANGE; ffs_update(vp, 0); vput(vp); } /* * Inode de-allocation dependencies. * * When an inode's link count is reduced to zero, it can be de-allocated. We * found it convenient to postpone de-allocation until after the inode is * written to disk with its new link count (zero). At this point, all of the * on-disk inode's block pointers are nullified and, with careful dependency * list ordering, all dependencies related to the inode will be satisfied and * the corresponding dependency structures de-allocated. So, if/when the * inode is reused, there will be no mixing of old dependencies with new * ones. This artificial dependency is set up by the block de-allocation * procedure above (softdep_setup_freeblocks) and completed by the * following procedure. */ static void handle_workitem_freefile(freefile) struct freefile *freefile; { struct fs *fs; struct inodedep *idp; struct ufsmount *ump; int error; ump = VFSTOUFS(freefile->fx_list.wk_mp); fs = ump->um_fs; #ifdef DEBUG ACQUIRE_LOCK(&lk); error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); FREE_LOCK(&lk); if (error) panic("handle_workitem_freefile: inodedep survived"); #endif UFS_LOCK(ump); fs->fs_pendinginodes -= 1; UFS_UNLOCK(ump); if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, freefile->fx_oldinum, freefile->fx_mode)) != 0) softdep_error("handle_workitem_freefile", error); ACQUIRE_LOCK(&lk); WORKITEM_FREE(freefile, D_FREEFILE); FREE_LOCK(&lk); } /* * Helper function which unlinks marker element from work list and returns * the next element on the list. */ static __inline struct worklist * markernext(struct worklist *marker) { struct worklist *next; next = LIST_NEXT(marker, wk_list); LIST_REMOVE(marker, wk_list); return next; } /* * Disk writes. * * The dependency structures constructed above are most actively used when file * system blocks are written to disk. No constraints are placed on when a * block can be written, but unsatisfied update dependencies are made safe by * modifying (or replacing) the source memory for the duration of the disk * write. When the disk write completes, the memory block is again brought * up-to-date. * * In-core inode structure reclamation. * * Because there are a finite number of "in-core" inode structures, they are * reused regularly. By transferring all inode-related dependencies to the * in-memory inode block and indexing them separately (via "inodedep"s), we * can allow "in-core" inode structures to be reused at any time and avoid * any increase in contention. * * Called just before entering the device driver to initiate a new disk I/O. * The buffer must be locked, thus, no I/O completion operations can occur * while we are manipulating its associated dependencies. */ static void softdep_disk_io_initiation(bp) struct buf *bp; /* structure describing disk write to occur */ { struct worklist *wk; struct worklist marker; struct indirdep *indirdep; struct inodedep *inodedep; /* * We only care about write operations. There should never * be dependencies for reads. */ if (bp->b_iocmd != BIO_WRITE) panic("softdep_disk_io_initiation: not write"); marker.wk_type = D_LAST + 1; /* Not a normal workitem */ PHOLD(curproc); /* Don't swap out kernel stack */ ACQUIRE_LOCK(&lk); /* * Do any necessary pre-I/O processing. */ for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; wk = markernext(&marker)) { LIST_INSERT_AFTER(wk, &marker, wk_list); switch (wk->wk_type) { case D_PAGEDEP: initiate_write_filepage(WK_PAGEDEP(wk), bp); continue; case D_INODEDEP: inodedep = WK_INODEDEP(wk); if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) initiate_write_inodeblock_ufs1(inodedep, bp); else initiate_write_inodeblock_ufs2(inodedep, bp); continue; case D_INDIRDEP: indirdep = WK_INDIRDEP(wk); if (indirdep->ir_state & GOINGAWAY) panic("disk_io_initiation: indirdep gone"); /* * If there are no remaining dependencies, this * will be writing the real pointers, so the * dependency can be freed. */ if (LIST_EMPTY(&indirdep->ir_deplisthd)) { struct buf *bp; bp = indirdep->ir_savebp; bp->b_flags |= B_INVAL | B_NOCACHE; /* inline expand WORKLIST_REMOVE(wk); */ wk->wk_state &= ~ONWORKLIST; LIST_REMOVE(wk, wk_list); WORKITEM_FREE(indirdep, D_INDIRDEP); FREE_LOCK(&lk); brelse(bp); ACQUIRE_LOCK(&lk); continue; } /* * Replace up-to-date version with safe version. */ FREE_LOCK(&lk); indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, M_SOFTDEP_FLAGS); ACQUIRE_LOCK(&lk); indirdep->ir_state &= ~ATTACHED; indirdep->ir_state |= UNDONE; bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); bcopy(indirdep->ir_savebp->b_data, bp->b_data, bp->b_bcount); continue; case D_MKDIR: case D_BMSAFEMAP: case D_ALLOCDIRECT: case D_ALLOCINDIR: continue; default: panic("handle_disk_io_initiation: Unexpected type %s", TYPENAME(wk->wk_type)); /* NOTREACHED */ } } FREE_LOCK(&lk); PRELE(curproc); /* Allow swapout of kernel stack */ } /* * Called from within the procedure above to deal with unsatisfied * allocation dependencies in a directory. The buffer must be locked, * thus, no I/O completion operations can occur while we are * manipulating its associated dependencies. */ static void initiate_write_filepage(pagedep, bp) struct pagedep *pagedep; struct buf *bp; { struct diradd *dap; struct direct *ep; int i; if (pagedep->pd_state & IOSTARTED) { /* * This can only happen if there is a driver that does not * understand chaining. Here biodone will reissue the call * to strategy for the incomplete buffers. */ printf("initiate_write_filepage: already started\n"); return; } pagedep->pd_state |= IOSTARTED; for (i = 0; i < DAHASHSZ; i++) { LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { ep = (struct direct *) ((char *)bp->b_data + dap->da_offset); if (ep->d_ino != dap->da_newinum) panic("%s: dir inum %d != new %d", "initiate_write_filepage", ep->d_ino, dap->da_newinum); if (dap->da_state & DIRCHG) ep->d_ino = dap->da_previous->dm_oldinum; else ep->d_ino = 0; dap->da_state &= ~ATTACHED; dap->da_state |= UNDONE; } } } /* * Version of initiate_write_inodeblock that handles UFS1 dinodes. * Note that any bug fixes made to this routine must be done in the * version found below. * * Called from within the procedure above to deal with unsatisfied * allocation dependencies in an inodeblock. The buffer must be * locked, thus, no I/O completion operations can occur while we * are manipulating its associated dependencies. */ static void initiate_write_inodeblock_ufs1(inodedep, bp) struct inodedep *inodedep; struct buf *bp; /* The inode block */ { struct allocdirect *adp, *lastadp; struct ufs1_dinode *dp; struct ufs1_dinode *sip; struct fs *fs; ufs_lbn_t i; #ifdef INVARIANTS ufs_lbn_t prevlbn = 0; #endif int deplist; if (inodedep->id_state & IOSTARTED) panic("initiate_write_inodeblock_ufs1: already started"); inodedep->id_state |= IOSTARTED; fs = inodedep->id_fs; dp = (struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, inodedep->id_ino); /* * If the bitmap is not yet written, then the allocated * inode cannot be written to disk. */ if ((inodedep->id_state & DEPCOMPLETE) == 0) { if (inodedep->id_savedino1 != NULL) panic("initiate_write_inodeblock_ufs1: I/O underway"); FREE_LOCK(&lk); sip = malloc(sizeof(struct ufs1_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); ACQUIRE_LOCK(&lk); inodedep->id_savedino1 = sip; *inodedep->id_savedino1 = *dp; bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); dp->di_gen = inodedep->id_savedino1->di_gen; return; } /* * If no dependencies, then there is nothing to roll back. */ inodedep->id_savedsize = dp->di_size; inodedep->id_savedextsize = 0; if (TAILQ_EMPTY(&inodedep->id_inoupdt)) return; /* * Set the dependencies to busy. */ for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = TAILQ_NEXT(adp, ad_next)) { #ifdef INVARIANTS if (deplist != 0 && prevlbn >= adp->ad_lbn) panic("softdep_write_inodeblock: lbn order"); prevlbn = adp->ad_lbn; if (adp->ad_lbn < NDADDR && dp->di_db[adp->ad_lbn] != adp->ad_newblkno) panic("%s: direct pointer #%jd mismatch %d != %jd", "softdep_write_inodeblock", (intmax_t)adp->ad_lbn, dp->di_db[adp->ad_lbn], (intmax_t)adp->ad_newblkno); if (adp->ad_lbn >= NDADDR && dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) panic("%s: indirect pointer #%jd mismatch %d != %jd", "softdep_write_inodeblock", (intmax_t)adp->ad_lbn - NDADDR, dp->di_ib[adp->ad_lbn - NDADDR], (intmax_t)adp->ad_newblkno); deplist |= 1 << adp->ad_lbn; if ((adp->ad_state & ATTACHED) == 0) panic("softdep_write_inodeblock: Unknown state 0x%x", adp->ad_state); #endif /* INVARIANTS */ adp->ad_state &= ~ATTACHED; adp->ad_state |= UNDONE; } /* * The on-disk inode cannot claim to be any larger than the last * fragment that has been written. Otherwise, the on-disk inode * might have fragments that were not the last block in the file * which would corrupt the filesystem. */ for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { if (adp->ad_lbn >= NDADDR) break; dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; /* keep going until hitting a rollback to a frag */ if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) continue; dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; for (i = adp->ad_lbn + 1; i < NDADDR; i++) { #ifdef INVARIANTS if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) panic("softdep_write_inodeblock: lost dep1"); #endif /* INVARIANTS */ dp->di_db[i] = 0; } for (i = 0; i < NIADDR; i++) { #ifdef INVARIANTS if (dp->di_ib[i] != 0 && (deplist & ((1 << NDADDR) << i)) == 0) panic("softdep_write_inodeblock: lost dep2"); #endif /* INVARIANTS */ dp->di_ib[i] = 0; } return; } /* * If we have zero'ed out the last allocated block of the file, * roll back the size to the last currently allocated block. * We know that this last allocated block is a full-sized as * we already checked for fragments in the loop above. */ if (lastadp != NULL && dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { for (i = lastadp->ad_lbn; i >= 0; i--) if (dp->di_db[i] != 0) break; dp->di_size = (i + 1) * fs->fs_bsize; } /* * The only dependencies are for indirect blocks. * * The file size for indirect block additions is not guaranteed. * Such a guarantee would be non-trivial to achieve. The conventional * synchronous write implementation also does not make this guarantee. * Fsck should catch and fix discrepancies. Arguably, the file size * can be over-estimated without destroying integrity when the file * moves into the indirect blocks (i.e., is large). If we want to * postpone fsck, we are stuck with this argument. */ for (; adp; adp = TAILQ_NEXT(adp, ad_next)) dp->di_ib[adp->ad_lbn - NDADDR] = 0; } /* * Version of initiate_write_inodeblock that handles UFS2 dinodes. * Note that any bug fixes made to this routine must be done in the * version found above. * * Called from within the procedure above to deal with unsatisfied * allocation dependencies in an inodeblock. The buffer must be * locked, thus, no I/O completion operations can occur while we * are manipulating its associated dependencies. */ static void initiate_write_inodeblock_ufs2(inodedep, bp) struct inodedep *inodedep; struct buf *bp; /* The inode block */ { struct allocdirect *adp, *lastadp; struct ufs2_dinode *dp; struct ufs2_dinode *sip; struct fs *fs; ufs_lbn_t i; #ifdef INVARIANTS ufs_lbn_t prevlbn = 0; #endif int deplist; if (inodedep->id_state & IOSTARTED) panic("initiate_write_inodeblock_ufs2: already started"); inodedep->id_state |= IOSTARTED; fs = inodedep->id_fs; dp = (struct ufs2_dinode *)bp->b_data + ino_to_fsbo(fs, inodedep->id_ino); /* * If the bitmap is not yet written, then the allocated * inode cannot be written to disk. */ if ((inodedep->id_state & DEPCOMPLETE) == 0) { if (inodedep->id_savedino2 != NULL) panic("initiate_write_inodeblock_ufs2: I/O underway"); FREE_LOCK(&lk); sip = malloc(sizeof(struct ufs2_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); ACQUIRE_LOCK(&lk); inodedep->id_savedino2 = sip; *inodedep->id_savedino2 = *dp; bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); dp->di_gen = inodedep->id_savedino2->di_gen; return; } /* * If no dependencies, then there is nothing to roll back. */ inodedep->id_savedsize = dp->di_size; inodedep->id_savedextsize = dp->di_extsize; if (TAILQ_EMPTY(&inodedep->id_inoupdt) && TAILQ_EMPTY(&inodedep->id_extupdt)) return; /* * Set the ext data dependencies to busy. */ for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = TAILQ_NEXT(adp, ad_next)) { #ifdef INVARIANTS if (deplist != 0 && prevlbn >= adp->ad_lbn) panic("softdep_write_inodeblock: lbn order"); prevlbn = adp->ad_lbn; if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) panic("%s: direct pointer #%jd mismatch %jd != %jd", "softdep_write_inodeblock", (intmax_t)adp->ad_lbn, (intmax_t)dp->di_extb[adp->ad_lbn], (intmax_t)adp->ad_newblkno); deplist |= 1 << adp->ad_lbn; if ((adp->ad_state & ATTACHED) == 0) panic("softdep_write_inodeblock: Unknown state 0x%x", adp->ad_state); #endif /* INVARIANTS */ adp->ad_state &= ~ATTACHED; adp->ad_state |= UNDONE; } /* * The on-disk inode cannot claim to be any larger than the last * fragment that has been written. Otherwise, the on-disk inode * might have fragments that were not the last block in the ext * data which would corrupt the filesystem. */ for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno; /* keep going until hitting a rollback to a frag */ if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) continue; dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; for (i = adp->ad_lbn + 1; i < NXADDR; i++) { #ifdef INVARIANTS if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) panic("softdep_write_inodeblock: lost dep1"); #endif /* INVARIANTS */ dp->di_extb[i] = 0; } lastadp = NULL; break; } /* * If we have zero'ed out the last allocated block of the ext * data, roll back the size to the last currently allocated block. * We know that this last allocated block is a full-sized as * we already checked for fragments in the loop above. */ if (lastadp != NULL && dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { for (i = lastadp->ad_lbn; i >= 0; i--) if (dp->di_extb[i] != 0) break; dp->di_extsize = (i + 1) * fs->fs_bsize; } /* * Set the file data dependencies to busy. */ for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = TAILQ_NEXT(adp, ad_next)) { #ifdef INVARIANTS if (deplist != 0 && prevlbn >= adp->ad_lbn) panic("softdep_write_inodeblock: lbn order"); prevlbn = adp->ad_lbn; if (adp->ad_lbn < NDADDR && dp->di_db[adp->ad_lbn] != adp->ad_newblkno) panic("%s: direct pointer #%jd mismatch %jd != %jd", "softdep_write_inodeblock", (intmax_t)adp->ad_lbn, (intmax_t)dp->di_db[adp->ad_lbn], (intmax_t)adp->ad_newblkno); if (adp->ad_lbn >= NDADDR && dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) panic("%s indirect pointer #%jd mismatch %jd != %jd", "softdep_write_inodeblock:", (intmax_t)adp->ad_lbn - NDADDR, (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR], (intmax_t)adp->ad_newblkno); deplist |= 1 << adp->ad_lbn; if ((adp->ad_state & ATTACHED) == 0) panic("softdep_write_inodeblock: Unknown state 0x%x", adp->ad_state); #endif /* INVARIANTS */ adp->ad_state &= ~ATTACHED; adp->ad_state |= UNDONE; } /* * The on-disk inode cannot claim to be any larger than the last * fragment that has been written. Otherwise, the on-disk inode * might have fragments that were not the last block in the file * which would corrupt the filesystem. */ for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { if (adp->ad_lbn >= NDADDR) break; dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; /* keep going until hitting a rollback to a frag */ if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) continue; dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; for (i = adp->ad_lbn + 1; i < NDADDR; i++) { #ifdef INVARIANTS if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) panic("softdep_write_inodeblock: lost dep2"); #endif /* INVARIANTS */ dp->di_db[i] = 0; } for (i = 0; i < NIADDR; i++) { #ifdef INVARIANTS if (dp->di_ib[i] != 0 && (deplist & ((1 << NDADDR) << i)) == 0) panic("softdep_write_inodeblock: lost dep3"); #endif /* INVARIANTS */ dp->di_ib[i] = 0; } return; } /* * If we have zero'ed out the last allocated block of the file, * roll back the size to the last currently allocated block. * We know that this last allocated block is a full-sized as * we already checked for fragments in the loop above. */ if (lastadp != NULL && dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { for (i = lastadp->ad_lbn; i >= 0; i--) if (dp->di_db[i] != 0) break; dp->di_size = (i + 1) * fs->fs_bsize; } /* * The only dependencies are for indirect blocks. * * The file size for indirect block additions is not guaranteed. * Such a guarantee would be non-trivial to achieve. The conventional * synchronous write implementation also does not make this guarantee. * Fsck should catch and fix discrepancies. Arguably, the file size * can be over-estimated without destroying integrity when the file * moves into the indirect blocks (i.e., is large). If we want to * postpone fsck, we are stuck with this argument. */ for (; adp; adp = TAILQ_NEXT(adp, ad_next)) dp->di_ib[adp->ad_lbn - NDADDR] = 0; } /* * This routine is called during the completion interrupt * service routine for a disk write (from the procedure called * by the device driver to inform the filesystem caches of * a request completion). It should be called early in this * procedure, before the block is made available to other * processes or other routines are called. */ static void softdep_disk_write_complete(bp) struct buf *bp; /* describes the completed disk write */ { struct worklist *wk; struct worklist *owk; struct workhead reattach; struct newblk *newblk; struct allocindir *aip; struct allocdirect *adp; struct indirdep *indirdep; struct inodedep *inodedep; struct bmsafemap *bmsafemap; /* * If an error occurred while doing the write, then the data * has not hit the disk and the dependencies cannot be unrolled. */ if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) return; LIST_INIT(&reattach); /* * This lock must not be released anywhere in this code segment. */ ACQUIRE_LOCK(&lk); owk = NULL; while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { WORKLIST_REMOVE(wk); if (wk == owk) panic("duplicate worklist: %p\n", wk); owk = wk; switch (wk->wk_type) { case D_PAGEDEP: if (handle_written_filepage(WK_PAGEDEP(wk), bp)) WORKLIST_INSERT(&reattach, wk); continue; case D_INODEDEP: if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) WORKLIST_INSERT(&reattach, wk); continue; case D_BMSAFEMAP: bmsafemap = WK_BMSAFEMAP(wk); while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { newblk->nb_state |= DEPCOMPLETE; newblk->nb_bmsafemap = NULL; LIST_REMOVE(newblk, nb_deps); } while ((adp = LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { adp->ad_state |= DEPCOMPLETE; adp->ad_buf = NULL; LIST_REMOVE(adp, ad_deps); handle_allocdirect_partdone(adp); } while ((aip = LIST_FIRST(&bmsafemap->sm_allocindirhd))) { aip->ai_state |= DEPCOMPLETE; aip->ai_buf = NULL; LIST_REMOVE(aip, ai_deps); handle_allocindir_partdone(aip); } while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { inodedep->id_state |= DEPCOMPLETE; LIST_REMOVE(inodedep, id_deps); inodedep->id_buf = NULL; } WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); continue; case D_MKDIR: handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); continue; case D_ALLOCDIRECT: adp = WK_ALLOCDIRECT(wk); adp->ad_state |= COMPLETE; handle_allocdirect_partdone(adp); continue; case D_ALLOCINDIR: aip = WK_ALLOCINDIR(wk); aip->ai_state |= COMPLETE; handle_allocindir_partdone(aip); continue; case D_INDIRDEP: indirdep = WK_INDIRDEP(wk); if (indirdep->ir_state & GOINGAWAY) panic("disk_write_complete: indirdep gone"); bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); free(indirdep->ir_saveddata, M_INDIRDEP); indirdep->ir_saveddata = 0; indirdep->ir_state &= ~UNDONE; indirdep->ir_state |= ATTACHED; while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { handle_allocindir_partdone(aip); if (aip == LIST_FIRST(&indirdep->ir_donehd)) panic("disk_write_complete: not gone"); } WORKLIST_INSERT(&reattach, wk); if ((bp->b_flags & B_DELWRI) == 0) stat_indir_blk_ptrs++; bdirty(bp); continue; default: panic("handle_disk_write_complete: Unknown type %s", TYPENAME(wk->wk_type)); /* NOTREACHED */ } } /* * Reattach any requests that must be redone. */ while ((wk = LIST_FIRST(&reattach)) != NULL) { WORKLIST_REMOVE(wk); WORKLIST_INSERT(&bp->b_dep, wk); } FREE_LOCK(&lk); } /* * Called from within softdep_disk_write_complete above. Note that * this routine is always called from interrupt level with further * splbio interrupts blocked. */ static void handle_allocdirect_partdone(adp) struct allocdirect *adp; /* the completed allocdirect */ { struct allocdirectlst *listhead; struct allocdirect *listadp; struct inodedep *inodedep; long bsize, delay; if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) return; if (adp->ad_buf != NULL) panic("handle_allocdirect_partdone: dangling dep"); /* * The on-disk inode cannot claim to be any larger than the last * fragment that has been written. Otherwise, the on-disk inode * might have fragments that were not the last block in the file * which would corrupt the filesystem. Thus, we cannot free any * allocdirects after one whose ad_oldblkno claims a fragment as * these blocks must be rolled back to zero before writing the inode. * We check the currently active set of allocdirects in id_inoupdt * or id_extupdt as appropriate. */ inodedep = adp->ad_inodedep; bsize = inodedep->id_fs->fs_bsize; if (adp->ad_state & EXTDATA) listhead = &inodedep->id_extupdt; else listhead = &inodedep->id_inoupdt; TAILQ_FOREACH(listadp, listhead, ad_next) { /* found our block */ if (listadp == adp) break; /* continue if ad_oldlbn is not a fragment */ if (listadp->ad_oldsize == 0 || listadp->ad_oldsize == bsize) continue; /* hit a fragment */ return; } /* * If we have reached the end of the current list without * finding the just finished dependency, then it must be * on the future dependency list. Future dependencies cannot * be freed until they are moved to the current list. */ if (listadp == NULL) { #ifdef DEBUG if (adp->ad_state & EXTDATA) listhead = &inodedep->id_newextupdt; else listhead = &inodedep->id_newinoupdt; TAILQ_FOREACH(listadp, listhead, ad_next) /* found our block */ if (listadp == adp) break; if (listadp == NULL) panic("handle_allocdirect_partdone: lost dep"); #endif /* DEBUG */ return; } /* * If we have found the just finished dependency, then free * it along with anything that follows it that is complete. * If the inode still has a bitmap dependency, then it has * never been written to disk, hence the on-disk inode cannot * reference the old fragment so we can free it without delay. */ delay = (inodedep->id_state & DEPCOMPLETE); for (; adp; adp = listadp) { listadp = TAILQ_NEXT(adp, ad_next); if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) return; free_allocdirect(listhead, adp, delay); } } /* * Called from within softdep_disk_write_complete above. Note that * this routine is always called from interrupt level with further * splbio interrupts blocked. */ static void handle_allocindir_partdone(aip) struct allocindir *aip; /* the completed allocindir */ { struct indirdep *indirdep; if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) return; if (aip->ai_buf != NULL) panic("handle_allocindir_partdone: dangling dependency"); indirdep = aip->ai_indirdep; if (indirdep->ir_state & UNDONE) { LIST_REMOVE(aip, ai_next); LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); return; } if (indirdep->ir_state & UFS1FMT) ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = aip->ai_newblkno; else ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = aip->ai_newblkno; LIST_REMOVE(aip, ai_next); if (aip->ai_freefrag != NULL) add_to_worklist(&aip->ai_freefrag->ff_list); WORKITEM_FREE(aip, D_ALLOCINDIR); } /* * Called from within softdep_disk_write_complete above to restore * in-memory inode block contents to their most up-to-date state. Note * that this routine is always called from interrupt level with further * splbio interrupts blocked. */ static int handle_written_inodeblock(inodedep, bp) struct inodedep *inodedep; struct buf *bp; /* buffer containing the inode block */ { struct worklist *wk, *filefree; struct allocdirect *adp, *nextadp; struct ufs1_dinode *dp1 = NULL; struct ufs2_dinode *dp2 = NULL; int hadchanges, fstype; if ((inodedep->id_state & IOSTARTED) == 0) panic("handle_written_inodeblock: not started"); inodedep->id_state &= ~IOSTARTED; if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { fstype = UFS1; dp1 = (struct ufs1_dinode *)bp->b_data + ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); } else { fstype = UFS2; dp2 = (struct ufs2_dinode *)bp->b_data + ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); } /* * If we had to rollback the inode allocation because of * bitmaps being incomplete, then simply restore it. * Keep the block dirty so that it will not be reclaimed until * all associated dependencies have been cleared and the * corresponding updates written to disk. */ if (inodedep->id_savedino1 != NULL) { if (fstype == UFS1) *dp1 = *inodedep->id_savedino1; else *dp2 = *inodedep->id_savedino2; free(inodedep->id_savedino1, M_SAVEDINO); inodedep->id_savedino1 = NULL; if ((bp->b_flags & B_DELWRI) == 0) stat_inode_bitmap++; bdirty(bp); return (1); } inodedep->id_state |= COMPLETE; /* * Roll forward anything that had to be rolled back before * the inode could be updated. */ hadchanges = 0; for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { nextadp = TAILQ_NEXT(adp, ad_next); if (adp->ad_state & ATTACHED) panic("handle_written_inodeblock: new entry"); if (fstype == UFS1) { if (adp->ad_lbn < NDADDR) { if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) panic("%s %s #%jd mismatch %d != %jd", "handle_written_inodeblock:", "direct pointer", (intmax_t)adp->ad_lbn, dp1->di_db[adp->ad_lbn], (intmax_t)adp->ad_oldblkno); dp1->di_db[adp->ad_lbn] = adp->ad_newblkno; } else { if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) panic("%s: %s #%jd allocated as %d", "handle_written_inodeblock", "indirect pointer", (intmax_t)adp->ad_lbn - NDADDR, dp1->di_ib[adp->ad_lbn - NDADDR]); dp1->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; } } else { if (adp->ad_lbn < NDADDR) { if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) panic("%s: %s #%jd %s %jd != %jd", "handle_written_inodeblock", "direct pointer", (intmax_t)adp->ad_lbn, "mismatch", (intmax_t)dp2->di_db[adp->ad_lbn], (intmax_t)adp->ad_oldblkno); dp2->di_db[adp->ad_lbn] = adp->ad_newblkno; } else { if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) panic("%s: %s #%jd allocated as %jd", "handle_written_inodeblock", "indirect pointer", (intmax_t)adp->ad_lbn - NDADDR, (intmax_t) dp2->di_ib[adp->ad_lbn - NDADDR]); dp2->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; } } adp->ad_state &= ~UNDONE; adp->ad_state |= ATTACHED; hadchanges = 1; } for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { nextadp = TAILQ_NEXT(adp, ad_next); if (adp->ad_state & ATTACHED) panic("handle_written_inodeblock: new entry"); if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) panic("%s: direct pointers #%jd %s %jd != %jd", "handle_written_inodeblock", (intmax_t)adp->ad_lbn, "mismatch", (intmax_t)dp2->di_extb[adp->ad_lbn], (intmax_t)adp->ad_oldblkno); dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno; adp->ad_state &= ~UNDONE; adp->ad_state |= ATTACHED; hadchanges = 1; } if (hadchanges && (bp->b_flags & B_DELWRI) == 0) stat_direct_blk_ptrs++; /* * Reset the file size to its most up-to-date value. */ if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) panic("handle_written_inodeblock: bad size"); if (fstype == UFS1) { if (dp1->di_size != inodedep->id_savedsize) { dp1->di_size = inodedep->id_savedsize; hadchanges = 1; } } else { if (dp2->di_size != inodedep->id_savedsize) { dp2->di_size = inodedep->id_savedsize; hadchanges = 1; } if (dp2->di_extsize != inodedep->id_savedextsize) { dp2->di_extsize = inodedep->id_savedextsize; hadchanges = 1; } } inodedep->id_savedsize = -1; inodedep->id_savedextsize = -1; /* * If there were any rollbacks in the inode block, then it must be * marked dirty so that its will eventually get written back in * its correct form. */ if (hadchanges) bdirty(bp); /* * Process any allocdirects that completed during the update. */ if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) handle_allocdirect_partdone(adp); if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) handle_allocdirect_partdone(adp); /* * Process deallocations that were held pending until the * inode had been written to disk. Freeing of the inode * is delayed until after all blocks have been freed to * avoid creation of new triples * before the old ones have been deleted. */ filefree = NULL; while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { WORKLIST_REMOVE(wk); switch (wk->wk_type) { case D_FREEFILE: /* * We defer adding filefree to the worklist until * all other additions have been made to ensure * that it will be done after all the old blocks * have been freed. */ if (filefree != NULL) panic("handle_written_inodeblock: filefree"); filefree = wk; continue; case D_MKDIR: handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); continue; case D_DIRADD: diradd_inode_written(WK_DIRADD(wk), inodedep); continue; case D_FREEBLKS: wk->wk_state |= COMPLETE; if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) continue; /* -- fall through -- */ case D_FREEFRAG: case D_DIRREM: add_to_worklist(wk); continue; case D_NEWDIRBLK: free_newdirblk(WK_NEWDIRBLK(wk)); continue; default: panic("handle_written_inodeblock: Unknown type %s", TYPENAME(wk->wk_type)); /* NOTREACHED */ } } if (filefree != NULL) { if (free_inodedep(inodedep) == 0) panic("handle_written_inodeblock: live inodedep"); add_to_worklist(filefree); return (0); } /* * If no outstanding dependencies, free it. */ if (free_inodedep(inodedep) || (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && TAILQ_FIRST(&inodedep->id_extupdt) == 0)) return (0); return (hadchanges); } /* * Process a diradd entry after its dependent inode has been written. * This routine must be called with splbio interrupts blocked. */ static void diradd_inode_written(dap, inodedep) struct diradd *dap; struct inodedep *inodedep; { struct pagedep *pagedep; dap->da_state |= COMPLETE; if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { if (dap->da_state & DIRCHG) pagedep = dap->da_previous->dm_pagedep; else pagedep = dap->da_pagedep; LIST_REMOVE(dap, da_pdlist); LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); } WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); } /* * Handle the completion of a mkdir dependency. */ static void handle_written_mkdir(mkdir, type) struct mkdir *mkdir; int type; { struct diradd *dap; struct pagedep *pagedep; if (mkdir->md_state != type) panic("handle_written_mkdir: bad type"); dap = mkdir->md_diradd; dap->da_state &= ~type; if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) dap->da_state |= DEPCOMPLETE; if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { if (dap->da_state & DIRCHG) pagedep = dap->da_previous->dm_pagedep; else pagedep = dap->da_pagedep; LIST_REMOVE(dap, da_pdlist); LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); } LIST_REMOVE(mkdir, md_mkdirs); WORKITEM_FREE(mkdir, D_MKDIR); } /* * Called from within softdep_disk_write_complete above. * A write operation was just completed. Removed inodes can * now be freed and associated block pointers may be committed. * Note that this routine is always called from interrupt level * with further splbio interrupts blocked. */ static int handle_written_filepage(pagedep, bp) struct pagedep *pagedep; struct buf *bp; /* buffer containing the written page */ { struct dirrem *dirrem; struct diradd *dap, *nextdap; struct direct *ep; int i, chgs; if ((pagedep->pd_state & IOSTARTED) == 0) panic("handle_written_filepage: not started"); pagedep->pd_state &= ~IOSTARTED; /* * Process any directory removals that have been committed. */ while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { LIST_REMOVE(dirrem, dm_next); dirrem->dm_dirinum = pagedep->pd_ino; add_to_worklist(&dirrem->dm_list); } /* * Free any directory additions that have been committed. * If it is a newly allocated block, we have to wait until * the on-disk directory inode claims the new block. */ if ((pagedep->pd_state & NEWBLOCK) == 0) while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) free_diradd(dap); /* * Uncommitted directory entries must be restored. */ for (chgs = 0, i = 0; i < DAHASHSZ; i++) { for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; dap = nextdap) { nextdap = LIST_NEXT(dap, da_pdlist); if (dap->da_state & ATTACHED) panic("handle_written_filepage: attached"); ep = (struct direct *) ((char *)bp->b_data + dap->da_offset); ep->d_ino = dap->da_newinum; dap->da_state &= ~UNDONE; dap->da_state |= ATTACHED; chgs = 1; /* * If the inode referenced by the directory has * been written out, then the dependency can be * moved to the pending list. */ if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { LIST_REMOVE(dap, da_pdlist); LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); } } } /* * If there were any rollbacks in the directory, then it must be * marked dirty so that its will eventually get written back in * its correct form. */ if (chgs) { if ((bp->b_flags & B_DELWRI) == 0) stat_dir_entry++; bdirty(bp); return (1); } /* * If we are not waiting for a new directory block to be * claimed by its inode, then the pagedep will be freed. * Otherwise it will remain to track any new entries on * the page in case they are fsync'ed. */ if ((pagedep->pd_state & NEWBLOCK) == 0) { LIST_REMOVE(pagedep, pd_hash); WORKITEM_FREE(pagedep, D_PAGEDEP); } return (0); } /* * Writing back in-core inode structures. * * The filesystem only accesses an inode's contents when it occupies an * "in-core" inode structure. These "in-core" structures are separate from * the page frames used to cache inode blocks. Only the latter are * transferred to/from the disk. So, when the updated contents of the * "in-core" inode structure are copied to the corresponding in-memory inode * block, the dependencies are also transferred. The following procedure is * called when copying a dirty "in-core" inode to a cached inode block. */ /* * Called when an inode is loaded from disk. If the effective link count * differed from the actual link count when it was last flushed, then we * need to ensure that the correct effective link count is put back. */ void softdep_load_inodeblock(ip) struct inode *ip; /* the "in_core" copy of the inode */ { struct inodedep *inodedep; /* * Check for alternate nlink count. */ ip->i_effnlink = ip->i_nlink; ACQUIRE_LOCK(&lk); if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep) == 0) { FREE_LOCK(&lk); return; } ip->i_effnlink -= inodedep->id_nlinkdelta; if (inodedep->id_state & SPACECOUNTED) ip->i_flag |= IN_SPACECOUNTED; FREE_LOCK(&lk); } /* * This routine is called just before the "in-core" inode * information is to be copied to the in-memory inode block. * Recall that an inode block contains several inodes. If * the force flag is set, then the dependencies will be * cleared so that the update can always be made. Note that * the buffer is locked when this routine is called, so we * will never be in the middle of writing the inode block * to disk. */ void softdep_update_inodeblock(ip, bp, waitfor) struct inode *ip; /* the "in_core" copy of the inode */ struct buf *bp; /* the buffer containing the inode block */ int waitfor; /* nonzero => update must be allowed */ { struct inodedep *inodedep; struct worklist *wk; struct mount *mp; struct buf *ibp; int error; /* * If the effective link count is not equal to the actual link * count, then we must track the difference in an inodedep while * the inode is (potentially) tossed out of the cache. Otherwise, * if there is no existing inodedep, then there are no dependencies * to track. */ mp = UFSTOVFS(ip->i_ump); ACQUIRE_LOCK(&lk); if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { FREE_LOCK(&lk); if (ip->i_effnlink != ip->i_nlink) panic("softdep_update_inodeblock: bad link count"); return; } if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) panic("softdep_update_inodeblock: bad delta"); /* * Changes have been initiated. Anything depending on these * changes cannot occur until this inode has been written. */ inodedep->id_state &= ~COMPLETE; if ((inodedep->id_state & ONWORKLIST) == 0) WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); /* * Any new dependencies associated with the incore inode must * now be moved to the list associated with the buffer holding * the in-memory copy of the inode. Once merged process any * allocdirects that are completed by the merger. */ merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); if (!TAILQ_EMPTY(&inodedep->id_extupdt)) handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt)); /* * Now that the inode has been pushed into the buffer, the * operations dependent on the inode being written to disk * can be moved to the id_bufwait so that they will be * processed when the buffer I/O completes. */ while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { WORKLIST_REMOVE(wk); WORKLIST_INSERT(&inodedep->id_bufwait, wk); } /* * Newly allocated inodes cannot be written until the bitmap * that allocates them have been written (indicated by * DEPCOMPLETE being set in id_state). If we are doing a * forced sync (e.g., an fsync on a file), we force the bitmap * to be written so that the update can be done. */ if (waitfor == 0) { FREE_LOCK(&lk); return; } retry: if ((inodedep->id_state & DEPCOMPLETE) != 0) { FREE_LOCK(&lk); return; } ibp = inodedep->id_buf; ibp = getdirtybuf(ibp, &lk, MNT_WAIT); if (ibp == NULL) { /* * If ibp came back as NULL, the dependency could have been * freed while we slept. Look it up again, and check to see * that it has completed. */ if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) goto retry; FREE_LOCK(&lk); return; } FREE_LOCK(&lk); if ((error = bwrite(ibp)) != 0) softdep_error("softdep_update_inodeblock: bwrite", error); } /* * Merge the a new inode dependency list (such as id_newinoupdt) into an * old inode dependency list (such as id_inoupdt). This routine must be * called with splbio interrupts blocked. */ static void merge_inode_lists(newlisthead, oldlisthead) struct allocdirectlst *newlisthead; struct allocdirectlst *oldlisthead; { struct allocdirect *listadp, *newadp; newadp = TAILQ_FIRST(newlisthead); for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { if (listadp->ad_lbn < newadp->ad_lbn) { listadp = TAILQ_NEXT(listadp, ad_next); continue; } TAILQ_REMOVE(newlisthead, newadp, ad_next); TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); if (listadp->ad_lbn == newadp->ad_lbn) { allocdirect_merge(oldlisthead, newadp, listadp); listadp = newadp; } newadp = TAILQ_FIRST(newlisthead); } while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { TAILQ_REMOVE(newlisthead, newadp, ad_next); TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); } } /* * If we are doing an fsync, then we must ensure that any directory * entries for the inode have been written after the inode gets to disk. */ int softdep_fsync(vp) struct vnode *vp; /* the "in_core" copy of the inode */ { struct inodedep *inodedep; struct pagedep *pagedep; struct worklist *wk; struct diradd *dap; struct mount *mp; struct vnode *pvp; struct inode *ip; struct buf *bp; struct fs *fs; struct thread *td = curthread; int error, flushparent, pagedep_new_block; ino_t parentino; ufs_lbn_t lbn; ip = VTOI(vp); fs = ip->i_fs; mp = vp->v_mount; ACQUIRE_LOCK(&lk); if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { FREE_LOCK(&lk); return (0); } if (!LIST_EMPTY(&inodedep->id_inowait) || !LIST_EMPTY(&inodedep->id_bufwait) || !TAILQ_EMPTY(&inodedep->id_extupdt) || !TAILQ_EMPTY(&inodedep->id_newextupdt) || !TAILQ_EMPTY(&inodedep->id_inoupdt) || !TAILQ_EMPTY(&inodedep->id_newinoupdt)) panic("softdep_fsync: pending ops"); for (error = 0, flushparent = 0; ; ) { if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) break; if (wk->wk_type != D_DIRADD) panic("softdep_fsync: Unexpected type %s", TYPENAME(wk->wk_type)); dap = WK_DIRADD(wk); /* * Flush our parent if this directory entry has a MKDIR_PARENT * dependency or is contained in a newly allocated block. */ if (dap->da_state & DIRCHG) pagedep = dap->da_previous->dm_pagedep; else pagedep = dap->da_pagedep; parentino = pagedep->pd_ino; lbn = pagedep->pd_lbn; if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) panic("softdep_fsync: dirty"); if ((dap->da_state & MKDIR_PARENT) || (pagedep->pd_state & NEWBLOCK)) flushparent = 1; else flushparent = 0; /* * If we are being fsync'ed as part of vgone'ing this vnode, * then we will not be able to release and recover the * vnode below, so we just have to give up on writing its * directory entry out. It will eventually be written, just * not now, but then the user was not asking to have it * written, so we are not breaking any promises. */ if (vp->v_iflag & VI_DOOMED) break; /* * We prevent deadlock by always fetching inodes from the * root, moving down the directory tree. Thus, when fetching * our parent directory, we first try to get the lock. If * that fails, we must unlock ourselves before requesting * the lock on our parent. See the comment in ufs_lookup * for details on possible races. */ FREE_LOCK(&lk); if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, FFSV_FORCEINSMQ)) { error = vfs_busy(mp, MBF_NOWAIT); if (error != 0) { + vfs_ref(mp); VOP_UNLOCK(vp, 0); error = vfs_busy(mp, 0); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); + vfs_rel(mp); if (error != 0) return (ENOENT); if (vp->v_iflag & VI_DOOMED) { vfs_unbusy(mp); return (ENOENT); } } VOP_UNLOCK(vp, 0); error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, &pvp, FFSV_FORCEINSMQ); vfs_unbusy(mp); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); if (vp->v_iflag & VI_DOOMED) { if (error == 0) vput(pvp); error = ENOENT; } if (error != 0) return (error); } /* * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps * that are contained in direct blocks will be resolved by * doing a ffs_update. Pagedeps contained in indirect blocks * may require a complete sync'ing of the directory. So, we * try the cheap and fast ffs_update first, and if that fails, * then we do the slower ffs_syncvnode of the directory. */ if (flushparent) { int locked; if ((error = ffs_update(pvp, 1)) != 0) { vput(pvp); return (error); } ACQUIRE_LOCK(&lk); locked = 1; if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { if (wk->wk_type != D_DIRADD) panic("softdep_fsync: Unexpected type %s", TYPENAME(wk->wk_type)); dap = WK_DIRADD(wk); if (dap->da_state & DIRCHG) pagedep = dap->da_previous->dm_pagedep; else pagedep = dap->da_pagedep; pagedep_new_block = pagedep->pd_state & NEWBLOCK; FREE_LOCK(&lk); locked = 0; if (pagedep_new_block && (error = ffs_syncvnode(pvp, MNT_WAIT))) { vput(pvp); return (error); } } } if (locked) FREE_LOCK(&lk); } /* * Flush directory page containing the inode's name. */ error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, &bp); if (error == 0) error = bwrite(bp); else brelse(bp); vput(pvp); if (error != 0) return (error); ACQUIRE_LOCK(&lk); if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) break; } FREE_LOCK(&lk); return (0); } /* * Flush all the dirty bitmaps associated with the block device * before flushing the rest of the dirty blocks so as to reduce * the number of dependencies that will have to be rolled back. */ void softdep_fsync_mountdev(vp) struct vnode *vp; { struct buf *bp, *nbp; struct worklist *wk; struct bufobj *bo; if (!vn_isdisk(vp, NULL)) panic("softdep_fsync_mountdev: vnode not a disk"); bo = &vp->v_bufobj; restart: BO_LOCK(bo); ACQUIRE_LOCK(&lk); TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { /* * If it is already scheduled, skip to the next buffer. */ if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) continue; if ((bp->b_flags & B_DELWRI) == 0) panic("softdep_fsync_mountdev: not dirty"); /* * We are only interested in bitmaps with outstanding * dependencies. */ if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || wk->wk_type != D_BMSAFEMAP || (bp->b_vflags & BV_BKGRDINPROG)) { BUF_UNLOCK(bp); continue; } FREE_LOCK(&lk); BO_UNLOCK(bo); bremfree(bp); (void) bawrite(bp); goto restart; } FREE_LOCK(&lk); drain_output(vp); BO_UNLOCK(bo); } /* * This routine is called when we are trying to synchronously flush a * file. This routine must eliminate any filesystem metadata dependencies * so that the syncing routine can succeed by pushing the dirty blocks * associated with the file. If any I/O errors occur, they are returned. */ int softdep_sync_metadata(struct vnode *vp) { struct pagedep *pagedep; struct allocdirect *adp; struct allocindir *aip; struct buf *bp, *nbp; struct worklist *wk; struct bufobj *bo; int i, error, waitfor; if (!DOINGSOFTDEP(vp)) return (0); /* * Ensure that any direct block dependencies have been cleared. */ ACQUIRE_LOCK(&lk); if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) { FREE_LOCK(&lk); return (error); } FREE_LOCK(&lk); /* * For most files, the only metadata dependencies are the * cylinder group maps that allocate their inode or blocks. * The block allocation dependencies can be found by traversing * the dependency lists for any buffers that remain on their * dirty buffer list. The inode allocation dependency will * be resolved when the inode is updated with MNT_WAIT. * This work is done in two passes. The first pass grabs most * of the buffers and begins asynchronously writing them. The * only way to wait for these asynchronous writes is to sleep * on the filesystem vnode which may stay busy for a long time * if the filesystem is active. So, instead, we make a second * pass over the dependencies blocking on each write. In the * usual case we will be blocking against a write that we * initiated, so when it is done the dependency will have been * resolved. Thus the second pass is expected to end quickly. */ waitfor = MNT_NOWAIT; bo = &vp->v_bufobj; top: /* * We must wait for any I/O in progress to finish so that * all potential buffers on the dirty list will be visible. */ BO_LOCK(bo); drain_output(vp); while ((bp = TAILQ_FIRST(&bo->bo_dirty.bv_hd)) != NULL) { bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT); if (bp) break; } BO_UNLOCK(bo); if (bp == NULL) return (0); loop: /* While syncing snapshots, we must allow recursive lookups */ BUF_AREC(bp); ACQUIRE_LOCK(&lk); /* * As we hold the buffer locked, none of its dependencies * will disappear. */ LIST_FOREACH(wk, &bp->b_dep, wk_list) { switch (wk->wk_type) { case D_ALLOCDIRECT: adp = WK_ALLOCDIRECT(wk); if (adp->ad_state & DEPCOMPLETE) continue; nbp = adp->ad_buf; nbp = getdirtybuf(nbp, &lk, waitfor); if (nbp == NULL) continue; FREE_LOCK(&lk); if (waitfor == MNT_NOWAIT) { bawrite(nbp); } else if ((error = bwrite(nbp)) != 0) { break; } ACQUIRE_LOCK(&lk); continue; case D_ALLOCINDIR: aip = WK_ALLOCINDIR(wk); if (aip->ai_state & DEPCOMPLETE) continue; nbp = aip->ai_buf; nbp = getdirtybuf(nbp, &lk, waitfor); if (nbp == NULL) continue; FREE_LOCK(&lk); if (waitfor == MNT_NOWAIT) { bawrite(nbp); } else if ((error = bwrite(nbp)) != 0) { break; } ACQUIRE_LOCK(&lk); continue; case D_INDIRDEP: restart: LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { if (aip->ai_state & DEPCOMPLETE) continue; nbp = aip->ai_buf; nbp = getdirtybuf(nbp, &lk, MNT_WAIT); if (nbp == NULL) goto restart; FREE_LOCK(&lk); if ((error = bwrite(nbp)) != 0) { goto loop_end; } ACQUIRE_LOCK(&lk); goto restart; } continue; case D_INODEDEP: if ((error = flush_inodedep_deps(wk->wk_mp, WK_INODEDEP(wk)->id_ino)) != 0) { FREE_LOCK(&lk); break; } continue; case D_PAGEDEP: /* * We are trying to sync a directory that may * have dependencies on both its own metadata * and/or dependencies on the inodes of any * recently allocated files. We walk its diradd * lists pushing out the associated inode. */ pagedep = WK_PAGEDEP(wk); for (i = 0; i < DAHASHSZ; i++) { if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) continue; if ((error = flush_pagedep_deps(vp, wk->wk_mp, &pagedep->pd_diraddhd[i]))) { FREE_LOCK(&lk); goto loop_end; } } continue; case D_MKDIR: /* * This case should never happen if the vnode has * been properly sync'ed. However, if this function * is used at a place where the vnode has not yet * been sync'ed, this dependency can show up. So, * rather than panic, just flush it. */ nbp = WK_MKDIR(wk)->md_buf; nbp = getdirtybuf(nbp, &lk, waitfor); if (nbp == NULL) continue; FREE_LOCK(&lk); if (waitfor == MNT_NOWAIT) { bawrite(nbp); } else if ((error = bwrite(nbp)) != 0) { break; } ACQUIRE_LOCK(&lk); continue; case D_BMSAFEMAP: /* * This case should never happen if the vnode has * been properly sync'ed. However, if this function * is used at a place where the vnode has not yet * been sync'ed, this dependency can show up. So, * rather than panic, just flush it. */ nbp = WK_BMSAFEMAP(wk)->sm_buf; nbp = getdirtybuf(nbp, &lk, waitfor); if (nbp == NULL) continue; FREE_LOCK(&lk); if (waitfor == MNT_NOWAIT) { bawrite(nbp); } else if ((error = bwrite(nbp)) != 0) { break; } ACQUIRE_LOCK(&lk); continue; default: panic("softdep_sync_metadata: Unknown type %s", TYPENAME(wk->wk_type)); /* NOTREACHED */ } loop_end: /* We reach here only in error and unlocked */ if (error == 0) panic("softdep_sync_metadata: zero error"); BUF_NOREC(bp); bawrite(bp); return (error); } FREE_LOCK(&lk); BO_LOCK(bo); while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) { nbp = getdirtybuf(nbp, BO_MTX(bo), MNT_WAIT); if (nbp) break; } BO_UNLOCK(bo); BUF_NOREC(bp); bawrite(bp); if (nbp != NULL) { bp = nbp; goto loop; } /* * The brief unlock is to allow any pent up dependency * processing to be done. Then proceed with the second pass. */ if (waitfor == MNT_NOWAIT) { waitfor = MNT_WAIT; goto top; } /* * If we have managed to get rid of all the dirty buffers, * then we are done. For certain directories and block * devices, we may need to do further work. * * We must wait for any I/O in progress to finish so that * all potential buffers on the dirty list will be visible. */ BO_LOCK(bo); drain_output(vp); BO_UNLOCK(bo); return (0); } /* * Flush the dependencies associated with an inodedep. * Called with splbio blocked. */ static int flush_inodedep_deps(mp, ino) struct mount *mp; ino_t ino; { struct inodedep *inodedep; int error, waitfor; /* * This work is done in two passes. The first pass grabs most * of the buffers and begins asynchronously writing them. The * only way to wait for these asynchronous writes is to sleep * on the filesystem vnode which may stay busy for a long time * if the filesystem is active. So, instead, we make a second * pass over the dependencies blocking on each write. In the * usual case we will be blocking against a write that we * initiated, so when it is done the dependency will have been * resolved. Thus the second pass is expected to end quickly. * We give a brief window at the top of the loop to allow * any pending I/O to complete. */ for (error = 0, waitfor = MNT_NOWAIT; ; ) { if (error) return (error); FREE_LOCK(&lk); ACQUIRE_LOCK(&lk); if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) return (0); if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || flush_deplist(&inodedep->id_extupdt, waitfor, &error) || flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) continue; /* * If pass2, we are done, otherwise do pass 2. */ if (waitfor == MNT_WAIT) break; waitfor = MNT_WAIT; } /* * Try freeing inodedep in case all dependencies have been removed. */ if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) (void) free_inodedep(inodedep); return (0); } /* * Flush an inode dependency list. * Called with splbio blocked. */ static int flush_deplist(listhead, waitfor, errorp) struct allocdirectlst *listhead; int waitfor; int *errorp; { struct allocdirect *adp; struct buf *bp; mtx_assert(&lk, MA_OWNED); TAILQ_FOREACH(adp, listhead, ad_next) { if (adp->ad_state & DEPCOMPLETE) continue; bp = adp->ad_buf; bp = getdirtybuf(bp, &lk, waitfor); if (bp == NULL) { if (waitfor == MNT_NOWAIT) continue; return (1); } FREE_LOCK(&lk); if (waitfor == MNT_NOWAIT) { bawrite(bp); } else if ((*errorp = bwrite(bp)) != 0) { ACQUIRE_LOCK(&lk); return (1); } ACQUIRE_LOCK(&lk); return (1); } return (0); } /* * Eliminate a pagedep dependency by flushing out all its diradd dependencies. * Called with splbio blocked. */ static int flush_pagedep_deps(pvp, mp, diraddhdp) struct vnode *pvp; struct mount *mp; struct diraddhd *diraddhdp; { struct inodedep *inodedep; struct ufsmount *ump; struct diradd *dap; struct vnode *vp; struct bufobj *bo; int error = 0; struct buf *bp; ino_t inum; struct worklist *wk; ump = VFSTOUFS(mp); while ((dap = LIST_FIRST(diraddhdp)) != NULL) { /* * Flush ourselves if this directory entry * has a MKDIR_PARENT dependency. */ if (dap->da_state & MKDIR_PARENT) { FREE_LOCK(&lk); if ((error = ffs_update(pvp, 1)) != 0) break; ACQUIRE_LOCK(&lk); /* * If that cleared dependencies, go on to next. */ if (dap != LIST_FIRST(diraddhdp)) continue; if (dap->da_state & MKDIR_PARENT) panic("flush_pagedep_deps: MKDIR_PARENT"); } /* * A newly allocated directory must have its "." and * ".." entries written out before its name can be * committed in its parent. We do not want or need * the full semantics of a synchronous ffs_syncvnode as * that may end up here again, once for each directory * level in the filesystem. Instead, we push the blocks * and wait for them to clear. We have to fsync twice * because the first call may choose to defer blocks * that still have dependencies, but deferral will * happen at most once. */ inum = dap->da_newinum; if (dap->da_state & MKDIR_BODY) { FREE_LOCK(&lk); if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ))) break; if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) || (error=ffs_syncvnode(vp, MNT_NOWAIT))) { vput(vp); break; } bo = &vp->v_bufobj; BO_LOCK(bo); drain_output(vp); /* * If first block is still dirty with a D_MKDIR * dependency then it needs to be written now. */ for (;;) { error = 0; bp = gbincore(bo, 0); if (bp == NULL) break; /* First block not present */ error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo)); BO_LOCK(bo); if (error == ENOLCK) continue; /* Slept, retry */ if (error != 0) break; /* Failed */ if ((bp->b_flags & B_DELWRI) == 0) { BUF_UNLOCK(bp); break; /* Buffer not dirty */ } for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; wk = LIST_NEXT(wk, wk_list)) if (wk->wk_type == D_MKDIR) break; if (wk == NULL) BUF_UNLOCK(bp); /* Dependency gone */ else { /* * D_MKDIR dependency remains, * must write buffer to stable * storage. */ BO_UNLOCK(bo); bremfree(bp); error = bwrite(bp); BO_LOCK(bo); } break; } BO_UNLOCK(bo); vput(vp); if (error != 0) break; /* Flushing of first block failed */ ACQUIRE_LOCK(&lk); /* * If that cleared dependencies, go on to next. */ if (dap != LIST_FIRST(diraddhdp)) continue; if (dap->da_state & MKDIR_BODY) panic("flush_pagedep_deps: MKDIR_BODY"); } /* * Flush the inode on which the directory entry depends. * Having accounted for MKDIR_PARENT and MKDIR_BODY above, * the only remaining dependency is that the updated inode * count must get pushed to disk. The inode has already * been pushed into its inode buffer (via VOP_UPDATE) at * the time of the reference count change. So we need only * locate that buffer, ensure that there will be no rollback * caused by a bitmap dependency, then write the inode buffer. */ retry: if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) panic("flush_pagedep_deps: lost inode"); /* * If the inode still has bitmap dependencies, * push them to disk. */ if ((inodedep->id_state & DEPCOMPLETE) == 0) { bp = inodedep->id_buf; bp = getdirtybuf(bp, &lk, MNT_WAIT); if (bp == NULL) goto retry; FREE_LOCK(&lk); if ((error = bwrite(bp)) != 0) break; ACQUIRE_LOCK(&lk); if (dap != LIST_FIRST(diraddhdp)) continue; } /* * If the inode is still sitting in a buffer waiting * to be written, push it to disk. */ FREE_LOCK(&lk); if ((error = bread(ump->um_devvp, fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) { brelse(bp); break; } if ((error = bwrite(bp)) != 0) break; ACQUIRE_LOCK(&lk); /* * If we have failed to get rid of all the dependencies * then something is seriously wrong. */ if (dap == LIST_FIRST(diraddhdp)) panic("flush_pagedep_deps: flush failed"); } if (error) ACQUIRE_LOCK(&lk); return (error); } /* * A large burst of file addition or deletion activity can drive the * memory load excessively high. First attempt to slow things down * using the techniques below. If that fails, this routine requests * the offending operations to fall back to running synchronously * until the memory load returns to a reasonable level. */ int softdep_slowdown(vp) struct vnode *vp; { int max_softdeps_hard; ACQUIRE_LOCK(&lk); max_softdeps_hard = max_softdeps * 11 / 10; if (num_dirrem < max_softdeps_hard / 2 && num_inodedep < max_softdeps_hard && VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) { FREE_LOCK(&lk); return (0); } if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps) softdep_speedup(); stat_sync_limit_hit += 1; FREE_LOCK(&lk); return (1); } /* * Called by the allocation routines when they are about to fail * in the hope that we can free up some disk space. * * First check to see if the work list has anything on it. If it has, * clean up entries until we successfully free some space. Because this * process holds inodes locked, we cannot handle any remove requests * that might block on a locked inode as that could lead to deadlock. * If the worklist yields no free space, encourage the syncer daemon * to help us. In no event will we try for longer than tickdelay seconds. */ int softdep_request_cleanup(fs, vp) struct fs *fs; struct vnode *vp; { struct ufsmount *ump; long starttime; ufs2_daddr_t needed; int error; ump = VTOI(vp)->i_ump; mtx_assert(UFS_MTX(ump), MA_OWNED); needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize; starttime = time_second + tickdelay; /* * If we are being called because of a process doing a * copy-on-write, then it is not safe to update the vnode * as we may recurse into the copy-on-write routine. */ if (!(curthread->td_pflags & TDP_COWINPROGRESS)) { UFS_UNLOCK(ump); error = ffs_update(vp, 1); UFS_LOCK(ump); if (error != 0) return (0); } while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) { if (time_second > starttime) return (0); UFS_UNLOCK(ump); ACQUIRE_LOCK(&lk); if (ump->softdep_on_worklist > 0 && process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) { stat_worklist_push += 1; FREE_LOCK(&lk); UFS_LOCK(ump); continue; } request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT); FREE_LOCK(&lk); UFS_LOCK(ump); } return (1); } /* * If memory utilization has gotten too high, deliberately slow things * down and speed up the I/O processing. */ extern struct thread *syncertd; static int request_cleanup(mp, resource) struct mount *mp; int resource; { struct thread *td = curthread; struct ufsmount *ump; mtx_assert(&lk, MA_OWNED); /* * We never hold up the filesystem syncer or buf daemon. */ if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) return (0); ump = VFSTOUFS(mp); /* * First check to see if the work list has gotten backlogged. * If it has, co-opt this process to help clean up two entries. * Because this process may hold inodes locked, we cannot * handle any remove requests that might block on a locked * inode as that could lead to deadlock. We set TDP_SOFTDEP * to avoid recursively processing the worklist. */ if (ump->softdep_on_worklist > max_softdeps / 10) { td->td_pflags |= TDP_SOFTDEP; process_worklist_item(mp, LK_NOWAIT); process_worklist_item(mp, LK_NOWAIT); td->td_pflags &= ~TDP_SOFTDEP; stat_worklist_push += 2; return(1); } /* * Next, we attempt to speed up the syncer process. If that * is successful, then we allow the process to continue. */ if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT) return(0); /* * If we are resource constrained on inode dependencies, try * flushing some dirty inodes. Otherwise, we are constrained * by file deletions, so try accelerating flushes of directories * with removal dependencies. We would like to do the cleanup * here, but we probably hold an inode locked at this point and * that might deadlock against one that we try to clean. So, * the best that we can do is request the syncer daemon to do * the cleanup for us. */ switch (resource) { case FLUSH_INODES: stat_ino_limit_push += 1; req_clear_inodedeps += 1; stat_countp = &stat_ino_limit_hit; break; case FLUSH_REMOVE: case FLUSH_REMOVE_WAIT: stat_blk_limit_push += 1; req_clear_remove += 1; stat_countp = &stat_blk_limit_hit; break; default: panic("request_cleanup: unknown type"); } /* * Hopefully the syncer daemon will catch up and awaken us. * We wait at most tickdelay before proceeding in any case. */ proc_waiting += 1; if (callout_pending(&softdep_callout) == FALSE) callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, pause_timer, 0); msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); proc_waiting -= 1; return (1); } /* * Awaken processes pausing in request_cleanup and clear proc_waiting * to indicate that there is no longer a timer running. */ static void pause_timer(arg) void *arg; { /* * The callout_ API has acquired mtx and will hold it around this * function call. */ *stat_countp += 1; wakeup_one(&proc_waiting); if (proc_waiting > 0) callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, pause_timer, 0); } /* * Flush out a directory with at least one removal dependency in an effort to * reduce the number of dirrem, freefile, and freeblks dependency structures. */ static void clear_remove(td) struct thread *td; { struct pagedep_hashhead *pagedephd; struct pagedep *pagedep; static int next = 0; struct mount *mp; struct vnode *vp; struct bufobj *bo; int error, cnt; ino_t ino; mtx_assert(&lk, MA_OWNED); for (cnt = 0; cnt < pagedep_hash; cnt++) { pagedephd = &pagedep_hashtbl[next++]; if (next >= pagedep_hash) next = 0; LIST_FOREACH(pagedep, pagedephd, pd_hash) { if (LIST_EMPTY(&pagedep->pd_dirremhd)) continue; mp = pagedep->pd_list.wk_mp; ino = pagedep->pd_ino; if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) continue; FREE_LOCK(&lk); if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ))) { softdep_error("clear_remove: vget", error); vn_finished_write(mp); ACQUIRE_LOCK(&lk); return; } if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) softdep_error("clear_remove: fsync", error); bo = &vp->v_bufobj; BO_LOCK(bo); drain_output(vp); BO_UNLOCK(bo); vput(vp); vn_finished_write(mp); ACQUIRE_LOCK(&lk); return; } } } /* * Clear out a block of dirty inodes in an effort to reduce * the number of inodedep dependency structures. */ static void clear_inodedeps(td) struct thread *td; { struct inodedep_hashhead *inodedephd; struct inodedep *inodedep; static int next = 0; struct mount *mp; struct vnode *vp; struct fs *fs; int error, cnt; ino_t firstino, lastino, ino; mtx_assert(&lk, MA_OWNED); /* * Pick a random inode dependency to be cleared. * We will then gather up all the inodes in its block * that have dependencies and flush them out. */ for (cnt = 0; cnt < inodedep_hash; cnt++) { inodedephd = &inodedep_hashtbl[next++]; if (next >= inodedep_hash) next = 0; if ((inodedep = LIST_FIRST(inodedephd)) != NULL) break; } if (inodedep == NULL) return; fs = inodedep->id_fs; mp = inodedep->id_list.wk_mp; /* * Find the last inode in the block with dependencies. */ firstino = inodedep->id_ino & ~(INOPB(fs) - 1); for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) break; /* * Asynchronously push all but the last inode with dependencies. * Synchronously push the last inode with dependencies to ensure * that the inode block gets written to free up the inodedeps. */ for (ino = firstino; ino <= lastino; ino++) { if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) continue; if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) continue; FREE_LOCK(&lk); if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ)) != 0) { softdep_error("clear_inodedeps: vget", error); vn_finished_write(mp); ACQUIRE_LOCK(&lk); return; } if (ino == lastino) { if ((error = ffs_syncvnode(vp, MNT_WAIT))) softdep_error("clear_inodedeps: fsync1", error); } else { if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) softdep_error("clear_inodedeps: fsync2", error); BO_LOCK(&vp->v_bufobj); drain_output(vp); BO_UNLOCK(&vp->v_bufobj); } vput(vp); vn_finished_write(mp); ACQUIRE_LOCK(&lk); } } /* * Function to determine if the buffer has outstanding dependencies * that will cause a roll-back if the buffer is written. If wantcount * is set, return number of dependencies, otherwise just yes or no. */ static int softdep_count_dependencies(bp, wantcount) struct buf *bp; int wantcount; { struct worklist *wk; struct inodedep *inodedep; struct indirdep *indirdep; struct allocindir *aip; struct pagedep *pagedep; struct diradd *dap; int i, retval; retval = 0; ACQUIRE_LOCK(&lk); LIST_FOREACH(wk, &bp->b_dep, wk_list) { switch (wk->wk_type) { case D_INODEDEP: inodedep = WK_INODEDEP(wk); if ((inodedep->id_state & DEPCOMPLETE) == 0) { /* bitmap allocation dependency */ retval += 1; if (!wantcount) goto out; } if (TAILQ_FIRST(&inodedep->id_inoupdt)) { /* direct block pointer dependency */ retval += 1; if (!wantcount) goto out; } if (TAILQ_FIRST(&inodedep->id_extupdt)) { /* direct block pointer dependency */ retval += 1; if (!wantcount) goto out; } continue; case D_INDIRDEP: indirdep = WK_INDIRDEP(wk); LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { /* indirect block pointer dependency */ retval += 1; if (!wantcount) goto out; } continue; case D_PAGEDEP: pagedep = WK_PAGEDEP(wk); for (i = 0; i < DAHASHSZ; i++) { LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { /* directory entry dependency */ retval += 1; if (!wantcount) goto out; } } continue; case D_BMSAFEMAP: case D_ALLOCDIRECT: case D_ALLOCINDIR: case D_MKDIR: /* never a dependency on these blocks */ continue; default: panic("softdep_check_for_rollback: Unexpected type %s", TYPENAME(wk->wk_type)); /* NOTREACHED */ } } out: FREE_LOCK(&lk); return retval; } /* * Acquire exclusive access to a buffer. * Must be called with a locked mtx parameter. * Return acquired buffer or NULL on failure. */ static struct buf * getdirtybuf(bp, mtx, waitfor) struct buf *bp; struct mtx *mtx; int waitfor; { int error; mtx_assert(mtx, MA_OWNED); if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { if (waitfor != MNT_WAIT) return (NULL); error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx); /* * Even if we sucessfully acquire bp here, we have dropped * mtx, which may violates our guarantee. */ if (error == 0) BUF_UNLOCK(bp); else if (error != ENOLCK) panic("getdirtybuf: inconsistent lock: %d", error); mtx_lock(mtx); return (NULL); } if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { if (mtx == &lk && waitfor == MNT_WAIT) { mtx_unlock(mtx); BO_LOCK(bp->b_bufobj); BUF_UNLOCK(bp); if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { bp->b_vflags |= BV_BKGRDWAIT; msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), PRIBIO | PDROP, "getbuf", 0); } else BO_UNLOCK(bp->b_bufobj); mtx_lock(mtx); return (NULL); } BUF_UNLOCK(bp); if (waitfor != MNT_WAIT) return (NULL); /* * The mtx argument must be bp->b_vp's mutex in * this case. */ #ifdef DEBUG_VFS_LOCKS if (bp->b_vp->v_type != VCHR) ASSERT_BO_LOCKED(bp->b_bufobj); #endif bp->b_vflags |= BV_BKGRDWAIT; msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0); return (NULL); } if ((bp->b_flags & B_DELWRI) == 0) { BUF_UNLOCK(bp); return (NULL); } bremfree(bp); return (bp); } /* * Check if it is safe to suspend the file system now. On entry, * the vnode interlock for devvp should be held. Return 0 with * the mount interlock held if the file system can be suspended now, * otherwise return EAGAIN with the mount interlock held. */ int softdep_check_suspend(struct mount *mp, struct vnode *devvp, int softdep_deps, int softdep_accdeps, int secondary_writes, int secondary_accwrites) { struct bufobj *bo; struct ufsmount *ump; int error; ump = VFSTOUFS(mp); bo = &devvp->v_bufobj; ASSERT_BO_LOCKED(bo); for (;;) { if (!TRY_ACQUIRE_LOCK(&lk)) { BO_UNLOCK(bo); ACQUIRE_LOCK(&lk); FREE_LOCK(&lk); BO_LOCK(bo); continue; } MNT_ILOCK(mp); if (mp->mnt_secondary_writes != 0) { FREE_LOCK(&lk); BO_UNLOCK(bo); msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), (PUSER - 1) | PDROP, "secwr", 0); BO_LOCK(bo); continue; } break; } /* * Reasons for needing more work before suspend: * - Dirty buffers on devvp. * - Softdep activity occurred after start of vnode sync loop * - Secondary writes occurred after start of vnode sync loop */ error = 0; if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0 || softdep_deps != 0 || ump->softdep_deps != 0 || softdep_accdeps != ump->softdep_accdeps || secondary_writes != 0 || mp->mnt_secondary_writes != 0 || secondary_accwrites != mp->mnt_secondary_accwrites) error = EAGAIN; FREE_LOCK(&lk); BO_UNLOCK(bo); return (error); } /* * Get the number of dependency structures for the file system, both * the current number and the total number allocated. These will * later be used to detect that softdep processing has occurred. */ void softdep_get_depcounts(struct mount *mp, int *softdep_depsp, int *softdep_accdepsp) { struct ufsmount *ump; ump = VFSTOUFS(mp); ACQUIRE_LOCK(&lk); *softdep_depsp = ump->softdep_deps; *softdep_accdepsp = ump->softdep_accdeps; FREE_LOCK(&lk); } /* * Wait for pending output on a vnode to complete. * Must be called with vnode lock and interlock locked. * * XXX: Should just be a call to bufobj_wwait(). */ static void drain_output(vp) struct vnode *vp; { struct bufobj *bo; bo = &vp->v_bufobj; ASSERT_VOP_LOCKED(vp, "drain_output"); ASSERT_BO_LOCKED(bo); while (bo->bo_numoutput) { bo->bo_flag |= BO_WWAIT; msleep((caddr_t)&bo->bo_numoutput, BO_MTX(bo), PRIBIO + 1, "drainvp", 0); } } /* * Called whenever a buffer that is being invalidated or reallocated * contains dependencies. This should only happen if an I/O error has * occurred. The routine is called with the buffer locked. */ static void softdep_deallocate_dependencies(bp) struct buf *bp; { if ((bp->b_ioflags & BIO_ERROR) == 0) panic("softdep_deallocate_dependencies: dangling deps"); softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); panic("softdep_deallocate_dependencies: unrecovered I/O error"); } /* * Function to handle asynchronous write errors in the filesystem. */ static void softdep_error(func, error) char *func; int error; { /* XXX should do something better! */ printf("%s: got error %d while accessing filesystem\n", func, error); } #ifdef DDB DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) { struct inodedep_hashhead *inodedephd; struct inodedep *inodedep; struct fs *fs; int cnt; fs = have_addr ? (struct fs *)addr : NULL; for (cnt = 0; cnt < inodedep_hash; cnt++) { inodedephd = &inodedep_hashtbl[cnt]; LIST_FOREACH(inodedep, inodedephd, id_hash) { if (fs != NULL && fs != inodedep->id_fs) continue; db_printf("%p fs %p st %x ino %jd inoblk %jd\n", inodedep, inodedep->id_fs, inodedep->id_state, (intmax_t)inodedep->id_ino, (intmax_t)fsbtodb(inodedep->id_fs, ino_to_fsba(inodedep->id_fs, inodedep->id_ino))); } } } #endif /* DDB */ #endif /* SOFTUPDATES */