| This trivial patch allows qpolicies to be moved to userspace as demonstrated | by attached C# code (requires mono). This way userspace-kernelspace interface | is dead simple allowing for quite a lot flexibility in designing the way | packet buffer is organized. | The basic idea is that we set send buffer size so that only one packet fits in | it. Sending second packet while first is inside buffer (has not been sent | yet) causes send() to block until ccid allows the first packet to be 'send'. One difficulty that I see is what to do when the send socket is non-blocking. | For each real packet we call send() twice: one 'fake' (will be discarded just | before sending) and one 'real' (contains user data). This makes all the | waiting to be done on the 'fake' packet making sending the real one really | fast. This allows for late data choice without difficult to implement ring | buffers shared by user and kernel space or complicated qpolicies code on | kernel side. Maybe purists will not like this since the complexity is now shifted to the user-kernel interface. But it may well be a start to develop some kind of new interface which achieves the same in a cleaner way. In any case, if people would like to test this feature, I'd be happy to fork a new qpolicy subtree from the test tree. | Results on CCID2 are very promising (near ideal), but CCID3 is somewhat | disappointing to the extent I'm starting to think that I'm doing something | seriously wrong while testing it. | Can you please narrow down what is happening with CCID-3. CCID-3 is very complicated and sensitive, so it is possible that the problem is within CCID-3 and not the qpolicy code - need more detailed information. | Therefore I have two questions: | 1. What do you think about the idea in general? | 2. Could anybody test the attached code on both CCID2 and CCID3 and send | results/observations to the list? | At the moment this is just enough code to allow a what seems to me working proof-of-concept implementation. I don't have mono installed and need to prioritise with time - the highest priority at the moment is to keep the test tree up to date and bug-free. If a stable interface evolves out of this development work, it would be great to integrate it with the existing (C flat) code http://www.erg.abdn.ac.uk/users/gerrit/dccp/apps/#DCCP_Library | --- a/net/dccp/qpolicy.c | +++ b/net/dccp/qpolicy.c | @@ -120,11 +122,11 @@ struct sk_buff *dccp_qpolicy_pop(struct sock *sk) | { | struct sk_buff *skb = dccp_qpolicy_top(sk); | | - /* Clear any skb fields that we used internally */ | - skb->priority = 0; | - | - if (skb) | + if (skb) { | + /* Clear any skb fields that we used internally */ | + skb->priority = 0; | skb_unlink(skb, &sk->sk_write_queue); | + } | return skb; | } Thank you for the above hunk -- since it fixes in fact a potential source of trouble, I have merged it into the patch entitled "dccp: Policy-based packet dequeueing infrastructure" and updated the test tree to use this patch. http://eden-feed.erg.abdn.ac.uk/cgi-bin/gitweb.cgi?p=dccp_exp.git;a=commitdiff;h=ffc6589ad8fb51edb37f31cbad9b913e9b47f630 -- To unsubscribe from this list: send the line "unsubscribe dccp" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
