Re: Adding glob support to remotes

Andy Parkins [off-list ref] writes:

I started to add code to git-parse-remote.sh:canon_refs_list_for_fetch() to 
preprocess the reflist to catch lines with a "*" in them then use the remote 
pattern to filter the output of from "git-ls-remote -h", blah, blah, you get 
the idea...

However, git-ls-remote needs the name of the remote repository (of course), 
but that isn't directly available in git-parse-remote.sh.  

Is it really the case?  I do not remember the details offhand,
but I do not think canon_refs_list_for_fetch is the function you
should be messing with to implement the remote."origin".fetch
stuff.  It should be get_remote_default_refs_for_fetch().  The
function returns the list based on which remote, so it surely
knows which remote the caller is talking about.

However, I would recommend against actually running ls-remote to
help "git-fetch" inside git-parse-remote.sh.  I think you should
run ls-remote upfront early in git-fetch because there are at
least two other parts in git-fetch that wants the same ls-remote
output:

 (1) dumb protocols currently cannot deal with a remote that has
     run "packed-ref --prune" because git-fetch.sh first uses
     curl executable to download the loose ref, read it and then
     use the object name read from that to drive git-http-fetch.
     We can and should get rid of the /max_depth=5/,/done/ loop
     there and replace it with a grep of ls-remote output to
     make them work against such a remote.  When tracking many
     branches from the remote, this would reduce the number of
     http requests (one per branch vs a ls-remote which is just
     a single download of info/refs).

 (2) when doing a fetch with tracking branches (which is what
     your change is about), we would need to run ls-remote to
     find out the remote tags for tag following purposes anyway.
     Running "ls-remote -h" once for your purpose and then
     "ls-remote -t" for tag following later is obviously very
     wasteful.

On Wednesday 2006 November 22 12:56, Junio C Hamano wrote:

quoted

However, git-ls-remote needs the name of the remote repository (of
course), but that isn't directly available in git-parse-remote.sh.

Is it really the case?  I do not remember the details offhand,
but I do not think canon_refs_list_for_fetch is the function you
should be messing with to implement the remote."origin".fetch
stuff.  It should be get_remote_default_refs_for_fetch().  The
function returns the list based on which remote, so it surely
knows which remote the caller is talking about.

The problem is that canon_refs_list_for_fetch bombs out too early because "*" 
is not an acceptable name for a ref.

However, I would recommend against actually running ls-remote to
help "git-fetch" inside git-parse-remote.sh.  I think you should
run ls-remote upfront early in git-fetch because there are at
least two other parts in git-fetch that wants the same ls-remote
output:

Okay.  That's what I'll do.  It means altering git-check-ref-format to prevent 
the early bomb out.  Perhaps I should move this check to somewhere after I've 
done the reflist expansion?

 (1) dumb protocols currently cannot deal with a remote that has

I'm not sure I've understood this point.  I shall look at git-fetch.sh more 
closely to try and address this though.

 (2) when doing a fetch with tracking branches (which is what

Accepted.



Andy

-- 
Dr Andy Parkins, M Eng (hons), MIEE