Re: [PATCH v2 1/3] refs: keep track of unresolved reference value in iterator

[John Cai <johncai86@gmail.com>]

Hi Junio

On 8 Apr 2024, at 19:02, Junio C Hamano wrote:

>> diff --git a/refs.h b/refs.h
>> index 298caf6c618..2e740c692ac 100644
>> --- a/refs.h
>> +++ b/refs.h
>> @@ -71,9 +71,10 @@ struct pack_refs_opts {
>>  	struct ref_exclusions *exclusions;
>>  	struct string_list *includes;
>>  };
>> -
>>  const char *refs_resolve_ref_unsafe(struct ref_store *refs,
>> +
>>  				    const char *refname,
>> +				    char **referent,
>>  				    int resolve_flags,
>>  				    struct object_id *oid,
>>  				    int *flags);
>
> If referent is meant to be an out-parameter, it should sit next to
> oid that is also an out-parameter.  And as a late-comer sibling, it
> should sit after its elder brother.
>
> Also, I do not see the reason for the shuffling of blank lines.
> Shouldn't it be the other way around, even?  After an unrelated
> definition of "struct pack_refs_opts", there is (and should be)
> a blank line, then the first line of the declaration of function.
>
> Perhaps some fat-fingering.

This was indeed a case of fat-fingering.

>
>> @@ -1928,6 +1928,7 @@ int refs_read_symbolic_ref(struct ref_store *ref_store, const char *refname,
>>
>>  const char *refs_resolve_ref_unsafe(struct ref_store *refs,
>>  				    const char *refname,
>> +				    char **referent,
>>  				    int resolve_flags,
>>  				    struct object_id *oid,
>>  				    int *flags)
>> @@ -1989,6 +1990,8 @@ const char *refs_resolve_ref_unsafe(struct ref_store *refs,
>>  		}
>>
>>  		*flags |= read_flags;
>> +		if (referent && read_flags & REF_ISSYMREF && sb_refname.len > 0)
>> +			*referent = sb_refname.buf;
>
> Is this safe?  After this assignment, which "return" in this loop
> are you expecting to return from this function?  If you fail to
> return from the function during this iteration, you'll clobber the
> same strbuf with the next refs_read_raw_ref(), but I do not see how
> you are ensuring that you'll return from the function without such
> corruption happening.
>
> This assignment happens only when read_flags has REF_ISSYMREF set,
> so the next "if it is not, then return refname" would not even
> trigger.  If RESOLVE_REF_NO_RECURSE bit is on in resolve_flags,
> then we'd return without further dereferencing, but if that is the
> only safe exit from the function, shouldn't you be guarding the
> function with something like
>
> 	if (referent && !(resolve_flags & RESOLVE_REF_NO_RECURSE))
> 		BUG("recursive dereference can will clobber *referent");
>
> to protect its callers from their own mistakes?
>
> Another return before we start the next iteration of the loop and
> clobber sb_refname with another call to refs_read_raw_ref() is the
> error return codepath at the end of the loop, but that is a totally
> uninteresting case.
>
> Or am I totally confused?

Yeah good point. This probably not a good idea. In fact, perhaps adding another
argument to this function is unnecessary. We already have refs_read_symbolic_ref
and we can just make a separate call to that once we know that a ref is a
symbolic reference. Though a separate call is less efficient, I'm not sure it's
worth adding this parameter.

thanks
John