It would be nice if the RCS file format (which is implemented by a great many tools, both free and non-free, both by calling GNU RCS and by reimplementing access to RCS files) were documented in some standard separate from any one tool. But as far as I know no such standard exists. Hence this file. The place to start is the rcsfile.5 manpage in the GNU RCS 5.7 distribution. Then look at the diff at the end of this file (which contains a few fixes and clarifications to that manpage). If you are interested in MKS RCS, src/ci.c in GNU RCS 5.7 has a comment about their date format. However, as far as we know there isn't really any document describing MKS's changes to the RCS file format. The rcsfile.5 manpage does not document what goes in the "text" field for each revision. The answer is that the head revision contains the contents of that revision and every other revision contain a bunch of edits to produce that revision ("a" and "d" lines). The GNU diff manual (the version I looked at was for GNU diff 2.4) documents this format somewhat (as the "RCS output format"), but the presentation is a bit confusing as it is all tangled up with the documentation of several other output formats. If you just want some source code to look at, the part of CVS which applies these is RCS_deltas in src/rcs.c. The rcsfile.5 documentation only _very_ briefly touches on the order of the revisions. The order _is_ important and CVS relies on it. Here is an example of what I was able to find, based on the join3 sanity.sh testcase (and the behavior I am documenting here seems to be the same for RCS 5.7 and CVS 1.9.27): 1.1 -----------------> 1.2 \---> 1.1.2.1 \---> 1.2.2.1 Here is how this shows up in the RCS file (omitting irrelevant parts): admin: head 1.2; deltas: 1.2 branches 1.2.2.1; next 1.1; 1.1 branches 1.1.2.1; next; 1.1.2.1 branches; next; 1.2.2.1 branches; next; deltatexts: 1.2 1.2.2.1 1.1 1.1.2.1 Yes, the order seems to differ between the deltas and the deltatexts. I have no idea how much of this should actually be considered part of the RCS file format, and how much programs reading it should expect to encounter any order. The rcsfile.5 grammar shows the {num} after "next" as optional; if it is omitted then there is no next delta node (for example 1.1 or the head of a branch will typically have no next). There is one case where CVS uses CVS-specific, non-compatible changes to the RCS file format, and this is magic branches. See cvs.texinfo for more information on them. CVS also sets the RCS state to "dead" to indicate that a file does not exist in a given revision (this is stored just as any other RCS state is). The RCS file format allows quite a variety of extensions to be added in a compatible manner by use of the "newphrase" feature documented in rcsfile.5. We won't try to document extensions not used by CVS in any detail, but we will briefly list them. Each occurrence of a newphrase begins with an identifier, which is what we list here. Future designers of extensions are strongly encouraged to pick non-conflicting identifiers. Note that newphrase occurs several places in the RCS grammar, and a given extension may not be legal in all locations. However, it seems better to reserve a particular identifier for all locations, to avoid confusion and complicated rules. Identifier Used by ---------- ------- namespace RCS library done at Silicon Graphics Inc. (SGI) in 1996 (a modified RCS 5.7--not sure it has any other name). dead A set of RCS patches developed by Rich Pixley at Cygnus about 1992. These were for CVS, and predated the current CVS death support, which uses a state "dead" rather than a "dead" newphrase. CVS does use newphrases to implement the `PreservePermissions' extension introduced in CVS 1.9.26. The following new keywords are defined when PreservePermissions=yes: owner group permissions special symlink hardlinks The contents of the `owner' and `group' field should be a numeric uid and a numeric gid, respectively, representing the user and group who own the file. The `permissions' field contains an octal integer, representing the permissions that should be applied to the file. The `special' field contains two words; the first must be either `block' or `character', and the second is the file's device number. The `symlink' field should be present only in files which are symbolic links to other files, and absent on all regular files. The `hardlinks' field contains a list of filenames to which the current file is linked, in alphabetical order. Because files often contain characters special to RCS, like `.' and sometimes even contain spaces or eight-bit characters, the filenames in the hardlinks field will usually be enclosed in RCS strings. For example: hardlinks README @install.txt@ @Installation Notes@; The hardlinks field should always include the name of the current file. That is, in the repository file README,v, any hardlinks fields in the delta nodes should include `README'; CVS will not operate properly if this is not done. Newphrases are also used to implement the 'commitid' feature. The following new keyword is defined: commitid The rules regarding keyword expansion are not documented along with the rest of the RCS file format; they are documented in the co(1) manpage in the RCS 5.7 distribution. See also the "Keyword substitution" chapter of cvs.texinfo. The co(1) manpage refers to special behavior if the log prefix for the $Log keyword is /* or (*. RCS 5.7 produces a warning whenever it behaves that way, and current versions of CVS do not handle this case in a special way (CVS 1.9 and earlier invoke RCS to perform keyword expansion). Note that if the "expand" keyword is omitted from the RCS file, the default is "kv". Note that the "comment {string};" syntax from rcsfile.5 specifies a comment leader, which affects expansion of the $Log keyword for old versions of RCS. The comment leader is not used by RCS 5.7 or current versions of CVS. Both RCS 5.7 and current versions of CVS handle the $Log keyword in a different way if the log message starts with "checked in with -k by ". I don't think this behavior is documented anywhere. Here is a clarification regarding characters versus bytes in certain character sets like JIS and Big5: The RCS file format, as described in the rcsfile(5) man page, is actually byte-oriented, not character-oriented, despite hints to the contrary in the man page. This distinction is important for multibyte characters. For example, if a multibyte character contains a `@' byte, the `@' must be doubled within strings in RCS files, since RCS uses `@' bytes as escapes. This point is not an issue for encodings like ISO 8859, which do not have multibyte characters. Nor is it an issue for encodings like UTF-8 and EUC-JIS, which never uses ASCII bytes within a multibyte character. It is an issue only for multibyte encodings like JIS and BIG5, which _do_ usurp ASCII bytes. If `@' doubling occurs within a multibyte char, the resulting RCS file is not a properly encoded text file. Instead, it is a byte stream that does not use a consistent character encoding that can be understood by the usual text tools, since doubling `@' messes up the encoding. This point affects only programs that examine the RCS files -- it doesn't affect the external RCS interface, as the RCS commands always give you the properly encoded text files and logs (assuming that you always check in properly encoded text). CVS 1.10 (and earlier) probably has some bugs in this area on systems where a C "char" is signed and where the data contains bytes with the eighth bit set. One common concern about the RCS file format is the fact that to get the head of a branch, one must apply deltas from the head of the trunk to the branchpoint, and then from the branchpoint to the head of the branch. While more detailed analyses might be worth doing, we will note: * The performance bottleneck for CVS generally is figuring out which files to operate on and that sort of thing, not applying deltas. * Here is one quick test (probably not a very good test; a better test would use a normally sized file (say 50-200K) instead of a small one): I just did a quick test with a small file (on a Sun Ultra 1/170E running Solaris 5.5.1), with 1000 revisions on the main branch and 1000 revisions on branch that forked at the root (i.e., RCS revisions 1.1, 1.2, ..., 1.1000, and branch revisions 1.1.1.1, 1.1.1.2, ..., 1.1.1.1000). It took about 0.15 seconds real time to check in the first revision, and about 0.6 seconds to check in and 0.3 seconds to retrieve revision 1.1.1.1000 (the worst case). * Any attempt to "fix" this problem should be careful not to interfere with other features, such as lightweight creation of branches (particularly using CVS magic branches). Diff follows: (Note that in the following diff the old value for the Id keyword was: Id: rcsfile.5in,v 5.6 1995/06/05 08:28:35 eggert Exp and the new one was: Id: rcsfile.5in,v 5.7 1996/12/09 17:31:44 eggert Exp but since this file itself might be subject to keyword expansion I haven't included a diff for that fact). =================================================================== RCS file: RCS/rcsfile.5in,v retrieving revision 5.6 retrieving revision 5.7 diff -u -r5.6 -r5.7 --- rcsfile.5in 1995/06/05 08:28:35 5.6 +++ rcsfile.5in 1996/12/09 17:31:44 5.7 @@ -85,7 +85,8 @@ .LP \f2sym\fP ::= {\f2digit\fP}* \f2idchar\fP {\f2idchar\fP | \f2digit\fP}* .LP -\f2idchar\fP ::= any visible graphic character except \f2special\fP +\f2idchar\fP ::= any visible graphic character, + except \f2digit\fP or \f2special\fP .LP \f2special\fP ::= \f3$\fP | \f3,\fP | \f3.\fP | \f3:\fP | \f3;\fP | \f3@\fP .LP @@ -119,12 +120,23 @@ the minute (00\-59), and .I ss -the second (00\-60). +the second (00\-59). +If .I Y -contains just the last two digits of the year -for years from 1900 through 1999, -and all the digits of years thereafter. -Dates use the Gregorian calendar; times use UTC. +contains exactly two digits, +they are the last two digits of a year from 1900 through 1999; +otherwise, +.I Y +contains all the digits of the year. +Dates use the Gregorian calendar. +Times use UTC, except that for portability's sake leap seconds are not allowed; +implementations that support leap seconds should output +.B 59 +for +.I ss +during an inserted leap second, and should accept +.B 59 +for a deleted leap second. .PP The .I newphrase @@ -144,16 +156,23 @@ field in order of decreasing numbers. The .B head -field in the -.I admin -node points to the head of that sequence (i.e., contains +field points to the head of that sequence (i.e., contains the highest pair). The .B branch -node in the admin node indicates the default +field indicates the default branch (or revision) for most \*r operations. If empty, the default branch is the highest branch on the trunk. +The +.B symbols +field associates symbolic names with revisions. +For example, if the file contains +.B "symbols rr:1.1;" +then +.B rr +is a name for revision +.BR 1.1 . .PP All .I delta