Re: [PATCH v4 1/6] revision: add mark_tree_uninteresting_sparse

From: Junio C Hamano <gitster@pobox.com>
Date: Jan 11, 2019, 20:25 UTC
Message-ID: <xmqqva2vx7p2.fsf@gitster-ct.c.googlers.com>
In-Reply-To: <xmqqzhs7x9nm.fsf@gitster-ct.c.googlers.com>

Junio C Hamano <gitster@pobox.com> writes:

Show 9 quoted lines

>> In preparation for a new algorithm that walks fewer trees when
>> creating a pack from a set of revisions, create a method that
>> takes an oidset of tree oids and marks reachable objects as
>> UNINTERESTING.
>> ...
>> There is one new assumption in this approach: we are also given
>> the oids of the interesting trees. This implementation does not
>> use those trees at the moment, but we will use them in a later
>> rewrite of this method.

Ahh....

Show 15 quoted lines

> The proposed log message claims that the method takes an oidset and
> marks reachable objects as uninteresting, but the implementation
> only marks those that are reachable from already uninteresting
> trees.  Either one of them must be wrong.
>
> Did you mean to have this instead?
>
> 		if (!tree)
> 			continue;
> 		/*
> 		 * Force traversal of the tree, even if it has been
> 		 * already marked as UNINTERESTING.
> 		 */
> 		tree->object.flags &= ~UNINTERESTING;
> 		mark_tree_uninteresting(r, tree);

So, I assumed that the implementation was wrong, but it is the other way around. You do mean to pick only already uninteresting trees out of "set" and mark its reachables.

One thing that would make me worried is what help the callers of this function will get (or they may have to devise the way themselves) to avoid having to traverse the same tree number of times. A tree can be made uninteresting after a traversal of another tree that contains it, but the logic in this function

Show 9 quoted lines

> +		if (tree->object.flags & UNINTERESTING) {
> +			/*
> +			 * Remove the flag so the next call
> +			 * is not a no-op. The flag is added
> +			 * in mark_tree_unintersting().
> +			 */
> +			tree->object.flags ^= UNINTERESTING;
> +			mark_tree_uninteresting(r, tree);
> +		}

ignores the fact that it is already UNINTERESTING (in fact, in a sense it is even worse---it cannot be used to make a not-yet UNINTERESTING one into UNINTERESTING), drops the UNINTERESING bit and forces the traversal of that tree. The only thing I see that would work as a saving grace is that mark_tree_uninteresting() itself would honor existing UNINTERESING bit and refuses to recurse into its subtrees, but that means blobs at the top-level of such a tree would be marked UNINTERESING while those in its subtrees can be left not-UNINTERESING, which sounds like inviting a mess.

It does *not* immediately mean this function is misdesigned. It just means that the caller needs to carefully follow whatever calling convention this series will establish in the later patches (which we haven't seen yet at this point).

Show 7 quoted lines

> By the way, one of the bigger reasons why I have to ask, instead of
> making an educated guess, is because "struct oidset *set" parameter
> does not give any useful information with the variable name to the
> readers.  We know it is a set because its type is oidset; readers
> need to know what meaning the 'set' has, what it is used for, why
> the caller wants to (or decides not to) place a tree object in the
> set when it calls it.  None of that can be read from its name.