From: Ramsay Jones Date: Thu, 31 Jul 2014 11:37:02 GMT Subject: Re: [PATCH v4 0/5] git_config callers rewritten with the new config cache API Message-ID: <53DA2A5E.4020101@ramsay1.demon.co.uk> In-Reply-To: On 30/07/14 17:45, Matthieu Moy wrote: > Tanay Abhra writes: >> Also, any thoughts on what to do with git_default_config()? We can, >> >> 1> make a new function git_load_default_config(), use it for the rewrites. > > That seems the most sensible option. It could be called it git.c before > the command-dependant part, so that any call to git loads this. > > I didn't check if it was correct (e.g. do some command rely on the fact > that the default config is not loaded?) > Hmm, here be dragons ... :-P I don't know that there has actually been any kind of policy regarding the reading of config files/variables in git (or if there is a different policy for plumbing vs porcelain), but it has always seemed to be somewhat ad-hoc; each command decides for itself what it wants to read. However, with increased use of common code which _uses_ certain config variables for correct operation, the 'choice' is much harder to make (and may change after the fact!). For example, about a year ago I submitted a couple of patches which added a call to git_config(git_default_config, NULL) to both 'git pack-refs' and 'git show-refs'. This was as a result of the 'mh/ref-races' branch which introduced a 'stat_validity' api for checking if the packed-refs file had changed on the filesystem since last you looked. This re-used some of the same code used to handle index updates that used config variables like core.checkstat and core.trustctime. These config variables can affect the correctness and/or the efficiency of the code on some platforms (e.g. cygwin, mingw). [Note: 'stat_validity' has since been re-used again (why not?) in some shallow clone code, so similar comments may apply ... I haven't looked.] However, those patches were dropped, because it resulted in an (unwanted) change in behaviour. In particular, 'git show-refs' changed behaviour because it now 'listened' to core.abbrev! I started to look at splitting the 'core config variables' into two groups; essential variables that _all_ git commands should read for correct/efficient/consistent behaviour and everything else (mainly UI related variables). However, something else came up ... Just an FYI. ATB, Ramsay Jones