Re: [PATCH v9 1/2] add `config_set` API for caching config-like files

[Tanay Abhra <tanayabh@gmail.com>]

On 7/15/2014 9:16 PM, Junio C Hamano wrote:
> Tanay Abhra <tanayabh@gmail.com> writes:
> 
>> diff --git a/config.c b/config.c
>> index ba882a1..89e2d67 100644
>> --- a/config.c
>> +++ b/config.c
>> ...
>> +const struct string_list *git_configset_get_value_multi(struct config_set *cs, const char *key)
>> +{
>> +	struct config_set_element *e = configset_find_element(cs, key);
>> +	return e ? &e->value_list : NULL;
>> +}
> 
> Somehow I find the placement of this function out of place.  Didn't
> you get the same impression while proofreading the entire file after
> you are done writing this patch?
> 
> The right place for it would probably be immediately before
> git_configset_get_value(), I would think.
>

Noted. My fault, will correct in the reroll.

>> +int git_configset_add_file(struct config_set *cs, const char *filename)
>> +{
>> +	int ret = 0;
>> +	ret = git_config_from_file(config_hash_callback, filename, cs);
>> +	if (!ret)
>> +		return 0;
>> +	else {
>> +		git_configset_clear(cs);
>> +		return -1;
>> +	}
>> +}
> 
> If I add contents from file "A" successfully and then attempt to
> further add contents from file "B" which fails, I would lose
> contents I read from "A"?
> 
> It would not be worth trying to make it transactional (i.e. making
> sure cs contains what was there before the config-from-file was
> called if it failed), so in such a case we may end up seeing a
> mixture of full contents from A and partial contents from B, which
> is just as useless as a cleared configset, so it is not wrong to
> clear it per-se.  An alternative might be to return without
> clearing, as we are leaving the configset in a useless state either
> way and give the caller a choice between clearing it and continue,
> and dying without even spending unnecessary cycles to clear.  That
> might be more consistent with what git_config_check_init() does,
> where you die without bothering to clear what was half-read into the
> configset.
> 
> Whatever behaviour is chosen in the error codepath, it needs to be
> documented.
>

Since syntactical errors in reading the file will cause a die when we
use either `git_config_from_file` or `git_config`, I don't think
incomplete parsing would trigger the error path, as it will die before
reaching there.

So, the best way would be just to return without clearing and let the
user take the decision if he wants to go on with the incomplete
config_set or clear it when he sees that configset_add_file() failed.

Die in git_config_check_init() is never triggered expect in rare race
conditions like between access_or_die() and git_config_from_file() in

	if (git_config_system() && !access_or_die(git_etc_gitconfig(), R_OK, 0)) {
		ret += git_config_from_file(fn, git_etc_gitconfig(),
					    data);
		found += 1;
	}
which I think would never happen in reality, dunno. I think I will take
out the die() in git_config_check_init(). Thus, the behavior of both
functions git_config_check_init() & git_configset_add_file will be consistent.

Thanks.