X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/0b4c1975fb5e4eccf1012a35081f7e7799b81046..db6096698656d32db7df630594bd9617ee54f828:/bsd/hfs/hfs_vnops.c diff --git a/bsd/hfs/hfs_vnops.c b/bsd/hfs/hfs_vnops.c index 49973d29c..1d0497a0f 100644 --- a/bsd/hfs/hfs_vnops.c +++ b/bsd/hfs/hfs_vnops.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2000-2009 Apple Inc. All rights reserved. + * Copyright (c) 2000-2013 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * @@ -33,6 +33,7 @@ #include #include #include +#include #include #include #include @@ -45,6 +46,10 @@ #include #include #include +#include +#include +#include +#include #include #include @@ -72,20 +77,25 @@ /* Always F_FULLFSYNC? 1=yes,0=no (default due to "various" reasons is 'no') */ int always_do_fullfsync = 0; SYSCTL_DECL(_vfs_generic); -SYSCTL_INT (_vfs_generic, OID_AUTO, always_do_fullfsync, CTLFLAG_RW, &always_do_fullfsync, 0, "always F_FULLFSYNC when fsync is called"); +SYSCTL_INT (_vfs_generic, OID_AUTO, always_do_fullfsync, CTLFLAG_RW | CTLFLAG_LOCKED, &always_do_fullfsync, 0, "always F_FULLFSYNC when fsync is called"); -static int hfs_makenode(struct vnode *dvp, struct vnode **vpp, +int hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct vnode_attr *vap, vfs_context_t ctx); +int hfs_metasync(struct hfsmount *hfsmp, daddr64_t node, __unused struct proc *p); +int hfs_metasync_all(struct hfsmount *hfsmp); -static int hfs_metasync(struct hfsmount *hfsmp, daddr64_t node, __unused struct proc *p); -static int hfs_metasync_all(struct hfsmount *hfsmp); +int hfs_removedir(struct vnode *, struct vnode *, struct componentname *, + int, int); +int hfs_removefile(struct vnode *, struct vnode *, struct componentname *, + int, int, int, struct vnode *, int); -static int hfs_removedir(struct vnode *, struct vnode *, struct componentname *, - int); +/* Used here and in cnode teardown -- for symlinks */ +int hfs_removefile_callback(struct buf *bp, void *hfsmp); -static int hfs_removefile(struct vnode *, struct vnode *, struct componentname *, - int, int, int, struct vnode *); +int hfs_movedata (struct vnode *, struct vnode*); +static int hfs_move_fork (struct filefork *srcfork, struct cnode *src, + struct filefork *dstfork, struct cnode *dst); #if FIFO static int hfsfifo_read(struct vnop_read_args *); @@ -95,26 +105,27 @@ static int hfsfifo_close(struct vnop_close_args *); extern int (**fifo_vnodeop_p)(void *); #endif /* FIFO */ -static int hfs_vnop_close(struct vnop_close_args*); -static int hfs_vnop_create(struct vnop_create_args*); -static int hfs_vnop_exchange(struct vnop_exchange_args*); -static int hfs_vnop_fsync(struct vnop_fsync_args*); -static int hfs_vnop_mkdir(struct vnop_mkdir_args*); -static int hfs_vnop_mknod(struct vnop_mknod_args*); -static int hfs_vnop_getattr(struct vnop_getattr_args*); -static int hfs_vnop_open(struct vnop_open_args*); -static int hfs_vnop_readdir(struct vnop_readdir_args*); -static int hfs_vnop_remove(struct vnop_remove_args*); -static int hfs_vnop_rename(struct vnop_rename_args*); -static int hfs_vnop_rmdir(struct vnop_rmdir_args*); -static int hfs_vnop_symlink(struct vnop_symlink_args*); -static int hfs_vnop_setattr(struct vnop_setattr_args*); -static int hfs_vnop_readlink(struct vnop_readlink_args *); -static int hfs_vnop_pathconf(struct vnop_pathconf_args *); -static int hfs_vnop_whiteout(struct vnop_whiteout_args *); -static int hfsspec_read(struct vnop_read_args *); -static int hfsspec_write(struct vnop_write_args *); -static int hfsspec_close(struct vnop_close_args *); +int hfs_vnop_close(struct vnop_close_args*); +int hfs_vnop_create(struct vnop_create_args*); +int hfs_vnop_exchange(struct vnop_exchange_args*); +int hfs_vnop_fsync(struct vnop_fsync_args*); +int hfs_vnop_mkdir(struct vnop_mkdir_args*); +int hfs_vnop_mknod(struct vnop_mknod_args*); +int hfs_vnop_getattr(struct vnop_getattr_args*); +int hfs_vnop_open(struct vnop_open_args*); +int hfs_vnop_readdir(struct vnop_readdir_args*); +int hfs_vnop_remove(struct vnop_remove_args*); +int hfs_vnop_rename(struct vnop_rename_args*); +int hfs_vnop_rmdir(struct vnop_rmdir_args*); +int hfs_vnop_symlink(struct vnop_symlink_args*); +int hfs_vnop_setattr(struct vnop_setattr_args*); +int hfs_vnop_readlink(struct vnop_readlink_args *); +int hfs_vnop_pathconf(struct vnop_pathconf_args *); +int hfs_vnop_whiteout(struct vnop_whiteout_args *); +int hfs_vnop_mmap(struct vnop_mmap_args *ap); +int hfsspec_read(struct vnop_read_args *); +int hfsspec_write(struct vnop_write_args *); +int hfsspec_close(struct vnop_close_args *); /* Options for hfs_removedir and hfs_removefile */ #define HFSRM_SKIP_RESERVE 0x01 @@ -128,10 +139,31 @@ static int hfsspec_close(struct vnop_close_args *); * *****************************************************************************/ +/* + * Is the given cnode either the .journal or .journal_info_block file on + * a volume with an active journal? Many VNOPs use this to deny access + * to those files. + * + * Note: the .journal file on a volume with an external journal still + * returns true here, even though it does not actually hold the contents + * of the volume's journal. + */ +static _Bool +hfs_is_journal_file(struct hfsmount *hfsmp, struct cnode *cp) +{ + if (hfsmp->jnl != NULL && + (cp->c_fileid == hfsmp->hfs_jnlinfoblkid || + cp->c_fileid == hfsmp->hfs_jnlfileid)) { + return true; + } else { + return false; + } +} + /* * Create a regular file. */ -static int +int hfs_vnop_create(struct vnop_create_args *ap) { int error; @@ -164,6 +196,7 @@ again: /* Make sure it was file. */ if ((error == 0) && !vnode_isreg(*args.a_vpp)) { vnode_put(*args.a_vpp); + *args.a_vpp = NULLVP; error = EEXIST; } args.a_cnp->cn_nameiop = CREATE; @@ -174,7 +207,7 @@ again: /* * Make device special file. */ -static int +int hfs_vnop_mknod(struct vnop_mknod_args *ap) { struct vnode_attr *vap = ap->a_vap; @@ -245,7 +278,7 @@ hfs_ref_data_vp(struct cnode *cp, struct vnode **data_vp, int skiplock) return EINVAL; } - if (0 == hfs_vget(VTOHFS(cp->c_rsrc_vp), cp->c_cnid, data_vp, 1) && + if (0 == hfs_vget(VTOHFS(cp->c_rsrc_vp), cp->c_cnid, data_vp, 1, 0) && 0 != data_vp) { vref = vnode_ref(*data_vp); vnode_put(*data_vp); @@ -298,7 +331,7 @@ hfs_file_is_compressed(struct cnode *cp, int skiplock) int ret = 0; /* fast check to see if file is compressed. If flag is clear, just answer no */ - if (!(cp->c_flags & UF_COMPRESSED)) { + if (!(cp->c_bsdflags & UF_COMPRESSED)) { return 0; } @@ -334,6 +367,8 @@ hfs_file_is_compressed(struct cnode *cp, int skiplock) * if the caller has passed a valid vnode (has a ref count > 0), then hfsmp and fid are not required. * if the caller doesn't have a vnode, pass NULL in vp, and pass valid hfsmp and fid. * files size is returned in size (required) + * if the indicated file is a directory (or something that doesn't have a data fork), then this call + * will return an error and the caller should fall back to treating the item as an uncompressed file */ int hfs_uncompressed_size_of_compressed_file(struct hfsmount *hfsmp, struct vnode *vp, cnid_t fid, off_t *size, int skiplock) @@ -349,7 +384,7 @@ hfs_uncompressed_size_of_compressed_file(struct hfsmount *hfsmp, struct vnode *v if (!hfsmp || !fid) { /* make sure we have the required parameters */ return EINVAL; } - if (0 != hfs_vget(hfsmp, fid, &vp, skiplock)) { /* vnode is null, use hfs_vget() to get it */ + if (0 != hfs_vget(hfsmp, fid, &vp, skiplock, 0)) { /* vnode is null, use hfs_vget() to get it */ vp = NULL; } else { putaway = 1; /* note that hfs_vget() was used to aquire the vnode */ @@ -359,10 +394,27 @@ hfs_uncompressed_size_of_compressed_file(struct hfsmount *hfsmp, struct vnode *v * ensures the cached size is present in case decmpfs hasn't * encountered this node yet. */ - if ( ( NULL != vp ) && hfs_file_is_compressed(VTOC(vp), skiplock) ) { - *size = decmpfs_cnode_get_vnode_cached_size(VTOCMP(vp)); /* file info will be cached now, so get size */ - } else { - ret = EINVAL; + if (vp) { + if (hfs_file_is_compressed(VTOC(vp), skiplock) ) { + *size = decmpfs_cnode_get_vnode_cached_size(VTOCMP(vp)); /* file info will be cached now, so get size */ + } else { + if (VTOCMP(vp) && VTOCMP(vp)->cmp_type >= CMP_MAX) { + if (VTOCMP(vp)->cmp_type != DATALESS_CMPFS_TYPE) { + // if we don't recognize this type, just use the real data fork size + if (VTOC(vp)->c_datafork) { + *size = VTOC(vp)->c_datafork->ff_size; + ret = 0; + } else { + ret = EINVAL; + } + } else { + *size = decmpfs_cnode_get_vnode_cached_size(VTOCMP(vp)); /* file info will be cached now, so get size */ + ret = 0; + } + } else { + ret = EINVAL; + } + } } if (putaway) { /* did we use hfs_vget() to get this vnode? */ @@ -396,7 +448,7 @@ hfs_hides_xattr(vfs_context_t ctx, struct cnode *cp, const char *name, int skipl /* * Open a file/directory. */ -static int +int hfs_vnop_open(struct vnop_open_args *ap) { struct vnode *vp = ap->a_vp; @@ -443,27 +495,17 @@ hfs_vnop_open(struct vnop_open_args *ap) /* * Files marked append-only must be opened for appending. */ - if ((cp->c_flags & APPEND) && !vnode_isdir(vp) && + if ((cp->c_bsdflags & APPEND) && !vnode_isdir(vp) && (ap->a_mode & (FWRITE | O_APPEND)) == FWRITE) return (EPERM); if (vnode_isreg(vp) && !UBCINFOEXISTS(vp)) return (EBUSY); /* file is in use by the kernel */ - /* Don't allow journal file to be opened externally. */ - if (cp->c_fileid == hfsmp->hfs_jnlfileid) + /* Don't allow journal to be opened externally. */ + if (hfs_is_journal_file(hfsmp, cp)) return (EPERM); - /* If we're going to write to the file, initialize quotas. */ -#if QUOTA - if ((ap->a_mode & FWRITE) && (hfsmp->hfs_flags & HFS_QUOTAS)) - (void)hfs_getinoquota(cp); -#endif /* QUOTA */ - - /* - * On the first (non-busy) open of a fragmented - * file attempt to de-frag it (if its less than 20MB). - */ if ((hfsmp->hfs_flags & HFS_READ_ONLY) || (hfsmp->jnl == NULL) || #if NAMEDSTREAMS @@ -476,6 +518,17 @@ hfs_vnop_open(struct vnop_open_args *ap) if ((error = hfs_lock(cp, HFS_EXCLUSIVE_LOCK))) return (error); + +#if QUOTA + /* If we're going to write to the file, initialize quotas. */ + if ((ap->a_mode & FWRITE) && (hfsmp->hfs_flags & HFS_QUOTAS)) + (void)hfs_getinoquota(cp); +#endif /* QUOTA */ + + /* + * On the first (non-busy) open of a fragmented + * file attempt to de-frag it (if its less than 20MB). + */ fp = VTOF(vp); if (fp->ff_blocks && fp->ff_extents[7].blockCount != 0 && @@ -507,6 +560,7 @@ hfs_vnop_open(struct vnop_open_args *ap) vfs_context_proc(ap->a_context)); } } + hfs_unlock(cp); return (0); @@ -516,7 +570,7 @@ hfs_vnop_open(struct vnop_open_args *ap) /* * Close a file/directory. */ -static int +int hfs_vnop_close(ap) struct vnop_close_args /* { struct vnode *a_vp; @@ -559,11 +613,11 @@ hfs_vnop_close(ap) // release cnode lock; must acquire truncate lock BEFORE cnode lock hfs_unlock(cp); - hfs_lock_truncate(cp, TRUE); + hfs_lock_truncate(cp, HFS_EXCLUSIVE_LOCK); tooktrunclock = 1; if (hfs_lock(VTOC(vp), HFS_EXCLUSIVE_LOCK) != 0) { - hfs_unlock_truncate(cp, TRUE); + hfs_unlock_truncate(cp, 0); // bail out if we can't re-acquire cnode lock return 0; } @@ -585,8 +639,8 @@ hfs_vnop_close(ap) // if we froze the fs and we're exiting, then "thaw" the fs if (hfsmp->hfs_freezing_proc == p && proc_exiting(p)) { hfsmp->hfs_freezing_proc = NULL; - hfs_global_exclusive_lock_release(hfsmp); - lck_rw_unlock_exclusive(&hfsmp->hfs_insync); + hfs_unlock_global (hfsmp); + lck_rw_unlock_exclusive(&hfsmp->hfs_insync); } busy = vnode_isinuse(vp, 1); @@ -601,7 +655,7 @@ hfs_vnop_close(ap) } if (tooktrunclock){ - hfs_unlock_truncate(cp, TRUE); + hfs_unlock_truncate(cp, 0); } hfs_unlock(cp); @@ -615,7 +669,7 @@ hfs_vnop_close(ap) /* * Get basic attributes. */ -static int +int hfs_vnop_getattr(struct vnop_getattr_args *ap) { #define VNODE_ATTR_TIMES \ @@ -648,10 +702,16 @@ hfs_vnop_getattr(struct vnop_getattr_args *ap) /* if it's a data fork, we need to know if it was compressed so we can report the uncompressed size */ compressed = hfs_file_is_compressed(cp, 0); } - if (compressed && (VATTR_IS_ACTIVE(vap, va_data_size) || VATTR_IS_ACTIVE(vap, va_total_size))) { - if (0 != hfs_uncompressed_size_of_compressed_file(NULL, vp, 0, &uncompressed_size, 0)) { - /* failed to get the uncompressed size, we'll check for this later */ - uncompressed_size = -1; + if ((VATTR_IS_ACTIVE(vap, va_data_size) || VATTR_IS_ACTIVE(vap, va_total_size))) { + // if it's compressed + if (compressed || (!VNODE_IS_RSRC(vp) && cp->c_decmp && cp->c_decmp->cmp_type >= CMP_MAX)) { + if (0 != hfs_uncompressed_size_of_compressed_file(NULL, vp, 0, &uncompressed_size, 0)) { + /* failed to get the uncompressed size, we'll check for this later */ + uncompressed_size = -1; + } else { + // fake that it's compressed + compressed = 1; + } } } } @@ -670,7 +730,7 @@ hfs_vnop_getattr(struct vnop_getattr_args *ap) vap->va_uid = cp->c_uid; vap->va_gid = cp->c_gid; vap->va_mode = cp->c_mode; - vap->va_flags = cp->c_flags; + vap->va_flags = cp->c_bsdflags; vap->va_supported |= VNODE_ATTR_AUTH & ~VNODE_ATTR_va_acl; if ((cp->c_attr.ca_recflags & kHFSHasSecurityMask) == 0) { @@ -812,13 +872,17 @@ hfs_vnop_getattr(struct vnop_getattr_args *ap) if (cp->c_blocks - VTOF(vp)->ff_blocks) { /* We deal with rsrc fork vnode iocount at the end of the function */ - error = hfs_vgetrsrc(hfsmp, vp, &rvp, TRUE, TRUE); + error = hfs_vgetrsrc(hfsmp, vp, &rvp, TRUE, FALSE); if (error) { - /* - * hfs_vgetrsrc may have returned a vnode in rvp even though - * we got an error, because we specified error_on_unlinked. - * We need to drop the iocount after we release the cnode lock, so - * it will be taken care of at the end of the function if it's needed. + /* + * Note that we call hfs_vgetrsrc with error_on_unlinked + * set to FALSE. This is because we may be invoked via + * fstat() on an open-unlinked file descriptor and we must + * continue to support access to the rsrc fork until it disappears. + * The code at the end of this function will be + * responsible for releasing the iocount generated by + * hfs_vgetrsrc. This is because we can't drop the iocount + * without unlocking the cnode first. */ goto out; } @@ -876,12 +940,22 @@ hfs_vnop_getattr(struct vnop_getattr_args *ap) vap->va_backup_time.tv_sec = cp->c_btime; vap->va_backup_time.tv_nsec = 0; + /* See if we need to emit the date added field to the user */ + if (VATTR_IS_ACTIVE(vap, va_addedtime)) { + u_int32_t dateadded = hfs_get_dateadded (cp); + if (dateadded) { + vap->va_addedtime.tv_sec = dateadded; + vap->va_addedtime.tv_nsec = 0; + VATTR_SET_SUPPORTED (vap, va_addedtime); + } + } + /* XXX is this really a good 'optimal I/O size'? */ vap->va_iosize = hfsmp->hfs_logBlockSize; vap->va_uid = cp->c_uid; vap->va_gid = cp->c_gid; vap->va_mode = cp->c_mode; - vap->va_flags = cp->c_flags; + vap->va_flags = cp->c_bsdflags; /* * Exporting file IDs from HFS Plus: @@ -972,7 +1046,7 @@ hfs_vnop_getattr(struct vnop_getattr_args *ap) * have an open-unlinked file. Go to the next link in this case. */ if ((cp->c_desc.cd_namelen == 0) && (vap->va_linkid == cp->c_fileid)) { - if ((error = hfs_lookuplink(hfsmp, vap->va_linkid, &prevlinkid, &nextlinkid))){ + if ((error = hfs_lookup_siblinglinks(hfsmp, vap->va_linkid, &prevlinkid, &nextlinkid))){ goto out; } } @@ -1029,7 +1103,7 @@ out: return (error); } -static int +int hfs_vnop_setattr(ap) struct vnop_setattr_args /* { struct vnode *a_vp; @@ -1046,7 +1120,10 @@ hfs_vnop_setattr(ap) int error = 0; uid_t nuid; gid_t ngid; + time_t orig_ctime; + orig_ctime = VTOC(vp)->c_ctime; + #if HFS_COMPRESSION int decmpfs_reset_state = 0; /* @@ -1056,12 +1133,27 @@ hfs_vnop_setattr(ap) error = decmpfs_update_attributes(vp, vap); if (error) return error; + + // + // if this is not a size-changing setattr and it is not just + // an atime update, then check for a snapshot. + // + if (!VATTR_IS_ACTIVE(vap, va_data_size) && !(vap->va_active == VNODE_ATTR_va_access_time)) { + check_for_tracked_file(vp, orig_ctime, NAMESPACE_HANDLER_METADATA_MOD, NULL); + } #endif + +#if CONFIG_PROTECT + if ((error = cp_handle_vnop(vp, CP_WRITE_ACCESS, 0)) != 0) { + return (error); + } +#endif /* CONFIG_PROTECT */ + hfsmp = VTOHFS(vp); - /* Don't allow modification of the journal file. */ - if (hfsmp->hfs_jnlfileid == VTOC(vp)->c_fileid) { + /* Don't allow modification of the journal. */ + if (hfs_is_journal_file(hfsmp, VTOC(vp))) { return (EPERM); } @@ -1090,6 +1182,8 @@ hfs_vnop_setattr(ap) } } + check_for_tracked_file(vp, orig_ctime, vap->va_data_size == 0 ? NAMESPACE_HANDLER_TRUNCATE_OP|NAMESPACE_HANDLER_DELETE_OP : NAMESPACE_HANDLER_TRUNCATE_OP, NULL); + decmpfs_lock_compressed_data(dp, 1); if (hfs_file_is_compressed(VTOC(vp), 1)) { error = decmpfs_decompress_file(vp, dp, -1/*vap->va_data_size*/, 0, 1); @@ -1101,13 +1195,13 @@ hfs_vnop_setattr(ap) #endif /* Take truncate lock before taking cnode lock. */ - hfs_lock_truncate(VTOC(vp), TRUE); + hfs_lock_truncate(VTOC(vp), HFS_EXCLUSIVE_LOCK); /* Perform the ubc_setsize before taking the cnode lock. */ ubc_setsize(vp, vap->va_data_size); if ((error = hfs_lock(VTOC(vp), HFS_EXCLUSIVE_LOCK))) { - hfs_unlock_truncate(VTOC(vp), TRUE); + hfs_unlock_truncate(VTOC(vp), 0); #if HFS_COMPRESSION decmpfs_unlock_compressed_data(dp, 1); #endif @@ -1117,7 +1211,7 @@ hfs_vnop_setattr(ap) error = hfs_truncate(vp, vap->va_data_size, vap->va_vaflags & 0xffff, 1, 0, ap->a_context); - hfs_unlock_truncate(cp, TRUE); + hfs_unlock_truncate(cp, 0); #if HFS_COMPRESSION decmpfs_unlock_compressed_data(dp, 1); #endif @@ -1177,7 +1271,7 @@ hfs_vnop_setattr(ap) u_int16_t *fdFlags; #if HFS_COMPRESSION - if ((cp->c_flags ^ vap->va_flags) & UF_COMPRESSED) { + if ((cp->c_bsdflags ^ vap->va_flags) & UF_COMPRESSED) { /* * the UF_COMPRESSED was toggled, so reset our cached compressed state * but we don't want to actually do the update until we've released the cnode lock down below @@ -1188,7 +1282,7 @@ hfs_vnop_setattr(ap) } #endif - cp->c_flags = vap->va_flags; + cp->c_bsdflags = vap->va_flags; cp->c_touch_chgtime = TRUE; /* @@ -1297,7 +1391,6 @@ out: * Change the mode on a file. * cnode must be locked before calling. */ -__private_extern__ int hfs_chmod(struct vnode *vp, int mode, __unused kauth_cred_t cred, __unused struct proc *p) { @@ -1306,14 +1399,9 @@ hfs_chmod(struct vnode *vp, int mode, __unused kauth_cred_t cred, __unused struc if (VTOVCB(vp)->vcbSigWord != kHFSPlusSigWord) return (0); - // XXXdbg - don't allow modification of the journal or journal_info_block - if (VTOHFS(vp)->jnl && cp && cp->c_datafork) { - struct HFSPlusExtentDescriptor *extd; - - extd = &cp->c_datafork->ff_extents[0]; - if (extd->startBlock == VTOVCB(vp)->vcbJinfoBlock || extd->startBlock == VTOHFS(vp)->jnl_start) { - return EPERM; - } + // Don't allow modification of the journal or journal_info_block + if (hfs_is_journal_file(VTOHFS(vp), cp)) { + return EPERM; } #if OVERRIDE_UNKNOWN_PERMISSIONS @@ -1328,7 +1416,6 @@ hfs_chmod(struct vnode *vp, int mode, __unused kauth_cred_t cred, __unused struc } -__private_extern__ int hfs_write_access(struct vnode *vp, kauth_cred_t cred, struct proc *p, Boolean considerFlags) { @@ -1353,7 +1440,7 @@ hfs_write_access(struct vnode *vp, kauth_cred_t cred, struct proc *p, Boolean co } /* If immutable bit set, nobody gets to write it. */ - if (considerFlags && (cp->c_flags & IMMUTABLE)) + if (considerFlags && (cp->c_bsdflags & IMMUTABLE)) return (EPERM); /* Otherwise, user id 0 always gets access. */ @@ -1378,7 +1465,6 @@ hfs_write_access(struct vnode *vp, kauth_cred_t cred, struct proc *p, Boolean co * Perform chown operation on cnode cp; * code must be locked prior to call. */ -__private_extern__ int #if !QUOTA hfs_chown(struct vnode *vp, uid_t uid, gid_t gid, __unused kauth_cred_t cred, @@ -1512,7 +1598,7 @@ good: * case the file is being tracked through its file ID. Typically * its used after creating a new file during a safe-save. */ -static int +int hfs_vnop_exchange(ap) struct vnop_exchange_args /* { struct vnode *a_fvp; @@ -1532,12 +1618,14 @@ hfs_vnop_exchange(ap) const unsigned char *to_nameptr; char from_iname[32]; char to_iname[32]; - u_int32_t tempflag; + uint32_t to_flag_special; + uint32_t from_flag_special; cnid_t from_parid; cnid_t to_parid; int lockflags; int error = 0, started_tr = 0, got_cookie = 0; cat_cookie_t cookie; + time_t orig_from_ctime, orig_to_ctime; /* The files must be on the same volume. */ if (vnode_mount(from_vp) != vnode_mount(to_vp)) @@ -1546,6 +1634,21 @@ hfs_vnop_exchange(ap) if (from_vp == to_vp) return (EINVAL); + orig_from_ctime = VTOC(from_vp)->c_ctime; + orig_to_ctime = VTOC(to_vp)->c_ctime; + + +#if CONFIG_PROTECT + /* + * Do not allow exchangedata/F_MOVEDATAEXTENTS on data-protected filesystems + * because the EAs will not be swapped. As a result, the persistent keys would not + * match and the files will be garbage. + */ + if (cp_fs_protected (vnode_mount(from_vp))) { + return EINVAL; + } +#endif + #if HFS_COMPRESSION if ( hfs_file_is_compressed(VTOC(from_vp), 0) ) { if ( 0 != ( error = decmpfs_decompress_file(from_vp, VTOCMP(from_vp), -1, 0, 1) ) ) { @@ -1560,6 +1663,48 @@ hfs_vnop_exchange(ap) } #endif // HFS_COMPRESSION + /* + * Normally, we want to notify the user handlers about the event, + * except if it's a handler driving the event. + */ + if ((ap->a_options & FSOPT_EXCHANGE_DATA_ONLY) == 0) { + check_for_tracked_file(from_vp, orig_from_ctime, NAMESPACE_HANDLER_WRITE_OP, NULL); + check_for_tracked_file(to_vp, orig_to_ctime, NAMESPACE_HANDLER_WRITE_OP, NULL); + } else { + /* + * We're doing a data-swap. + * Take the truncate lock/cnode lock, then verify there are no mmap references. + * Issue a hfs_filedone to flush out all of the remaining state for this file. + * Allow the rest of the codeflow to re-acquire the cnode locks in order. + */ + + hfs_lock_truncate (VTOC(from_vp), HFS_SHARED_LOCK); + + if ((error = hfs_lock(VTOC(from_vp), HFS_EXCLUSIVE_LOCK))) { + hfs_unlock_truncate (VTOC(from_vp), 0); + return error; + } + + /* Verify the source file is not in use by anyone besides us (including mmap refs) */ + if (vnode_isinuse(from_vp, 1)) { + error = EBUSY; + hfs_unlock(VTOC(from_vp)); + hfs_unlock_truncate (VTOC(from_vp), 0); + return error; + } + + /* Flush out the data in the source file */ + VTOC(from_vp)->c_flag |= C_SWAPINPROGRESS; + error = hfs_filedone (from_vp, ap->a_context); + VTOC(from_vp)->c_flag &= ~C_SWAPINPROGRESS; + hfs_unlock(VTOC(from_vp)); + hfs_unlock_truncate(VTOC(from_vp), 0); + + if (error) { + return error; + } + } + if ((error = hfs_lockpair(VTOC(from_vp), VTOC(to_vp), HFS_EXCLUSIVE_LOCK))) return (error); @@ -1567,33 +1712,28 @@ hfs_vnop_exchange(ap) to_cp = VTOC(to_vp); hfsmp = VTOHFS(from_vp); - /* Only normal files can be exchanged. */ - if (!vnode_isreg(from_vp) || !vnode_isreg(to_vp) || - VNODE_IS_RSRC(from_vp) || VNODE_IS_RSRC(to_vp)) { + /* Resource forks cannot be exchanged. */ + if (VNODE_IS_RSRC(from_vp) || VNODE_IS_RSRC(to_vp)) { error = EINVAL; goto exit; } - // XXXdbg - don't allow modification of the journal or journal_info_block - if (hfsmp->jnl) { - struct HFSPlusExtentDescriptor *extd; - - if (from_cp->c_datafork) { - extd = &from_cp->c_datafork->ff_extents[0]; - if (extd->startBlock == VTOVCB(from_vp)->vcbJinfoBlock || extd->startBlock == hfsmp->jnl_start) { - error = EPERM; - goto exit; - } - } - - if (to_cp->c_datafork) { - extd = &to_cp->c_datafork->ff_extents[0]; - if (extd->startBlock == VTOVCB(to_vp)->vcbJinfoBlock || extd->startBlock == hfsmp->jnl_start) { - error = EPERM; - goto exit; - } - } + // Don't allow modification of the journal or journal_info_block + if (hfs_is_journal_file(hfsmp, from_cp) || + hfs_is_journal_file(hfsmp, to_cp)) { + error = EPERM; + goto exit; } + + /* + * Ok, now that all of the pre-flighting is done, call the underlying + * function if needed. + */ + if (ap->a_options & FSOPT_EXCHANGE_DATA_ONLY) { + error = hfs_movedata(from_vp, to_vp); + goto exit; + } + if ((error = hfs_start_transaction(hfsmp)) != 0) { goto exit; @@ -1660,7 +1800,14 @@ hfs_vnop_exchange(ap) /* Save a copy of from attributes before swapping. */ bcopy(&from_cp->c_desc, &tempdesc, sizeof(struct cat_desc)); bcopy(&from_cp->c_attr, &tempattr, sizeof(struct cat_attr)); - tempflag = from_cp->c_flag & (C_HARDLINK | C_HASXATTRS); + + /* Save whether or not each cnode is a hardlink or has EAs */ + from_flag_special = from_cp->c_flag & (C_HARDLINK | C_HASXATTRS); + to_flag_special = to_cp->c_flag & (C_HARDLINK | C_HASXATTRS); + + /* Drop the special bits from each cnode */ + from_cp->c_flag &= ~(C_HARDLINK | C_HASXATTRS); + to_cp->c_flag &= ~(C_HARDLINK | C_HASXATTRS); /* * Swap the descriptors and all non-fork related attributes. @@ -1669,33 +1816,79 @@ hfs_vnop_exchange(ap) bcopy(&to_cp->c_desc, &from_cp->c_desc, sizeof(struct cat_desc)); from_cp->c_hint = 0; - from_cp->c_fileid = from_cp->c_cnid; + /* + * If 'to' was a hardlink, then we copied over its link ID/CNID/(namespace ID) + * when we bcopy'd the descriptor above. However, we need to be careful + * when setting up the fileID below, because we cannot assume that the + * file ID is the same as the CNID if either one was a hardlink. + * The file ID is stored in the c_attr as the ca_fileid. So it needs + * to be pulled explicitly; we cannot just use the CNID. + */ + from_cp->c_fileid = to_cp->c_attr.ca_fileid; + from_cp->c_itime = to_cp->c_itime; from_cp->c_btime = to_cp->c_btime; from_cp->c_atime = to_cp->c_atime; from_cp->c_ctime = to_cp->c_ctime; from_cp->c_gid = to_cp->c_gid; from_cp->c_uid = to_cp->c_uid; - from_cp->c_flags = to_cp->c_flags; + from_cp->c_bsdflags = to_cp->c_bsdflags; from_cp->c_mode = to_cp->c_mode; from_cp->c_linkcount = to_cp->c_linkcount; - from_cp->c_flag = to_cp->c_flag & (C_HARDLINK | C_HASXATTRS); + from_cp->c_attr.ca_linkref = to_cp->c_attr.ca_linkref; + from_cp->c_attr.ca_firstlink = to_cp->c_attr.ca_firstlink; + + /* + * The cnode flags need to stay with the cnode and not get transferred + * over along with everything else because they describe the content; they are + * not attributes that reflect changes specific to the file ID. In general, + * fields that are tied to the file ID are the ones that will move. + * + * This reflects the fact that the file may have borrowed blocks, dirty metadata, + * or other extents, which may not yet have been written to the catalog. If + * they were, they would have been transferred above in the ExchangeFileIDs call above... + * + * The flags that are special are: + * C_HARDLINK, C_HASXATTRS + * + * These flags move with the item and file ID in the namespace since their + * state is tied to that of the file ID. + * + * So to transfer the flags, we have to take the following steps + * 1) Store in a localvar whether or not the special bits are set. + * 2) Drop the special bits from the current flags + * 3) swap the special flag bits to their destination + */ + from_cp->c_flag |= to_flag_special; + from_cp->c_attr.ca_recflags = to_cp->c_attr.ca_recflags; bcopy(to_cp->c_finderinfo, from_cp->c_finderinfo, 32); bcopy(&tempdesc, &to_cp->c_desc, sizeof(struct cat_desc)); to_cp->c_hint = 0; - to_cp->c_fileid = to_cp->c_cnid; + /* + * Pull the file ID from the tempattr we copied above. We can't assume + * it is the same as the CNID. + */ + to_cp->c_fileid = tempattr.ca_fileid; to_cp->c_itime = tempattr.ca_itime; to_cp->c_btime = tempattr.ca_btime; to_cp->c_atime = tempattr.ca_atime; to_cp->c_ctime = tempattr.ca_ctime; to_cp->c_gid = tempattr.ca_gid; to_cp->c_uid = tempattr.ca_uid; - to_cp->c_flags = tempattr.ca_flags; + to_cp->c_bsdflags = tempattr.ca_flags; to_cp->c_mode = tempattr.ca_mode; to_cp->c_linkcount = tempattr.ca_linkcount; - to_cp->c_flag = tempflag; + to_cp->c_attr.ca_linkref = tempattr.ca_linkref; + to_cp->c_attr.ca_firstlink = tempattr.ca_firstlink; + + /* + * Only OR in the "from" flags into our cnode flags below. + * Leave the rest of the flags alone. + */ + to_cp->c_flag |= from_flag_special; + to_cp->c_attr.ca_recflags = tempattr.ca_recflags; bcopy(tempattr.ca_finderinfo, to_cp->c_finderinfo, 32); @@ -1706,14 +1899,14 @@ hfs_vnop_exchange(ap) * When a file moves out of "Cleanup At Startup" * we can drop its NODUMP status. */ - if ((from_cp->c_flags & UF_NODUMP) && + if ((from_cp->c_bsdflags & UF_NODUMP) && (from_cp->c_parentcnid != to_cp->c_parentcnid)) { - from_cp->c_flags &= ~UF_NODUMP; + from_cp->c_bsdflags &= ~UF_NODUMP; from_cp->c_touch_chgtime = TRUE; } - if ((to_cp->c_flags & UF_NODUMP) && + if ((to_cp->c_bsdflags & UF_NODUMP) && (to_cp->c_parentcnid != from_cp->c_parentcnid)) { - to_cp->c_flags &= ~UF_NODUMP; + to_cp->c_bsdflags &= ~UF_NODUMP; to_cp->c_touch_chgtime = TRUE; } @@ -1729,11 +1922,337 @@ exit: return (error); } +int +hfs_vnop_mmap(struct vnop_mmap_args *ap) +{ + struct vnode *vp = ap->a_vp; + int error; + + if (VNODE_IS_RSRC(vp)) { + /* allow pageins of the resource fork */ + } else { + int compressed = hfs_file_is_compressed(VTOC(vp), 1); /* 1 == don't take the cnode lock */ + time_t orig_ctime = VTOC(vp)->c_ctime; + + if (!compressed && (VTOC(vp)->c_bsdflags & UF_COMPRESSED)) { + error = check_for_dataless_file(vp, NAMESPACE_HANDLER_READ_OP); + if (error != 0) { + return error; + } + } + + if (ap->a_fflags & PROT_WRITE) { + check_for_tracked_file(vp, orig_ctime, NAMESPACE_HANDLER_WRITE_OP, NULL); + } + } + + // + // NOTE: we return ENOTSUP because we want the cluster layer + // to actually do all the real work. + // + return (ENOTSUP); +} + +/* + * hfs_movedata + * + * This is a non-symmetric variant of exchangedata. In this function, + * the contents of the fork in from_vp are moved to the fork + * specified by to_vp. + * + * The cnodes pointed to by 'from_vp' and 'to_vp' must be locked. + * + * The vnode pointed to by 'to_vp' *must* be empty prior to invoking this function. + * We impose this restriction because we may not be able to fully delete the entire + * file's contents in a single transaction, particularly if it has a lot of extents. + * In the normal file deletion codepath, the file is screened for two conditions: + * 1) bigger than 400MB, and 2) more than 8 extents. If so, the file is relocated to + * the hidden directory and the deletion is broken up into multiple truncates. We can't + * do that here because both files need to exist in the namespace. The main reason this + * is imposed is that we may have to touch a whole lot of bitmap blocks if there are + * many extents. + * + * Any data written to 'from_vp' after this call completes is not guaranteed + * to be moved. + * + * Arguments: + * vnode from_vp: source file + * vnode to_vp: destination file; must be empty + * + * Returns: + * EFBIG - Destination file was not empty + * 0 - success + * + * + */ +int hfs_movedata (struct vnode *from_vp, struct vnode *to_vp) { + + struct cnode *from_cp; + struct cnode *to_cp; + struct hfsmount *hfsmp = NULL; + int error = 0; + int started_tr = 0; + int lockflags = 0; + int overflow_blocks; + int rsrc = 0; + + + /* Get the HFS pointers */ + from_cp = VTOC(from_vp); + to_cp = VTOC(to_vp); + hfsmp = VTOHFS(from_vp); + + /* Verify that neither source/dest file is open-unlinked */ + if (from_cp->c_flag & (C_DELETED | C_NOEXISTS)) { + error = EBUSY; + goto movedata_exit; + } + + if (to_cp->c_flag & (C_DELETED | C_NOEXISTS)) { + error = EBUSY; + goto movedata_exit; + } + + /* + * Verify the source file is not in use by anyone besides us. + * + * This function is typically invoked by a namespace handler + * process responding to a temporarily stalled system call. + * The FD that it is working off of is opened O_EVTONLY, so + * it really has no active usecounts (the kusecount from O_EVTONLY + * is subtracted from the total usecounts). + * + * As a result, we shouldn't have any active usecounts against + * this vnode when we go to check it below. + */ + if (vnode_isinuse(from_vp, 0)) { + error = EBUSY; + goto movedata_exit; + } + + if (from_cp->c_rsrc_vp == from_vp) { + rsrc = 1; + } + + /* + * We assume that the destination file is already empty. + * Verify that it is. + */ + if (rsrc) { + if (to_cp->c_rsrcfork->ff_size > 0) { + error = EFBIG; + goto movedata_exit; + } + } + else { + if (to_cp->c_datafork->ff_size > 0) { + error = EFBIG; + goto movedata_exit; + } + } + + /* If the source has the rsrc open, make sure the destination is also the rsrc */ + if (rsrc) { + if (to_vp != to_cp->c_rsrc_vp) { + error = EINVAL; + goto movedata_exit; + } + } + else { + /* Verify that both forks are data forks */ + if (to_vp != to_cp->c_vp) { + error = EINVAL; + goto movedata_exit; + } + } + + /* + * See if the source file has overflow extents. If it doesn't, we don't + * need to call into MoveData, and the catalog will be enough. + */ + if (rsrc) { + overflow_blocks = overflow_extents(from_cp->c_rsrcfork); + } + else { + overflow_blocks = overflow_extents(from_cp->c_datafork); + } + + if ((error = hfs_start_transaction (hfsmp)) != 0) { + goto movedata_exit; + } + started_tr = 1; + + /* Lock the system files: catalog, extents, attributes */ + lockflags = hfs_systemfile_lock(hfsmp, SFL_CATALOG | SFL_EXTENTS | SFL_ATTRIBUTE, HFS_EXCLUSIVE_LOCK); + + /* Copy over any catalog allocation data into the new spot. */ + if (rsrc) { + if ((error = hfs_move_fork (from_cp->c_rsrcfork, from_cp, to_cp->c_rsrcfork, to_cp))){ + hfs_systemfile_unlock(hfsmp, lockflags); + goto movedata_exit; + } + } + else { + if ((error = hfs_move_fork (from_cp->c_datafork, from_cp, to_cp->c_datafork, to_cp))) { + hfs_systemfile_unlock(hfsmp, lockflags); + goto movedata_exit; + } + } + + /* + * Note that because all we're doing is moving the extents around, we can + * probably do this in a single transaction: Each extent record (group of 8) + * is 64 bytes. A extent overflow B-Tree node is typically 4k. This means + * each node can hold roughly ~60 extent records == (480 extents). + * + * If a file was massively fragmented and had 20k extents, this means we'd + * roughly touch 20k/480 == 41 to 42 nodes, plus the index nodes, for half + * of the operation. (inserting or deleting). So if we're manipulating 80-100 + * nodes, this is basically 320k of data to write to the journal in + * a bad case. + */ + if (overflow_blocks != 0) { + if (rsrc) { + error = MoveData(hfsmp, from_cp->c_cnid, to_cp->c_cnid, 1); + } + else { + error = MoveData (hfsmp, from_cp->c_cnid, to_cp->c_cnid, 0); + } + } + + if (error) { + /* Reverse the operation. Copy the fork data back into the source */ + if (rsrc) { + hfs_move_fork (to_cp->c_rsrcfork, to_cp, from_cp->c_rsrcfork, from_cp); + } + else { + hfs_move_fork (to_cp->c_datafork, to_cp, from_cp->c_datafork, from_cp); + } + } + else { + struct cat_fork *src_data = NULL; + struct cat_fork *src_rsrc = NULL; + struct cat_fork *dst_data = NULL; + struct cat_fork *dst_rsrc = NULL; + + /* Touch the times*/ + to_cp->c_touch_acctime = TRUE; + to_cp->c_touch_chgtime = TRUE; + to_cp->c_touch_modtime = TRUE; + + from_cp->c_touch_acctime = TRUE; + from_cp->c_touch_chgtime = TRUE; + from_cp->c_touch_modtime = TRUE; + + hfs_touchtimes(hfsmp, to_cp); + hfs_touchtimes(hfsmp, from_cp); + + if (from_cp->c_datafork) { + src_data = &from_cp->c_datafork->ff_data; + } + if (from_cp->c_rsrcfork) { + src_rsrc = &from_cp->c_rsrcfork->ff_data; + } + + if (to_cp->c_datafork) { + dst_data = &to_cp->c_datafork->ff_data; + } + if (to_cp->c_rsrcfork) { + dst_rsrc = &to_cp->c_rsrcfork->ff_data; + } + + /* Update the catalog nodes */ + (void) cat_update(hfsmp, &from_cp->c_desc, &from_cp->c_attr, + src_data, src_rsrc); + + (void) cat_update(hfsmp, &to_cp->c_desc, &to_cp->c_attr, + dst_data, dst_rsrc); + + } + /* unlock the system files */ + hfs_systemfile_unlock(hfsmp, lockflags); + + +movedata_exit: + if (started_tr) { + hfs_end_transaction(hfsmp); + } + + return error; + +} + +/* + * Copy all of the catalog and runtime data in srcfork to dstfork. + * + * This allows us to maintain the invalid ranges across the movedata operation so + * we don't need to force all of the pending IO right now. In addition, we move all + * non overflow-extent extents into the destination here. + */ +static int hfs_move_fork (struct filefork *srcfork, struct cnode *src_cp, + struct filefork *dstfork, struct cnode *dst_cp) { + struct rl_entry *invalid_range; + int size = sizeof(struct HFSPlusExtentDescriptor); + size = size * kHFSPlusExtentDensity; + + /* If the dstfork has any invalid ranges, bail out */ + invalid_range = TAILQ_FIRST(&dstfork->ff_invalidranges); + if (invalid_range != NULL) { + return EFBIG; + } + + if (dstfork->ff_data.cf_size != 0 || dstfork->ff_data.cf_new_size != 0) { + return EFBIG; + } + + /* First copy the invalid ranges */ + while ((invalid_range = TAILQ_FIRST(&srcfork->ff_invalidranges))) { + off_t start = invalid_range->rl_start; + off_t end = invalid_range->rl_end; + + /* Remove it from the srcfork and add it to dstfork */ + rl_remove(start, end, &srcfork->ff_invalidranges); + rl_add(start, end, &dstfork->ff_invalidranges); + } + + /* + * Ignore the ff_union. We don't move symlinks or system files. + * Now copy the in-catalog extent information + */ + dstfork->ff_data.cf_size = srcfork->ff_data.cf_size; + dstfork->ff_data.cf_new_size = srcfork->ff_data.cf_new_size; + dstfork->ff_data.cf_vblocks = srcfork->ff_data.cf_vblocks; + dstfork->ff_data.cf_blocks = srcfork->ff_data.cf_blocks; + + /* just memcpy the whole array of extents to the new location. */ + memcpy (dstfork->ff_data.cf_extents, srcfork->ff_data.cf_extents, size); + + /* + * Copy the cnode attribute data. + * + */ + src_cp->c_blocks -= srcfork->ff_data.cf_vblocks; + src_cp->c_blocks -= srcfork->ff_data.cf_blocks; + + dst_cp->c_blocks += srcfork->ff_data.cf_vblocks; + dst_cp->c_blocks += srcfork->ff_data.cf_blocks; + + /* Now delete the entries in the source fork */ + srcfork->ff_data.cf_size = 0; + srcfork->ff_data.cf_new_size = 0; + srcfork->ff_data.cf_union.cfu_bytesread = 0; + srcfork->ff_data.cf_vblocks = 0; + srcfork->ff_data.cf_blocks = 0; + + /* Zero out the old extents */ + bzero (srcfork->ff_data.cf_extents, size); + return 0; +} + /* * cnode must be locked */ -__private_extern__ int hfs_fsync(struct vnode *vp, int waitfor, int fullsync, struct proc *p) { @@ -1747,7 +2266,7 @@ hfs_fsync(struct vnode *vp, int waitfor, int fullsync, struct proc *p) int wait; /* all other attributes (e.g. atime, etc.) */ int lockflag; int took_trunc_lock = 0; - boolean_t trunc_lock_exclusive = FALSE; + int locked_buffers = 0; /* * Applications which only care about data integrity rather than full @@ -1777,14 +2296,13 @@ hfs_fsync(struct vnode *vp, int waitfor, int fullsync, struct proc *p) } } else if (UBCINFOEXISTS(vp)) { hfs_unlock(cp); - hfs_lock_truncate(cp, trunc_lock_exclusive); + hfs_lock_truncate(cp, HFS_SHARED_LOCK); took_trunc_lock = 1; if (fp->ff_unallocblocks != 0) { - hfs_unlock_truncate(cp, trunc_lock_exclusive); + hfs_unlock_truncate(cp, 0); - trunc_lock_exclusive = TRUE; - hfs_lock_truncate(cp, trunc_lock_exclusive); + hfs_lock_truncate(cp, HFS_EXCLUSIVE_LOCK); } /* Don't hold cnode lock when calling into cluster layer. */ (void) cluster_push(vp, waitdata ? IO_SYNC : 0); @@ -1800,7 +2318,7 @@ hfs_fsync(struct vnode *vp, int waitfor, int fullsync, struct proc *p) */ if (fp && (((cp->c_flag & C_ALWAYS_ZEROFILL) && !TAILQ_EMPTY(&fp->ff_invalidranges)) || ((wait || (cp->c_flag & C_ZFWANTSYNC)) && - ((cp->c_flags & UF_NODUMP) == 0) && + ((cp->c_bsdflags & UF_NODUMP) == 0) && UBCINFOEXISTS(vp) && (vnode_issystem(vp) ==0) && cp->c_zftimeout != 0))) { @@ -1811,13 +2329,12 @@ hfs_fsync(struct vnode *vp, int waitfor, int fullsync, struct proc *p) goto datasync; } if (!TAILQ_EMPTY(&fp->ff_invalidranges)) { - if (!took_trunc_lock || trunc_lock_exclusive == FALSE) { + if (!took_trunc_lock || (cp->c_truncatelockowner == HFS_SHARED_OWNER)) { hfs_unlock(cp); - if (took_trunc_lock) - hfs_unlock_truncate(cp, trunc_lock_exclusive); - - trunc_lock_exclusive = TRUE; - hfs_lock_truncate(cp, trunc_lock_exclusive); + if (took_trunc_lock) { + hfs_unlock_truncate(cp, 0); + } + hfs_lock_truncate(cp, HFS_EXCLUSIVE_LOCK); hfs_lock(cp, HFS_FORCE_LOCK); took_trunc_lock = 1; } @@ -1848,7 +2365,7 @@ hfs_fsync(struct vnode *vp, int waitfor, int fullsync, struct proc *p) } datasync: if (took_trunc_lock) { - hfs_unlock_truncate(cp, trunc_lock_exclusive); + hfs_unlock_truncate(cp, 0); took_trunc_lock = 0; } /* @@ -1868,8 +2385,32 @@ datasync: /* * Flush all dirty buffers associated with a vnode. + * Record how many of them were dirty AND locked (if necessary). */ - buf_flushdirtyblks(vp, waitdata, lockflag, "hfs_fsync"); + locked_buffers = buf_flushdirtyblks_skipinfo(vp, waitdata, lockflag, "hfs_fsync"); + if ((lockflag & BUF_SKIP_LOCKED) && (locked_buffers) && (vnode_vtype(vp) == VLNK)) { + /* + * If there are dirty symlink buffers, then we may need to take action + * to prevent issues later on if we are journaled. If we're fsyncing a + * symlink vnode then we are in one of three cases: + * + * 1) automatic sync has fired. In this case, we don't want the behavior to change. + * + * 2) Someone has opened the FD for the symlink (not what it points to) + * and has issued an fsync against it. This should be rare, and we don't + * want the behavior to change. + * + * 3) We are being called by a vclean which is trying to reclaim this + * symlink vnode. If this is the case, then allowing this fsync to + * proceed WITHOUT flushing the journal could result in the vclean + * invalidating the buffer's blocks before the journal transaction is + * written to disk. To prevent this, we force a journal flush + * if the vnode is in the middle of a recycle (VL_TERMINATE or VL_DEAD is set). + */ + if (vnode_isrecycled(vp)) { + fullsync = 1; + } + } metasync: if (vnode_isreg(vp) && vnode_issystem(vp)) { @@ -1899,13 +2440,23 @@ metasync: * changes get to stable storage. */ if (fullsync) { - if (hfsmp->jnl) { - hfs_journal_flush(hfsmp); - } else { - retval = hfs_metasync_all(hfsmp); - /* XXX need to pass context! */ - VNOP_IOCTL(hfsmp->hfs_devvp, DKIOCSYNCHRONIZECACHE, NULL, FWRITE, NULL); - } + if (hfsmp->jnl) { + hfs_journal_flush(hfsmp, FALSE); + + if (journal_uses_fua(hfsmp->jnl)) { + /* + * the journal_flush did NOT issue a sync track cache command, + * and the fullsync indicates we are supposed to flush all cached + * data to the media, so issue the sync track cache command + * explicitly + */ + VNOP_IOCTL(hfsmp->hfs_devvp, DKIOCSYNCHRONIZECACHE, NULL, FWRITE, NULL); + } + } else { + retval = hfs_metasync_all(hfsmp); + /* XXX need to pass context! */ + VNOP_IOCTL(hfsmp->hfs_devvp, DKIOCSYNCHRONIZECACHE, NULL, FWRITE, NULL); + } } } @@ -1914,7 +2465,7 @@ metasync: /* Sync an hfs catalog b-tree node */ -static int +int hfs_metasync(struct hfsmount *hfsmp, daddr64_t node, __unused struct proc *p) { vnode_t vp; @@ -1960,7 +2511,7 @@ hfs_metasync(struct hfsmount *hfsmp, daddr64_t node, __unused struct proc *p) * we rely on fsck_hfs to fix that up (which it can do without any loss * of data). */ -static int +int hfs_metasync_all(struct hfsmount *hfsmp) { int lockflags; @@ -2002,7 +2553,6 @@ hfs_btsync_callback(struct buf *bp, __unused void *dummy) } -__private_extern__ int hfs_btsync(struct vnode *vp, int sync_transaction) { @@ -2030,7 +2580,7 @@ hfs_btsync(struct vnode *vp, int sync_transaction) /* * Remove a directory. */ -static int +int hfs_vnop_rmdir(ap) struct vnop_rmdir_args /* { struct vnode *a_dvp; @@ -2044,6 +2594,9 @@ hfs_vnop_rmdir(ap) struct cnode *dcp = VTOC(dvp); struct cnode *cp = VTOC(vp); int error; + time_t orig_ctime; + + orig_ctime = VTOC(vp)->c_ctime; if (!S_ISDIR(cp->c_mode)) { return (ENOTDIR); @@ -2051,6 +2604,10 @@ hfs_vnop_rmdir(ap) if (dvp == vp) { return (EINVAL); } + + check_for_tracked_file(vp, orig_ctime, NAMESPACE_HANDLER_DELETE_OP, NULL); + cp = VTOC(vp); + if ((error = hfs_lockpair(dcp, cp, HFS_EXCLUSIVE_LOCK))) { return (error); } @@ -2060,7 +2617,7 @@ hfs_vnop_rmdir(ap) hfs_unlockpair (dcp, cp); return ENOENT; } - error = hfs_removedir(dvp, vp, ap->a_cnp, 0); + error = hfs_removedir(dvp, vp, ap->a_cnp, 0, 0); hfs_unlockpair(dcp, cp); @@ -2072,9 +2629,9 @@ hfs_vnop_rmdir(ap) * * Both dvp and vp cnodes are locked */ -static int +int hfs_removedir(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, - int skip_reserve) + int skip_reserve, int only_unlink) { struct cnode *cp; struct cnode *dcp; @@ -2096,24 +2653,77 @@ hfs_removedir(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, if (cp->c_entries != 0) { return (ENOTEMPTY); } + + /* + * If the directory is open or in use (e.g. opendir() or current working + * directory for some process); wait for inactive/reclaim to actually + * remove cnode from the catalog. Both inactive and reclaim codepaths are capable + * of removing open-unlinked directories from the catalog, as well as getting rid + * of EAs still on the element. So change only_unlink to true, so that it will get + * cleaned up below. + * + * Otherwise, we can get into a weird old mess where the directory has C_DELETED, + * but it really means C_NOEXISTS because the item was actually removed from the + * catalog. Then when we try to remove the entry from the catalog later on, it won't + * really be there anymore. + */ + if (vnode_isinuse(vp, 0)) { + only_unlink = 1; + } - /* Check if we're removing the last link to an empty directory. */ + /* Deal with directory hardlinks */ if (cp->c_flag & C_HARDLINK) { - /* We could also return EBUSY here */ + /* + * Note that if we have a directory which was a hardlink at any point, + * its actual directory data is stored in the directory inode in the hidden + * directory rather than the leaf element(s) present in the namespace. + * + * If there are still other hardlinks to this directory, + * then we'll just eliminate this particular link and the vnode will still exist. + * If this is the last link to an empty directory, then we'll open-unlink the + * directory and it will be only tagged with C_DELETED (as opposed to C_NOEXISTS). + * + * We could also return EBUSY here. + */ + return hfs_unlink(hfsmp, dvp, vp, cnp, skip_reserve); } /* - * We want to make sure that if the directory has a lot of attributes, we process them - * in separate transactions to ensure we don't panic in the journal with a gigantic - * transaction. This means we'll let hfs_removefile deal with the directory, which generally - * follows the same codepath as open-unlinked files. Note that the last argument to - * hfs_removefile specifies that it is supposed to handle directories for this case. - */ - if ((hfsmp->hfs_attribute_vp != NULL) && - (cp->c_attr.ca_recflags & kHFSHasAttributesMask) != 0) { - - return hfs_removefile(dvp, vp, cnp, 0, 0, 1, NULL); + * In a few cases, we may want to allow the directory to persist in an + * open-unlinked state. If the directory is being open-unlinked (still has usecount + * references), or if it has EAs, or if it was being deleted as part of a rename, + * then we go ahead and move it to the hidden directory. + * + * If the directory is being open-unlinked, then we want to keep the catalog entry + * alive so that future EA calls and fchmod/fstat etc. do not cause issues later. + * + * If the directory had EAs, then we want to use the open-unlink trick so that the + * EA removal is not done in one giant transaction. Otherwise, it could cause a panic + * due to overflowing the journal. + * + * Finally, if it was deleted as part of a rename, we move it to the hidden directory + * in order to maintain rename atomicity. + * + * Note that the allow_dirs argument to hfs_removefile specifies that it is + * supposed to handle directories for this case. + */ + + if (((hfsmp->hfs_attribute_vp != NULL) && + ((cp->c_attr.ca_recflags & kHFSHasAttributesMask) != 0)) || + (only_unlink != 0)) { + + int ret = hfs_removefile(dvp, vp, cnp, 0, 0, 1, NULL, only_unlink); + /* + * Even though hfs_vnop_rename calls vnode_recycle for us on tvp we call + * it here just in case we were invoked by rmdir() on a directory that had + * EAs. To ensure that we start reclaiming the space as soon as possible, + * we call vnode_recycle on the directory. + */ + vnode_recycle(vp); + + return ret; + } dcp->c_flag |= C_DIR_MODIFICATION; @@ -2134,7 +2744,7 @@ hfs_removedir(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, * the current directory and thus be * non-empty.) */ - if ((dcp->c_flags & APPEND) || (cp->c_flags & (IMMUTABLE | APPEND))) { + if ((dcp->c_bsdflags & APPEND) || (cp->c_bsdflags & (IMMUTABLE | APPEND))) { error = EPERM; goto out; } @@ -2155,7 +2765,7 @@ hfs_removedir(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, desc.cd_encoding = cp->c_encoding; desc.cd_hint = 0; - if (!hfs_valid_cnode(hfsmp, dvp, cnp, cp->c_fileid)) { + if (!hfs_valid_cnode(hfsmp, dvp, cnp, cp->c_fileid, NULL, &error)) { error = 0; goto out; } @@ -2199,16 +2809,8 @@ hfs_removedir(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, hfs_volupdate(hfsmp, VOL_RMDIR, (dcp->c_cnid == kHFSRootFolderID)); - /* - * directory open or in use (e.g. opendir() or current working - * directory for some process); wait for inactive to actually - * remove catalog entry - */ - if (vnode_isinuse(vp, 0)) { - cp->c_flag |= C_DELETED; - } else { - cp->c_flag |= C_NOEXISTS; - } + /* Mark C_NOEXISTS since the catalog entry is now gone */ + cp->c_flag |= C_NOEXISTS; out: dcp->c_flag &= ~C_DIR_MODIFICATION; wakeup((caddr_t)&dcp->c_flag); @@ -2224,7 +2826,7 @@ out: /* * Remove a file or link. */ -static int +int hfs_vnop_remove(ap) struct vnop_remove_args /* { struct vnode *a_dvp; @@ -2237,60 +2839,78 @@ hfs_vnop_remove(ap) struct vnode *dvp = ap->a_dvp; struct vnode *vp = ap->a_vp; struct cnode *dcp = VTOC(dvp); - struct cnode *cp = VTOC(vp); + struct cnode *cp; struct vnode *rvp = NULL; - struct hfsmount *hfsmp = VTOHFS(vp); int error=0, recycle_rsrc=0; - int drop_rsrc_vnode = 0; - int vref; + time_t orig_ctime; + uint32_t rsrc_vid = 0; if (dvp == vp) { return (EINVAL); } - /* - * We need to grab the cnode lock on 'cp' before the lockpair() - * to get an iocount on the rsrc fork BEFORE we enter hfs_removefile. - * To prevent other deadlocks, it's best to call hfs_vgetrsrc in a way that - * allows it to drop the cnode lock that it expects to be held coming in. - * If we don't, we could commit a lock order violation, causing a deadlock. - * In order to safely get the rsrc vnode with an iocount, we need to only hold the - * lock on the file temporarily. Unlike hfs_vnop_rename, we don't have to worry - * about one rsrc fork getting recycled for another, but we do want to ensure - * that there are no deadlocks due to lock ordering issues. - * + orig_ctime = VTOC(vp)->c_ctime; + if ( (!vnode_isnamedstream(vp)) && ((ap->a_flags & VNODE_REMOVE_SKIP_NAMESPACE_EVENT) == 0)) { + error = check_for_tracked_file(vp, orig_ctime, NAMESPACE_HANDLER_DELETE_OP, NULL); + if (error) { + // XXXdbg - decide on a policy for handling namespace handler failures! + // for now we just let them proceed. + } + } + error = 0; + + cp = VTOC(vp); + +relock: + + hfs_lock_truncate(cp, HFS_EXCLUSIVE_LOCK); + + if ((error = hfs_lockpair(dcp, cp, HFS_EXCLUSIVE_LOCK))) { + hfs_unlock_truncate(cp, 0); + if (rvp) { + vnode_put (rvp); + } + return (error); + } + + /* + * Lazily respond to determining if there is a valid resource fork + * vnode attached to 'cp' if it is a regular file or symlink. + * If the vnode does not exist, then we may proceed without having to + * create it. + * + * If, however, it does exist, then we need to acquire an iocount on the + * vnode after acquiring its vid. This ensures that if we have to do I/O + * against it, it can't get recycled from underneath us in the middle + * of this call. + * * Note: this function may be invoked for directory hardlinks, so just skip these * steps if 'vp' is a directory. */ if ((vp->v_type == VLNK) || (vp->v_type == VREG)) { + if ((cp->c_rsrc_vp) && (rvp == NULL)) { + /* We need to acquire the rsrc vnode */ + rvp = cp->c_rsrc_vp; + rsrc_vid = vnode_vid (rvp); + + /* Unlock everything to acquire iocount on the rsrc vnode */ + hfs_unlock_truncate (cp, 0); + hfs_unlockpair (dcp, cp); - if ((error = hfs_lock (cp, HFS_EXCLUSIVE_LOCK))) { - return (error); - } - error = hfs_vgetrsrc(hfsmp, vp, &rvp, TRUE, TRUE); - hfs_unlock(cp); - if (error) { - /* We may have gotten a rsrc vp out even though we got an error back. */ - if (rvp) { - vnode_put(rvp); + /* Use the vid to maintain identity on rvp */ + if (vnode_getwithvid(rvp, rsrc_vid)) { + /* + * If this fails, then it was recycled or + * reclaimed in the interim. Reset fields and + * start over. + */ rvp = NULL; + rsrc_vid = 0; } - return error; + goto relock; } - drop_rsrc_vnode = 1; - } - /* Now that we may have an iocount on rvp, do the lock pair */ - hfs_lock_truncate(cp, TRUE); - - if ((error = hfs_lockpair(dcp, cp, HFS_EXCLUSIVE_LOCK))) { - hfs_unlock_truncate(cp, TRUE); - /* drop the iocount on rvp if necessary */ - if (drop_rsrc_vnode) { - vnode_put (rvp); - } - return (error); } /* @@ -2302,20 +2922,27 @@ hfs_vnop_remove(ap) goto rm_done; } - error = hfs_removefile(dvp, vp, ap->a_cnp, ap->a_flags, 0, 0, rvp); - - // - // If the remove succeeded and it's an open-unlinked file that has - // a resource fork vnode that's not in use, we will want to recycle - // the rvp *after* we're done unlocking everything. Otherwise the - // resource vnode will keep a v_parent reference on this vnode which - // prevents it from going through inactive/reclaim which means that - // the disk space associated with this file won't get free'd until - // something forces the resource vnode to get recycled (and that can - // take a very long time). - // - if (error == 0 && (cp->c_flag & C_DELETED) && - (rvp) && !vnode_isinuse(rvp, 0)) { + error = hfs_removefile(dvp, vp, ap->a_cnp, ap->a_flags, 0, 0, NULL, 0); + + /* + * If the remove succeeded in deleting the file, then we may need to mark + * the resource fork for recycle so that it is reclaimed as quickly + * as possible. If it were not recycled quickly, then this resource fork + * vnode could keep a v_parent reference on the data fork, which prevents it + * from going through reclaim (by giving it extra usecounts), except in the force- + * unmount case. + * + * However, a caveat: we need to continue to supply resource fork + * access to open-unlinked files even if the resource fork is not open. This is + * a requirement for the compressed files work. Luckily, hfs_vgetrsrc will handle + * this already if the data fork has been re-parented to the hidden directory. + * + * As a result, all we really need to do here is mark the resource fork vnode + * for recycle. If it goes out of core, it can be brought in again if needed. + * If the cnode was instead marked C_NOEXISTS, then there wouldn't be any + * more work. + */ + if ((error == 0) && (rvp)) { recycle_rsrc = 1; } @@ -2326,19 +2953,15 @@ hfs_vnop_remove(ap) * truncate lock) */ rm_done: - hfs_unlock_truncate(cp, TRUE); + hfs_unlock_truncate(cp, 0); hfs_unlockpair(dcp, cp); if (recycle_rsrc) { - vref = vnode_ref(rvp); - if (vref == 0) { - /* vnode_ref could return an error, only release if we got a ref */ - vnode_rele(rvp); - } + /* inactive or reclaim on rvp will clean up the blocks from the rsrc fork */ vnode_recycle(rvp); } - if (drop_rsrc_vnode) { + if (rvp) { /* drop iocount on rsrc fork, was obtained at beginning of fxn */ vnode_put(rvp); } @@ -2347,7 +2970,7 @@ rm_done: } -static int +int hfs_removefile_callback(struct buf *bp, void *hfsmp) { if ( !(buf_flags(bp) & B_META)) @@ -2367,32 +2990,37 @@ hfs_removefile_callback(struct buf *bp, void *hfsmp) { * This function may be used to remove directories if they have * lots of EA's -- note the 'allow_dirs' argument. * - * The 'rvp' argument is used to pass in a resource fork vnode with - * an iocount to prevent it from getting recycled during usage. If it - * is NULL, then it is assumed the caller is a VNOP that cannot operate - * on resource forks, like hfs_vnop_symlink or hfs_removedir. Otherwise in - * a VNOP that takes multiple vnodes, we could violate lock order and - * cause a deadlock. + * This function is able to delete blocks & fork data for the resource + * fork even if it does not exist in core (and have a backing vnode). + * It should infer the correct behavior based on the number of blocks + * in the cnode and whether or not the resource fork pointer exists or + * not. As a result, one only need pass in the 'vp' corresponding to the + * data fork of this file (or main vnode in the case of a directory). + * Passing in a resource fork will result in an error. + * + * Because we do not create any vnodes in this function, we are not at + * risk of deadlocking against ourselves by double-locking. * * Requires cnode and truncate locks to be held. */ -static int +int hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, - int flags, int skip_reserve, int allow_dirs, struct vnode *rvp) + int flags, int skip_reserve, int allow_dirs, + __unused struct vnode *rvp, int only_unlink) { struct cnode *cp; struct cnode *dcp; + struct vnode *rsrc_vp = NULL; struct hfsmount *hfsmp; struct cat_desc desc; struct timeval tv; - vfs_context_t ctx = cnp->cn_context; int dataforkbusy = 0; int rsrcforkbusy = 0; - int truncated = 0; int lockflags; int error = 0; int started_tr = 0; int isbigfile = 0, defer_remove=0, isdir=0; + int update_vh = 0; cp = VTOC(vp); dcp = VTOC(dvp); @@ -2403,7 +3031,7 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, return (0); } - if (!hfs_valid_cnode(hfsmp, dvp, cnp, cp->c_fileid)) { + if (!hfs_valid_cnode(hfsmp, dvp, cnp, cp->c_fileid, NULL, &error)) { return 0; } @@ -2411,11 +3039,37 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, if (VNODE_IS_RSRC(vp)) { return (EPERM); } + else { + /* + * We know it's a data fork. + * Probe the cnode to see if we have a valid resource fork + * in hand or not. + */ + rsrc_vp = cp->c_rsrc_vp; + } + /* Don't allow deleting the journal or journal_info_block. */ - if (hfsmp->jnl && - (cp->c_fileid == hfsmp->hfs_jnlfileid || cp->c_fileid == hfsmp->hfs_jnlinfoblkid)) { + if (hfs_is_journal_file(hfsmp, cp)) { return (EPERM); } + + /* + * If removing a symlink, then we need to ensure that the + * data blocks for the symlink are not still in-flight or pending. + * If so, we will unlink the symlink here, making its blocks + * available for re-allocation by a subsequent transaction. That is OK, but + * then the I/O for the data blocks could then go out before the journal + * transaction that created it was flushed, leading to I/O ordering issues. + */ + if (vp->v_type == VLNK) { + /* + * This will block if the asynchronous journal flush is in progress. + * If this symlink is not being renamed over and doesn't have any open FDs, + * then we'll remove it from the journal's bufs below in kill_block. + */ + buf_wait_for_shadow_io (vp, 0); + } + /* * Hard links require special handling. */ @@ -2434,6 +3088,7 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, return hfs_unlink(hfsmp, dvp, vp, cnp, skip_reserve); } } + /* Directories should call hfs_rmdir! (unless they have a lot of attributes) */ if (vnode_isdir(vp)) { if (allow_dirs == 0) @@ -2454,23 +3109,30 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, /* Remove our entry from the namei cache. */ cache_purge(vp); - + /* - * We expect the caller, if operating on files, - * will have passed in a resource fork vnode with - * an iocount, even if there was no content. - * We only do the hfs_truncate on the rsrc fork - * if we know that it DID have content, however. - * This has the bonus of not requiring us to defer - * its removal, unless it is in use. + * If the caller was operating on a file (as opposed to a + * directory with EAs), then we need to figure out + * whether or not it has a valid resource fork vnode. + * + * If there was a valid resource fork vnode, then we need + * to use hfs_truncate to eliminate its data. If there is + * no vnode, then we hold the cnode lock which would + * prevent it from being created. As a result, + * we can use the data deletion functions which do not + * require that a cnode/vnode pair exist. */ /* Check if this file is being used. */ if (isdir == 0) { dataforkbusy = vnode_isinuse(vp, 0); - /* Only need to defer resource fork removal if in use and has content */ - if (rvp && (cp->c_blocks - VTOF(vp)->ff_blocks)) { - rsrcforkbusy = vnode_isinuse(rvp, 0); + /* + * At this point, we know that 'vp' points to the + * a data fork because we checked it up front. And if + * there is no rsrc fork, rsrc_vp will be NULL. + */ + if (rsrc_vp && (cp->c_blocks - VTOF(vp)->ff_blocks)) { + rsrcforkbusy = vnode_isinuse(rsrc_vp, 0); } } @@ -2485,6 +3147,11 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, (cp->c_attr.ca_recflags & kHFSHasAttributesMask) != 0) { defer_remove = 1; } + + /* If we are explicitly told to only unlink item and move to hidden dir, then do it */ + if (only_unlink) { + defer_remove = 1; + } /* * Carbon semantics prohibit deleting busy files. @@ -2502,9 +3169,16 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, if (hfsmp->hfs_flags & HFS_QUOTAS) (void)hfs_getinoquota(cp); #endif /* QUOTA */ - - /* Check if we need a ubc_setsize. */ - if (isdir == 0 && (!dataforkbusy || !rsrcforkbusy)) { + + /* + * Do a ubc_setsize to indicate we need to wipe contents if: + * 1) item is a regular file. + * 2) Neither fork is busy AND we are not told to unlink this. + * + * We need to check for the defer_remove since it can be set without + * having a busy data or rsrc fork + */ + if (isdir == 0 && (!dataforkbusy || !rsrcforkbusy) && (defer_remove == 0)) { /* * A ubc_setsize can cause a pagein so defer it * until after the cnode lock is dropped. The @@ -2514,7 +3188,7 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, if (!dataforkbusy && cp->c_datafork->ff_blocks && !isbigfile) { cp->c_flag |= C_NEED_DATA_SETSIZE; } - if (!rsrcforkbusy && rvp) { + if (!rsrcforkbusy && rsrc_vp) { cp->c_flag |= C_NEED_RSRC_SETSIZE; } } @@ -2525,40 +3199,50 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, started_tr = 1; // XXXdbg - if we're journaled, kill any dirty symlink buffers - if (hfsmp->jnl && vnode_islnk(vp)) + if (hfsmp->jnl && vnode_islnk(vp) && (defer_remove == 0)) { buf_iterate(vp, hfs_removefile_callback, BUF_SKIP_NONLOCKED, (void *)hfsmp); + } /* - * Truncate any non-busy forks. Busy forks will + * Prepare to truncate any non-busy forks. Busy forks will * get truncated when their vnode goes inactive. * Note that we will only enter this region if we * can avoid creating an open-unlinked file. If * either region is busy, we will have to create an open * unlinked file. - * Since we're already inside a transaction, - * tell hfs_truncate to skip the ubc_setsize. + * + * Since we are deleting the file, we need to stagger the runtime + * modifications to do things in such a way that a crash won't + * result in us getting overlapped extents or any other + * bad inconsistencies. As such, we call prepare_release_storage + * which updates the UBC, updates quota information, and releases + * any loaned blocks that belong to this file. No actual + * truncation or bitmap manipulation is done until *AFTER* + * the catalog record is removed. */ - if (isdir == 0 && (!dataforkbusy && !rsrcforkbusy)) { - /* - * Note that 5th argument to hfs_truncate indicates whether or not - * hfs_update calls should be suppressed in call to do_hfs_truncate - */ + if (isdir == 0 && (!dataforkbusy && !rsrcforkbusy) && (only_unlink == 0)) { + if (!dataforkbusy && !isbigfile && cp->c_datafork->ff_blocks != 0) { - /* skip update in hfs_truncate */ - error = hfs_truncate(vp, (off_t)0, IO_NDELAY, 1, 1, ctx); - if (error) + + error = hfs_prepare_release_storage (hfsmp, vp); + if (error) { goto out; - truncated = 1; + } + update_vh = 1; } - if (!rsrcforkbusy && rvp) { - /* skip update in hfs_truncate */ - error = hfs_truncate(rvp, (off_t)0, IO_NDELAY, 1, 1, ctx); - if (error) + + /* + * If the resource fork vnode does not exist, we can skip this step. + */ + if (!rsrcforkbusy && rsrc_vp) { + error = hfs_prepare_release_storage (hfsmp, rsrc_vp); + if (error) { goto out; - truncated = 1; + } + update_vh = 1; } } - + /* * Protect against a race with rename by using the component * name passed in and parent id from dvp (instead of using @@ -2658,46 +3342,68 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, if (error) goto out; - } else /* Not busy */ { + } + else { + /* + * Nobody is using this item; we can safely remove everything. + */ + struct filefork *temp_rsrc_fork = NULL; +#if QUOTA + off_t savedbytes; + int blksize = hfsmp->blockSize; +#endif + u_int32_t fileid = cp->c_fileid; + + /* + * Figure out if we need to read the resource fork data into + * core before wiping out the catalog record. + * + * 1) Must not be a directory + * 2) cnode's c_rsrcfork ptr must be NULL. + * 3) rsrc fork must have actual blocks + */ + if ((isdir == 0) && (cp->c_rsrcfork == NULL) && + (cp->c_blocks - VTOF(vp)->ff_blocks)) { + /* + * The resource fork vnode & filefork did not exist. + * Create a temporary one for use in this function only. + */ + MALLOC_ZONE (temp_rsrc_fork, struct filefork *, sizeof (struct filefork), M_HFSFORK, M_WAITOK); + bzero(temp_rsrc_fork, sizeof(struct filefork)); + temp_rsrc_fork->ff_cp = cp; + rl_init(&temp_rsrc_fork->ff_invalidranges); + } + + lockflags = hfs_systemfile_lock(hfsmp, SFL_CATALOG | SFL_ATTRIBUTE | SFL_BITMAP, HFS_EXCLUSIVE_LOCK); - if (cp->c_blocks > 0) { - printf("hfs_remove: attempting to delete a non-empty file %s\n", - cp->c_desc.cd_nameptr); - error = EBUSY; - goto out; + /* Look up the resource fork first, if necessary */ + if (temp_rsrc_fork) { + error = cat_lookup (hfsmp, &desc, 1, (struct cat_desc*) NULL, + (struct cat_attr*) NULL, &temp_rsrc_fork->ff_data, NULL); + if (error) { + FREE_ZONE (temp_rsrc_fork, sizeof(struct filefork), M_HFSFORK); + hfs_systemfile_unlock (hfsmp, lockflags); + goto out; + } } - lockflags = hfs_systemfile_lock(hfsmp, SFL_CATALOG | SFL_ATTRIBUTE | SFL_BITMAP, HFS_EXCLUSIVE_LOCK); if (!skip_reserve) { if ((error = cat_preflight(hfsmp, CAT_DELETE, NULL, 0))) { + if (temp_rsrc_fork) { + FREE_ZONE (temp_rsrc_fork, sizeof(struct filefork), M_HFSFORK); + } hfs_systemfile_unlock(hfsmp, lockflags); goto out; } } - + error = cat_delete(hfsmp, &desc, &cp->c_attr); - - if (error && error != ENXIO && error != ENOENT && truncated) { - if ((cp->c_datafork && cp->c_datafork->ff_size != 0) || - (cp->c_rsrcfork && cp->c_rsrcfork->ff_size != 0)) { - off_t data_size = 0; - off_t rsrc_size = 0; - if (cp->c_datafork) { - data_size = cp->c_datafork->ff_size; - } - if (cp->c_rsrcfork) { - rsrc_size = cp->c_rsrcfork->ff_size; - } - printf("hfs: remove: couldn't delete a truncated file (%s)" - "(error %d, data sz %lld; rsrc sz %lld)", - cp->c_desc.cd_nameptr, error, data_size, rsrc_size); - hfs_mark_volume_inconsistent(hfsmp); - } else { - printf("hfs: remove: strangely enough, deleting truncated file %s (%d) got err %d\n", - cp->c_desc.cd_nameptr, cp->c_attr.ca_fileid, error); - } + + if (error && error != ENXIO && error != ENOENT) { + printf("hfs_removefile: deleting file %s (%d), err: %d\n", + cp->c_desc.cd_nameptr, cp->c_attr.ca_fileid, error); } - + if (error == 0) { /* Update the parent directory */ if (dcp->c_entries > 0) @@ -2708,26 +3414,92 @@ hfs_removefile(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, (void) cat_update(hfsmp, &dcp->c_desc, &dcp->c_attr, NULL, NULL); } hfs_systemfile_unlock(hfsmp, lockflags); - if (error) - goto out; + if (error) { + if (temp_rsrc_fork) { + FREE_ZONE (temp_rsrc_fork, sizeof(struct filefork), M_HFSFORK); + } + goto out; + } + + /* + * Now that we've wiped out the catalog record, the file effectively doesn't + * exist anymore. So update the quota records to reflect the loss of the + * data fork and the resource fork. + */ #if QUOTA - if (hfsmp->hfs_flags & HFS_QUOTAS) + if (cp->c_datafork->ff_blocks > 0) { + savedbytes = ((off_t)cp->c_datafork->ff_blocks * (off_t)blksize); + (void) hfs_chkdq(cp, (int64_t)-(savedbytes), NOCRED, 0); + } + + /* + * We may have just deleted the catalog record for a resource fork even + * though it did not exist in core as a vnode. However, just because there + * was a resource fork pointer in the cnode does not mean that it had any blocks. + */ + if (temp_rsrc_fork || cp->c_rsrcfork) { + if (cp->c_rsrcfork) { + if (cp->c_rsrcfork->ff_blocks > 0) { + savedbytes = ((off_t)cp->c_rsrcfork->ff_blocks * (off_t)blksize); + (void) hfs_chkdq(cp, (int64_t)-(savedbytes), NOCRED, 0); + } + } + else { + /* we must have used a temporary fork */ + savedbytes = ((off_t)temp_rsrc_fork->ff_blocks * (off_t)blksize); + (void) hfs_chkdq(cp, (int64_t)-(savedbytes), NOCRED, 0); + } + } + + if (hfsmp->hfs_flags & HFS_QUOTAS) { (void)hfs_chkiq(cp, -1, NOCRED, 0); -#endif /* QUOTA */ + } +#endif + + /* + * If we didn't get any errors deleting the catalog entry, then go ahead + * and release the backing store now. The filefork pointers are still valid. + */ + if (temp_rsrc_fork) { + error = hfs_release_storage (hfsmp, cp->c_datafork, temp_rsrc_fork, fileid); + } + else { + /* if cp->c_rsrcfork == NULL, hfs_release_storage will skip over it. */ + error = hfs_release_storage (hfsmp, cp->c_datafork, cp->c_rsrcfork, fileid); + } + if (error) { + /* + * If we encountered an error updating the extents and bitmap, + * mark the volume inconsistent. At this point, the catalog record has + * already been deleted, so we can't recover it at this point. We need + * to proceed and update the volume header and mark the cnode C_NOEXISTS. + * The subsequent fsck should be able to recover the free space for us. + */ + hfs_mark_volume_inconsistent(hfsmp); + } + else { + /* reset update_vh to 0, since hfs_release_storage should have done it for us */ + update_vh = 0; + } + + /* Get rid of the temporary rsrc fork */ + if (temp_rsrc_fork) { + FREE_ZONE (temp_rsrc_fork, sizeof(struct filefork), M_HFSFORK); + } cp->c_flag |= C_NOEXISTS; cp->c_flag &= ~C_DELETED; - truncated = 0; // because the catalog entry is gone - + cp->c_touch_chgtime = TRUE; /* XXX needed ? */ --cp->c_linkcount; - + /* * We must never get a directory if we're in this else block. We could * accidentally drop the number of files in the volume header if we did. */ hfs_volupdate(hfsmp, VOL_RMFILE, (dcp->c_cnid == kHFSRootFolderID)); + } /* @@ -2744,14 +3516,14 @@ out: if (error) { cp->c_flag &= ~C_DELETED; } - - /* Commit the truncation to the catalog record */ - if (truncated) { - cp->c_flag |= C_FORCEUPDATE; - cp->c_touch_chgtime = TRUE; - cp->c_touch_modtime = TRUE; - (void) hfs_update(vp, 0); - } + + if (update_vh) { + /* + * If we bailed out earlier, we may need to update the volume header + * to deal with the borrowed blocks accounting. + */ + hfs_volupdate (hfsmp, VOL_UPDATE, 0); + } if (started_tr) { hfs_end_transaction(hfsmp); @@ -2813,7 +3585,7 @@ replace_desc(struct cnode *cp, struct cat_desc *cdp) * been locked. By taking the rsrc fork vnodes up front we ensure that they * cannot be recycled, and that the situation mentioned above cannot happen. */ -static int +int hfs_vnop_rename(ap) struct vnop_rename_args /* { struct vnode *a_fdvp; @@ -2829,8 +3601,13 @@ hfs_vnop_rename(ap) struct vnode *tdvp = ap->a_tdvp; struct vnode *fvp = ap->a_fvp; struct vnode *fdvp = ap->a_fdvp; - struct vnode *fvp_rsrc = NULLVP; + /* + * Note that we only need locals for the target/destination's + * resource fork vnode (and only if necessary). We don't care if the + * source has a resource fork vnode or not. + */ struct vnode *tvp_rsrc = NULLVP; + uint32_t tvp_rsrc_vid = 0; struct componentname *tcnp = ap->a_tcnp; struct componentname *fcnp = ap->a_fcnp; struct proc *p = vfs_context_proc(ap->a_context); @@ -2849,88 +3626,96 @@ hfs_vnop_rename(ap) int took_trunc_lock = 0; int lockflags; int error; - int recycle_rsrc = 0; + time_t orig_from_ctime, orig_to_ctime; + int emit_rename = 1; + int emit_delete = 1; + orig_from_ctime = VTOC(fvp)->c_ctime; + if (tvp && VTOC(tvp)) { + orig_to_ctime = VTOC(tvp)->c_ctime; + } else { + orig_to_ctime = ~0; + } + hfsmp = VTOHFS(tdvp); /* - * Before grabbing the four locks, we may need to get an iocount on the resource fork - * vnodes in question, just like hfs_vnop_remove. If fvp and tvp are not - * directories, then go ahead and grab the resource fork vnodes now - * one at a time. We don't actively need the fvp_rsrc to do the rename operation, - * but we need the iocount to prevent the vnode from getting recycled/reclaimed - * during the middle of the VNOP. + * Do special case checks here. If fvp == tvp then we need to check the + * cnode with locks held. */ - - - if ((vnode_isreg(fvp)) || (vnode_islnk(fvp))) { - - if ((error = hfs_lock (VTOC(fvp), HFS_EXCLUSIVE_LOCK))) { - return (error); - } - - /* - * We care if we race against rename/delete with this cnode, so we'll - * error out if this file becomes open-unlinked during this call. - */ - error = hfs_vgetrsrc(VTOHFS(fvp), fvp, &fvp_rsrc, TRUE, TRUE); - hfs_unlock (VTOC(fvp)); - if (error) { - if (fvp_rsrc) { - vnode_put (fvp_rsrc); - } + if (fvp == tvp) { + int is_hardlink = 0; + /* + * In this case, we do *NOT* ever emit a DELETE event. + * We may not necessarily emit a RENAME event + */ + emit_delete = 0; + if ((error = hfs_lock(VTOC(fvp), HFS_SHARED_LOCK))) { return error; } - } + /* Check to see if the item is a hardlink or not */ + is_hardlink = (VTOC(fvp)->c_flag & C_HARDLINK); + hfs_unlock (VTOC(fvp)); - if (tvp && (vnode_isreg(tvp) || vnode_islnk(tvp))) { /* - * Lock failure is OK on tvp, since we may race with a remove on the dst. - * But this shouldn't stop rename from proceeding, so only try to - * grab the resource fork if the lock succeeded. + * If the item is not a hardlink, then case sensitivity must be off, otherwise + * two names should not resolve to the same cnode unless they were case variants. */ - if (hfs_lock (VTOC(tvp), HFS_EXCLUSIVE_LOCK) == 0) { - tcp = VTOC(tvp); - - /* - * We only care if we get an open-unlinked file on the dst so we - * know to null out tvp/tcp to make the rename operation act - * as if they never existed. Because they're effectively out of the - * namespace already it's fine to do this. If this is true, then - * make sure to unlock the cnode and drop the iocount only after the unlock. + if (is_hardlink) { + emit_rename = 0; + /* + * Hardlinks are a little trickier. We only want to emit a rename event + * if the item is a hardlink, the parent directories are the same, case sensitivity + * is off, and the case folded names are the same. See the fvp == tvp case below for more + * info. */ - error = hfs_vgetrsrc(VTOHFS(tvp), tvp, &tvp_rsrc, TRUE, TRUE); - hfs_unlock (tcp); - if (error) { - /* - * Since we specify TRUE for error-on-unlinked in hfs_vgetrsrc, - * we can get a rsrc fork vp even if it returns an error. - */ - tcp = NULL; - tvp = NULL; - if (tvp_rsrc) { - vnode_put (tvp_rsrc); - tvp_rsrc = NULLVP; + + if ((fdvp == tdvp) && ((hfsmp->hfs_flags & HFS_CASE_SENSITIVE) == 0)) { + if (hfs_namecmp((const u_int8_t *)fcnp->cn_nameptr, fcnp->cn_namelen, + (const u_int8_t *)tcnp->cn_nameptr, tcnp->cn_namelen) == 0) { + /* Then in this case only it is ok to emit a rename */ + emit_rename = 1; } - /* just bypass truncate lock and act as if we never got tcp/tvp */ - goto retry; } } } + if (emit_rename) { + check_for_tracked_file(fvp, orig_from_ctime, NAMESPACE_HANDLER_RENAME_OP, NULL); + } + if (tvp && VTOC(tvp)) { + if (emit_delete) { + check_for_tracked_file(tvp, orig_to_ctime, NAMESPACE_HANDLER_DELETE_OP, NULL); + } + } + +retry: /* When tvp exists, take the truncate lock for hfs_removefile(). */ if (tvp && (vnode_isreg(tvp) || vnode_islnk(tvp))) { - hfs_lock_truncate(VTOC(tvp), TRUE); + hfs_lock_truncate(VTOC(tvp), HFS_EXCLUSIVE_LOCK); took_trunc_lock = 1; } - retry: error = hfs_lockfour(VTOC(fdvp), VTOC(fvp), VTOC(tdvp), tvp ? VTOC(tvp) : NULL, HFS_EXCLUSIVE_LOCK, &error_cnode); if (error) { if (took_trunc_lock) { - hfs_unlock_truncate(VTOC(tvp), TRUE); + hfs_unlock_truncate(VTOC(tvp), 0); took_trunc_lock = 0; } + + /* + * We hit an error path. If we were trying to re-acquire the locks + * after coming through here once, we might have already obtained + * an iocount on tvp's resource fork vnode. Drop that before dealing + * with the failure. Note this is safe -- since we are in an + * error handling path, we can't be holding the cnode locks. + */ + if (tvp_rsrc) { + vnode_put (tvp_rsrc); + tvp_rsrc_vid = 0; + tvp_rsrc = NULL; + } + /* * tvp might no longer exist. If the cause of the lock failure * was tvp, then we can try again with tvp/tcp set to NULL. @@ -2942,13 +3727,7 @@ hfs_vnop_rename(ap) tvp = NULL; goto retry; } - /* otherwise, drop iocounts on the rsrc forks and bail out */ - if (fvp_rsrc) { - vnode_put (fvp_rsrc); - } - if (tvp_rsrc) { - vnode_put (tvp_rsrc); - } + return (error); } @@ -2956,7 +3735,37 @@ hfs_vnop_rename(ap) fcp = VTOC(fvp); tdcp = VTOC(tdvp); tcp = tvp ? VTOC(tvp) : NULL; - hfsmp = VTOHFS(tdvp); + + /* + * Acquire iocounts on the destination's resource fork vnode + * if necessary. If dst/src are files and the dst has a resource + * fork vnode, then we need to try and acquire an iocount on the rsrc vnode. + * If it does not exist, then we don't care and can skip it. + */ + if ((vnode_isreg(fvp)) || (vnode_islnk(fvp))) { + if ((tvp) && (tcp->c_rsrc_vp) && (tvp_rsrc == NULL)) { + tvp_rsrc = tcp->c_rsrc_vp; + /* + * We can look at the vid here because we're holding the + * cnode lock on the underlying cnode for this rsrc vnode. + */ + tvp_rsrc_vid = vnode_vid (tvp_rsrc); + + /* Unlock everything to acquire iocount on this rsrc vnode */ + if (took_trunc_lock) { + hfs_unlock_truncate (VTOC(tvp), 0); + took_trunc_lock = 0; + } + hfs_unlockfour(fdcp, fcp, tdcp, tcp); + + if (vnode_getwithvid (tvp_rsrc, tvp_rsrc_vid)) { + /* iocount acquisition failed. Reset fields and start over.. */ + tvp_rsrc_vid = 0; + tvp_rsrc = NULL; + } + goto retry; + } + } /* Ensure we didn't race src or dst parent directories with rmdir. */ if (fdcp->c_flag & (C_NOEXISTS | C_DELETED)) { @@ -2974,21 +3783,22 @@ hfs_vnop_rename(ap) * the parent/child relationship with fdcp and tdcp, as well as the * component name of the target cnodes. */ - if ((fcp->c_flag & (C_NOEXISTS | C_DELETED)) || !hfs_valid_cnode(hfsmp, fdvp, fcnp, fcp->c_fileid)) { + if ((fcp->c_flag & (C_NOEXISTS | C_DELETED)) || !hfs_valid_cnode(hfsmp, fdvp, fcnp, fcp->c_fileid, NULL, &error)) { error = ENOENT; goto out; } - if (tcp && ((tcp->c_flag & (C_NOEXISTS | C_DELETED)) || !hfs_valid_cnode(hfsmp, tdvp, tcnp, tcp->c_fileid))) { + if (tcp && ((tcp->c_flag & (C_NOEXISTS | C_DELETED)) || !hfs_valid_cnode(hfsmp, tdvp, tcnp, tcp->c_fileid, NULL, &error))) { // // hmm, the destination vnode isn't valid any more. // in this case we can just drop him and pretend he // never existed in the first place. // if (took_trunc_lock) { - hfs_unlock_truncate(VTOC(tvp), TRUE); - took_trunc_lock = 0; + hfs_unlock_truncate(VTOC(tvp), 0); + took_trunc_lock = 0; } + error = 0; hfs_unlockfour(fdcp, fcp, tdcp, tcp); @@ -3083,7 +3893,7 @@ hfs_vnop_rename(ap) /* * Make sure "from" vnode and its parent are changeable. */ - if ((fcp->c_flags & (IMMUTABLE | APPEND)) || (fdcp->c_flags & APPEND)) { + if ((fcp->c_bsdflags & (IMMUTABLE | APPEND)) || (fdcp->c_bsdflags & APPEND)) { error = EPERM; goto out; } @@ -3105,6 +3915,13 @@ hfs_vnop_rename(ap) goto out; } + /* Don't allow modification of the journal or journal_info_block */ + if (hfs_is_journal_file(hfsmp, fcp) || + (tcp && hfs_is_journal_file(hfsmp, tcp))) { + error = EPERM; + goto out; + } + #if QUOTA if (tvp) (void)hfs_getinoquota(tcp); @@ -3186,7 +4003,33 @@ hfs_vnop_rename(ap) got_cookie = 1; /* - * If the destination exists then it may need to be removed. + * If the destination exists then it may need to be removed. + * + * Due to HFS's locking system, we should always move the + * existing 'tvp' element to the hidden directory in hfs_vnop_rename. + * Because the VNOP_LOOKUP call enters and exits the filesystem independently + * of the actual vnop that it was trying to do (stat, link, readlink), + * we must release the cnode lock of that element during the interim to + * do MAC checking, vnode authorization, and other calls. In that time, + * the item can be deleted (or renamed over). However, only in the rename + * case is it inappropriate to return ENOENT from any of those calls. Either + * the call should return information about the old element (stale), or get + * information about the newer element that we are about to write in its place. + * + * HFS lookup has been modified to detect a rename and re-drive its + * lookup internally. For other calls that have already succeeded in + * their lookup call and are waiting to acquire the cnode lock in order + * to proceed, that cnode lock will not fail due to the cnode being marked + * C_NOEXISTS, because it won't have been marked as such. It will only + * have C_DELETED. Thus, they will simply act on the stale open-unlinked + * element. All future callers will get the new element. + * + * To implement this behavior, we pass the "only_unlink" argument to + * hfs_removefile and hfs_removedir. This will result in the vnode acting + * as though it is open-unlinked. Additionally, when we are done moving the + * element to the hidden directory, we vnode_recycle the target so that it is + * reclaimed as soon as possible. Reclaim and inactive are both + * capable of clearing out unused blocks for an open-unlinked file or dir. */ if (tvp) { /* @@ -3195,42 +4038,102 @@ hfs_vnop_rename(ap) */ if (fvp == tvp) { if (!(fcp->c_flag & C_HARDLINK)) { + /* + * If they're not hardlinks, then fvp == tvp must mean we + * are using case-insensitive HFS because case-sensitive would + * not use the same vnode for both. In this case we just update + * the catalog for: a -> A + */ goto skip_rm; /* simple case variant */ - } else if ((fdvp != tdvp) || + } + /* For all cases below, we must be using hardlinks */ + else if ((fdvp != tdvp) || (hfsmp->hfs_flags & HFS_CASE_SENSITIVE)) { + /* + * If the parent directories are not the same, AND the two items + * are hardlinks, posix says to do nothing: + * dir1/fred <-> dir2/bob and the op was mv dir1/fred -> dir2/bob + * We just return 0 in this case. + * + * If case sensitivity is on, and we are using hardlinks + * then renaming is supposed to do nothing. + * dir1/fred <-> dir2/FRED, and op == mv dir1/fred -> dir2/FRED + */ goto out; /* matching hardlinks, nothing to do */ } else if (hfs_namecmp((const u_int8_t *)fcnp->cn_nameptr, fcnp->cn_namelen, (const u_int8_t *)tcnp->cn_nameptr, tcnp->cn_namelen) == 0) { + /* + * If we get here, then the following must be true: + * a) We are running case-insensitive HFS+. + * b) Both paths 'fvp' and 'tvp' are in the same parent directory. + * c) the two names are case-variants of each other. + * + * In this case, we are really only dealing with a single catalog record + * whose name is being updated. + * + * op is dir1/fred -> dir1/FRED + * + * We need to special case the name matching, because if + * dir1/fred <-> dir1/bob were the two links, and the + * op was dir1/fred -> dir1/bob + * That would fail/do nothing. + */ goto skip_rm; /* case-variant hardlink in the same dir */ } else { goto out; /* matching hardlink, nothing to do */ } } - if (vnode_isdir(tvp)) - error = hfs_removedir(tdvp, tvp, tcnp, HFSRM_SKIP_RESERVE); - else { - error = hfs_removefile(tdvp, tvp, tcnp, 0, HFSRM_SKIP_RESERVE, 0, tvp_rsrc); - - /* - * If the destination file had a rsrc fork vnode, it may have been cleaned up - * in hfs_removefile if it was not busy (had no usecounts). This is possible - * because we grabbed the iocount on the rsrc fork safely at the beginning - * of the function before we did the lockfour. However, we may still need - * to take action to prevent block leaks, so aggressively recycle the vnode - * if possible. The vnode cannot be recycled because we hold an iocount on it. + + if (vnode_isdir(tvp)) { + /* + * hfs_removedir will eventually call hfs_removefile on the directory + * we're working on, because only hfs_removefile does the renaming of the + * item to the hidden directory. The directory will stay around in the + * hidden directory with C_DELETED until it gets an inactive or a reclaim. + * That way, we can destroy all of the EAs as needed and allow new ones to be + * written. */ - - if ((error == 0) && (tcp->c_flag & C_DELETED) && tvp_rsrc && !vnode_isinuse(tvp_rsrc, 0)) { - recycle_rsrc = 1; - } + error = hfs_removedir(tdvp, tvp, tcnp, HFSRM_SKIP_RESERVE, 1); + } + else { + error = hfs_removefile(tdvp, tvp, tcnp, 0, HFSRM_SKIP_RESERVE, 0, NULL, 1); + + /* + * If the destination file had a resource fork vnode, then we need to get rid of + * its blocks when there are no more references to it. Because the call to + * hfs_removefile above always open-unlinks things, we need to force an inactive/reclaim + * on the resource fork vnode, in order to prevent block leaks. Otherwise, + * the resource fork vnode could prevent the data fork vnode from going out of scope + * because it holds a v_parent reference on it. So we mark it for termination + * with a call to vnode_recycle. hfs_vnop_reclaim has been modified so that it + * can clean up the blocks of open-unlinked files and resource forks. + * + * We can safely call vnode_recycle on the resource fork because we took an iocount + * reference on it at the beginning of the function. + */ + + if ((error == 0) && (tcp->c_flag & C_DELETED) && (tvp_rsrc)) { + vnode_recycle(tvp_rsrc); + } } - if (error) + if (error) { goto out; + } + tvp_deleted = 1; + + /* Mark 'tcp' as being deleted due to a rename */ + tcp->c_flag |= C_RENAMED; + + /* + * Aggressively mark tvp/tcp for termination to ensure that we recover all blocks + * as quickly as possible. + */ + vnode_recycle(tvp); } skip_rm: /* @@ -3269,6 +4172,11 @@ skip_rm: fcp->c_parentcnid = tdcp->c_fileid; fcp->c_hint = 0; + /* Now indicate this cnode needs to have date-added written to the finderinfo */ + fcp->c_flag |= C_NEEDS_DATEADDED; + (void) hfs_update (fvp, 0); + + hfs_volupdate(hfsmp, vnode_isdir(fvp) ? VOL_RMDIR : VOL_RMFILE, (fdcp->c_cnid == kHFSRootFolderID)); hfs_volupdate(hfsmp, vnode_isdir(fvp) ? VOL_MKDIR : VOL_MKFILE, @@ -3312,6 +4220,41 @@ skip_rm: tdcp->c_flag |= C_FORCEUPDATE; // XXXdbg - force it out! (void) hfs_update(tdvp, 0); + + /* Update the vnode's name now that the rename has completed. */ + vnode_update_identity(fvp, tdvp, tcnp->cn_nameptr, tcnp->cn_namelen, + tcnp->cn_hash, (VNODE_UPDATE_PARENT | VNODE_UPDATE_NAME)); + + /* + * At this point, we may have a resource fork vnode attached to the + * 'from' vnode. If it exists, we will want to update its name, because + * it contains the old name + _PATH_RSRCFORKSPEC. ("/..namedfork/rsrc"). + * + * Note that the only thing we need to update here is the name attached to + * the vnode, since a resource fork vnode does not have a separate resource + * cnode -- it's still 'fcp'. + */ + if (fcp->c_rsrc_vp) { + char* rsrc_path = NULL; + int len; + + /* Create a new temporary buffer that's going to hold the new name */ + MALLOC_ZONE (rsrc_path, caddr_t, MAXPATHLEN, M_NAMEI, M_WAITOK); + len = snprintf (rsrc_path, MAXPATHLEN, "%s%s", tcnp->cn_nameptr, _PATH_RSRCFORKSPEC); + len = MIN(len, MAXPATHLEN); + + /* + * vnode_update_identity will do the following for us: + * 1) release reference on the existing rsrc vnode's name. + * 2) copy/insert new name into the name cache + * 3) attach the new name to the resource vnode + * 4) update the vnode's vid + */ + vnode_update_identity (fcp->c_rsrc_vp, fvp, rsrc_path, len, 0, (VNODE_UPDATE_NAME | VNODE_UPDATE_CACHE)); + + /* Free the memory associated with the resource fork's name */ + FREE_ZONE (rsrc_path, MAXPATHLEN, M_NAMEI); + } out: if (got_cookie) { cat_postflight(hfsmp, &cookie, p); @@ -3327,34 +4270,16 @@ out: wakeup((caddr_t)&tdcp->c_flag); } - if (took_trunc_lock) - hfs_unlock_truncate(VTOC(tvp), TRUE); + if (took_trunc_lock) { + hfs_unlock_truncate(VTOC(tvp), 0); + } hfs_unlockfour(fdcp, fcp, tdcp, tcp); - /* - * Now that we've dropped all of the locks, we need to force an inactive and a recycle - * on the old destination's rsrc fork to prevent a leak of its blocks. Note that - * doing the ref/rele is to twiddle the VL_NEEDINACTIVE bit of the vnode's flags, so that - * on the last vnode_put for this vnode, we will force inactive to get triggered. - * We hold an iocount from the beginning of this function so we know it couldn't have been - * recycled already. - */ - if (recycle_rsrc) { - int vref; - vref = vnode_ref(tvp_rsrc); - if (vref == 0) { - vnode_rele(tvp_rsrc); - } - vnode_recycle(tvp_rsrc); - } - - /* Now vnode_put the resource forks vnodes if necessary */ + /* Now vnode_put the resource fork vnode if necessary */ if (tvp_rsrc) { vnode_put(tvp_rsrc); - } - if (fvp_rsrc) { - vnode_put(fvp_rsrc); + tvp_rsrc = NULL; } /* After tvp is removed the only acceptable error is EIO */ @@ -3368,7 +4293,7 @@ out: /* * Make a directory. */ -static int +int hfs_vnop_mkdir(struct vnop_mkdir_args *ap) { /***** HACK ALERT ********/ @@ -3380,7 +4305,7 @@ hfs_vnop_mkdir(struct vnop_mkdir_args *ap) /* * Create a symbolic link. */ -static int +int hfs_vnop_symlink(struct vnop_symlink_args *ap) { struct vnode **vpp = ap->a_vpp; @@ -3456,7 +4381,7 @@ hfs_vnop_symlink(struct vnop_symlink_args *ap) /* hfs_removefile() requires holding the truncate lock */ hfs_unlock(cp); - hfs_lock_truncate(cp, TRUE); + hfs_lock_truncate(cp, HFS_EXCLUSIVE_LOCK); hfs_lock(cp, HFS_FORCE_LOCK); if (hfs_start_transaction(hfsmp) != 0) { @@ -3465,8 +4390,8 @@ hfs_vnop_symlink(struct vnop_symlink_args *ap) goto out; } - (void) hfs_removefile(dvp, vp, ap->a_cnp, 0, 0, 0, NULL); - hfs_unlock_truncate(cp, TRUE); + (void) hfs_removefile(dvp, vp, ap->a_cnp, 0, 0, 0, NULL, 0); + hfs_unlock_truncate(cp, 0); goto out; } @@ -3562,7 +4487,7 @@ typedef union { * If the directory is marked as deleted-but-in-use (cp->c_flag & C_DELETED), * do NOT synthesize entries for "." and "..". */ -static int +int hfs_vnop_readdir(ap) struct vnop_readdir_args /* { vnode_t a_vp; @@ -3601,11 +4526,23 @@ hfs_vnop_readdir(ap) /* Sanity check the uio data. */ if (uio_iovcnt(uio) > 1) return (EINVAL); + + if (VTOC(vp)->c_bsdflags & UF_COMPRESSED) { + int compressed = hfs_file_is_compressed(VTOC(vp), 0); /* 0 == take the cnode lock */ + if (VTOCMP(vp) != NULL && !compressed) { + error = check_for_dataless_file(vp, NAMESPACE_HANDLER_READ_OP); + if (error) { + return error; + } + } + } + + cp = VTOC(vp); + hfsmp = VTOHFS(vp); + /* Note that the dirhint calls require an exclusive lock. */ if ((error = hfs_lock(VTOC(vp), HFS_EXCLUSIVE_LOCK))) return (error); - cp = VTOC(vp); - hfsmp = VTOHFS(vp); /* Pick up cnid hint (if any). */ if (nfs_cookies) { @@ -3741,7 +4678,7 @@ hfs_vnop_readdir(ap) if (index == 0) { dirhint->dh_threadhint = cp->c_dirthreadhint; - } + } else { /* * If we have a non-zero index, there is a possibility that during the last @@ -3763,7 +4700,7 @@ hfs_vnop_readdir(ap) } /* Pack the buffer with dirent entries. */ - error = cat_getdirentries(hfsmp, cp->c_entries, dirhint, uio, extended, &items, &eofflag); + error = cat_getdirentries(hfsmp, cp->c_entries, dirhint, uio, ap->a_flags, &items, &eofflag); if (index == 0 && error == 0) { cp->c_dirthreadhint = dirhint->dh_threadhint; @@ -3822,7 +4759,7 @@ out: /* * Read contents of a symbolic link. */ -static int +int hfs_vnop_readlink(ap) struct vnop_readlink_args /* { struct vnode *a_vp; @@ -3845,7 +4782,6 @@ hfs_vnop_readlink(ap) /* Zero length sym links are not allowed */ if (fp->ff_size == 0 || fp->ff_size > MAXPATHLEN) { - printf("hfs: zero length symlink on fileid %d\n", cp->c_fileid); error = EINVAL; goto exit; } @@ -3907,7 +4843,7 @@ exit: /* * Get configurable pathname variables. */ -static int +int hfs_vnop_pathconf(ap) struct vnop_pathconf_args /* { struct vnode *a_vp; @@ -3925,9 +4861,9 @@ hfs_vnop_pathconf(ap) break; case _PC_NAME_MAX: if (VTOHFS(ap->a_vp)->hfs_flags & HFS_STANDARD) - *ap->a_retval = kHFSMaxFileNameChars; /* 255 */ + *ap->a_retval = kHFSMaxFileNameChars; /* 31 */ else - *ap->a_retval = kHFSPlusMaxFileNameChars; /* 31 */ + *ap->a_retval = kHFSPlusMaxFileNameChars; /* 255 */ break; case _PC_PATH_MAX: *ap->a_retval = PATH_MAX; /* 1024 */ @@ -3942,7 +4878,10 @@ hfs_vnop_pathconf(ap) *ap->a_retval = 200112; /* _POSIX_NO_TRUNC */ break; case _PC_NAME_CHARS_MAX: - *ap->a_retval = kHFSPlusMaxFileNameChars; + if (VTOHFS(ap->a_vp)->hfs_flags & HFS_STANDARD) + *ap->a_retval = kHFSMaxFileNameChars; /* 31 */ + else + *ap->a_retval = kHFSPlusMaxFileNameChars; /* 255 */ break; case _PC_CASE_SENSITIVE: if (VTOHFS(ap->a_vp)->hfs_flags & HFS_CASE_SENSITIVE) @@ -3959,6 +4898,10 @@ hfs_vnop_pathconf(ap) else *ap->a_retval = 64; /* number of bits to store max file size */ break; + case _PC_XATTR_SIZE_BITS: + /* Number of bits to store maximum extended attribute size */ + *ap->a_retval = HFS_XATTR_SIZE_BITS; + break; default: return (EINVAL); } @@ -3975,7 +4918,6 @@ hfs_vnop_pathconf(ap) * * The cnode must be locked exclusive */ -__private_extern__ int hfs_update(struct vnode *vp, __unused int waitfor) { @@ -4040,30 +4982,52 @@ hfs_update(struct vnode *vp, __unused int waitfor) return error; } - /* - * For files with invalid ranges (holes) the on-disk - * field representing the size of the file (cf_size) - * must be no larger than the start of the first hole. - */ - if (dataforkp && !TAILQ_EMPTY(&cp->c_datafork->ff_invalidranges)) { - bcopy(dataforkp, &datafork, sizeof(datafork)); - datafork.cf_size = TAILQ_FIRST(&cp->c_datafork->ff_invalidranges)->rl_start; - dataforkp = &datafork; - } else if (dataforkp && (cp->c_datafork->ff_unallocblocks != 0)) { - // always make sure the block count and the size - // of the file match the number of blocks actually - // allocated to the file on disk - bcopy(dataforkp, &datafork, sizeof(datafork)); - // make sure that we don't assign a negative block count - if (cp->c_datafork->ff_blocks < cp->c_datafork->ff_unallocblocks) { - panic("hfs: ff_blocks %d is less than unalloc blocks %d\n", - cp->c_datafork->ff_blocks, cp->c_datafork->ff_unallocblocks); - } - datafork.cf_blocks = (cp->c_datafork->ff_blocks - cp->c_datafork->ff_unallocblocks); - datafork.cf_size = datafork.cf_blocks * HFSTOVCB(hfsmp)->blockSize; - dataforkp = &datafork; - } - + /* + * Modify the values passed to cat_update based on whether or not + * the file has invalid ranges or borrowed blocks. + */ + if (dataforkp) { + off_t numbytes = 0; + + /* copy the datafork into a temporary copy so we don't pollute the cnode's */ + bcopy(dataforkp, &datafork, sizeof(datafork)); + dataforkp = &datafork; + + /* + * If there are borrowed blocks, ensure that they are subtracted + * from the total block count before writing the cnode entry to disk. + * Only extents that have actually been marked allocated in the bitmap + * should be reflected in the total block count for this fork. + */ + if (cp->c_datafork->ff_unallocblocks != 0) { + // make sure that we don't assign a negative block count + if (cp->c_datafork->ff_blocks < cp->c_datafork->ff_unallocblocks) { + panic("hfs: ff_blocks %d is less than unalloc blocks %d\n", + cp->c_datafork->ff_blocks, cp->c_datafork->ff_unallocblocks); + } + + /* Also cap the LEOF to the total number of bytes that are allocated. */ + datafork.cf_blocks = (cp->c_datafork->ff_blocks - cp->c_datafork->ff_unallocblocks); + datafork.cf_size = datafork.cf_blocks * HFSTOVCB(hfsmp)->blockSize; + } + + /* + * For files with invalid ranges (holes) the on-disk + * field representing the size of the file (cf_size) + * must be no larger than the start of the first hole. + * However, note that if the first invalid range exists + * solely within borrowed blocks, then our LEOF and block + * count should both be zero. As a result, set it to the + * min of the current cf_size and the start of the first + * invalid range, because it may have already been reduced + * to zero by the borrowed blocks check above. + */ + if (!TAILQ_EMPTY(&cp->c_datafork->ff_invalidranges)) { + numbytes = TAILQ_FIRST(&cp->c_datafork->ff_invalidranges)->rl_start; + datafork.cf_size = MIN((numbytes), (datafork.cf_size)); + } + } + /* * For resource forks with delayed allocations, make sure * the block count and file size match the number of blocks @@ -4098,7 +5062,7 @@ hfs_update(struct vnode *vp, __unused int waitfor) * Allocate a new node * Note - Function does not create and return a vnode for whiteout creation. */ -static int +int hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct vnode_attr *vap, vfs_context_t ctx) { @@ -4113,22 +5077,36 @@ hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, int error, started_tr = 0; enum vtype vnodetype; int mode; + int newvnode_flags = 0; + u_int32_t gnv_flags = 0; + int protectable_target = 0; + +#if CONFIG_PROTECT + struct cprotect *entry = NULL; + uint32_t cp_class = 0; + if (VATTR_IS_ACTIVE(vap, va_dataprotect_class)) { + cp_class = vap->va_dataprotect_class; + } + int protected_mount = 0; +#endif + if ((error = hfs_lock(VTOC(dvp), HFS_EXCLUSIVE_LOCK))) return (error); /* set the cnode pointer only after successfully acquiring lock */ dcp = VTOC(dvp); - + /* Don't allow creation of new entries in open-unlinked directories */ - if ((error = hfs_checkdeleted (dcp))) { - hfs_unlock (dcp); + if ((error = hfs_checkdeleted(dcp))) { + hfs_unlock(dcp); return error; } dcp->c_flag |= C_DIR_MODIFICATION; - + hfsmp = VTOHFS(dvp); + *vpp = NULL; tvp = NULL; out_desc.cd_flags = 0; @@ -4139,6 +5117,11 @@ hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, vnodetype = VREG; mode = MAKEIMODE(vnodetype, vap->va_mode); + if (S_ISDIR (mode) || S_ISREG (mode)) { + protectable_target = 1; + } + + /* Check if were out of usable disk space. */ if ((hfs_freeblks(hfsmp, 1) == 0) && (vfs_context_suser(ctx) != 0)) { error = ENOSPC; @@ -4183,6 +5166,31 @@ hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, } } +#if CONFIG_PROTECT + if (cp_fs_protected(hfsmp->hfs_mp)) { + protected_mount = 1; + } + /* + * On a content-protected HFS+/HFSX filesystem, files and directories + * cannot be created without atomically setting/creating the EA that + * contains the protection class metadata and keys at the same time, in + * the same transaction. As a result, pre-set the "EAs exist" flag + * on the cat_attr for protectable catalog record creations. This will + * cause the cnode creation routine in hfs_getnewvnode to mark the cnode + * as having EAs. + */ + if ((protected_mount) && (protectable_target)) { + attr.ca_recflags |= kHFSHasAttributesMask; + } +#endif + + + /* + * Add the date added to the item. See above, as + * all of the dates are set to the itime. + */ + hfs_write_dateadded (&attr, attr.ca_atime); + attr.ca_uid = vap->va_uid; attr.ca_gid = vap->va_gid; VATTR_SET_SUPPORTED(vap, va_mode); @@ -4221,6 +5229,22 @@ hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, in_desc.cd_hint = dcp->c_childhint; in_desc.cd_encoding = 0; +#if CONFIG_PROTECT + /* + * To preserve file creation atomicity with regards to the content protection EA, + * we must create the file in the catalog and then write out the EA in the same + * transaction. Pre-flight any operations that we can (such as allocating/preparing + * the buffer, wrapping the keys) before we start the txn and take the requisite + * b-tree locks. We pass '0' as the fileid because we do not know it yet. + */ + if ((protected_mount) && (protectable_target)) { + error = cp_entry_create_keys (&entry, dcp, hfsmp, cp_class, 0, attr.ca_mode); + if (error) { + goto exit; + } + } +#endif + if ((error = hfs_start_transaction(hfsmp)) != 0) { goto exit; } @@ -4248,6 +5272,40 @@ hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, dcp->c_ctime = tv.tv_sec; dcp->c_mtime = tv.tv_sec; (void) cat_update(hfsmp, &dcp->c_desc, &dcp->c_attr, NULL, NULL); + +#if CONFIG_PROTECT + /* + * If we are creating a content protected file, now is when + * we create the EA. We must create it in the same transaction + * that creates the file. We can also guarantee that the file + * MUST exist because we are still holding the catalog lock + * at this point. + */ + if ((attr.ca_fileid != 0) && (protected_mount) && (protectable_target)) { + error = cp_setxattr (NULL, entry, hfsmp, attr.ca_fileid, XATTR_CREATE); + + if (error) { + int delete_err; + /* + * If we fail the EA creation, then we need to delete the file. + * Luckily, we are still holding all of the right locks. + */ + delete_err = cat_delete (hfsmp, &out_desc, &attr); + if (delete_err == 0) { + /* Update the parent directory */ + if (dcp->c_entries > 0) + dcp->c_entries--; + dcp->c_dirchangecnt++; + dcp->c_ctime = tv.tv_sec; + dcp->c_mtime = tv.tv_sec; + (void) cat_update(hfsmp, &dcp->c_desc, &dcp->c_attr, NULL, NULL); + } + + /* Emit EINVAL if we fail to create EA*/ + error = EINVAL; + } + } +#endif } hfs_systemfile_unlock(hfsmp, lockflags); if (error) @@ -4279,10 +5337,26 @@ hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, started_tr = 0; } +#if CONFIG_PROTECT + /* + * At this point, we must have encountered success with writing the EA. + * Update MKB with the data for the cached key, then destroy it. This may + * prevent information leakage by ensuring the cache key is only unwrapped + * to perform file I/O and it is allowed. + */ + + if ((attr.ca_fileid != 0) && (protected_mount) && (protectable_target)) { + cp_update_mkb (entry, attr.ca_fileid); + cp_entry_destroy (&entry); + } +#endif + /* Do not create vnode for whiteouts */ if (S_ISWHT(mode)) { goto exit; - } + } + + gnv_flags |= GNV_CREATE; /* * Create a vnode for the object just created. @@ -4297,18 +5371,44 @@ hfs_makenode(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, * try to create a new vnode, and then end up reclaiming another shadow vnode to * create the new one. However, if everything is working properly, this should * be a non-issue as we would never enter that reclaim codepath. - * + * * The cnode is locked on successful return. */ - error = hfs_getnewvnode(hfsmp, dvp, cnp, &out_desc, GNV_CREATE, &attr, NULL, &tvp); + error = hfs_getnewvnode(hfsmp, dvp, cnp, &out_desc, gnv_flags, &attr, + NULL, &tvp, &newvnode_flags); if (error) goto exit; cp = VTOC(tvp); *vpp = tvp; + +#if QUOTA + /* + * Once we create this vnode, we need to initialize its quota data + * structures, if necessary. We know that it is OK to just go ahead and + * initialize because we've already validated earlier (through the hfs_quotacheck + * function) to see if creating this cnode/vnode would cause us to go over quota. + */ + if (hfsmp->hfs_flags & HFS_QUOTAS) { + (void) hfs_getinoquota(cp); + } +#endif + exit: cat_releasedesc(&out_desc); +#if CONFIG_PROTECT + /* + * We may have jumped here in error-handling various situations above. + * If we haven't already dumped the temporary CP used to initialize + * the file atomically, then free it now. cp_entry_destroy should null + * out the pointer if it was called already. + */ + if (entry) { + cp_entry_destroy (&entry); + } +#endif + /* * Make sure we release cnode lock on dcp. */ @@ -4330,8 +5430,8 @@ exit: } - -/* hfs_vgetrsrc acquires a resource fork vnode corresponding to the cnode that is +/* + * hfs_vgetrsrc acquires a resource fork vnode corresponding to the cnode that is * found in 'vp'. The rsrc fork vnode is returned with the cnode locked and iocount * on the rsrc vnode. * @@ -4351,10 +5451,9 @@ exit: * there's really no reason to double-check for errors on the cnode. */ -__private_extern__ int -hfs_vgetrsrc(struct hfsmount *hfsmp, struct vnode *vp, - struct vnode **rvpp, int can_drop_lock, int error_on_unlinked) +hfs_vgetrsrc(struct hfsmount *hfsmp, struct vnode *vp, struct vnode **rvpp, + int can_drop_lock, int error_on_unlinked) { struct vnode *rvp; struct vnode *dvp = NULLVP; @@ -4363,18 +5462,21 @@ hfs_vgetrsrc(struct hfsmount *hfsmp, struct vnode *vp, int vid; int delete_status = 0; - + if (vnode_vtype(vp) == VDIR) { + return EINVAL; + } + /* - * Need to check the status of the cnode to validate it hasn't - * gone open-unlinked on us before we can actually do work with it. + * Need to check the status of the cnode to validate it hasn't gone + * open-unlinked on us before we can actually do work with it. */ - delete_status = hfs_checkdeleted (cp); + delete_status = hfs_checkdeleted(cp); if ((delete_status) && (error_on_unlinked)) { return delete_status; } restart: - /* Attempt to use exising vnode */ + /* Attempt to use existing vnode */ if ((rvp = cp->c_rsrc_vp)) { vid = vnode_vid(rvp); @@ -4410,11 +5512,10 @@ restart: if ((delete_status = hfs_checkdeleted(cp))) { /* * If error == 0, this means that we succeeded in acquiring an iocount on the - * rsrc fork vnode. However, if we're in this block of code, that - * means that we noticed that the cnode has gone open-unlinked. In - * this case, the caller requested that we not do any other work and - * return an errno. The caller will be responsible for dropping the - * iocount we just acquired because we can't do it until we've released + * rsrc fork vnode. However, if we're in this block of code, that means that we noticed + * that the cnode has gone open-unlinked. In this case, the caller requested that we + * not do any other work and return an errno. The caller will be responsible for + * dropping the iocount we just acquired because we can't do it until we've released * the cnode lock. */ if (error == 0) { @@ -4447,7 +5548,8 @@ restart: struct cat_desc to_desc; char delname[32]; int lockflags; - + int newvnode_flags = 0; + /* * Make sure cnode lock is exclusive, if not upgrade it. * @@ -4478,7 +5580,7 @@ restart: */ if ((error_on_unlinked) && (can_drop_lock)) { - if ((error = hfs_checkdeleted (cp))) { + if ((error = hfs_checkdeleted(cp))) { return error; } } @@ -4502,14 +5604,43 @@ restart: lockflags = hfs_systemfile_lock(hfsmp, SFL_CATALOG, HFS_SHARED_LOCK); - /* Get resource fork data */ - error = cat_lookup(hfsmp, descptr, 1, (struct cat_desc *)0, - (struct cat_attr *)0, &rsrcfork, NULL); + /* + * Get resource fork data + * + * We call cat_idlookup (instead of cat_lookup) below because we can't + * trust the descriptor in the provided cnode for lookups at this point. + * Between the time of the original lookup of this vnode and now, the + * descriptor could have gotten swapped or replaced. If this occurred, + * the parent/name combo originally desired may not necessarily be provided + * if we use the descriptor. Even worse, if the vnode represents + * a hardlink, we could have removed one of the links from the namespace + * but left the descriptor alone, since hfs_unlink does not invalidate + * the descriptor in the cnode if other links still point to the inode. + * + * Consider the following (slightly contrived) scenario: + * /tmp/a <--> /tmp/b (hardlinks). + * 1. Thread A: open rsrc fork on /tmp/b. + * 1a. Thread A: does lookup, goes out to lunch right before calling getnamedstream. + * 2. Thread B does 'mv /foo/b /tmp/b' + * 2. Thread B succeeds. + * 3. Thread A comes back and wants rsrc fork info for /tmp/b. + * + * Even though the hardlink backing /tmp/b is now eliminated, the descriptor + * is not removed/updated during the unlink process. So, if you were to + * do a lookup on /tmp/b, you'd acquire an entirely different record's resource + * fork. + * + * As a result, we use the fileid, which should be invariant for the lifetime + * of the cnode (possibly barring calls to exchangedata). + */ + + error = cat_idlookup (hfsmp, cp->c_attr.ca_fileid, 0, 1, NULL, NULL, &rsrcfork); hfs_systemfile_unlock(hfsmp, lockflags); if (error) { return (error); } + /* * Supply hfs_getnewvnode with a component name. */ @@ -4530,7 +5661,7 @@ restart: dvp = vnode_getparent(vp); error = hfs_getnewvnode(hfsmp, dvp, cn.cn_pnbuf ? &cn : NULL, descptr, GNV_WANTRSRC | GNV_SKIPLOCK, &cp->c_attr, - &rsrcfork, &rvp); + &rsrcfork, &rvp, &newvnode_flags); if (dvp) vnode_put(dvp); if (cn.cn_pnbuf) @@ -4546,7 +5677,7 @@ restart: /* * Wrapper for special device reads */ -static int +int hfsspec_read(ap) struct vnop_read_args /* { struct vnode *a_vp; @@ -4565,7 +5696,7 @@ hfsspec_read(ap) /* * Wrapper for special device writes */ -static int +int hfsspec_write(ap) struct vnop_write_args /* { struct vnode *a_vp; @@ -4587,7 +5718,7 @@ hfsspec_write(ap) * * Update the times on the cnode then do device close. */ -static int +int hfsspec_close(ap) struct vnop_close_args /* { struct vnode *a_vp; @@ -4680,7 +5811,7 @@ hfsfifo_close(ap) /* * Synchronize a file's in-core state with that on disk. */ -static int +int hfs_vnop_fsync(ap) struct vnop_fsync_args /* { struct vnode *a_vp; @@ -4691,6 +5822,21 @@ hfs_vnop_fsync(ap) struct vnode* vp = ap->a_vp; int error; + /* Note: We check hfs flags instead of vfs mount flag because during + * read-write update, hfs marks itself read-write much earlier than + * the vfs, and hence won't result in skipping of certain writes like + * zero'ing out of unused nodes, creation of hotfiles btree, etc. + */ + if (VTOHFS(vp)->hfs_flags & HFS_READ_ONLY) { + return 0; + } + +#if CONFIG_PROTECT + if ((error = cp_handle_vnop(vp, CP_WRITE_ACCESS, 0)) != 0) { + return (error); + } +#endif /* CONFIG_PROTECT */ + /* * We need to allow ENOENT lock errors since unlink * systenm call can call VNOP_FSYNC during vclean. @@ -4706,7 +5852,7 @@ hfs_vnop_fsync(ap) } -static int +int hfs_vnop_whiteout(ap) struct vnop_whiteout_args /* { struct vnode *a_dvp; @@ -4817,7 +5963,11 @@ struct vnodeopv_entry_desc hfs_standard_vnodeop_entries[] = { { &vnop_pathconf_desc, (VOPFUNC)hfs_vnop_pathconf }, /* pathconf */ { &vnop_advlock_desc, (VOPFUNC)err_advlock }, /* advlock */ { &vnop_allocate_desc, (VOPFUNC)hfs_readonly_op }, /* allocate (READONLY) */ +#if CONFIG_SEARCHFS { &vnop_searchfs_desc, (VOPFUNC)hfs_vnop_search }, /* search fs */ +#else + { &vnop_searchfs_desc, (VOPFUNC)err_searchfs }, /* search fs */ +#endif { &vnop_bwrite_desc, (VOPFUNC)hfs_readonly_op }, /* bwrite (READONLY) */ { &vnop_pagein_desc, (VOPFUNC)hfs_vnop_pagein }, /* pagein */ { &vnop_pageout_desc,(VOPFUNC) hfs_readonly_op }, /* pageout (READONLY) */ @@ -4858,7 +6008,7 @@ struct vnodeopv_entry_desc hfs_vnodeop_entries[] = { { &vnop_select_desc, (VOPFUNC)hfs_vnop_select }, /* select */ { &vnop_revoke_desc, (VOPFUNC)nop_revoke }, /* revoke */ { &vnop_exchange_desc, (VOPFUNC)hfs_vnop_exchange }, /* exchange */ - { &vnop_mmap_desc, (VOPFUNC)err_mmap }, /* mmap */ + { &vnop_mmap_desc, (VOPFUNC)hfs_vnop_mmap }, /* mmap */ { &vnop_fsync_desc, (VOPFUNC)hfs_vnop_fsync }, /* fsync */ { &vnop_remove_desc, (VOPFUNC)hfs_vnop_remove }, /* remove */ { &vnop_link_desc, (VOPFUNC)hfs_vnop_link }, /* link */ @@ -4875,7 +6025,11 @@ struct vnodeopv_entry_desc hfs_vnodeop_entries[] = { { &vnop_pathconf_desc, (VOPFUNC)hfs_vnop_pathconf }, /* pathconf */ { &vnop_advlock_desc, (VOPFUNC)err_advlock }, /* advlock */ { &vnop_allocate_desc, (VOPFUNC)hfs_vnop_allocate }, /* allocate */ +#if CONFIG_SEARCHFS { &vnop_searchfs_desc, (VOPFUNC)hfs_vnop_search }, /* search fs */ +#else + { &vnop_searchfs_desc, (VOPFUNC)err_searchfs }, /* search fs */ +#endif { &vnop_bwrite_desc, (VOPFUNC)hfs_vnop_bwrite }, /* bwrite */ { &vnop_pagein_desc, (VOPFUNC)hfs_vnop_pagein }, /* pagein */ { &vnop_pageout_desc,(VOPFUNC) hfs_vnop_pageout }, /* pageout */