On Mon, Apr 30, 2012 at 3:53 PM, Ralph Castain <r...@open-mpi.org> wrote: > Hi folks > > I'm working with priorities in events using libevent 2.0.13. Since I'm not > using the most current release, I thought I'd ask about a behavior we are > seeing that might be a potential bug.
It's not the cause of your problem, but "event_add()"ing these events is nonsensical. If an event has fd -1, it is meaningless (and undefined!) to give it EV_READ or EV_WRITE and assign. It could cause bugs down the line if that's your standard practice. Looking at the issue, I think the problem occurs when you use event_active() to activate another event of priority equal to the currently executing event: since the event will happen in the currently executing instance of the event loop rather than the next one, it'll keep executing all events of that priority, including the new one, before it scans for new events and executes higher priority events again. Or from another point of view, the problem is that using event_active() on a higher-priority event doesn't tell event to stop executing events at the current priority and go to the next event loop iteration. Option 1: Call this a bug. Say that if you use event_active() to activate an event of higher (that is, numerically lower) priority than the currently executing event, the event loop needs to check for new events and then execute higher-priority events. Option 1a: apply the fix in 2.0, since it's probably right. Option 1b: apply the fix in 2.1, since it might potentially break something if anybody was relying on the old behavior. Option 2: Call this a regrettable wart, but say that option 1 is too destabilizing. Add an event_base_loopcontinue() that causes libevent to re-scan and start its dispatch phase again. Option 3: Call this an intended feature; tell everybody that they need to use event_config_set_max_dispatch_interval() as a workaround. Thoughts? -- Nick *********************************************************************** To unsubscribe, send an e-mail to majord...@freehaven.net with unsubscribe libevent-users in the body.
