Date:	Wed, 9 Jul 2014 14:35:08 -0400
From:	Neil Horman <nhorman@...driver.com>
To:	Vlad Yasevich <vyasevich@...il.com>
Cc:	Daniel Borkmann <dborkman@...hat.com>, davem@...emloft.net,
	geirola@...il.com, netdev@...r.kernel.org,
	linux-sctp@...r.kernel.org
Subject: Re: [PATCH net-next 0/5] SCTP updates

On Wed, Jul 09, 2014 at 12:32:55PM -0400, Vlad Yasevich wrote:
> On 07/09/2014 11:44 AM, Neil Horman wrote:
> > On Wed, Jul 09, 2014 at 11:36:07AM -0400, Vlad Yasevich wrote:
> >> On 07/09/2014 11:13 AM, Neil Horman wrote:
> >>> On Wed, Jul 09, 2014 at 03:28:03PM +0200, Daniel Borkmann wrote:
> >>>> On 07/09/2014 12:49 PM, Neil Horman wrote:
> >>>>> On Wed, Jul 09, 2014 at 11:57:37AM +0200, Daniel Borkmann wrote:
> >>>>>> On 07/08/2014 04:41 PM, Neil Horman wrote:
> >>>>>>> On Tue, Jul 08, 2014 at 04:05:26PM +0200, Daniel Borkmann wrote:
> >>>>>>>> On 07/08/2014 01:14 PM, Neil Horman wrote:
> >>>>>>>> ...
> >>>>>>>>> since you're adding cmsg's from rfc6458, do you also want to add some
> >>>>>>>>> deprecation warnings around the use of SCTP_SNDRCVINFO too, so we can start to
> >>>>>>>>> schedule its eventual removal?
> >>>>>>>>
> >>>>>>>> Sure, we can do that. Do you want me to include it into the set?
> >>>>>>>
> >>>>>>> If you're plan is to implement 6458, then yes, I think that would be good.
> >>>>>>
> >>>>>> Looking a bit closer at it, all our pr_warn_ratelimited(DEPRECATED ...)
> >>>>>> warnings in SCTP are being done in 'slowpath' {set,get}sockopt(2) operations
> >>>>>> only, which is fine. What you're suggesting is to place similar ratelimited
> >>>>>> warnings (due to different possible pids on the machine) into the 'fastpath'
> >>>>>> where we get and set cmsg message headers.
> >>>>>>
> >>>>>> While that may be fine for {set,get}sockopt(2) that's called once or very few
> >>>>>> times, I'm not sure this is a good idea in SCTP_SNDRCVINFO as it will yield
> >>>>>> to unnecessary spamming the klog since up to now this is the only way our
> >>>>>> users can set or receive this info. I'm not sure we want to annoy users like
> >>>>>> that ...
> >>>>>>
> >>>>> Then we wrap it in a ONCE macro so that it only triggers on the first use
> >>>>> instance.
> >>>>
> >>>> I'm not convinced about this so far. The whole point is that we also provide the
> >>>> pid just as we currently do, so that we give the user a chance to possibly pin
> >>>> point the process that needs code change to not use the deprecated API anymore.
> >>>>
> >>>>>> In how many years do you plan a removal ... I think we're stuck with uapi
> >>>>>> basically forever as we don't want to break old binaries, no? ;/
> >>>>>>
> >>>>> I thought we could remove things on a schedule if we followed the deprecation
> >>>>> process, but that may just be for sysfs.  Regardless, it would still be nice to
> >>>>> inform people they are using an older api.
> >>>>
> >>>> I think we rather might be stuck with also the deprecated stuff forever, just
> >>>> as AF_PACKET still has to carry around the old spkt stuff. :( So if we don't
> >>>> remove anything, there's actually also no point to spam the log about it, if
> >>>> everyone can/should read it from the RFC anyway.
> >>>>
> >>> So this begs the question as to why we have deprecation warnings to begin with.
> >>> At what point do we draw the line where we can change some aspect of the user
> >>> api.  I agree if the answer is never, then yeah we're stuck, but then, why
> >>> bother announcing deprecation warnings at all?
> >>
> >> I think we can deprecate user API after a significantly long period of time
> >> (like 5 or 6 years).  That's why we also have a deprecation schedule that can
> >> be updated and hopefully that's something people pay attention to.
> >>
> >> The problem here is deprecation of ancillary data and that's is a lot tougher
> >> then socket options.  In this particular case (SCTP_SNDRCVINFO vs SCTP_RCVINFO),
> >> I don't think there is any way to deprecate the SCTP_SNDRCVINFO since the event
> >> enabling it is the same as the one for SCTP_RCVINFO.  This was a mistake in the
> >> spec.  Ancillary data should not have been enabled using even notification api,
> >> as it is not an event, but we now have to live with it.
> >>
> > Ugh I didn't even consider cmsg type overlap.  Thats probably it then, we can't
> > deprecate it.  Though that does call the question up as to how to differentiate
> > expectations of the data format for each cmsg, if they use the same type.  Does
> > the SCTP_RCVINFO data struct overlay the SNDRCVINFO struct exactly?  (sorry I've
> > not checked myself yet).
> 
> No there is not direct overlap between the two.  However, as Michael pointed out,
> there is a new option to control SCTP_RCVINFO.  So would could add a deprecation
> warning to the over SCTP_EVENTS option and carry SCTP_SNDRCVINFO with it.
> Once SCTP_EVENTS goes away so can SCTP_SNDRCVINFO.
> 
Ok, so we should still consider deprecation warnings then.  Daniel, what about
ratelimited warnings with pids included then?
Neil

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