ikiwiki instance is holding a lock, a web user clicking on "add comment" (for example) will have to wait for the lock to be released. However, all they are then presented with is a web form. Perhaps CGI requests that are read-only (such as generating a comment form, or perhaps certain types of edits) should ignore locks? Of course, I'd understand that the submission would need to wait for a lock. — Jon
Ikiwiki has what I think of as the Big Wiki Lock (remembering the "Big Kernel Lock"). It takes the exclusive lock before loading any state, to ensure that any changes to that state are made safely.
A few CGI actions that don't need that info loaded do avoid taking the lock.
In the case of showing the comment form, the comments plugin needs CGI session information to be loaded, so it can check if the user is logged in, and so it can add XSRF prevention tokens based on the session ID. (Actually, it might be possible to rely on
CGI::Session's own locking of the sessions file, and have a hook that runs with a session but before the indexdb is loaded.)
But, the comment form also needs to load the indexdb, in order to call
check_canedit, which matches a pagespec, which can need to look things up in the indexdb. (Though the pagespecs that can do that are unlikely to be relevant when posting a comment.)
I've thought about trying to get rid of the Big Wiki Lock from time to time. It's difficult though; if two ikiwikis are both making changes to the stored state, it's hard to see a way to reconcile them. (There could be a daemon that all changes are fed thru using a protocol, but that's really complicated, and it'd almost be better to have a single daemon that just runs ikiwiki; a major architectural change.)
One way that almost seems it could work is to have a entry path that loads everything read-only, without a lock. And then in read-only mode,
saveindexwould be an error to run. However, both the commenting code and the page edit code currently have the same entry path for drawing the form as is used for handling the posted form, so they would need to be adapted to separate that into two code paths. --Joey