FIles transferred over SHell protocol (V 0.0.2) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ This protocol was designed for transferring files over a remote shell connection (rsh and compatibles). It can be as well used for transfers over rsh, and there may be other uses. Client sends requests of following form: #FISH_COMMAND equivalent shell commands, which may be multiline Only fish commands are defined here, shell equivalents are for your information only and will probably vary from implementation to implementation. Fish commands always have priority: server is expected to execute fish command if it understands it. If it does not, however, it can try the luck and execute shell command. Server's reply is multiline, but always ends with ### 000 line. ### is prefix to mark this line, 000 is return code. Return codes are superset to those used in ftp. There are few new exit codes defined: 000 don't know; if there were no previous lines, this marks COMPLETE success, if they were, it marks failure. 001 don't know; if there were no previous lines, this marks PRELIMinary success, if they were, it marks failure Connecting ~~~~~~~~~~ Client uses "echo FISH:;/bin/sh" as command executed on remote machine. This should make it possible for server to distinguish FISH connections from normal rsh/ssh. Commands ~~~~~~~~ #FISH echo; start_fish_server; echo '### 200' This command is sent at the beginning. It marks that client wishes to talk via FISH protocol. #VER command must follow. If server understands FISH protocol, it has option to put FISH server somewhere on system path and name it start_fish_server. #VER 0.0.2 <...> echo '### 000' This command is the second one. It sends client version and extensions to the server. Server should reply with protocol version to be used, and list of extensions accepted. VER 0.0.0 ### 200 #PWD pwd; echo '### 200' Server should reply with current directory (in form /abc/def/ghi) followed by line indicating success. #LIST /directory ls -lLa $1 | grep '^[^cbt]' | ( while read p x u g s m d y n; do echo "P$p $u.$g S$s d$m $d $y :$n "; done ) ls -lLa $1 | grep '^[cb]' | ( while read p x u g a i m d y n; do echo "P$p $u.$g E$a$i dD$m $d $y :$n "; done ) echo '### 200' This allows client to list directory or get status information about single file. Output is in following form (any line except : may be omitted): P . S d<3-letters month name> D [.1234] E, : L Unix permissions are of form X--------- where X is type of file. Currently, '-' means regular file, 'd' means directory, 'c', 'b' means character and block device, 'l' means symbolic link, 'p' means FIFO and 's' means socket. 'd' has three fields: month (one of strings Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec), day of month, and third is either single number indicating year, or HH:MM field (assume current year in such case). As you've probably noticed, this is pretty broken; it is for compatibility with ls listing. #RETR /some/name ls -l /some/name | ( read a b c d x e; echo $x ); echo '### 100'; cat /some/name; echo '### 200' Server sends line with filesize on it, followed by line with ### 100 indicating partial success, then it sends binary data (exactly filesize bytes) and follows them with (with no preceding newline) ### 200. Note that there's no way to abort running RETR command - except closing the connection. #STOR /file/name > /file/name; echo '### 001'; ( dd bs=4096 count=; dd bs= count=1 ) 2>/dev/null | ( cat > %s; cat > /dev/null ); echo '### 200' This command is for storing /file/name, which is exactly size bytes big. You probably think I went crazy. Well, I did not: that strange cat > /dev/null has purpose to discard any extra data which was not written to disk (due to for example out of space condition). [Why? Imagine uploading file with "rm -rf /" line in it.] #CWD /somewhere cd /somewhere; echo '### 000' It is specified here, but I'm not sure how wise idea is to use this one: it breaks stateless-ness of the protocol. Following commands should be rather self-explanatory: #CHMOD 1234 file chmod 1234 file; echo '### 000' #DELE /some/path rm -f /some/path; echo '### 000' #MKD /some/path mkdir /some/path; echo '### 000' #RMD /some/path rmdir /some/path; echo '### 000' #RENAME /path/a /path/b mv /path/a /path/b; echo '### 000' #LINK /path/a /path/b ln /path/a /path/b; echo '### 000' #SYMLINK /path/a /path/b ln -s /path/a /path/b; echo '### 000' #CHOWN user /file/name chown user /file/name; echo '### 000' #CHGRP group /file/name chgrp group /file/name; echo '### 000' #READ /path/and/filename cat /path/and/filename | ( dd bs=4096 count= > /dev/null; dd bs= count=1 > /dev/null; dd bs=4096 count=; dd bs= count=1; ) Returns ### 200 on successful exit, ### 291 on successful exit when reading ended at eof, ### 292 on successfull exit when reading did not end at eof. #WRITE /path/and/filename Hmm, shall we define these ones if we know our client is not going to use them? That's all, folks! pavel@ucw.cz