The bugs page, at least before I commit this, has a bug at the top that has been modified to link to done, and ikiwiki's dependency calculations failed to notice and update the bugs page. Looking at the indexdb, I saw that the page was not included in the depends_simple of the bugs page.

I was able to replicate the problem locally by starting off with the page marked done (when it did appear in the bugs page depends_simple (appropriatly as a link dependency, since a change to the page removing the link would make it match)), then removing the done link.

At that point, it vanished from depends_simple. Presumably because the main (pagespec) depends for the bugs page now matched it, as a content dependency. But, it seems to me it should still be listed in depends_simple here. This, I think, is the cause of the bug.

Then re-add the done link, and the dependency calc code breaks down, not noticing that bugs dependeded on the page and needs to be updated.

Ok.. Turns out this was not a problem with the actual influences calculation or dependency calculation code. Whew! match_link just didn't set the influence correctly when failing. fixed


Update: Reopening this because the fix for it rather sucks.

I made match_link return on failure an influence of type DEPEND_LINKS. So, a tag page that inlines tagged(foo) gets a depends_simple built up that contains link dependencies for every page in the wiki. A very bloaty way to represent the dependency!

Per dependency types, link(done) only needs to list in depends_simple the pages that currently match. If a page is modified to add the link, the regular dependency calculation code notices that a new page matches. If a page that had the link is modified to remove it, the depends_simple lets ikiwiki remember that the now non-matching page matched before.

Where that fell down was !link(done). A page matching that was not added to depends_simple, because the link(done) did not match it. If the page is modified to add the link, the regular dependency calculation code didn't notice, since the pagespec no longer matched.

In this case, depends_simple needs to contain all pages that do not match link(done), but before my change, it contained all pages that do match. After my change, it contained all pages.

So, seems what is needed is a way for influence info to be manipulated by the boolean operations that are applied. One way would be to have two sets of influences be returned, one for successful matches, and one for failed matches. Normally, these would be the same. For successful match_link, the successful influence would be the page. For failed match_link, the failed influence would be the page.

Then, when NOTting a *Reason, swap the two sets of influences. When ANDing/ORing, combine the individual sets. Querying the object for influences should return only the successful influences.

Would it be possible to avoid the complication of maintianing two sets of influence info?

Well, notice that the influence of pagespec_match($page, "link(done)") is $page. Iff the match succeeds.

Also, the influence of pagespec_match($page, "!link(done)") is $page. Iff the (overall) match succeeds.

Does that hold for all cases? If so, the code that populates depends_simple could just test if the pagespec was successful, and if not, avoid adding $page influences, while still adding any other, non-$page influences.

Hmm, commit f2b3d1341447cbf29189ab490daae418fbe5d02d seems thuroughly wrong. So, what about influence info for other matches like !author(foo) etc? Currently, none is returned, but it should be a content influence. (Backlink influence data seems ok.)

done again!