route/arp lifetime (Re: it's the output, not ack coalescing (Re: TSO and FreeBSD vs Linux))

[Alexander V. Chernikov]

On 14.08.2013 16:40, Luigi Rizzo wrote:
> On Wed, Aug 14, 2013 at 04:15:25PM +0400, Alexander V. Chernikov wrote:
>> On 14.08.2013 16:05, Luigi Rizzo wrote:
>>> On Wed, Aug 14, 2013 at 03:47:13PM +0400, Lev Serebryakov wrote:
>>>> Hello, Luigi.
>>>> You wrote 14 ?????????????? 2013 ??., 14:21:09:
>>>>
>>>> LR> Then the problem remains that we should keep a copy of route and
>>>> LR> arp information in the socket instead of redoing the lookups on
>>>> LR> every single transmission, as they consume some 25% of the time of
>>>> LR> a sendto(), and probably even more when it comes to large tcp
>>>> LR> segments, sendfile() and the like.
>>>>     And we should invalidate this info on ARP/route changes, or connection
>>>>    will be lost in such cases, am I right?.. So, on each such event code
>>>>    should look into all sockets and check, if routing/ARP information is still
>>>>    valid for them. Or we should store lists of sockets in routing and ARP
>>>>    tables... I don't know, what is worse.
>>> I think we should start by acknowledging that routing and ARP
>>> information is inherently stale, and changes unfrequently.
>>> So it is not a disaster if we have incorrect information for some
>>> short amount of time (milliseconds) because in the end the remote
>>> party that decides to change it and inform us may take much longer
>>> than that to distribute the update.
>> You can save rte&arp, however doing this
>> gives you perfect chance to crash your kernel if egress interface is
>> destroyed (like vlan or ng or tun).
> I hope I learned not to follow a stale ifp pointer :)
Well, currently we have no locks (or other means)  to ensure all other 
cores has "current" pointer to ifp or its fields (or am I wrong?)
> anyways ARP is really just the mac address so there is no
> dandling pointer issue.
>
> For the ifp associated to the route,
> i do not see a huge problem in marking the route/ifp as
> zombie and destroy it when the last reference goes away.
Yes, but references requires some synchronization primitives. One 
possible solution is using pcpu counters, but it does not play well on 
!amd64.
>
> Not that the current way is any better -- you need to lock/unlock
> the rte while you do the lookup, and hold a refcount to the ifp
> until the packet is queued. So how does my suggestion make
> things worse ?
>
> cheers
> luigi
>
>
>>>
>>> Considering that each lookup takes between 100..300ns if you are
>>> lucky (not many misses, relatively empty table etc.), one could
>>> reasonably do the lookup at most once per millisecond or so (just
>>> reading 'ticks', no need for a nanotime() if you have a slow clock),
>>> or whenever we get an error related to the socket, either in the
>>> forward path (e.g. ifp points to an interface that is down) or in
>>> the reverse path (e.g. a dupack because we sent a packet to the
>>> wrong place).
>> This sounds like "Hey, the kernel lookup is slow (which is true), let's
>> make a hack and don't bother lookups".
>> This approach gives us mtx-locked rte refcounts which are used (misused)
>> in many places making things worse and decreasing the ability to fix the
>> things up..
>>> cheers
>>> luigi
>>> _______________________________________________
>>> freebsd-net at freebsd.org mailing list
>>> http://lists.freebsd.org/mailman/listinfo/freebsd-net
>>> To unsubscribe, send any mail to "freebsd-net-unsubscribe at freebsd.org"
>>>