On Wed, Nov 23, 2016 at 09:03 +0100, Stefan Fritsch wrote: > On Wed, 23 Nov 2016, Mike Belopuhov wrote: > > > I guess we could do that. But then we cannot free the mbufs on DOWN > > > until the device has used them. > > > > Diff to this effect is below. Works on vmd and qemu (original > > one didn't because I kept the virtio_reset). > > > > > That sounds like an unnecessary waste of memory to me. > > > > > > > This is not so much memory we lose and then if you up it again > > you're going to have it all back. We can revert to the present > > behavior once vmd matures, in the meantime people won't have to > > juggle diffs around in their trees :) > > I am not convinced. Doing a reset allows to recover from all kinds of > problems with DOWN/UP. That was useful when we had bugs in the event_idx > implementation. >
I don't think this justifies it since bugs need to be fixed regardless. > Also, I don't like to change code that is known to work with at least 4 > independent device implementations to work around problems in one > incomplete implementation that we can easily change. > > Maybe something like this is enough already (untested): > The diff didn't help.