Internals/NamedPatch

Named patch

We call “named patch” what users would directly call “patch” in darcs. A named patch is what we create when we run record in a repository. The history of a repository is made of named patches.

A named patch is a collection of primitive patches, recorded together with the following attributes:

name
a single line of text
date
the timestamp taken at record-time. See Date for more information.
author
an arbitrary text, usually in the format “John De Bar <john@debar.fam>
log
a longer description of the patch, possibly multiline. When parsing a patch, the first space character is removed from each line of the log. When computing a patch’s hash, trailing newlines are removed from each log line before concatenating.
patch salt
the hexadecimal representation of some random number from 0 to 2^128. It’s random noise inserted into the log to help patch hashes stay unique. (see issue27, and the function Darcs.Patch.Info.addJunk). This is actually stored inside the log, but not shown in most of the user interface (you may see it preceded by Ignore-this:).

All these attributes are specified or determined when the patch is recorded or amended and are meant to be read-only after then.

There is one other attribute on which the user has no impact:

inverted
a boolean flag, that indicates if the patch is a result of a rollback (NB. obsolete since DarcsTwo uses a different way of doing rollbacks, but the flag is still there for backward compatibility). The inverted state is indicated by a ‘*’ or ‘-’ character immediately preceding the patch’s date.

Hash computation

The unique hash id of any patch is computed by the following Haskell code in src/Darcs/Patch/Info.hs:

makePatchname :: PatchInfo -> SHA1
makePatchname pi = sha1PS sha1_me
        where b2ps True = BC.pack "t"
              b2ps False = BC.pack "f"
              sha1_me = B.concat [_piName pi,
                                  _piAuthor pi,
                                  _piDate pi,
                                  B.concat $ _piLog pi,
                                  b2ps $ isInverted pi]

In other words, the hash is the SHA1 code of the tuple (name, author, date, log, inverted).

As you can notice, the hunks (that is the PrimitivePatch collection, the actual changes the NamedPatch is carrying) aren’t included in the formula. This is a strong assumption in darcs: there cannot be different patches with the exact same set of attributes, and in turn with the same hash.

On-disk patch name

Named patches are stored in the meta directory _darcs/patches.

A HashedFormat repository stores its patches in filenames composed by two parts:

  1. a ten digits zero-padded number, the uncompressed size of the whole patch
  2. the SHA256 64-digits signature of the whole patch

This has several benefits:

  1. darcs is able to assert that the mean of a patch wasn’t mangled in any way, because the filename is basically a checksum of the whole patch, hunks included
  2. it allows a better way of sharing patches between repositories, see CacheSystem

Since, as darcs commutes the patch with the others in the repository, the patch itself (its hunks, I mean) may change at any time:

  1. there’s no one-to-one relation between a particular hash name and the name of the file containing the patch
  2. there can be multiple instances of the same patch, collected under the _darcs/patches metadir: of course, at any given time, at most one of them are effectively listed in the RepositoryInventory

Note also that even if the patch is compressed on disk, its filename does not end with .gz.

See also