On Fri, May 15, 2026 at 11:06:42AM +0200, Stefano Garzarella wrote: > On Fri, May 15, 2026 at 04:57:41AM -0400, Michael S. Tsirkin wrote: > > On Fri, May 15, 2026 at 10:29:55AM +0200, Stefano Garzarella wrote: > > > On Thu, May 14, 2026 at 01:44:53PM -0400, Michael S. Tsirkin wrote: > > > > On Thu, May 14, 2026 at 06:45:00PM +0200, Stefano Garzarella wrote: > > > > > On Thu, 14 May 2026 at 17:16, Michael S. Tsirkin <[email protected]> > > > > > wrote: > > [...] > > > > > > > > > > > > > > > > > > > And so the bag of hacks grows. I feel this is energy not well spent. > > > > > > Please, let us fix this properly *first*. And then worry about how > > > > > > to > > > > > > backport. Maybe it will not be so terrible to backport after all. > > > > > > > > > > > > > > > > TBH I don't think this is an hack, but an issue we should fix in any > > > > > case. > > > > > Regarding the second patch, I see your point, but it's a big change > > > > > that worries me. I'd like some more time to fix it properly without > > > > > rushing. Staying calm without realizing that userspace is broken like > > > > > we are now without this series :-( > > > > > > > > > > That said, evaluating further, I think we have a similar issue also > > > > > with STREAM on the host side where the skb usually doesn't free space, > > > > > so we need a merge strategy also there. > > > > > > > > > > So, I'd like to have time to fix both definitely. If you have time and > > > > > want to go ahead, please do. > > > > > > > > > > Thanks, > > > > > Stefano > > > > > > > > Well my patch was a start, we just need a strategy how to avoid copying > > > > everything, right? > > > > > > Yep, and then there's the question of how to handle EOM without a payload, > > > but I think that's a special case. In theory, we don't support sending it, > > > but I'm not sure if POSIX allows it or not. > > > > It seems to, but given we didn't allow it in the past, we probably > > should not start now without a good solution. > > Agree. > > > Really we should add a feature bit for EOM to steal a byte from > > buf_alloc. Or several bytes) > > Yes, I agree. At this point, though, could we define a new protocol that > also takes overhead into account, or would it be too complicated to > synchronize both?
Well isn't it just X bytes per EOM? The rest can be fixed with merging? > > > > > That said, is it okay if I send a v4 of this series? > > > > > > (I'm not sure if I'll be able to work on the merging next week) > > > > > > Stefano > > > > > > I do worry we are piling up hacks and we'll end up with races > > for all our troubles. That said, up to you. > > > > I see it differently; Patch 1 should have been there from the start. > Patch 2, unless we completely remove the overhead, we should keep it, or use > it to trigger merging (e.g., when the overhead reaches a threshold that > depends on `buf_alloc`). I'm not sure we need to merge retroactively though. Maybe just start merging new stuff when over a threshold. Should be simpler. > I prefer to send a v4. > > Thanks, > Stefano
