Also, don't attempt xattr ops on special files on the BSD likes.
TODO: Should have a way to allow restore without special files,
otherwise very cumbersome for a regular user.
Special files are device nodes and named pipes. The necessity of the
former is clear, the latter is debatable.
In order to preserve backward compatibility, the device number is
encoded in the StartChunk/StartOffset fields of the entry.
This tracks inode/device from the stat info and creates backward
compatible snapshots that allow preserving hardlinks. Backwards
compatibility is preserved by saving a virtual inode number index in the
Link field of the file entry. Since this field was previously only used
for symlinks, this won't break old versions. Additionally, the entry
data is cloned so restoration with an old version works.
Current limitations are primarility with restore. They include:
- no command line option to prevent hard link restore
- if a file has the immutable or append only flag it will be set before
hardlinks are restored, so hardlinking will fail.
- if a partial restore includes a hardlink but not the parent
directories the hardlink will fail.
These will be solved by grouping restore of hardlinks together
with file, prior to applying final metadata.
- if a file is changed and is being rewritten by a restore hardlinks are
not preserved.
Basic support for BSD and Darwin style chflags (stat flags). Applies
these flags at the end of file restore.
Supports linux style ioctl_iflags(2) in a 2 step process. Flags that
need to be applied prior to writes such as compress and especially no-COW
are applied immediately upon file open.
The flags format is backwards compatible. An attribute starting with a
null byte is used to store flags in the entry attributes table. With
an old version of duplicacy the restore of this attribute should silently
fail (effectively be ignored).
Fixes xattr restore to use O_NOFOLLOW so attributes are applied to symlink.
TODO: Tests, possible option to switch off mutable/append prior to
restore of existing file similar to rsync. Does not apply attributes
or flags to the top most directory.
Main changes:
* Change the listing order of files/directories so that the local and remote
snapshots can be compared on-the-fly.
* Introduce a new struct called EntryList that maintains a list of
files/directories, which are kept in memory when the number is lower, and
serialized into a file when there are too many.
* EntryList can also be turned into an on-disk incomplete snapshot quickly,
to support fast-resume on next run.
* ChunkOperator can now download and upload chunks, thus replacing original
ChunkDownloader and ChunkUploader. The new ChunkDownloader is only used
to prefetch chunks during the restore operation.
This fix isn't probably necessary since filepath.Join can now produce UNC
paths too with the latest versions of go. However, we still want to keep
it for consistency.
Directories containing a file with this name will not be backed up. I find it easier to drop a .nobackup file in directories I don't want backed up instead of maintaining a file of exclusions. This is also useful for scripts that create data in the repository but don't want it to be backed up.
