Date:	Sun, 5 Feb 2012 20:32:47 -0500
From:	Neil Horman <nhorman@...driver.com>
To:	Hagen Paul Pfeifer <hagen@...u.net>
Cc:	"Frank Ch. Eigler" <fche@...stic.org>,
	Satoru Moriya <satoru.moriya@....com>,
	David Miller <davem@...emloft.net>,
	"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
	"tgraf@...radead.org" <tgraf@...radead.org>,
	"stephen.hemminger@...tta.com" <stephen.hemminger@...tta.com>,
	"eric.dumazet@...il.com" <eric.dumazet@...il.com>,
	Seiji Aguchi <seiji.aguchi@....com>, fche@...rceware.org,
	mathieu.desnoyers@...ymtl.ca, rostedt@...dmis.org
Subject: Re: [PATCH v2 0/2] Tracepoint for tcp retransmission

On Sun, Feb 05, 2012 at 09:04:29PM +0100, Hagen Paul Pfeifer wrote:
> * Neil Horman | 2012-02-05 14:17:08 [-0500]:
> 
> >Generally the fact that probe points may shift underneath a script between
> >kernels.  While its often ok, theres no guaranteed fixed set of probe points.
> >While it often works, theres always the possibility that functional changes,
> >even if they don't reults in behavioral changes to userspace, require re-working
> >of stap scripts.  Its not the sort of thing that works will for people trying to
> >build applications.
> 
> Hey Neil, I though about this more deeply. In my first iteration I had the
> same answer. But IMHO if you think about this more deeply, I think this is not
> a scalable approach (implementing in userspace as systemtap scripts compared
> to implement them in kernelspace (tracepoints). Why? Because at a deep TCP
> level the API (and thus the context) is not static and the meaning of return
> values (or arguments) change over time. No matter _where_ it is implemented or
> coded.
> 
Yes, this is why I'm suggesting alternatives to the tracepoint.

> Lets imaging the following (I construct one example, but this may universal):
> imagine we have two retransmission functions, one for RTO triggered
> retransmission and one for dup-ack fast-retransmission. Both functions where
> instrumented as tracepoints for years. After some years someone say "hey both
> functions are similar, lets fix this and make one function". What should
> happened with the tracepoints? Should they merged into one? What should be the
> name of the tracepoint?
> 
Yes, this is a perenial problem with tracepoints (and, in a simmilar fashion,
stap scripts).

> In the end you will end up fixing userspace analysis tools as well! IMHO there
> is - in my opinion - no clean way to add tracepoints and think you are free of
> maintaining the whole chain, from kernelspace to userspace.
> 
Yes.

> Only "stable" in kernel API's may be a candidate for tracepoints (like kmalloc
> and friends). Highly dynamic and often changed code not.
> 
Yes.

> The biggest advantage from tracepoints is that the function is "annotated"
> with something like "this is a tracepoint, please be nice and don't break the
> behavioral meaning of this function". On the other side this is also a
> disadvantage because it elevates a kernel function to a "userspace API": Tools
> will start using this tracepoint, you may end up in a situation where you
> can't change/delete tracepoints.
> 
Yes, tracepoints only make sense for API's that are clearly stable and well
defined.  This doesn't fit that bill.
 
The long and the short of it is, a tracepoint isn't going to be the solution for
the problem Satoru needs to solve.  Dave just made that abundantly clear.
Another solution needs to be found.  A perf socket layer seems to be a good long
term solution.  If something is needed in the short term for Satoru, I still
think a privately maintained module that registers a netfilter hook to watch
outgoing packets for retransmits offers a good solution.

Regards
Neil

> Hagen
> 
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists