From: Kirill A. Korinsky Subject: need help with test: sys/vfs: stop checkalias replacing vnodes To: OpenBSD tech Date: Sun, 08 Nov 2026 13:51:03 +0100 tech@, checkalias() mixed two contracts: it linked a special vnode into the alias list, but in the VT_NON block device case it returned an existing alias and forced the caller to discard the vnode it had just constructed. This is not a valid generic contract once filesystem private lookup caches can already name the provisional vnode; the node to vnode relation changes under concurrent lookup, and callers have to move v_data and vgone() a vnode which is no longer private. So, I had make checkalias() only register the supplied vnode in speclisth and mark alias peers with VALIASED. The function no longer returns a replacement vnode, and callers keep their filesystem vnode identity stable. I still keep the historical bdevvp() reuse case local to getdevvp(): look up an active VT_NON block device vnode before allocating a new vnode, and return it directly when present. This preserves device vnode sharing without making checkalias() consume caller owned vnodes. It should be safe, but I prefer to ask folks to test it. diff --git sys/isofs/cd9660/cd9660_vfsops.c sys/isofs/cd9660/cd9660_vfsops.c index 4e58ff91940..351dd51dd26 100644 --- sys/isofs/cd9660/cd9660_vfsops.c +++ sys/isofs/cd9660/cd9660_vfsops.c @@ -693,7 +693,7 @@ cd9660_vget_internal(struct mount *mp, cdino_t ino, struct vnode **vpp, struct iso_mnt *imp; struct iso_node *ip; struct buf *bp; - struct vnode *vp, *nvp; + struct vnode *vp; dev_t dev; int error; @@ -858,22 +858,7 @@ retry: * if device, look at device number table for translation */ vp->v_op = &cd9660_specvops; - if ((nvp = checkalias(vp, ip->inode.iso_rdev, mp)) != NULL) { - /* - * Discard unneeded vnode, but save its iso_node. - * Note that the lock is carried over in the iso_node - */ - nvp->v_data = vp->v_data; - vp->v_data = NULL; - vp->v_op = &spec_vops; - vrele(vp); - vgone(vp); - /* - * Reinitialize aliased inode. - */ - vp = nvp; - ip->i_vnode = vp; - } + checkalias(vp, ip->inode.iso_rdev); break; case VLNK: case VNON: diff --git sys/isofs/udf/udf_vfsops.c sys/isofs/udf/udf_vfsops.c index 9fc0479cf47..a80793a346f 100644 --- sys/isofs/udf/udf_vfsops.c +++ sys/isofs/udf/udf_vfsops.c @@ -547,7 +547,7 @@ udf_vget(struct mount *mp, ino_t ino, struct vnode **vpp) struct vnode *devvp; struct umount *ump; struct proc *p; - struct vnode *vp, *nvp; + struct vnode *vp; struct unode *up; struct extfile_entry *xfe; struct file_entry *fe; @@ -674,23 +674,7 @@ udf_vget(struct mount *mp, ino_t ino, struct vnode **vpp) } /* check if this is a vnode alias */ - if ((nvp = checkalias(vp, up->u_dev, ump->um_mountp)) != NULL) { - printf("found a vnode alias\n"); - /* - * Discard unneeded vnode, but save its udf_node. - * Note that the lock is carried over in the udf_node - */ - nvp->v_data = vp->v_data; - vp->v_data = NULL; - vp->v_op = &spec_vops; - vrele(vp); - vgone(vp); - /* - * Reinitialize aliased inode. - */ - vp = nvp; - ump->um_devvp = vp; - } + checkalias(vp, up->u_dev); *vpp = vp; diff --git sys/kern/vfs_default.c sys/kern/vfs_default.c index 8a01a2ea5b1..2627c665b22 100644 --- sys/kern/vfs_default.c +++ sys/kern/vfs_default.c @@ -54,6 +54,7 @@ vop_generic_revoke(void *v) struct vop_revoke_args *ap = v; struct vnode *vp, *vq; struct proc *p = curproc; + int referenced = 0; #ifdef DIAGNOSTIC if ((ap->a_flags & REVOKEALL) == 0) @@ -61,7 +62,21 @@ vop_generic_revoke(void *v) #endif vp = ap->a_vp; - + + /* Use mounted alias instead supplied by vdevgone() */ + if (vp->v_type == VBLK && vp->v_specinfo != NULL) { + SLIST_FOREACH(vq, vp->v_hashchain, v_specnext) { + if (vq->v_rdev != vp->v_rdev || + vq->v_type != VBLK || + vq->v_specmountpoint == NULL) + continue; + vp = vq; + vref(vp); + referenced = 1; + break; + } + } + while (vp->v_type == VBLK && vp->v_specinfo != NULL && vp->v_specmountpoint != NULL) { struct mount *mp = vp->v_specmountpoint; @@ -88,7 +103,7 @@ vop_generic_revoke(void *v) msleep_nsec(vp, &vnode_mtx, PINOD, "vop_generic_revokeall", INFSLP); mtx_leave(&vnode_mtx); - return(0); + goto out; } /* @@ -120,6 +135,9 @@ vop_generic_revoke(void *v) vgonel(vp, p); +out: + if (referenced) + vrele(vp); return (0); } diff --git sys/kern/vfs_subr.c sys/kern/vfs_subr.c index 2ab028d0307..57736c1bd29 100644 --- sys/kern/vfs_subr.c +++ sys/kern/vfs_subr.c @@ -112,6 +112,7 @@ void vclean(struct vnode *, int, struct proc *); void insmntque(struct vnode *, struct mount *); int getdevvp(dev_t, struct vnode **, enum vtype); +static struct vnode *getdevalias(dev_t, enum vtype); int vfs_hang_addrlist(struct mount *, struct netexport *, struct export_args *); @@ -526,24 +527,23 @@ int getdevvp(dev_t dev, struct vnode **vpp, enum vtype type) { struct vnode *vp; - struct vnode *nvp; int error; if (dev == NODEV) { *vpp = NULL; return (0); } - error = getnewvnode(VT_NON, NULL, &spec_vops, &nvp); + + if ((*vpp = getdevalias(dev, type)) != NULL) + return (0); + + error = getnewvnode(VT_NON, NULL, &spec_vops, &vp); if (error) { *vpp = NULL; return (error); } - vp = nvp; vp->v_type = type; - if ((nvp = checkalias(vp, dev, NULL)) != NULL) { - vput(vp); - vp = nvp; - } + checkalias(vp, dev); if (vp->v_type == VCHR && cdevsw[major(vp->v_rdev)].d_type == D_TTY) vp->v_flag |= VISTTY; *vpp = vp; @@ -553,13 +553,12 @@ getdevvp(dev_t dev, struct vnode **vpp, enum vtype type) /* * Check to see if the new vnode represents a special device * for which we already have a vnode (either because of - * bdevvp() or because of a different vnode representing - * the same block device). If such an alias exists, deallocate - * the existing contents and return the aliased vnode. The - * caller is responsible for filling it with its new contents. + * bdevvp() or because of a different vnode representing the + * same special device). If such an alias exists, link the new + * vnode into the alias list. */ -struct vnode * -checkalias(struct vnode *nvp, dev_t nvp_rdev, struct mount *mp) +void +checkalias(struct vnode *nvp, dev_t nvp_rdev) { struct proc *p = curproc; struct vnode *vp; @@ -567,7 +566,7 @@ checkalias(struct vnode *nvp, dev_t nvp_rdev, struct mount *mp) u_int vpid; if (nvp->v_type != VBLK && nvp->v_type != VCHR) - return (NULL); + return; vchain = &speclisth[SPECHASH(nvp_rdev)]; loop: @@ -593,53 +592,60 @@ loop: break; } - /* - * Common case is actually in the if statement - */ - if (vp == NULL || !(vp->v_tag == VT_NON && vp->v_type == VBLK)) { - nvp->v_specinfo = malloc(sizeof(struct specinfo), M_VNODE, - M_WAITOK); - nvp->v_rdev = nvp_rdev; - nvp->v_hashchain = vchain; - nvp->v_specmountpoint = NULL; - nvp->v_speclockf = NULL; - nvp->v_specbitmap = NULL; - if (nvp->v_type == VCHR && - (cdevsw[major(nvp_rdev)].d_flags & D_CLONE) && - (minor(nvp_rdev) >> CLONE_SHIFT == 0)) { - if (vp != NULL) - nvp->v_specbitmap = vp->v_specbitmap; - else - nvp->v_specbitmap = malloc(CLONE_MAPSZ, - M_VNODE, M_WAITOK | M_ZERO); - } - SLIST_INSERT_HEAD(vchain, nvp, v_specnext); - if (vp != NULL) { - nvp->v_flag |= VALIASED; - vp->v_flag |= VALIASED; - vput(vp); - } - return (NULL); + nvp->v_specinfo = malloc(sizeof(struct specinfo), M_VNODE, + M_WAITOK); + nvp->v_rdev = nvp_rdev; + nvp->v_hashchain = vchain; + nvp->v_specmountpoint = NULL; + nvp->v_speclockf = NULL; + nvp->v_specbitmap = NULL; + if (nvp->v_type == VCHR && + (cdevsw[major(nvp_rdev)].d_flags & D_CLONE) && + (minor(nvp_rdev) >> CLONE_SHIFT == 0)) { + if (vp != NULL) + nvp->v_specbitmap = vp->v_specbitmap; + else + nvp->v_specbitmap = malloc(CLONE_MAPSZ, + M_VNODE, M_WAITOK | M_ZERO); } + SLIST_INSERT_HEAD(vchain, nvp, v_specnext); + if (vp != NULL) { + nvp->v_flag |= VALIASED; + vp->v_flag |= VALIASED; + vput(vp); + } +} + +static struct vnode * +getdevalias(dev_t dev, enum vtype type) +{ + struct proc *p = curproc; + struct vnode *vp; + struct vnodechain *vchain; /* - * This code is the uncommon case. It is called in case - * we found an alias that was VT_NON && vtype of VBLK - * This means we found a block device that was created - * using bdevvp. - * An example of such a vnode is the root partition device vnode - * created in ffs_mountroot. - * - * The vnodes created by bdevvp should not be aliased (why?). + * Reuse active block device vnodes created by bdevvp(). */ + if (type != VBLK) + return (NULL); - VOP_UNLOCK(vp); - vclean(vp, 0, p); - vp->v_op = nvp->v_op; - vp->v_tag = nvp->v_tag; - nvp->v_type = VNON; - insmntque(vp, mp); - return (vp); + vchain = &speclisth[SPECHASH(dev)]; +loop: + SLIST_FOREACH(vp, vchain, v_specnext) { + if (dev != vp->v_rdev || type != vp->v_type || + vp->v_tag != VT_NON) + continue; + if (vp->v_usecount == 0) { + vgonel(vp, p); + goto loop; + } + if (vget(vp, LK_EXCLUSIVE)) + goto loop; + VOP_UNLOCK(vp); + return (vp); + } + + return (NULL); } /* @@ -1174,7 +1180,7 @@ vgonel(struct vnode *vp, struct proc *p) { struct vnode *vq; struct vnode *vx; - int s, flags = DOCLOSE; + int s; KASSERT(vp->v_uvcount == 0); @@ -1194,9 +1200,7 @@ vgonel(struct vnode *vp, struct proc *p) /* * Clean out the filesystem specific data. */ - if (vp->v_op == &spec_vops && vp->v_specinfo == NULL) - flags = 0; - vclean(vp, flags, p); + vclean(vp, DOCLOSE, p); /* * Delete from old mount point vnode list, if on one. */ diff --git sys/nfs/nfs_subs.c sys/nfs/nfs_subs.c index c22f3378bc6..9927f8e6ac3 100644 --- sys/nfs/nfs_subs.c +++ sys/nfs/nfs_subs.c @@ -943,7 +943,6 @@ nfs_loadattrcache(struct vnode **vpp, struct mbuf **mdp, caddr_t *dposp, enum vtype vtyp; mode_t vmode; struct timespec mtime; - struct vnode *nvp; int v3 = NFS_ISV3(vp); uid_t uid; gid_t gid; @@ -995,25 +994,7 @@ nfs_loadattrcache(struct vnode **vpp, struct mbuf **mdp, caddr_t *dposp, } if (vp->v_type == VCHR || vp->v_type == VBLK) { vp->v_op = &nfs_specvops; - nvp = checkalias(vp, (dev_t)rdev, vp->v_mount); - if (nvp) { - /* - * Discard unneeded vnode, but save its nfsnode. - * Since the nfsnode does not have a lock, its - * vnode lock has to be carried over. - */ - - nvp->v_data = vp->v_data; - vp->v_data = NULL; - vp->v_op = &spec_vops; - vrele(vp); - vgone(vp); - /* - * Reinitialize aliased node. - */ - np->n_vnode = nvp; - *vpp = vp = nvp; - } + checkalias(vp, (dev_t)rdev); } np->n_mtime = mtime; } diff --git sys/sys/vnode.h sys/sys/vnode.h index 6f1664ce18e..0dd0d754019 100644 --- sys/sys/vnode.h +++ sys/sys/vnode.h @@ -576,7 +576,7 @@ struct vnode; /* vfs_subr */ int bdevvp(dev_t, struct vnode **); int cdevvp(dev_t, struct vnode **); -struct vnode *checkalias(struct vnode *, dev_t, struct mount *); +void checkalias(struct vnode *, dev_t); int getnewvnode(enum vtagtype, struct mount *, const struct vops *, struct vnode **); int vaccess(enum vtype, mode_t, uid_t, gid_t, mode_t, struct ucred *); diff --git sys/tmpfs/tmpfs_subr.c sys/tmpfs/tmpfs_subr.c index 6b5322e1fb6..b30e3bdc916 100644 --- sys/tmpfs/tmpfs_subr.c +++ sys/tmpfs/tmpfs_subr.c @@ -271,7 +271,7 @@ tmpfs_free_node(tmpfs_mount_t *tmp, tmpfs_node_t *node) int tmpfs_vnode_get(struct mount *mp, tmpfs_node_t *node, struct vnode **vpp) { - struct vnode *vp, *nvp; + struct vnode *vp; /* kmutex_t *slock; */ int error; again: @@ -321,15 +321,7 @@ again: case VBLK: case VCHR: vp->v_op = &tmpfs_specvops; - if ((nvp = checkalias(vp, node->tn_spec.tn_dev.tn_rdev, mp))) { - nvp->v_data = vp->v_data; - vp->v_data = NULL; - vp->v_op = &spec_vops; - vrele(vp); - vgone(vp); - vp = nvp; - node->tn_vnode = vp; - } + checkalias(vp, node->tn_spec.tn_dev.tn_rdev); break; case VDIR: vp->v_flag |= node->tn_spec.tn_dir.tn_parent == node ? diff --git sys/ufs/ext2fs/ext2fs_subr.c sys/ufs/ext2fs/ext2fs_subr.c index 68769d8f28a..721ed0b5292 100644 --- sys/ufs/ext2fs/ext2fs_subr.c +++ sys/ufs/ext2fs/ext2fs_subr.c @@ -117,7 +117,7 @@ int ext2fs_vinit(struct mount *mp, struct vnode **vpp) { struct inode *ip; - struct vnode *vp, *nvp; + struct vnode *vp; struct timeval tv; vp = *vpp; @@ -128,26 +128,7 @@ ext2fs_vinit(struct mount *mp, struct vnode **vpp) case VCHR: case VBLK: vp->v_op = &ext2fs_specvops; - - nvp = checkalias(vp, letoh32(ip->i_e2din->e2di_rdev), mp); - if (nvp != NULL) { - /* - * Discard unneeded vnode, but save its inode. Note - * that the lock is carried over in the inode to the - * replacement vnode. - */ - nvp->v_data = vp->v_data; - vp->v_data = NULL; - vp->v_op = &spec_vops; -#ifdef VFSLCKDEBUG - vp->v_flag &= ~VLOCKSWORK; -#endif - vrele(vp); - vgone(vp); - /* Reinitialize aliased vnode. */ - vp = nvp; - ip->i_vnode = vp; - } + checkalias(vp, letoh32(ip->i_e2din->e2di_rdev)); break; diff --git sys/ufs/ffs/ffs_subr.c sys/ufs/ffs/ffs_subr.c index de81df8b619..2d8dbb22c1a 100644 --- sys/ufs/ffs/ffs_subr.c +++ sys/ufs/ffs/ffs_subr.c @@ -222,7 +222,7 @@ int ffs_vinit(struct mount *mntp, struct vnode **vpp) { struct inode *ip; - struct vnode *vp, *nvp; + struct vnode *vp; struct timeval mtv; vp = *vpp; @@ -231,26 +231,7 @@ ffs_vinit(struct mount *mntp, struct vnode **vpp) case VCHR: case VBLK: vp->v_op = &ffs_specvops; - if ((nvp = checkalias(vp, DIP(ip, rdev), mntp)) != NULL) { - /* - * Discard unneeded vnode, but save its inode. - * Note that the lock is carried over in the inode - * to the replacement vnode. - */ - nvp->v_data = vp->v_data; - vp->v_data = NULL; - vp->v_op = &spec_vops; -#ifdef VFSLCKDEBUG - vp->v_flag &= ~VLOCKSWORK; -#endif - vrele(vp); - vgone(vp); - /* - * Reinitialize aliased inode. - */ - vp = nvp; - ip->i_vnode = vp; - } + checkalias(vp, DIP(ip, rdev)); break; case VFIFO: #ifdef FIFO diff --git sys/ufs/mfs/mfs_vfsops.c sys/ufs/mfs/mfs_vfsops.c index 4ce1859b773..438c4846585 100644 --- sys/ufs/mfs/mfs_vfsops.c +++ sys/ufs/mfs/mfs_vfsops.c @@ -122,8 +122,7 @@ mfs_mount(struct mount *mp, const char *path, void *data, if (error) return (error); devvp->v_type = VBLK; - if (checkalias(devvp, makedev(255, mfs_minor), NULL)) - panic("mfs_mount: dup dev"); + checkalias(devvp, makedev(255, mfs_minor)); mfs_minor++; mfsp = malloc(sizeof *mfsp, M_MFSNODE, M_WAITOK | M_ZERO); devvp->v_data = mfsp; -- wbr, Kirill