/* * Lock a directory entry. A dirlock on <dzp, name> protects that name * in dzp's directory zap object. As long as you hold a dirlock, you can * assume two things: (1) dzp cannot be reaped, and (2) no other thread * can change the zap entry for (i.e. link or unlink) this name. * * Input arguments: * dzp - znode for directory * name - name of entry to lock * flag - ZNEW: if the entry already exists, fail with EEXIST. * ZEXISTS: if the entry does not exist, fail with ENOENT. * ZSHARED: allow concurrent access with other ZSHARED callers. * ZXATTR: we want dzp's xattr directory * ZCILOOK: On a mixed sensitivity file system, * this lookup should be case-insensitive. * ZCIEXACT: On a purely case-insensitive file system, * this lookup should be case-sensitive. * ZRENAMING: we are locking for renaming, force narrow locks * ZHAVELOCK: Don't grab the z_name_lock for this call. The * current thread already holds it. * * Output arguments: * zpp - pointer to the znode for the entry (NULL if there isn't one) * dlpp - pointer to the dirlock for this entry (NULL on error) * direntflags - (case-insensitive lookup only) * flags if multiple case-sensitive matches exist in directory * realpnp - (case-insensitive lookup only) * actual name matched within the directory * * Return value: 0 on success or errno on failure. * * NOTE: Always checks for, and rejects, '.' and '..'. * NOTE: For case-insensitive file systems we take wide locks (see below), * but return znode pointers to a single match. */ int zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp, int flag, int *direntflags, pathname_t *realpnp) { zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zfs_dirlock_t *dl; boolean_t update; boolean_t exact; uint64_t zoid; vnode_t *vp = NULL; int error = 0; int cmpflags; *zpp = NULL; *dlpp = NULL; /* * Verify that we are not trying to lock '.', '..', or '.zfs' */ if ((name[0] == '.' && (name[1] == '\0' || (name[1] == '.' && name[2] == '\0'))) || (zfs_has_ctldir(dzp) && strcmp(name, ZFS_CTLDIR_NAME) == 0)) return (EEXIST); /* * Case sensitivity and normalization preferences are set when * the file system is created. These are stored in the * zfsvfs->z_case and zfsvfs->z_norm fields. These choices * affect what vnodes can be cached in the DNLC, how we * perform zap lookups, and the "width" of our dirlocks. * * A normal dirlock locks a single name. Note that with * normalization a name can be composed multiple ways, but * when normalized, these names all compare equal. A wide * dirlock locks multiple names. We need these when the file * system is supporting mixed-mode access. It is sometimes * necessary to lock all case permutations of file name at * once so that simultaneous case-insensitive/case-sensitive * behaves as rationally as possible. */ /* * Decide if exact matches should be requested when performing * a zap lookup on file systems supporting case-insensitive * access. */ exact = ((zfsvfs->z_case == ZFS_CASE_INSENSITIVE) && (flag & ZCIEXACT)) || ((zfsvfs->z_case == ZFS_CASE_MIXED) && !(flag & ZCILOOK)); /* * Only look in or update the DNLC if we are looking for the * name on a file system that does not require normalization * or case folding. We can also look there if we happen to be * on a non-normalizing, mixed sensitivity file system IF we * are looking for the exact name. * * Maybe can add TO-UPPERed version of name to dnlc in ci-only * case for performance improvement? */ update = !zfsvfs->z_norm || ((zfsvfs->z_case == ZFS_CASE_MIXED) && !(zfsvfs->z_norm & ~U8_TEXTPREP_TOUPPER) && !(flag & ZCILOOK)); /* * ZRENAMING indicates we are in a situation where we should * take narrow locks regardless of the file system's * preferences for normalizing and case folding. This will * prevent us deadlocking trying to grab the same wide lock * twice if the two names happen to be case-insensitive * matches. */ if (flag & ZRENAMING) cmpflags = 0; else cmpflags = zfsvfs->z_norm; /* * Wait until there are no locks on this name. * * Don't grab the the lock if it is already held. However, cannot * have both ZSHARED and ZHAVELOCK together. */ ASSERT(!(flag & ZSHARED) || !(flag & ZHAVELOCK)); if (!(flag & ZHAVELOCK)) rw_enter(&dzp->z_name_lock, RW_READER); mutex_enter(&dzp->z_lock); for (;;) { if (dzp->z_unlinked) { mutex_exit(&dzp->z_lock); if (!(flag & ZHAVELOCK)) rw_exit(&dzp->z_name_lock); return (ENOENT); } for (dl = dzp->z_dirlocks; dl != NULL; dl = dl->dl_next) { if ((u8_strcmp(name, dl->dl_name, 0, cmpflags, U8_UNICODE_LATEST, &error) == 0) || error != 0) break; } if (error != 0) { mutex_exit(&dzp->z_lock); if (!(flag & ZHAVELOCK)) rw_exit(&dzp->z_name_lock); return (ENOENT); } if (dl == NULL) { /* * Allocate a new dirlock and add it to the list. */ dl = kmem_alloc(sizeof (zfs_dirlock_t), KM_SLEEP); cv_init(&dl->dl_cv, NULL, CV_DEFAULT, NULL); dl->dl_name = name; dl->dl_sharecnt = 0; dl->dl_namelock = 0; dl->dl_namesize = 0; dl->dl_dzp = dzp; dl->dl_next = dzp->z_dirlocks; dzp->z_dirlocks = dl; break; } if ((flag & ZSHARED) && dl->dl_sharecnt != 0) break; cv_wait(&dl->dl_cv, &dzp->z_lock); } /* * If the z_name_lock was NOT held for this dirlock record it. */ if (flag & ZHAVELOCK) dl->dl_namelock = 1; if ((flag & ZSHARED) && ++dl->dl_sharecnt > 1 && dl->dl_namesize == 0) { /* * We're the second shared reference to dl. Make a copy of * dl_name in case the first thread goes away before we do. * Note that we initialize the new name before storing its * pointer into dl_name, because the first thread may load * dl->dl_name at any time. He'll either see the old value, * which is his, or the new shared copy; either is OK. */ dl->dl_namesize = strlen(dl->dl_name) + 1; name = kmem_alloc(dl->dl_namesize, KM_SLEEP); bcopy(dl->dl_name, name, dl->dl_namesize); dl->dl_name = name; } mutex_exit(&dzp->z_lock); /* * We have a dirlock on the name. (Note that it is the dirlock, * not the dzp's z_lock, that protects the name in the zap object.) * See if there's an object by this name; if so, put a hold on it. */ if (flag & ZXATTR) { error = sa_lookup(dzp->z_sa_hdl, SA_ZPL_XATTR(zfsvfs), &zoid, sizeof (zoid)); if (error == 0) error = (zoid == 0 ? ENOENT : 0); } else { if (update) vp = dnlc_lookup(ZTOV(dzp), name); if (vp == DNLC_NO_VNODE) { VN_RELE(vp); error = ENOENT; } else if (vp) { if (flag & ZNEW) { zfs_dirent_unlock(dl); VN_RELE(vp); return (EEXIST); } *dlpp = dl; *zpp = VTOZ(vp); return (0); } else { error = zfs_match_find(zfsvfs, dzp, name, exact, update, direntflags, realpnp, &zoid); } } if (error) { if (error != ENOENT || (flag & ZEXISTS)) { zfs_dirent_unlock(dl); return (error); } } else { if (flag & ZNEW) { zfs_dirent_unlock(dl); return (EEXIST); } error = zfs_zget(zfsvfs, zoid, zpp); if (error) { zfs_dirent_unlock(dl); return (error); } if (!(flag & ZXATTR) && update) dnlc_update(ZTOV(dzp), name, ZTOV(*zpp)); } *dlpp = dl; return (0); }
zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp, int flag) #endif { zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zfs_dirlock_t *dl; uint64_t zoid; int error; vnode_t *vp; #ifdef __APPLE__ char *name; u_int8_t *nfc_name = NULL; /* NFC form of name */ int nfc_namesize = 0; #endif *zpp = NULL; *dlpp = NULL; #ifdef __APPLE__ /* Note: cnp will be NULL for ZXATTR case */ name = cnp ? cnp->cn_nameptr : ""; if (cnp) ASSERT(name[cnp->cn_namelen] == '\0'); #endif /* * Verify that we are not trying to lock '.', '..', or '.zfs' */ if ((name[0] == '.') && ((name[1] == '\0') || ((name[1] == '.') && (name[2] == '\0'))) || zfs_has_ctldir(dzp) && strcmp(name, ZFS_CTLDIR_NAME) == 0) return (EEXIST); #ifdef __APPLE__ /* * Mac OS X: store non-ascii names in UTF-8 NFC (pre-composed) on disk. * * The NFC name ptr is stored in dl->dl_name (allocated here) * and its freed by zfs_dirent_unlock (since dl_namesize != 0). * * Since NFC size will not expand, we can allocate the same sized buffer. */ if (!is_ascii_str(name)) { size_t outlen; nfc_namesize = strlen(name) + 1; nfc_name = kmem_alloc(nfc_namesize, KM_SLEEP); if (utf8_normalizestr((const u_int8_t *)name, nfc_namesize, nfc_name, &outlen, nfc_namesize, UTF_PRECOMPOSED) == 0) { /* Normalization succeeded, switch to NFC name. */ name = (char *)nfc_name; } else { /* Normalization failed, just use input name as-is. */ kmem_free(nfc_name, nfc_namesize); nfc_name = NULL; } } #endif /* * Wait until there are no locks on this name. */ rw_enter(&dzp->z_name_lock, RW_READER); mutex_enter(&dzp->z_lock); for (;;) { if (dzp->z_unlinked) { mutex_exit(&dzp->z_lock); rw_exit(&dzp->z_name_lock); #ifdef __APPLE__ /* Release any unused NFC name before returning */ if (nfc_name) { kmem_free(nfc_name, nfc_namesize); } #endif return (ENOENT); } for (dl = dzp->z_dirlocks; dl != NULL; dl = dl->dl_next) if (strcmp(name, dl->dl_name) == 0) break; if (dl == NULL) { /* * Allocate a new dirlock and add it to the list. */ dl = kmem_alloc(sizeof (zfs_dirlock_t), KM_SLEEP); cv_init(&dl->dl_cv, NULL, CV_DEFAULT, NULL); dl->dl_name = name; dl->dl_sharecnt = 0; dl->dl_namesize = 0; dl->dl_dzp = dzp; dl->dl_next = dzp->z_dirlocks; dzp->z_dirlocks = dl; #ifdef __APPLE__ /* * Keep the NFC name around in dir lock by tagging it * (setting nfc_namesize). */ if (nfc_name) { dl->dl_namesize = nfc_namesize; nfc_name = NULL; /* its now part of the dir lock */ } #endif break; } if ((flag & ZSHARED) && dl->dl_sharecnt != 0) break; cv_wait(&dl->dl_cv, &dzp->z_lock); dl=NULL; } #ifdef __APPLE__ /* * Release any unused NFC name (ie if we found a pre-existing lock entry) */ if (nfc_name) { kmem_free(nfc_name, nfc_namesize); nfc_name = NULL; } #endif if ((flag & ZSHARED) && ++dl->dl_sharecnt > 1 && dl->dl_namesize == 0) { /* * We're the second shared reference to dl. Make a copy of * dl_name in case the first thread goes away before we do. * Note that we initialize the new name before storing its * pointer into dl_name, because the first thread may load * dl->dl_name at any time. He'll either see the old value, * which is his, or the new shared copy; either is OK. */ dl->dl_namesize = strlen(dl->dl_name) + 1; name = kmem_alloc(dl->dl_namesize, KM_SLEEP); bcopy(dl->dl_name, name, dl->dl_namesize); dl->dl_name = name; } mutex_exit(&dzp->z_lock); /* * We have a dirlock on the name. (Note that it is the dirlock, * not the dzp's z_lock, that protects the name in the zap object.) * See if there's an object by this name; if so, put a hold on it. */ if (flag & ZXATTR) { error = sa_lookup(dzp->z_sa_hdl, SA_ZPL_XATTR(zfsvfs), &zoid, sizeof (zoid)); if (error == 0) error = (zoid == 0 ? ENOENT : 0); } else { #ifdef __APPLE__ /* * Lookup an entry in the vnode name cache * * If the lookup succeeds, the vnode is returned in *vpp, * and a status of -1 is returned. * * If the lookup determines that the name does not exist * (negative caching), a status of ENOENT is returned. * * If the lookup fails, a status of zero is returned. */ switch ( cache_lookup(ZTOV(dzp), &vp, cnp) ) { case -1: break; case ENOENT: vp = DNLC_NO_VNODE; break; default: vp = NULLVP; } #else vp = dnlc_lookup(ZTOV(dzp), name); #endif /* __APPLE__ */ if (vp == DNLC_NO_VNODE) { VN_RELE(vp); error = ENOENT; } else if (vp) { if (flag & ZNEW) { zfs_dirent_unlock(dl); VN_RELE(vp); return (EEXIST); } *dlpp = dl; *zpp = VTOZ(vp); return (0); } else { error = zap_lookup(zfsvfs->z_os, dzp->z_id, name, 8, 1, &zoid); zoid = ZFS_DIRENT_OBJ(zoid); if (error == ENOENT) #ifdef __APPLE__ /* * Add a negative entry into the VFS name cache */ if ((flag & ZNEW) == 0 && (dzp->z_pflags & ZFS_XATTR) == 0 && (cnp) && (cnp->cn_flags & MAKEENTRY) && (cnp->cn_nameiop != CREATE) && (cnp->cn_nameiop != RENAME)) { cache_enter(ZTOV(dzp), NULLVP, cnp); } #else dnlc_update(ZTOV(dzp), name, DNLC_NO_VNODE); #endif /* __APPLE__ */ } } if (error) { if (error != ENOENT || (flag & ZEXISTS)) { zfs_dirent_unlock(dl); return (error); } } else { if (flag & ZNEW) { zfs_dirent_unlock(dl); return (EEXIST); } //error = zfs_zget_sans_vnode(zfsvfs, zoid, zpp); error = zfs_zget(zfsvfs, zoid, zpp); if (error) { zfs_dirent_unlock(dl); return (error); } else { // Should this be here? //printf("zfs_dir attach 1\n"); //zfs_attach_vnode(*zpp); } if (!(flag & ZXATTR)) #ifdef __APPLE__ if (cnp && cnp->cn_flags & MAKEENTRY) cache_enter(ZTOV(dzp), ZTOV(*zpp), cnp); #else dnlc_update(ZTOV(dzp), name, ZTOV(*zpp)); #endif /* __APPLE__ */ } *dlpp = dl; return (0); }
/* * Lock a directory entry. A dirlock on <dzp, name> protects that name * in dzp's directory zap object. As long as you hold a dirlock, you can * assume two things: (1) dzp cannot be reaped, and (2) no other thread * can change the zap entry for (i.e. link or unlink) this name. * * Input arguments: * dzp - znode for directory * name - name of entry to lock * flag - ZNEW: if the entry already exists, fail with EEXIST. * ZEXISTS: if the entry does not exist, fail with ENOENT. * ZSHARED: allow concurrent access with other ZSHARED callers. * ZXATTR: we want dzp's xattr directory * * Output arguments: * zpp - pointer to the znode for the entry (NULL if there isn't one) * dlpp - pointer to the dirlock for this entry (NULL on error) * * Return value: 0 on success or errno on failure. * * NOTE: Always checks for, and rejects, '.' and '..'. */ int zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp, int flag) { zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zfs_dirlock_t *dl; uint64_t zoid; int error; vnode_t *vp; *zpp = NULL; *dlpp = NULL; /* * Verify that we are not trying to lock '.', '..', or '.zfs' */ if (name[0] == '.' && (name[1] == '\0' || (name[1] == '.' && name[2] == '\0')) || zfs_has_ctldir(dzp) && strcmp(name, ZFS_CTLDIR_NAME) == 0) return (EEXIST); /* * Wait until there are no locks on this name. */ mutex_enter(&dzp->z_lock); for (;;) { if (dzp->z_reap) { mutex_exit(&dzp->z_lock); return (ENOENT); } for (dl = dzp->z_dirlocks; dl != NULL; dl = dl->dl_next) if (strcmp(name, dl->dl_name) == 0) break; if (dl == NULL) { /* * Allocate a new dirlock and add it to the list. */ dl = kmem_alloc(sizeof (zfs_dirlock_t), KM_SLEEP); cv_init(&dl->dl_cv, NULL, CV_DEFAULT, NULL); dl->dl_name = name; dl->dl_sharecnt = 0; dl->dl_namesize = 0; dl->dl_dzp = dzp; dl->dl_next = dzp->z_dirlocks; dzp->z_dirlocks = dl; break; } if ((flag & ZSHARED) && dl->dl_sharecnt != 0) break; cv_wait(&dl->dl_cv, &dzp->z_lock); } if ((flag & ZSHARED) && ++dl->dl_sharecnt > 1 && dl->dl_namesize == 0) { /* * We're the second shared reference to dl. Make a copy of * dl_name in case the first thread goes away before we do. * Note that we initialize the new name before storing its * pointer into dl_name, because the first thread may load * dl->dl_name at any time. He'll either see the old value, * which is his, or the new shared copy; either is OK. */ dl->dl_namesize = strlen(dl->dl_name) + 1; name = kmem_alloc(dl->dl_namesize, KM_SLEEP); bcopy(dl->dl_name, name, dl->dl_namesize); dl->dl_name = name; } mutex_exit(&dzp->z_lock); /* * We have a dirlock on the name. (Note that it is the dirlock, * not the dzp's z_lock, that protects the name in the zap object.) * See if there's an object by this name; if so, put a hold on it. */ if (flag & ZXATTR) { zoid = dzp->z_phys->zp_xattr; error = (zoid == 0 ? ENOENT : 0); } else { vp = dnlc_lookup(ZTOV(dzp), name); if (vp == DNLC_NO_VNODE) { VN_RELE(vp); error = ENOENT; } else if (vp) { if (flag & ZNEW) { zfs_dirent_unlock(dl); VN_RELE(vp); return (EEXIST); } *dlpp = dl; *zpp = VTOZ(vp); return (0); } else { error = zap_lookup(zfsvfs->z_os, dzp->z_id, name, 8, 1, &zoid); if (error == ENOENT) dnlc_update(ZTOV(dzp), name, DNLC_NO_VNODE); } } if (error) { if (error != ENOENT || (flag & ZEXISTS)) { zfs_dirent_unlock(dl); return (error); } } else { if (flag & ZNEW) { zfs_dirent_unlock(dl); return (EEXIST); } error = zfs_zget(zfsvfs, zoid, zpp); if (error) { zfs_dirent_unlock(dl); return (error); } if (!(flag & ZXATTR)) dnlc_update(ZTOV(dzp), name, ZTOV(*zpp)); } *dlpp = dl; return (0); }
int ud_dirlook(struct ud_inode *dip, char *namep, struct ud_inode **ipp, struct cred *cr, int32_t skipdnlc) { struct udf_vfs *udf_vfsp; int32_t error = 0, namelen, adhoc_search; u_offset_t offset, adhoc_offset, dirsize, end; struct vnode *dvp, *vp; struct fbuf *fbp; struct file_id *fid; uint8_t *fname, dummy[3]; int32_t id_len, doingchk; uint32_t old_loc; uint16_t old_prn; uint8_t *dname; uint8_t *buf = NULL; ud_printf("ud_dirlook\n"); udf_vfsp = dip->i_udf; restart: doingchk = 0; old_prn = 0xFFFF; old_loc = 0; dvp = ITOV(dip); /* * Check accessibility of directory. */ if (dip->i_type != VDIR) { return (ENOTDIR); } if (error = ud_iaccess(dip, IEXEC, cr)) { return (error); } /* * Null component name is synonym for directory being searched. */ if (*namep == '\0') { VN_HOLD(dvp); *ipp = dip; return (0); } namelen = strlen(namep); if ((namelen == 1) && (namep[0] == '.') && (namep[1] == '\0')) { /* Current directory */ VN_HOLD(dvp); *ipp = dip; dnlc_enter(dvp, namep, ITOV(*ipp)); return (0); } if ((!skipdnlc) && (vp = dnlc_lookup(dvp, namep))) { /* vp is already held from dnlc_lookup */ *ipp = VTOI(vp); return (0); } dname = kmem_zalloc(1024, KM_SLEEP); buf = kmem_zalloc(udf_vfsp->udf_lbsize, KM_SLEEP); /* * Read lock the inode we are searching. You will notice that we * didn't hold the read lock while searching the dnlc. This means * that the entry could now be in the dnlc. This doesn't cause any * problems because dnlc_enter won't add an entry if it is already * there. */ rw_enter(&dip->i_rwlock, RW_READER); /* * Take care to look at dip->i_diroff only once, as it * may be changing due to other threads/cpus. */ recheck: offset = dip->i_diroff; end = dirsize = dip->i_size; if (offset > dirsize) { offset = 0; } adhoc_offset = offset; adhoc_search = (offset == 0) ? 1 : 2; fbp = NULL; while (adhoc_search--) { while (offset < end) { error = ud_get_next_fid(dip, &fbp, offset, &fid, &fname, buf); if (error != 0) { break; } if ((fid->fid_flags & FID_DELETED) == 0) { if (fid->fid_flags & FID_PARENT) { id_len = 2; fname = dummy; dummy[0] = '.'; dummy[1] = '.'; dummy[2] = '\0'; } else { if ((error = ud_uncompress( fid->fid_idlen, &id_len, fname, dname)) != 0) { break; } fname = (uint8_t *)dname; fname[id_len] = '\0'; } if ((namelen == id_len) && (strncmp(namep, (caddr_t)fname, namelen) == 0)) { uint32_t loc; uint16_t prn; loc = SWAP_32(fid->fid_icb.lad_ext_loc); prn = SWAP_16(fid->fid_icb.lad_ext_prn); dip->i_diroff = offset + FID_LEN(fid); if (doingchk) { if ((loc == old_loc) && (prn == old_prn)) { goto checkok; } else { if (fbp != NULL) { fbrelse(fbp, S_READ); fbp = NULL; } VN_RELE(ITOV(*ipp)); rw_exit(&dip->i_rwlock); goto restart; } /* NOTREACHED */ } if (namelen == 2 && fname[0] == '.' && fname[1] == '.') { struct timespec32 omtime; omtime = dip->i_mtime; rw_exit(&dip->i_rwlock); error = ud_iget(dip->i_vfs, prn, loc, ipp, NULL, cr); rw_enter(&dip->i_rwlock, RW_READER); if (error) { goto done; } if ((omtime.tv_sec != dip->i_mtime.tv_sec) || (omtime.tv_nsec != dip->i_mtime.tv_nsec)) { doingchk = 1; old_prn = prn; old_loc = loc; dip->i_diroff = 0; if (fbp != NULL) { fbrelse(fbp, S_READ); fbp = NULL; } goto recheck; } } else { error = ud_iget(dip->i_vfs, prn, loc, ipp, NULL, cr); } checkok: if (error == 0) { dnlc_enter(dvp, namep, ITOV(*ipp)); } goto done; } } offset += FID_LEN(fid); } if (fbp != NULL) { fbrelse(fbp, S_READ); fbp = NULL; } end = adhoc_offset; offset = 0; } error = ENOENT; done: kmem_free(buf, udf_vfsp->udf_lbsize); kmem_free(dname, 1024); if (fbp != NULL) { fbrelse(fbp, S_READ); } rw_exit(&dip->i_rwlock); return (error); }