[PATCH 1/1] scan-tree: handle error in git_config_from_file()

[John Keeping]

On Tue, Mar 08, 2016 at 04:01:59PM +0100, Christian Hesse wrote:
> John Keeping <john at keeping.me.uk> on Tue, 2016/03/08 14:59:
> > On Tue, Mar 08, 2016 at 03:54:22PM +0100, Christian Hesse wrote:
> > > John Keeping <john at keeping.me.uk> on Tue, 2016/03/08 14:36:  
> > > > On Tue, Mar 08, 2016 at 03:26:23PM +0100, Christian Hesse wrote:  
> > > > > John Keeping <john at keeping.me.uk> on Tue, 2016/03/08 14:11:    
> > > > > > On Tue, Mar 08, 2016 at 02:51:46PM +0100, Christian Hesse wrote:    
> > > > > > > From: Christian Hesse <mail at eworm.de>
> > > > > > > 
> > > > > > > Signed-off-by: Christian Hesse <mail at eworm.de>      
> > > > > > 
> > > > > > Is this solving a particular problem or did you just notice that the
> > > > > > return value is ignored?
> > > > > > 
> > > > > > I don't think returning when this fails is correct because we've
> > > > > > already added the repository to the list by this point and a lot of
> > > > > > the remaining code in this function will do something sensible even
> > > > > > if git_config_from_file() fails.
> > > > > > 
> > > > > > In fact, git_config_from_file() sets the die_on_error flag for
> > > > > > do_config_from() so the only case that gives us an error here is if
> > > > > > the config file cannot be opened.  I don't think it's unreasonable
> > > > > > to print an error if that happens but bailing out of the function
> > > > > > at this point is wrong.    
> > > > > 
> > > > > Ok, probably you are right...
> > > > > 
> > > > > Actually I do have a particular problem, but it is not solved by
> > > > > this patch. :-p Just stumbled and thought it is a good idea.
> > > > > 
> > > > > I have a repository that has a config with bad permissions, so http
> > > > > server's user can not read it. cgit does not print http headers and
> > > > > http server bails out with error 500. What path does it take?    
> > > > 
> > > > Hmm... bad permissions should result in fopen(2) failing, which would
> > > > take the path altered in your patch.
> > > > 
> > > > Can you run cgit as the http server's user from a terminal like this:
> > > > 
> > > > 	CGIT_CONFIG=/path/to/cgitrc QUERY_STRING=url=/ cgit
> > > > 
> > > > ?  It might produce an error message that your http server isn't
> > > > logging.  
> > > 
> > > It gives:
> > > 
> > > Error reading config /path/to/repository.git/config: Permission denied
> > > (13) fatal: unable to access '/path/to/repository.git/config': Permission
> > > denied
> > > 
> > > Looks like this is fatal in git...  
> > 
> > Is that when listing repositories or when trying to display that
> > specific repo?  I suspect that's a result of prepare_repo_cmd() setting
> > up for a specific repository.
> 
> Listing repositories works. This is when displaying the repository with
> bad/denied config file.

I don't think we can work around that without adding a load of extra
checks before calling setup_git_directory_gently(), which would allow us
to at least print a more obvious error message.  The problem is that
git_config_early() dies if the file cannot be accessed and there is no
way for us to read from a repository without going through that
function.