2. The Repository

The CVS repository stores a complete copy of all the files and directories which are under version control.

Normally, you never access any of the files in the repository directly. Instead, you use CVS commands to get your own copy of the files into a working directory, and then work on that copy. When you've finished a set of changes, you check (or commit) them back into the repository. The repository then contains the changes which you have made, as well as recording exactly what you changed, when you changed it, and other such information. Note that the repository is not a subdirectory of the working directory, or vice versa; they should be in separate locations.

CVS can access a repository by a variety of means. It might be on the local computer, or it might be on a computer across the room or across the world. To distinguish various ways to access a repository, the repository name can start with an access method. For example, the access method :local: means to access a repository directory, so the repository :local:/usr/local/cvsroot means that the repository is in `/usr/local/cvsroot' on the computer running CVS. For information on other access methods, see 2.9 Remote repositories.

If the access method is omitted, then if the repository does not contain `:', then :local: is assumed. If it does contain `:' then either :ext: or :server: is assumed. For example, if you have a local repository in `/usr/local/cvsroot', you can use /usr/local/cvsroot instead of :local:/usr/local/cvsroot. But if (under Windows NT, for example) your local repository is `c:\src\cvsroot', then you must specify the access method, as in :local:c:\src\cvsroot.

The repository is split in two parts. `$CVSROOT/CVSROOT' contains administrative files for CVS. The other directories contain the actual user-defined modules.

2.1 Telling CVS where your repository is

There are several ways to tell CVS where to find the repository. You can name the repository on the command line explicitly, with the -d (for "directory") option:

cvs -d /usr/local/cvsroot checkout yoyodyne/tc

Or you can set the $CVSROOT environment variable to an absolute path to the root of the repository, `/usr/local/cvsroot' in this example. To set $CVSROOT, csh and tcsh users should have this line in their `.cshrc' or `.tcshrc' files:

setenv CVSROOT /usr/local/cvsroot

sh and bash users should instead have these lines in their `.profile' or `.bashrc':

CVSROOT=/usr/local/cvsroot
export CVSROOT

A repository specified with -d will override the $CVSROOT environment variable. Once you've checked a working copy out from the repository, it will remember where its repository is (the information is recorded in the `CVS/Root' file in the working copy).

The -d option and the `CVS/Root' file both override the $CVSROOT environment variable. If -d option differs from `CVS/Root', the former is used. Of course, for proper operation they should be two ways of referring to the same repository.

2.2 How data is stored in the repository

For most purposes it isn't important how CVS stores information in the repository. In fact, the format has changed in the past, and is likely to change in the future. Since in almost all cases one accesses the repository via CVS commands, such changes need not be disruptive.

However, in some cases it may be necessary to understand how CVS stores data in the repository, for example you might need to track down CVS locks (see section 10.5 Several developers simultaneously attempting to run CVS) or you might need to deal with the file permissions appropriate for the repository.

2.2.1 Where files are stored within the repository

The overall structure of the repository is a directory tree corresponding to the directories in the working directory. For example, supposing the repository is in

/usr/local/cvsroot

here is a possible directory tree (showing only the directories):

/usr
 |
 +--local
 |   |
 |   +--cvsroot
 |   |    |
 |   |    +--CVSROOT
          |      (administrative files)
          |
          +--gnu
          |   |
          |   +--diff
          |   |   (source code to GNU diff)
          |   |
          |   +--rcs
          |   |   (source code to RCS)
          |   |
          |   +--cvs
          |       (source code to CVS)
          |
          +--yoyodyne
              |
              +--tc
              |    |
              |    +--man
              |    |
              |    +--testing
              |
              +--(other Yoyodyne software)

With the directories are history files for each file under version control. The name of the history file is the name of the corresponding file with `,v' appended to the end. Here is what the repository for the `yoyodyne/tc' directory might look like:

  $CVSROOT
    |
    +--yoyodyne
    |   |
    |   +--tc
    |   |   |
            +--Makefile,v
            +--backend.c,v
            +--driver.c,v
            +--frontend.c,v
            +--parser.c,v
            +--man
            |    |
            |    +--tc.1,v
            |
            +--testing
                 |
                 +--testpgm.t,v
                 +--test2.t,v

The history files contain, among other things, enough information to recreate any revision of the file, a log of all commit messages and the user-name of the person who committed the revision. The history files are known as RCS files, because the first program to store files in that format was a version control system known as RCS. For a full description of the file format, see the man page rcsfile(5), distributed with RCS, or the file `doc/RCSFILES' in the CVS source distribution. This file format has become very common--many systems other than CVS or RCS can at least import history files in this format.

The RCS files used in CVS differ in a few ways from the standard format. The biggest difference is magic branches; for more information see 5.5 Magic branch numbers. Also in CVS the valid tag names are a subset of what RCS accepts; for CVS's rules see 4.4 Tags--Symbolic revisions.

2.2.2 File permissions

This means that you can only control access to files on a per-directory basis.

Note that users must also have write access to check out files, because CVS needs to create lock files (see section 10.5 Several developers simultaneously attempting to run CVS).

Also note that users must have write access to the `CVSROOT/val-tags' file. CVS uses it to keep track of what tags are valid tag names (it is sometimes updated when tags are used, as well as when they are created).

Each RCS file will be owned by the user who last checked it in. This has little significance; what really matters is who owns the directories.

CVS tries to set up reasonable file permissions for new directories that are added inside the tree, but you must fix the permissions manually when a new directory should have different permissions than its parent directory. If you set the CVSUMASK environment variable that will control the file permissions which CVS uses in creating directories and/or files in the repository. CVSUMASK does not affect the file permissions in the working directory; such files have the permissions which are typical for newly created files, except that sometimes CVS creates them read-only (see the sections on watches, 10.6.1 Telling CVS to watch certain files; -r, A.4 Global options; or CVSREAD, D. All environment variables which affect CVS).

Note that using the client/server CVS (see section 2.9 Remote repositories), there is no good way to set CVSUMASK; the setting on the client machine has no effect. If you are connecting with rsh, you can set CVSUMASK in `.bashrc' or `.cshrc', as described in the documentation for your operating system. This behavior might change in future versions of CVS; do not rely on the setting of CVSUMASK on the client having no effect.

Using pserver, you will generally need stricter permissions on the CVSROOT directory and directories above it in the tree; see 2.9.3.3 Security considerations with password authentication.

Some operating systems have features which allow a particular program to run with the ability to perform operations which the caller of the program could not. For example, the set user ID (setuid) or set group ID (setgid) features of unix or the installed image feature of VMS. CVS was not written to use such features and therefore attempting to install CVS in this fashion will provide protection against only accidental lapses; anyone who is trying to circumvent the measure will be able to do so, and depending on how you have set it up may gain access to more than just CVS. You may wish to instead consider pserver. It shares some of the same attributes, in terms of possibly providing a false sense of security or opening security holes wider than the ones you are trying to fix, so read the documentation on pserver security carefully if you are considering this option (2.9.3.3 Security considerations with password authentication).

2.2.3 File Permission issues specific to Windows

Some file permission issues are specific to Windows operating systems (Windows 95, Windows NT, and presumably future operating systems in this family. Some of the following might apply to OS/2 but I'm not sure).

If you are using local CVS and the repository is on a networked file system which is served by the Samba SMB server, some people have reported problems with permissions. Enabling WRITE=YES in the samba configuration is said to fix/workaround it. Disclaimer: I haven't investigated enough to know the implications of enabling that option, nor do I know whether there is something which CVS could be doing differently in order to avoid the problem. If you find something out, please let us know as described in H. Dealing with bugs in CVS or this manual.

2.2.4 The attic

You will notice that sometimes CVS stores an RCS file in the Attic. For example, if the CVSROOT is `/usr/local/cvsroot' and we are talking about the file `backend.c' in the directory `yoyodyne/tc', then the file normally would be in

/usr/local/cvsroot/yoyodyne/tc/backend.c,v

but if it goes in the attic, it would be in

/usr/local/cvsroot/yoyodyne/tc/Attic/backend.c,v

instead. It should not matter from a user point of view whether a file is in the attic; CVS keeps track of this and looks in the attic when it needs to. But in case you want to know, the rule is that the RCS file is stored in the attic if and only if the head revision on the trunk has state dead. A dead state means that file has been removed, or never added, for that revision. For example, if you add a file on a branch, it will have a trunk revision in dead state, and a branch revision in a non-dead state.

2.2.5 The CVS directory in the repository

The `CVS' directory in each repository directory contains information such as file attributes (in a file called `CVS/fileattr'. In the future additional files may be added to this directory, so implementations should silently ignore additional files.

This behavior is implemented only by CVS 1.7 and later; for details see 10.6.5 Using watches with old versions of CVS.

The format of the fileattr file is a series of entries of the following form (where `{' and `}' means the text between the braces can be repeated zero or more times):

ent-type filename <tab> attrname = attrval {; attrname = attrval} <linefeed>

ent-type is `F' for a file, in which case the entry specifies the attributes for that file.

ent-type is `D', and filename empty, to specify default attributes to be used for newly added files.

Other ent-type are reserved for future expansion. CVS 1.9 and older will delete them any time it writes file attributes. CVS 1.10 and later will preserve them.

Note that the order of the lines is not significant; a program writing the fileattr file may rearrange them at its convenience.

There is currently no way of quoting tabs or linefeeds in the filename, `=' in attrname, `;' in attrval, etc. Note: some implementations also don't handle a NUL character in any of the fields, but implementations are encouraged to allow it.

By convention, attrname starting with `_' is for an attribute given special meaning by CVS; other attrnames are for user-defined attributes (or will be, once implementations start supporting user-defined attributes).

Builtin attributes:

_watched

Present means the file is watched and should be checked out read-only.

_watchers

Users with watches for this file. Value is watcher > type { , watcher > type } where watcher is a username, and type is zero or more of edit,unedit,commit separated by `+' (that is, nothing if none; there is no "none" or "all" keyword).

_editors

Users editing this file. Value is editor > val { , editor > val } where editor is a username, and val is time+hostname+pathname, where time is when the cvs edit command (or equivalent) happened, and hostname and pathname are for the working directory.

Example:

Ffile1 _watched=;_watchers=joe>edit,mary>commit
Ffile2 _watched=;_editors=sue>8 Jan 1975+workstn1+/home/sue/cvs
D _watched=

means that the file `file1' should be checked out read-only. Furthermore, joe is watching for edits and mary is watching for commits. The file `file2' should be checked out read-only; sue started editing it on 8 Jan 1975 in the directory `/home/sue/cvs' on the machine workstn1. Future files which are added should be checked out read-only. To represent this example here, we have shown a space after `D', `Ffile1', and `Ffile2', but in fact there must be a single tab character there and no spaces.

2.2.6 CVS locks in the repository

For an introduction to CVS locks focusing on user-visible behavior, see 10.5 Several developers simultaneously attempting to run CVS. The following section is aimed at people who are writing tools which want to access a CVS repository without interfering with other tools acessing the same repository. If you find yourself confused by concepts described here, like read lock, write lock, and deadlock, you might consult the literature on operating systems or databases.

Any file in the repository with a name starting with `#cvs.rfl.' is a read lock. Any file in the repository with a name starting with `#cvs.wfl' is a write lock. Old versions of CVS (before CVS 1.5) also created files with names starting with `#cvs.tfl', but they are not discussed here. The directory `#cvs.lock' serves as a master lock. That is, one must obtain this lock first before creating any of the other locks.

To obtain a readlock, first create the `#cvs.lock' directory. This operation must be atomic (which should be true for creating a directory under most operating systems). If it fails because the directory already existed, wait for a while and try again. After obtaining the `#cvs.lock' lock, create a file whose name is `#cvs.rfl.' followed by information of your choice (for example, hostname and process identification number). Then remove the `#cvs.lock' directory to release the master lock. Then proceed with reading the repository. When you are done, remove the `#cvs.rfl' file to release the read lock.

To obtain a writelock, first create the `#cvs.lock' directory, as with a readlock. Then check that there are no files whose names start with `#cvs.rfl.'. If there are, remove `#cvs.lock', wait for a while, and try again. If there are no readers, then create a file whose name is `#cvs.wfl' followed by information of your choice (for example, hostname and process identification number). Hang on to the `#cvs.lock' lock. Proceed with writing the repository. When you are done, first remove the `#cvs.wfl' file and then the `#cvs.lock' directory. Note that unlike the `#cvs.rfl' file, the `#cvs.wfl' file is just informational; it has no effect on the locking operation beyond what is provided by holding on to the `#cvs.lock' lock itself.

Note that each lock (writelock or readlock) only locks a single directory in the repository, including `Attic' and `CVS' but not including subdirectories which represent other directories under version control. To lock an entire tree, you need to lock each directory (note that if you fail to obtain any lock you need, you must release the whole tree before waiting and trying again, to avoid deadlocks).

Note also that CVS expects writelocks to control access to individual `foo,v' files. RCS has a scheme where the `,foo,' file serves as a lock, but CVS does not implement it and so taking out a CVS writelock is recommended. See the comments at rcs_internal_lockfile in the CVS source code for further discussion/rationale.

2.2.7 How files are stored in the CVSROOT directory

The `$CVSROOT/CVSROOT' directory contains the various administrative files. In some ways this directory is just like any other directory in the repository; it contains RCS files whose names end in `,v', and many of the CVS commands operate on it the same way. However, there are a few differences.

For each administrative file, in addition to the RCS file, there is also a checked out copy of the file. For example, there is an RCS file `loginfo,v' and a file `loginfo' which contains the latest revision contained in `loginfo,v'. When you check in an administrative file, CVS should print

cvs commit: Rebuilding administrative file database

and update the checked out copy in `$CVSROOT/CVSROOT'. If it does not, there is something wrong (see section H. Dealing with bugs in CVS or this manual). To add your own files to the files to be updated in this fashion, you can add them to the `checkoutlist' administrative file (see section C.10 The checkoutlist file).

By default, the `modules' file behaves as described above. If the modules file is very large, storing it as a flat text file may make looking up modules slow (I'm not sure whether this is as much of a concern now as when CVS first evolved this feature; I haven't seen benchmarks). Therefore, by making appropriate edits to the CVS source code one can store the modules file in a database which implements the ndbm interface, such as Berkeley db or GDBM. If this option is in use, then the modules database will be stored in the files `modules.db', `modules.pag', and/or `modules.dir'.

For information on the meaning of the various administrative files, see C. Reference manual for Administrative files.

2.3 How data is stored in the working directory

While we are discussing CVS internals which may become visible from time to time, we might as well talk about what CVS puts in the `CVS' directories in the working directories. As with the repository, CVS handles this information and one can usually access it via CVS commands. But in some cases it may be useful to look at it, and other programs, such as the jCVS graphical user interface or the VC package for emacs, may need to look at it. Such programs should follow the recommendations in this section if they hope to be able to work with other programs which use those files, including future versions of the programs just mentioned and the command-line CVS client.

The `CVS' directory contains several files. Programs which are reading this directory should silently ignore files which are in the directory but which are not documented here, to allow for future expansion.

The files are stored according to the text file convention for the system in question. This means that working directories are not portable between systems with differing conventions for storing text files. This is intentional, on the theory that the files being managed by CVS probably will not be portable between such systems either.

`Root'

This file contains the current CVS root, as described in 2.1 Telling CVS where your repository is.

`Repository'

This file contains the directory within the repository which the current directory corresponds with. It can be either an absolute pathname or a relative pathname; CVS has had the ability to read either format since at least version 1.3 or so. The relative pathname is relative to the root, and is the more sensible approach, but the absolute pathname is quite common and implementations should accept either. For example, after the command

cvs -d :local:/usr/local/cvsroot checkout yoyodyne/tc

`Root' will contain

:local:/usr/local/cvsroot

and `Repository' will contain either

/usr/local/cvsroot/yoyodyne/tc

or

yoyodyne/tc

If the particular working directory does not correspond to a directory in the repository, then `Repository' should contain `CVSROOT/Emptydir'.

`Entries'

This file lists the files and directories in the working directory. The first character of each line indicates what sort of line it is. If the character is unrecognized, programs reading the file should silently skip that line, to allow for future expansion.

If the first character is `/', then the format is:

/name/revision/timestamp[+conflict]/options/tagdate

where `[' and `]' are not part of the entry, but instead indicate that the `+' and conflict marker are optional. name is the name of the file within the directory. revision is the revision that the file in the working derives from, or `0' for an added file, or `-' followed by a revision for a removed file. timestamp is the timestamp of the file at the time that CVS created it; if the timestamp differs with the actual modification time of the file it means the file has been modified. It is stored in the format used by the ISO C asctime() function (for example, `Sun Apr 7 01:29:26 1996'). One may write a string which is not in that format, for example, `Result of merge', to indicate that the file should always be considered to be modified. This is not a special case; to see whether a file is modified a program should take the timestamp of the file and simply do a string compare with timestamp. If there was a conflict, conflict can be set to the modification time of the file after the file has been written with conflict markers (see section 10.3 Conflicts example). Thus if conflict is subsequently the same as the actual modification time of the file it means that the user has obviously not resolved the conflict. options contains sticky options (for example `-kb' for a binary file). tagdate contains `T' followed by a tag name, or `D' for a date, followed by a sticky tag or date. Note that if timestamp contains a pair of timestamps separated by a space, rather than a single timestamp, you are dealing with a version of CVS earlier than CVS 1.5 (not documented here).

The timezone on the timestamp in CVS/Entries (local or universal) should be the same as the operating system stores for the timestamp of the file itself. For example, on Unix the file's timestamp is in universal time (UT), so the timestamp in CVS/Entries should be too. On VMS, the file's timestamp is in local time, so CVS on VMS should use local time. This rule is so that files do not appear to be modified merely because the timezone changed (for example, to or from summer time).

If the first character of a line in `Entries' is `D', then it indicates a subdirectory. `D' on a line all by itself indicates that the program which wrote the `Entries' file does record subdirectories (therefore, if there is such a line and no other lines beginning with `D', one knows there are no subdirectories). Otherwise, the line looks like:

D/name/filler1/filler2/filler3/filler4

where name is the name of the subdirectory, and all the filler fields should be silently ignored, for future expansion. Programs which modify Entries files should preserve these fields.

The lines in the `Entries' file can be in any order.

`Entries.Log'

This file does not record any information beyond that in `Entries', but it does provide a way to update the information without having to rewrite the entire `Entries' file, including the ability to preserve the information even if the program writing `Entries' and `Entries.Log' abruptly aborts. Programs which are reading the `Entries' file should also check for `Entries.Log'. If the latter exists, they should read `Entries' and then apply the changes mentioned in `Entries.Log'. After applying the changes, the recommended practice is to rewrite `Entries' and then delete `Entries.Log'. The format of a line in `Entries.Log' is a single character command followed by a space followed by a line in the format specified for a line in `Entries'. The single character command is `A' to indicate that the entry is being added, `R' to indicate that the entry is being removed, or any other character to indicate that the entire line in `Entries.Log' should be silently ignored (for future expansion). If the second character of the line in `Entries.Log' is not a space, then it was written by an older version of CVS (not documented here).

Programs which are writing rather than reading can safely ignore `Entries.Log' if they so choose.

`Entries.Backup'

This is a temporary file. Recommended usage is to write a new entries file to `Entries.Backup', and then to rename it (atomically, where possible) to `Entries'.

`Entries.Static'

The only relevant thing about this file is whether it exists or not. If it exists, then it means that only part of a directory was gotten and CVS will not create additional files in that directory. To clear it, use the update command with the `-d' option, which will get the additional files and remove `Entries.Static'.

`Tag'

This file contains per-directory sticky tags or dates. The first character is `T' for a branch tag, `N' for a non-branch tag, or `D' for a date, or another character to mean the file should be silently ignored, for future expansion. This character is followed by the tag or date. Note that per-directory sticky tags or dates are used for things like applying to files which are newly added; they might not be the same as the sticky tags or dates on individual files. For general information on sticky tags and dates, see 4.9 Sticky tags.

`Checkin.prog'

`Update.prog'

These files store the programs specified by the `-i' and `-u' options in the modules file, respectively.

`Notify'

This file stores notifications (for example, for edit or unedit) which have not yet been sent to the server. Its format is not yet documented here.

`Notify.tmp'

This file is to `Notify' as `Entries.Backup' is to `Entries'. That is, to write `Notify', first write the new contents to `Notify.tmp' and then (atomically where possible), rename it to `Notify'.

`Base'

If watches are in use, then an edit command stores the original copy of the file in the `Base' directory. This allows the unedit command to operate even if it is unable to communicate with the server.

`Baserev'

The file lists the revision for each of the files in the `Base' directory. The format is:

Bname/rev/expansion

where expansion should be ignored, to allow for future expansion.

`Baserev.tmp'

This file is to `Baserev' as `Entries.Backup' is to `Entries'. That is, to write `Baserev', first write the new contents to `Baserev.tmp' and then (atomically where possible), rename it to `Baserev'.

`Template'

This file contains the template specified by the `rcsinfo' file (see section C.8 Rcsinfo). It is only used by the client; the non-client/server CVS consults `rcsinfo' directly.

2.4 The administrative files

The directory `$CVSROOT/CVSROOT' contains some administrative files. See section C. Reference manual for Administrative files, for a complete description. You can use CVS without any of these files, but some commands work better when at least the `modules' file is properly set up.

The most important of these files is the `modules' file. It defines all modules in the repository. This is a sample `modules' file.

CVSROOT         CVSROOT
modules         CVSROOT modules
cvs             gnu/cvs
rcs             gnu/rcs
diff            gnu/diff
tc              yoyodyne/tc

The `modules' file is line oriented. In its simplest form each line contains the name of the module, whitespace, and the directory where the module resides. The directory is a path relative to $CVSROOT. The last four lines in the example above are examples of such lines.

The line that defines the module called `modules' uses features that are not explained here. See section C.1 The modules file, for a full explanation of all the available features.

2.4.1 Editing administrative files

You edit the administrative files in the same way that you would edit any other module. Use `cvs checkout CVSROOT' to get a working copy, edit it, and commit your changes in the normal way.

It is possible to commit an erroneous administrative file. You can often fix the error and check in a new revision, but sometimes a particularly bad error in the administrative file makes it impossible to commit new revisions.

2.5 Multiple repositories

In some situations it is a good idea to have more than one repository, for instance if you have two development groups that work on separate projects without sharing any code. All you have to do to have several repositories is to specify the appropriate repository, using the CVSROOT environment variable, the `-d' option to CVS, or (once you have checked out a working directory) by simply allowing CVS to use the repository that was used to check out the working directory (see section 2.1 Telling CVS where your repository is).

The big advantage of having multiple repositories is that they can reside on different servers. With CVS version 1.10, a single command cannot recurse into directories from different repositories. With development versions of CVS, you can check out code from multiple servers into your working directory. CVS will recurse and handle all the details of making connections to as many server machines as necessary to perform the requested command. Here is an example of how to set up a working directory:

cvs -d server1:/cvs co dir1
cd dir1
cvs -d server2:/root co sdir
cvs update

The cvs co commands set up the working directory, and then the cvs update command will contact server2, to update the dir1/sdir subdirectory, and server1, to update everything else.

2.6 Creating a repository

To set up a CVS repository, first choose the machine and disk on which you want to store the revision history of the source files. CPU and memory requirements are modest, so most machines should be adequate. For details see 2.9.1 Server requirements.

To estimate disk space requirements, if you are importing RCS files from another system, the size of those files is the approximate initial size of your repository, or if you are starting without any version history, a rule of thumb is to allow for the server approximately three times the size of the code to be under CVS for the repository (you will eventually outgrow this, but not for a while). On the machines on which the developers will be working, you'll want disk space for approximately one working directory for each developer (either the entire tree or a portion of it, depending on what each developer uses).

The repository should be accessible (directly or via a networked file system) from all machines which want to use CVS in server or local mode; the client machines need not have any access to it other than via the CVS protocol. It is not possible to use CVS to read from a repository which one only has read access to; CVS needs to be able to create lock files (see section 10.5 Several developers simultaneously attempting to run CVS).

To create a repository, run the cvs init command. It will set up an empty repository in the CVS root specified in the usual way (see section 2. The Repository). For example,

cvs -d /usr/local/cvsroot init

cvs init is careful to never overwrite any existing files in the repository, so no harm is done if you run cvs init on an already set-up repository.

cvs init will enable history logging; if you don't want that, remove the history file after running cvs init. See section C.11 The history file.

2.7 Backing up a repository

There is nothing particularly magical about the files in the repository; for the most part it is possible to back them up just like any other files. However, there are a few issues to consider.

The first is that to be paranoid, one should either not use CVS during the backup, or have the backup program lock CVS while doing the backup. To not use CVS, you might forbid logins to machines which can access the repository, turn off your CVS server, or similar mechanisms. The details would depend on your operating system and how you have CVS set up. To lock CVS, you would create `#cvs.rfl' locks in each repository directory. See 10.5 Several developers simultaneously attempting to run CVS, for more on CVS locks. Having said all this, if you just back up without any of these precautions, the results are unlikely to be particularly dire. Restoring from backup, the repository might be in an inconsistent state, but this would not be particularly hard to fix manually.

When you restore a repository from backup, assuming that changes in the repository were made after the time of the backup, working directories which were not affected by the failure may refer to revisions which no longer exist in the repository. Trying to run CVS in such directories will typically produce an error message. One way to get those changes back into the repository is as follows:

• Get a new working directory.

• Copy the files from the working directory from before the failure over to the new working directory (do not copy the contents of the `CVS' directories, of course).

• Working in the new working directory, use commands such as cvs update and cvs diff to figure out what has changed, and then when you are ready, commit the changes into the repository.

2.8 Moving a repository

Just as backing up the files in the repository is pretty much like backing up any other files, if you need to move a repository from one place to another it is also pretty much like just moving any other collection of files.

The main thing to consider is that working directories point to the repository. The simplest way to deal with a moved repository is to just get a fresh working directory after the move. Of course, you'll want to make sure that the old working directory had been checked in before the move, or you figured out some other way to make sure that you don't lose any changes. If you really do want to reuse the existing working directory, it should be possible with manual surgery on the `CVS/Repository' files. You can see 2.3 How data is stored in the working directory, for information on the `CVS/Repository' and `CVS/Root' files, but unless you are sure you want to bother, it probably isn't worth it.

2.9 Remote repositories

Your working copy of the sources can be on a different machine than the repository. Using CVS in this manner is known as client/server operation. You run CVS on a machine which can mount your working directory, known as the client, and tell it to communicate to a machine which can mount the repository, known as the server. Generally, using a remote repository is just like using a local one, except that the format of the repository name is:

:method:[[user][:password]@]hostname[:[port]]/path/to/repository

Specifying a password in the repository name is not recommended during checkout, since this will cause CVS to store a cleartext copy of the password in each created directory. cvs login first instead (see section 2.9.3.2 Using the client with password authentication).

The details of exactly what needs to be set up depend on how you are connecting to the server.

If method is not specified, and the repository name contains `:', then the default is ext or server, depending on your platform; both are described in 2.9.2 Connecting with rsh.

2.9.1 Server requirements

The quick answer to what sort of machine is suitable as a server is that requirements are modest--a server with 32M of memory or even less can handle a fairly large source tree with a fair amount of activity.

The real answer, of course, is more complicated. Estimating the known areas of large memory consumption should be sufficient to estimate memory requirements. There are two such areas documented here; other memory consumption should be small by comparison (if you find that is not the case, let us know, as described in H. Dealing with bugs in CVS or this manual, so we can update this documentation).

The first area of big memory consumption is large checkouts, when using the CVS server. The server consists of two processes for each client that it is serving. Memory consumption on the child process should remain fairly small. Memory consumption on the parent process, particularly if the network connection to the client is slow, can be expected to grow to slightly more than the size of the sources in a single directory, or two megabytes, whichever is larger.

Multiplying the size of each CVS server by the number of servers which you expect to have active at one time should give an idea of memory requirements for the server. For the most part, the memory consumed by the parent process probably can be swap space rather than physical memory.

The second area of large memory consumption is diff, when checking in large files. This is required even for binary files. The rule of thumb is to allow about ten times the size of the largest file you will want to check in, although five times may be adequate. For example, if you want to check in a file which is 10 megabytes, you should have 100 megabytes of memory on the machine doing the checkin (the server machine for client/server, or the machine running CVS for non-client/server). This can be swap space rather than physical memory. Because the memory is only required briefly, there is no particular need to allow memory for more than one such checkin at a time.

Resource consumption for the client is even more modest--any machine with enough capacity to run the operating system in question should have little trouble.

For information on disk space requirements, see 2.6 Creating a repository.

2.9.2 Connecting with rsh

CVS uses the `rsh' protocol to perform these operations, so the remote user host needs to have a `.rhosts' file which grants access to the local user.

For example, suppose you are the user `mozart' on the local machine `toe.example.com', and the server machine is `faun.example.org'. On faun, put the following line into the file `.rhosts' in `bach''s home directory:

toe.example.com  mozart

Then test that rsh is working with

rsh -l bach faun.example.org 'echo $PATH'

Next you have to make sure that rsh will be able to find the server. Make sure that the path which rsh printed in the above example includes the directory containing a program named cvs which is the server. You need to set the path in `.bashrc', `.cshrc', etc., not `.login' or `.profile'. Alternately, you can set the environment variable CVS_SERVER on the client machine to the filename of the server you want to use, for example `/usr/local/bin/cvs-1.6'.

There is no need to edit `inetd.conf' or start a CVS server daemon.

There are two access methods that you use in CVSROOT for rsh. :server: specifies an internal rsh client, which is supported only by some CVS ports. :ext: specifies an external rsh program. By default this is rsh but you may set the CVS_RSH environment variable to invoke another program which can access the remote server (for example, remsh on HP-UX 9 because rsh is something different). It must be a program which can transmit data to and from the server without modifying it; for example the Windows NT rsh is not suitable since it by default translates between CRLF and LF. The OS/2 CVS port has a hack to pass `-b' to rsh to get around this, but since this could potentially cause problems for programs other than the standard rsh, it may change in the future. If you set CVS_RSH to SSH or some other rsh replacement, the instructions in the rest of this section concerning `.rhosts' and so on are likely to be inapplicable; consult the documentation for your rsh replacement.

Continuing our example, supposing you want to access the module `foo' in the repository `/usr/local/cvsroot/', on machine `faun.example.org', you are ready to go:

cvs -d :ext:[email protected]/usr/local/cvsroot checkout foo

(The `bach@' can be omitted if the username is the same on both the local and remote hosts.)

2.9.3 Direct connection with password authentication

The CVS client can also connect to the server using a password protocol. This is particularly useful if using rsh is not feasible (for example, the server is behind a firewall), and Kerberos also is not available.

To use this method, it is necessary to make some adjustments on both the server and client sides.

2.9.3.1 Setting up the server for password authentication

First of all, you probably want to tighten the permissions on the `$CVSROOT' and `$CVSROOT/CVSROOT' directories. See 2.9.3.3 Security considerations with password authentication, for more details.

On the server side, the file `/etc/inetd.conf' needs to be edited so inetd knows to run the command cvs pserver when it receives a connection on the right port. By default, the port number is 2401; it would be different if your client were compiled with CVS_AUTH_PORT defined to something else, though. This can also be sepcified in the CVSROOT variable (see section 2.9 Remote repositories) or overridden with the CVS_CLIENT_PORT environment variable (see section D. All environment variables which affect CVS).

If your inetd allows raw port numbers in `/etc/inetd.conf', then the following (all on a single line in `inetd.conf') should be sufficient:

2401  stream  tcp  nowait  root  /usr/local/bin/cvs
cvs -f --allow-root=/usr/cvsroot pserver

You could also use the `-T' option to specify a temporary directory.

The `--allow-root' option specifies the allowable CVSROOT directory. Clients which attempt to use a different CVSROOT directory will not be allowed to connect. If there is more than one CVSROOT directory which you want to allow, repeat the option. (Unfortunately, many versions of inetd have very small limits on the number of arguments and/or the total length of the command. The usual solution to this problem is to have inetd run a shell script which then invokes CVS with the necessary arguments.)

If your inetd wants a symbolic service name instead of a raw port number, then put this in `/etc/services':

cvspserver      2401/tcp

and put cvspserver instead of 2401 in `inetd.conf'.

Once the above is taken care of, restart your inetd, or do whatever is necessary to force it to reread its initialization files.

If you are having trouble setting this up, see F.2 Trouble making a connection to a CVS server.

Because the client stores and transmits passwords in cleartext (almost--see 2.9.3.3 Security considerations with password authentication, for details), a separate CVS password file is generally used, so people don't compromise their regular passwords when they access the repository. This file is `$CVSROOT/CVSROOT/passwd' (see section 2.4 The administrative files). It uses a colon-separated format, similar to `/etc/passwd' on Unix systems, except that it has fewer fields: CVS username, optional password, and an optional system username for CVS to run as if authentication succeeds. Here is an example `passwd' file with five entries:

anonymous:
bach:ULtgRLXo7NRxs
spwang:1sOp854gDF3DY
melissa:tGX1fS8sun6rY:pubcvs
qproj:XR4EZcEs0szik:pubcvs

(The passwords are encrypted according to the standard Unix crypt() function, so it is possible to paste in passwords directly from regular Unix `/etc/passwd' files.)

The first line in the example will grant access to any CVS client attempting to authenticate as user anonymous, no matter what password they use, including an empty password. (This is typical for sites granting anonymous read-only access; for information on how to do the "read-only" part, see 2.10 Read-only repository access.)

The second and third lines will grant access to bach and spwang if they supply their respective plaintext passwords.

The fourth line will grant access to melissa, if she supplies the correct password, but her CVS operations will actually run on the server side under the system user pubcvs. Thus, there need not be any system user named melissa, but there must be one named pubcvs.

The fifth line shows that system user identities can be shared: any client who successfully authenticates as qproj will actually run as pubcvs, just as melissa does. That way you could create a single, shared system user for each project in your repository, and give each developer their own line in the `$CVSROOT/CVSROOT/passwd' file. The CVS username on each line would be different, but the system username would be the same. The reason to have different CVS usernames is that CVS will log their actions under those names: when melissa commits a change to a project, the checkin is recorded in the project's history under the name melissa, not pubcvs. And the reason to have them share a system username is so that you can arrange permissions in the relevant area of the repository such that only that account has write-permission there.

If the system-user field is present, all password-authenticated CVS commands run as that user; if no system user is specified, CVS simply takes the CVS username as the system username and runs commands as that user. In either case, if there is no such user on the system, then the CVS operation will fail (regardless of whether the client supplied a valid password).

The password and system-user fields can both be omitted (and if the system-user field is omitted, then also omit the colon that would have separated it from the encrypted password). For example, this would be a valid `$CVSROOT/CVSROOT/passwd' file:

anonymous::pubcvs
fish:rKa5jzULzmhOo:kfogel
sussman:1sOp854gDF3DY

When the password field is omitted or empty, then the client's authentication attempt will succeed with any password, including the empty string. However, the colon after the CVS username is always necessary, even if the password is empty.

CVS can also fall back to use system authentication. When authenticating a password, the server first checks for the user in the `$CVSROOT/CVSROOT/passwd' file. If it finds the user, it will use that entry for authentication as described above. But if it does not find the user, or if the CVS `passwd' file does not exist, then the server can try to authenticate the username and password using the operating system's user-lookup routines (this "fallback" behavior can be disabled by setting SystemAuth=no in the CVS `config' file, see section C.13 The CVSROOT/config configuration file). Be aware, however, that falling back to system authentication might be a security risk: CVS operations would then be authenticated with that user's regular login password, and the password flies across the network in plaintext. See 2.9.3.3 Security considerations with password authentication for more on this.

Right now, the only way to put a password in the CVS `passwd' file is to paste it there from somewhere else. Someday, there may be a cvs passwd command.

Unlike many of the files in `$CVSROOT/CVSROOT', it is normal to edit the `passwd' file in-place, rather than via CVS. This is because of the possible security risks of having the `passwd' file checked out to people's working copies. If you do want to include the `passwd' file in checkouts of `$CVSROOT/CVSROOT', see C.10 The checkoutlist file.

2.9.3.2 Using the client with password authentication

cvs -d :pserver:faun.example.org:/usr/local/cvsroot checkout someproj

or

CVSROOT=:pserver:[email protected]:2401/usr/local/cvsroot
cvs checkout someproj

However, unless you're connecting to a public-access repository (i.e., one where that username doesn't require a password), you'll need to supply a password or log in first. Logging in verifies your password with the repository and stores it in a file. It's done with the login command, which will prompt you interactively for the password if you didn't supply one as part of $CVSROOT:

cvs -d :pserver:[email protected]:/usr/local/cvsroot login
CVS password:

or

cvs -d :pserver:bach:[email protected]:/usr/local/cvsroot login

After you enter the password, CVS verifies it with the server. If the verification succeeds, then that combination of username, host, repository, and password is permanently recorded, so future transactions with that repository won't require you to run cvs login. (If verification fails, CVS will exit complaining that the password was incorrect, and nothing will be recorded.)

The records are stored, by default, in the file `$HOME/.cvspass'. That file's format is human-readable, and to a degree human-editable, but note that the passwords are not stored in cleartext--they are trivially encoded to protect them from "innocent" compromise (i.e., inadvertent viewing by a system administrator or other non-malicious person).

You can change the default location of this file by setting the CVS_PASSFILE environment variable. If you use this variable, make sure you set it before cvs login is run. If you were to set it after running cvs login, then later CVS commands would be unable to look up the password for transmission to the server. Once you have logged in, all CVS commands using that remote repository and username will authenticate with the stored password. So, for example

cvs -d :pserver:[email protected]:/usr/local/cvsroot checkout foo

should just work (unless the password changes on the server side, in which case you'll have to re-run cvs login).

Note that if the `:pserver:' were not present in the repository specification, CVS would assume it should use rsh to connect with the server instead (see section 2.9.2 Connecting with rsh).

Of course, once you have a working copy checked out and are running CVS commands from within it, there is no longer any need to specify the repository explicitly, because CVS can deduce the repository from the working copy's `CVS' subdirectory.

The password for a given remote repository can be removed from the CVS_PASSFILE by using the cvs logout command.

2.9.3.3 Security considerations with password authentication

The passwords are stored on the client side in a trivial encoding of the cleartext, and transmitted in the same encoding. The encoding is done only to prevent inadvertent password compromises (i.e., a system administrator accidentally looking at the file), and will not prevent even a naive attacker from gaining the password.

The separate CVS password file (see section 2.9.3.1 Setting up the server for password authentication) allows people to use a different password for repository access than for login access. On the other hand, once a user has non-read-only access to the repository, she can execute programs on the server system through a variety of means. Thus, repository access implies fairly broad system access as well. It might be possible to modify CVS to prevent that, but no one has done so as of this writing.

Note that because the `$CVSROOT/CVSROOT' directory contains `passwd' and other files which are used to check security, you must control the permissions on this directory as tightly as the permissions on `/etc'. The same applies to the `$CVSROOT' directory itself and any directory above it in the tree. Anyone who has write access to such a directory will have the ability to become any user on the system. Note that these permissions are typically tighter than you would use if you are not using pserver.

In summary, anyone who gets the password gets repository access (which may imply some measure of general system access as well). The password is available to anyone who can sniff network packets or read a protected (i.e., user read-only) file. If you want real security, get Kerberos.

2.9.4 Direct connection with GSSAPI

GSSAPI is a generic interface to network security systems such as Kerberos 5. If you have a working GSSAPI library, you can have CVS connect via a direct TCP connection, authenticating with GSSAPI.

To do this, CVS needs to be compiled with GSSAPI support; when configuring CVS it tries to detect whether GSSAPI libraries using kerberos version 5 are present. You can also use the `--with-gssapi' flag to configure.

The connection is authenticated using GSSAPI, but the message stream is not authenticated by default. You must use the -a global option to request stream authentication.

The data transmitted is not encrypted by default. Encryption support must be compiled into both the client and the server; use the `--enable-encrypt' configure option to turn it on. You must then use the -x global option to request encryption.

GSSAPI connections are handled on the server side by the same server which handles the password authentication server; see 2.9.3.1 Setting up the server for password authentication. If you are using a GSSAPI mechanism such as Kerberos which provides for strong authentication, you will probably want to disable the ability to authenticate via cleartext passwords. To do so, create an empty `CVSROOT/passwd' password file, and set SystemAuth=no in the config file (see section C.13 The CVSROOT/config configuration file).

The GSSAPI server uses a principal name of cvs/hostname, where hostname is the canonical name of the server host. You will have to set this up as required by your GSSAPI mechanism.

To connect using GSSAPI, use `:gserver:'. For example,

cvs -d :gserver:faun.example.org:/usr/local/cvsroot checkout foo

2.9.5 Direct connection with kerberos

The easiest way to use kerberos is to use the kerberos rsh, as described in 2.9.2 Connecting with rsh. The main disadvantage of using rsh is that all the data needs to pass through additional programs, so it may be slower. So if you have kerberos installed you can connect via a direct TCP connection, authenticating with kerberos.

This section concerns the kerberos network security system, version 4. Kerberos version 5 is supported via the GSSAPI generic network security interface, as described in the previous section.

To do this, CVS needs to be compiled with kerberos support; when configuring CVS it tries to detect whether kerberos is present or you can use the `--with-krb4' flag to configure.

The data transmitted is not encrypted by default. Encryption support must be compiled into both the client and server; use the `--enable-encryption' configure option to turn it on. You must then use the -x global option to request encryption.

You need to edit `inetd.conf' on the server machine to run cvs kserver. The client uses port 1999 by default; if you want to use another port specify it in the CVSROOT (see section 2.9 Remote repositories) or the CVS_CLIENT_PORT environment variable on the client.

When you want to use CVS, get a ticket in the usual way (generally kinit); it must be a ticket which allows you to log into the server machine. Then you are ready to go:

cvs -d :kserver:faun.example.org:/usr/local/cvsroot checkout foo

Previous versions of CVS would fall back to a connection via rsh; this version will not do so.

2.9.6 Connecting with fork

This access method allows you to connect to a repository on your local disk via the remote protocol. In other words it does pretty much the same thing as :local:, but various quirks, bugs and the like are those of the remote CVS rather than the local CVS.

For day-to-day operations you might prefer either :local: or :fork:, depending on your preferences. Of course :fork: comes in particularly handy in testing or debugging cvs and the remote protocol. Specifically, we avoid all of the network-related setup/configuration, timeouts, and authentication inherent in the other remote access methods but still create a connection which uses the remote protocol.

To connect using the fork method, use `:fork:' and the pathname to your local repository. For example:

cvs -d :fork:/usr/local/cvsroot checkout foo

As with :ext:, the server is called `cvs' by default, or the value of the CVS_SERVER environment variable.

2.10 Read-only repository access

It is possible to grant read-only repository access to people using the password-authenticated server (see section 2.9.3 Direct connection with password authentication). (The other access methods do not have explicit support for read-only users because those methods all assume login access to the repository machine anyway, and therefore the user can do whatever local file permissions allow her to do.)

A user who has read-only access can do only those CVS operations which do not modify the repository, except for certain "administrative" files (such as lock files and the history file). It may be desirable to use this feature in conjunction with user-aliasing (see section 2.9.3.1 Setting up the server for password authentication).

Unlike with previous versions of CVS, read-only users should be able merely to read the repository, and not to execute programs on the server or otherwise gain unexpected levels of access. Or to be more accurate, the known holes have been plugged. Because this feature is new and has not received a comprehensive security audit, you should use whatever level of caution seems warranted given your attitude concerning security.

There are two ways to specify read-only access for a user: by inclusion, and by exclusion.

"Inclusion" means listing that user specifically in the `$CVSROOT/CVSROOT/readers' file, which is simply a newline-separated list of users. Here is a sample `readers' file:

melissa
splotnik
jrandom

(Don't forget the newline after the last user.)

"Exclusion" means explicitly listing everyone who has write access--if the file

$CVSROOT/CVSROOT/writers

exists, then only those users listed in it have write access, and everyone else has read-only access (of course, even the read-only users still need to be listed in the CVS `passwd' file). The `writers' file has the same format as the `readers' file.

Note: if your CVS `passwd' file maps cvs users onto system users (see section 2.9.3.1 Setting up the server for password authentication), make sure you deny or grant read-only access using the cvs usernames, not the system usernames. That is, the `readers' and `writers' files contain cvs usernames, which may or may not be the same as system usernames.

Here is a complete description of the server's behavior in deciding whether to grant read-only or read-write access:

If `readers' exists, and this user is listed in it, then she gets read-only access. Or if `writers' exists, and this user is NOT listed in it, then she also gets read-only access (this is true even if `readers' exists but she is not listed there). Otherwise, she gets full read-write access.

Of course there is a conflict if the user is listed in both files. This is resolved in the more conservative way, it being better to protect the repository too much than too little: such a user gets read-only access.

2.11 Temporary directories for the server

While running, the CVS server creates temporary directories. They are named

cvs-servpid

where pid is the process identification number of the server. They are located in the directory specified by the TMPDIR environment variable (see section D. All environment variables which affect CVS), the `-T' global option (see section A.4 Global options), or failing that `/tmp'.

In most cases the server will remove the temporary directory when it is done, whether it finishes normally or abnormally. However, there are a few cases in which the server does not or cannot remove the temporary directory, for example:

• If the server aborts due to an internal server error, it may preserve the directory to aid in debugging

• If the server is killed in a way that it has no way of cleaning up (most notably, `kill -KILL' on unix).

• If the system shuts down without an orderly shutdown, which tells the server to clean up.

In cases such as this, you will need to manually remove the `cvs-servpid' directories. As long as there is no server running with process identification number pid, it is safe to do so.