Diff for "fncacheRepoFormat"

fncache is a new repository layout (or format) for Mercurial that solves the following issues:

• http://www.selenic.com/mercurial/bts/issue839 - Hg local store creates paths too long for Windows

• http://www.selenic.com/mercurial/bts/issue793 - Can't clone repos that use Windows reserved names in paths

Status: Released in Mercurial 1.1 (introduced with [http://hg.intevation.org/mercurial/crew/rev/7946503ec76e 7946503ec76e])

TableOfContents

Usage

With this change, all new repositories on all platforms will be fncache repositories. You don't have to do anything (besides using a version of Mercurial containing this change).

The new layout does not affect the wire (or bundle) protocol(s) in any way. So you can push/pull/clone over the wire to/from any repo being in any layout using any Mercurial version on both ends.

New repositories are for example created by non-hardlink cloning of existing repos or if you clone over the wire (http, ssh).

For example, if you have a (pre 1.1) non-fncache repo and you do a local clone --pull you will end up with an fncache repo. If you do a plain local clone (without --pull) of a non-fncache repo, you will get a non-fncache repo with hardlinks to the existing repo.

In short, use clone --pull to convert repos (in case you want to convert repos to the fncache repo format, which will almost never be needed).

Of course old versions of Mercurial will not be able to read fncache repos. If you try to access an fncache repo with a version of Mercurial prior to 1.1 it will abort with:

abort: requirement 'fncache' not supported!

which tells you that the repo at hand requires knowledge of the fncache repo format in Mercurial.

(BTW, if, for whatever reason, the fnache file in the repo becomes corrupted, you can do a clone --pull to rebuild it. The fncache file contains a list of all revlog files in the repo).

Existing non-fncache repositories, that is, repositories created with Mercurial 1.0 (or older), will remain as they are, as Mercurial will still be able to read and write non-fncache repositories.

There is a hgrc option [format] usestore [1], which enables the 'store' format, which is the default for repos created with Mercurial 0.9.5 or newer.

The "store" format encodes filenames with uppercase chars X as _x (underbar + x). If you disable that, you will have to make sure that the repo is only used on a platform that does not fold case (that is, don't use or copy it to/on Windows).

The fncache repo layout is a descendant of the store format, so if you disable the store format on Mercurial 1.1 you implicitly disable the fncache layout.

The fncache repo format can be disabled with '[format] usefncache = False' in the hgrc or with --config format.usefncache=0 on the command line (see also [2])

• [1] http://www.selenic.com/mercurial/hgrc.5.html#format

• [2] http://selenic.com/pipermail/mercurial/2008-October/021980.html

History and background information

'fncache' didn't really change any single bit in the underlying data structures of Mercurial (revlog, indices, changelog, manifest, etc. are all the same).

We reorganized just the revlog data file names (and directory names) inside the store to work around some nasty file name limitations of Windows (long path names, names reserved by Windows).

This introduced a non-reversible encoding for long path names in the store (hybrid hashing of path names), which triggered the introduction of an additional file called 'fncache', which simply contains a redundant list of the so-called "direncoded" path names of all files in the repo, to allow for enumerating the data files at a reasonable speed (currently only used when doing a 'clone --uncompressed').

The 'fncache' file is entirely redundant information and built from scratch (i.e. all manifest revisions) if you do a 'clone --pull' (which is implied if you do a clone over the wire). As a side note, the wire protocol of course didn't change a single bit, so you can combine Mercurial versions on each end of the wire as you want.

Since this was a backwards-incompatible change in the way the files in the store are named, we introduced an new format specifier ('fncache') in the 'requires' file, thus telling old versions of Mercurial that it should keep its fingers off from new 'fncache' repositories (since we know those old versions of Mercurial won't know how to find the files in the store).

But the "format" is not really new, since the contents of the files in the store did not change. We just changed where we store the bits - not how we store them.

Technical details

The fncache repo layout uses a new encoding for path names inside the store.

1. Encoding of Windows reserved names

Path elements consisting of Windows reserved names are now encoded using ~xx where xx is the two digit ASCII hex code of the third character of that reserved name. For example "aux" is encoded as "au~78".

Windows reserved names are: 'con', 'prn', 'aux', 'nul', 'com1'..'com9' and 'lpt1'..'lpt9'.

For example the path

data/aux.bla/bla.aux/prn/PRN/lpt/com3/nul/coma/foo.NUL/normal.c.i

is encoded as

data/au~78.bla/bla.aux/pr~6e/_p_r_n/lpt/co~6d3/nu~6c/coma/foo._n_u_l/normal.c.i

Note that 'aux.bla' needs to be encoded, but not 'bla.aux'.

2. Hashing of long paths

Paths inside the store that would be longer than 120 chars are now hash encoded.

For the encoding used see the function mercurial.store.hybridencode.

Some encoding examples for paths that are hashed (A1→B1, A2→B2, ...):

(A1) data/AUX/SECOND/X.PRN/FOURTH/FI:FTH/SIXTH/SEVENTH/EIGHTH/NINETH/TENTH/ELEVENTH/LOREMIPSUM.TXT.i
(B1) dh/au~78/second/x.prn/fourth/fi~3afth/sixth/seventh/eighth/nineth/tenth/loremia20419e358ddff1bf8751e38288aff1d7c32ec05.i

(A2) data/enterprise/openesbaddons/contrib-imola/corba-bc/netbeansplugin/wsdlExtension/src/main/java/META-INF/services/org.netbeans.modules.xml.wsdl.bindingsupport.spi.ExtensibilityElementTemplateProvider.i
(B2) dh/enterpri/openesba/contrib-/corba-bc/netbeans/wsdlexte/src/main/java/org.net7018f27961fdf338a598a40c4683429e7ffb9743.i

(A3) data/AUX.THE-QUICK-BROWN-FOX-JU:MPS-OVER-THE-LAZY-DOG-THE-QUICK-BROWN-FOX-JUMPS-OVER-THE-LAZY-DOG.TXT.i
(B3) dh/au~78.the-quick-brown-fox-ju~3amps-over-the-lazy-dog-the-quick-brown-fox-jud4dcadd033000ab2b26eb66bae1906bcb15d4a70.i

All paths that are hashed are stored in the directory 'dh' inside '.hg/store'. Non-hashed paths are stored inside '.hg/store/data'.

The hashing used is the sha1 digest (40 characters) of the direncoded path below '.hg/store', as pre-encoded by mercurial.filelog.encodedir.

For the hashencoded path, the first eight characters of the first n directory levels are taken (converted to lowercase), where n is adapted slightly to use more levels if space allows (see store.hybridencode). If space allows, the filename before the hash value is filled up with to lowercase converted chars from the filename of the input path.

As you can see, the path encoding done may fold multiple files originating from different input path directories into the same encoded path directory. The sha1 digest part of the filename ensures that the filenames are distinct and no name clashes occur.

3. The fncache file

For the fncache repository format Mercurial maintains a new file 'fncache' (thus the name of the format) inside '.hg/store'. The fncache file contains the paths of all filelog files in the store as encoded by mercurial.filelog.encodedir. The paths are separated by '\n' (LF).

The fncache file is used to enumerate all filelog files in the store, for example when doing a clone --uncompressed. The fncache file may contain duplicates or inexistent entries (this can happen when using the strip or rollback commmands).

During a clone --uncompressed or a hg verify the fncache file is read and rewritten if duplicates or entries with missing filelog files are detected, so even operations that don't modify the history of the repository may lead to modifying the fncache file (this was a deliberate design decision as discussed with mpm).

The fncache file is not read by a hg clone --pull, so that command may be used to resurrect a damaged fncache file, since hg clone --pull rewrites the fncache file from the information found in all manifest revisions. That's also the reason why it is basically cached information.

The verify command checks the fncache file and removes non-existent or duplicate entries. If a filelog file referenced in a manifest revision is not found in the fncache file, hg verify reports an error.

4. New entry 'fncache' in the requires file

Mercurial writes a file named 'requires' in the .hg directory when creating a new repository. For an fncache repository, the requires file contains:

revlogv1
store
fncache

In a pre-fncache repository, the entry 'fncache' in the requires file is missing.

If Mercurial reads a repository, it first reads the requires file to see what is required to read the repository at hand.

An old (pre 1.1) Mercurial will abort when it finds 'fncache' in the requires files (it requires a Mercurial version that knows what to do when 'fncache' is required, that is, at least 1.1).

Mercurial 1.1 or newer of course are also able to read and write "old" non-fncache repositories. A non-fncache repository is identified by the missing 'fncache' string in the requires file.

See also

• [:FileFormats]