X-Git-Url: https://git.subgeniuskitty.com/unix-history/.git/blobdiff_plain/b991bc2d64308582c742c3119265c271e4f849ae..c93d29a8f5f38b8c0c60239a2e70bc231ae69e20:/usr/src/sys/miscfs/union/union_subr.c diff --git a/usr/src/sys/miscfs/union/union_subr.c b/usr/src/sys/miscfs/union/union_subr.c index 9199dcd391..f0cd58e1d5 100644 --- a/usr/src/sys/miscfs/union/union_subr.c +++ b/usr/src/sys/miscfs/union/union_subr.c @@ -8,7 +8,7 @@ * * %sccs.include.redist.c% * - * @(#)union_subr.c 1.1 (Berkeley) %G% + * @(#)union_subr.c 8.5 (Berkeley) %G% */ #include @@ -18,22 +18,157 @@ #include #include #include -#include "union.h" /**/ +#include +#include +#include +#include +#include -static struct union_node *unhead; -static int unvplock; +#ifdef DIAGNOSTIC +#include +#endif + +/* must be power of two, otherwise change UNION_HASH() */ +#define NHASH 32 + +/* unsigned int ... */ +#define UNION_HASH(u, l) \ + (((((unsigned long) (u)) + ((unsigned long) l)) >> 8) & (NHASH-1)) + +static LIST_HEAD(unhead, union_node) unhead[NHASH]; +static int unvplock[NHASH]; int union_init() +{ + int i; + + for (i = 0; i < NHASH; i++) + LIST_INIT(&unhead[i]); + bzero((caddr_t) unvplock, sizeof(unvplock)); +} + +static int +union_list_lock(ix) + int ix; +{ + + if (unvplock[ix] & UN_LOCKED) { + unvplock[ix] |= UN_WANT; + sleep((caddr_t) &unvplock[ix], PINOD); + return (1); + } + + unvplock[ix] |= UN_LOCKED; + + return (0); +} + +static void +union_list_unlock(ix) + int ix; +{ + + unvplock[ix] &= ~UN_LOCKED; + + if (unvplock[ix] & UN_WANT) { + unvplock[ix] &= ~UN_WANT; + wakeup((caddr_t) &unvplock[ix]); + } +} + +void +union_updatevp(un, uppervp, lowervp) + struct union_node *un; + struct vnode *uppervp; + struct vnode *lowervp; +{ + int ohash = UNION_HASH(un->un_uppervp, un->un_lowervp); + int nhash = UNION_HASH(uppervp, lowervp); + + if (ohash != nhash) { + /* + * Ensure locking is ordered from lower to higher + * to avoid deadlocks. + */ + if (nhash < ohash) { + int t = ohash; + ohash = nhash; + nhash = t; + } + + while (union_list_lock(ohash)) + continue; + + while (union_list_lock(nhash)) + continue; + + LIST_REMOVE(un, un_cache); + union_list_unlock(ohash); + } else { + while (union_list_lock(nhash)) + continue; + } + + if (un->un_lowervp != lowervp) { + if (un->un_lowervp) { + vrele(un->un_lowervp); + if (un->un_path) { + free(un->un_path, M_TEMP); + un->un_path = 0; + } + if (un->un_dirvp) { + vrele(un->un_dirvp); + un->un_dirvp = NULLVP; + } + } + un->un_lowervp = lowervp; + } + + if (un->un_uppervp != uppervp) { + if (un->un_uppervp) + vrele(un->un_uppervp); + + un->un_uppervp = uppervp; + } + + if (ohash != nhash) + LIST_INSERT_HEAD(&unhead[nhash], un, un_cache); + + union_list_unlock(nhash); +} + +void +union_newlower(un, lowervp) + struct union_node *un; + struct vnode *lowervp; +{ + + union_updatevp(un, un->un_uppervp, lowervp); +} + +void +union_newupper(un, uppervp) + struct union_node *un; + struct vnode *uppervp; { - unhead = 0; - unvplock = 0; + union_updatevp(un, uppervp, un->un_lowervp); } /* * allocate a union_node/vnode pair. the vnode is - * referenced and locked. + * referenced and locked. the new vnode is returned + * via (vpp). (mp) is the mountpoint of the union filesystem, + * (dvp) is the parent directory where the upper layer object + * should exist (but doesn't) and (cnp) is the componentname + * information which is partially copied to allow the upper + * layer object to be created at a later time. (uppervp) + * and (lowervp) reference the upper and lower layer objects + * being mapped. either, but not both, can be nil. + * if supplied, (uppervp) is locked. + * the reference is either maintained in the new union_node + * object which is allocated, or they are vrele'd. * * all union_nodes are maintained on a singly-linked * list. new nodes are only allocated when they cannot @@ -53,9 +188,10 @@ union_init() * the vnode free list. */ int -union_allocvp(vpp, mp, dvp, cnp, uppervp, lowervp) +union_allocvp(vpp, mp, undvp, dvp, cnp, uppervp, lowervp) struct vnode **vpp; struct mount *mp; + struct vnode *undvp; struct vnode *dvp; /* may be null */ struct componentname *cnp; /* may be null */ struct vnode *uppervp; /* may be null */ @@ -64,75 +200,229 @@ union_allocvp(vpp, mp, dvp, cnp, uppervp, lowervp) int error; struct union_node *un; struct union_node **pp; + struct vnode *xlowervp = NULLVP; + struct union_mount *um = MOUNTTOUNIONMOUNT(mp); + int hash; + int vflag; + int try; + + if (uppervp == NULLVP && lowervp == NULLVP) + panic("union: unidentifiable allocation"); + + if (uppervp && lowervp && (uppervp->v_type != lowervp->v_type)) { + xlowervp = lowervp; + lowervp = NULLVP; + } + + /* detect the root vnode (and aliases) */ + vflag = 0; + if ((uppervp == um->um_uppervp) && + ((lowervp == NULLVP) || lowervp == um->um_lowervp)) { + if (lowervp == NULLVP) { + lowervp = um->um_lowervp; + VREF(lowervp); + } + vflag = VROOT; + } loop: - for (un = unhead; un != 0; un = un->un_next) { - if ((un->un_lowervp == lowervp || - un->un_lowervp == 0) && - (un->un_uppervp == uppervp || - un->un_uppervp == 0) && - (UNIONTOV(un)->v_mount == mp)) { - if (vget(un->un_vnode, 1)) + for (try = 0; try < 3; try++) { + switch (try) { + case 0: + if (lowervp == NULLVP) + continue; + hash = UNION_HASH(uppervp, lowervp); + break; + + case 1: + if (uppervp == NULLVP) + continue; + hash = UNION_HASH(uppervp, NULLVP); + break; + + case 2: + if (lowervp == NULLVP) + continue; + hash = UNION_HASH(NULLVP, lowervp); + break; + } + + while (union_list_lock(hash)) + continue; + + for (un = unhead[hash].lh_first; un != 0; + un = un->un_cache.le_next) { + if ((un->un_lowervp == lowervp || + un->un_lowervp == NULLVP) && + (un->un_uppervp == uppervp || + un->un_uppervp == NULLVP) && + (UNIONTOV(un)->v_mount == mp)) { + if (vget(UNIONTOV(un), 0)) { + union_list_unlock(hash); + goto loop; + } + break; + } + } + + union_list_unlock(hash); + + if (un) + break; + } + + if (un) { + /* + * Obtain a lock on the union_node. + * uppervp is locked, though un->un_uppervp + * may not be. this doesn't break the locking + * hierarchy since in the case that un->un_uppervp + * is not yet locked it will be vrele'd and replaced + * with uppervp. + */ + + if ((dvp != NULLVP) && (uppervp == dvp)) { + /* + * Access ``.'', so (un) will already + * be locked. Since this process has + * the lock on (uppervp) no other + * process can hold the lock on (un). + */ +#ifdef DIAGNOSTIC + if ((un->un_flags & UN_LOCKED) == 0) + panic("union: . not locked"); + else if (curproc && un->un_pid != curproc->p_pid && + un->un_pid > -1 && curproc->p_pid > -1) + panic("union: allocvp not lock owner"); +#endif + } else { + if (un->un_flags & UN_LOCKED) { + vrele(UNIONTOV(un)); + un->un_flags |= UN_WANT; + sleep((caddr_t) &un->un_flags, PINOD); goto loop; - un->un_lowervp = lowervp; - un->un_uppervp = uppervp; - *vpp = un->un_vnode; - return (0); + } + un->un_flags |= UN_LOCKED; + +#ifdef DIAGNOSTIC + if (curproc) + un->un_pid = curproc->p_pid; + else + un->un_pid = -1; +#endif + } + + /* + * At this point, the union_node is locked, + * un->un_uppervp may not be locked, and uppervp + * is locked or nil. + */ + + /* + * Save information about the upper layer. + */ + if (uppervp != un->un_uppervp) { + union_newupper(un, uppervp); + } else if (uppervp) { + vrele(uppervp); + } + + if (un->un_uppervp) { + un->un_flags |= UN_ULOCK; + un->un_flags &= ~UN_KLOCK; + } + + /* + * Save information about the lower layer. + * This needs to keep track of pathname + * and directory information which union_vn_create + * might need. + */ + if (lowervp != un->un_lowervp) { + union_newlower(un, lowervp); + if (cnp && (lowervp != NULLVP) && + (lowervp->v_type == VREG)) { + un->un_hash = cnp->cn_hash; + un->un_path = malloc(cnp->cn_namelen+1, + M_TEMP, M_WAITOK); + bcopy(cnp->cn_nameptr, un->un_path, + cnp->cn_namelen); + un->un_path[cnp->cn_namelen] = '\0'; + VREF(dvp); + un->un_dirvp = dvp; + } + } else if (lowervp) { + vrele(lowervp); } + *vpp = UNIONTOV(un); + return (0); } /* * otherwise lock the vp list while we call getnewvnode * since that can block. */ - if (unvplock & UN_LOCKED) { - unvplock |= UN_WANT; - sleep((caddr_t) &unvplock, PINOD); + hash = UNION_HASH(uppervp, lowervp); + + if (union_list_lock(hash)) goto loop; - } - unvplock |= UN_LOCKED; error = getnewvnode(VT_UNION, mp, union_vnodeop_p, vpp); - if (error) + if (error) { + if (uppervp) { + if (dvp == uppervp) + vrele(uppervp); + else + vput(uppervp); + } + if (lowervp) + vrele(lowervp); + goto out; + } MALLOC((*vpp)->v_data, void *, sizeof(struct union_node), M_TEMP, M_WAITOK); + (*vpp)->v_flag |= vflag; + if (uppervp) + (*vpp)->v_type = uppervp->v_type; + else + (*vpp)->v_type = lowervp->v_type; un = VTOUNION(*vpp); - un->un_next = 0; - un->un_dirvp = dvp; + un->un_vnode = *vpp; un->un_uppervp = uppervp; un->un_lowervp = lowervp; - un->un_vnode = *vpp; - un->un_flags = 0; - if (cnp) { + un->un_openl = 0; + un->un_flags = UN_LOCKED; + if (un->un_uppervp) + un->un_flags |= UN_ULOCK; +#ifdef DIAGNOSTIC + if (curproc) + un->un_pid = curproc->p_pid; + else + un->un_pid = -1; +#endif + if (cnp && (lowervp != NULLVP) && (lowervp->v_type == VREG)) { + un->un_hash = cnp->cn_hash; un->un_path = malloc(cnp->cn_namelen+1, M_TEMP, M_WAITOK); bcopy(cnp->cn_nameptr, un->un_path, cnp->cn_namelen); un->un_path[cnp->cn_namelen] = '\0'; + VREF(dvp); + un->un_dirvp = dvp; } else { + un->un_hash = 0; un->un_path = 0; + un->un_dirvp = 0; } -#ifdef DIAGNOSTIC - un->un_pid = 0; -#endif + LIST_INSERT_HEAD(&unhead[hash], un, un_cache); - /* add to union vnode list */ - for (pp = &unhead; *pp; pp = &(*pp)->un_next) - continue; - *pp = un; + if (xlowervp) + vrele(xlowervp); out: - unvplock &= ~UN_LOCKED; - - if (unvplock & UN_WANT) { - unvplock &= ~UN_WANT; - wakeup((caddr_t) &unvplock); - } - - if (un) - VOP_LOCK(UNIONTOV(un)); + union_list_unlock(hash); return (error); } @@ -141,20 +431,303 @@ int union_freevp(vp) struct vnode *vp; { - struct union_node **unpp; struct union_node *un = VTOUNION(vp); - for (unpp = &unhead; *unpp != 0; unpp = &(*unpp)->un_next) { - if (*unpp == un) { - *unpp = un->un_next; - break; - } - } + LIST_REMOVE(un, un_cache); + if (un->un_uppervp) + vrele(un->un_uppervp); + if (un->un_lowervp) + vrele(un->un_lowervp); + if (un->un_dirvp) + vrele(un->un_dirvp); if (un->un_path) - FREE(un->un_path, M_TEMP); + free(un->un_path, M_TEMP); FREE(vp->v_data, M_TEMP); vp->v_data = 0; + + return (0); +} + +/* + * copyfile. copy the vnode (fvp) to the vnode (tvp) + * using a sequence of reads and writes. both (fvp) + * and (tvp) are locked on entry and exit. + */ +int +union_copyfile(p, cred, fvp, tvp) + struct proc *p; + struct ucred *cred; + struct vnode *fvp; + struct vnode *tvp; +{ + char *buf; + struct uio uio; + struct iovec iov; + int error = 0; + + /* + * strategy: + * allocate a buffer of size MAXBSIZE. + * loop doing reads and writes, keeping track + * of the current uio offset. + * give up at the first sign of trouble. + */ + + uio.uio_procp = p; + uio.uio_segflg = UIO_SYSSPACE; + uio.uio_offset = 0; + + VOP_UNLOCK(fvp); /* XXX */ + LEASE_CHECK(fvp, p, cred, LEASE_READ); + VOP_LOCK(fvp); /* XXX */ + VOP_UNLOCK(tvp); /* XXX */ + LEASE_CHECK(tvp, p, cred, LEASE_WRITE); + VOP_LOCK(tvp); /* XXX */ + + buf = malloc(MAXBSIZE, M_TEMP, M_WAITOK); + + /* ugly loop follows... */ + do { + off_t offset = uio.uio_offset; + + uio.uio_iov = &iov; + uio.uio_iovcnt = 1; + iov.iov_base = buf; + iov.iov_len = MAXBSIZE; + uio.uio_resid = iov.iov_len; + uio.uio_rw = UIO_READ; + error = VOP_READ(fvp, &uio, 0, cred); + + if (error == 0) { + uio.uio_iov = &iov; + uio.uio_iovcnt = 1; + iov.iov_base = buf; + iov.iov_len = MAXBSIZE - uio.uio_resid; + uio.uio_offset = offset; + uio.uio_rw = UIO_WRITE; + uio.uio_resid = iov.iov_len; + + if (uio.uio_resid == 0) + break; + + do { + error = VOP_WRITE(tvp, &uio, 0, cred); + } while ((uio.uio_resid > 0) && (error == 0)); + } + + } while (error == 0); + + free(buf, M_TEMP); + return (error); +} + +/* + * Create a shadow directory in the upper layer. + * The new vnode is returned locked. + * + * (um) points to the union mount structure for access to the + * the mounting process's credentials. + * (dvp) is the directory in which to create the shadow directory. + * it is unlocked on entry and exit. + * (cnp) is the componentname to be created. + * (vpp) is the returned newly created shadow directory, which + * is returned locked. + */ +int +union_mkshadow(um, dvp, cnp, vpp) + struct union_mount *um; + struct vnode *dvp; + struct componentname *cnp; + struct vnode **vpp; +{ + int error; + struct vattr va; + struct proc *p = cnp->cn_proc; + struct componentname cn; + + /* + * policy: when creating the shadow directory in the + * upper layer, create it owned by the user who did + * the mount, group from parent directory, and mode + * 777 modified by umask (ie mostly identical to the + * mkdir syscall). (jsp, kb) + */ + + /* + * A new componentname structure must be faked up because + * there is no way to know where the upper level cnp came + * from or what it is being used for. This must duplicate + * some of the work done by NDINIT, some of the work done + * by namei, some of the work done by lookup and some of + * the work done by VOP_LOOKUP when given a CREATE flag. + * Conclusion: Horrible. + * + * The pathname buffer will be FREEed by VOP_MKDIR. + */ + cn.cn_pnbuf = malloc(cnp->cn_namelen+1, M_NAMEI, M_WAITOK); + bcopy(cnp->cn_nameptr, cn.cn_pnbuf, cnp->cn_namelen); + cn.cn_pnbuf[cnp->cn_namelen] = '\0'; + + cn.cn_nameiop = CREATE; + cn.cn_flags = (LOCKPARENT|HASBUF|SAVENAME|SAVESTART|ISLASTCN); + cn.cn_proc = cnp->cn_proc; + if (um->um_op == UNMNT_ABOVE) + cn.cn_cred = cnp->cn_cred; + else + cn.cn_cred = um->um_cred; + cn.cn_nameptr = cn.cn_pnbuf; + cn.cn_namelen = cnp->cn_namelen; + cn.cn_hash = cnp->cn_hash; + cn.cn_consume = cnp->cn_consume; + + VREF(dvp); + if (error = relookup(dvp, vpp, &cn)) + return (error); + vrele(dvp); + + if (*vpp) { + VOP_ABORTOP(dvp, &cn); + VOP_UNLOCK(dvp); + vrele(*vpp); + *vpp = NULLVP; + return (EEXIST); + } + + VATTR_NULL(&va); + va.va_type = VDIR; + va.va_mode = um->um_cmode; + + /* LEASE_CHECK: dvp is locked */ + LEASE_CHECK(dvp, p, p->p_ucred, LEASE_WRITE); + + error = VOP_MKDIR(dvp, vpp, &cn, &va); + return (error); +} + +/* + * union_vn_create: creates and opens a new shadow file + * on the upper union layer. this function is similar + * in spirit to calling vn_open but it avoids calling namei(). + * the problem with calling namei is that a) it locks too many + * things, and b) it doesn't start at the "right" directory, + * whereas relookup is told where to start. + */ +int +union_vn_create(vpp, un, p) + struct vnode **vpp; + struct union_node *un; + struct proc *p; +{ + struct vnode *vp; + struct ucred *cred = p->p_ucred; + struct vattr vat; + struct vattr *vap = &vat; + int fmode = FFLAGS(O_WRONLY|O_CREAT|O_TRUNC|O_EXCL); + int error; + int cmode = UN_FILEMODE & ~p->p_fd->fd_cmask; + char *cp; + struct componentname cn; + + *vpp = NULLVP; + + /* + * Build a new componentname structure (for the same + * reasons outlines in union_mkshadow). + * The difference here is that the file is owned by + * the current user, rather than by the person who + * did the mount, since the current user needs to be + * able to write the file (that's why it is being + * copied in the first place). + */ + cn.cn_namelen = strlen(un->un_path); + cn.cn_pnbuf = (caddr_t) malloc(cn.cn_namelen, M_NAMEI, M_WAITOK); + bcopy(un->un_path, cn.cn_pnbuf, cn.cn_namelen+1); + cn.cn_nameiop = CREATE; + cn.cn_flags = (LOCKPARENT|HASBUF|SAVENAME|SAVESTART|ISLASTCN); + cn.cn_proc = p; + cn.cn_cred = p->p_ucred; + cn.cn_nameptr = cn.cn_pnbuf; + cn.cn_hash = un->un_hash; + cn.cn_consume = 0; + + VREF(un->un_dirvp); + if (error = relookup(un->un_dirvp, &vp, &cn)) + return (error); + vrele(un->un_dirvp); + + if (vp) { + VOP_ABORTOP(un->un_dirvp, &cn); + if (un->un_dirvp == vp) + vrele(un->un_dirvp); + else + vput(un->un_dirvp); + vrele(vp); + return (EEXIST); + } + + /* + * Good - there was no race to create the file + * so go ahead and create it. The permissions + * on the file will be 0666 modified by the + * current user's umask. Access to the file, while + * it is unioned, will require access to the top *and* + * bottom files. Access when not unioned will simply + * require access to the top-level file. + * TODO: confirm choice of access permissions. + */ + VATTR_NULL(vap); + vap->va_type = VREG; + vap->va_mode = cmode; + LEASE_CHECK(un->un_dirvp, p, cred, LEASE_WRITE); + if (error = VOP_CREATE(un->un_dirvp, &vp, &cn, vap)) + return (error); + + if (error = VOP_OPEN(vp, fmode, cred, p)) { + vput(vp); + return (error); + } + + vp->v_writecount++; + *vpp = vp; return (0); } + +int +union_vn_close(vp, fmode, cred, p) + struct vnode *vp; + int fmode; + struct ucred *cred; + struct proc *p; +{ + if (fmode & FWRITE) + --vp->v_writecount; + return (VOP_CLOSE(vp, fmode)); +} + +void +union_removed_upper(un) + struct union_node *un; +{ + if (un->un_flags & UN_ULOCK) { + un->un_flags &= ~UN_ULOCK; + VOP_UNLOCK(un->un_uppervp); + } + + union_newupper(un, NULLVP); +} + +struct vnode * +union_lowervp(vp) + struct vnode *vp; +{ + struct union_node *un = VTOUNION(vp); + + if (un->un_lowervp && (vp->v_type == un->un_lowervp->v_type)) { + if (vget(un->un_lowervp, 0)) + return (NULLVP); + } + + return (un->un_lowervp); +}