[RFC] Add support for hardware transmit rate limiting queues [WAS: Add support for changing the flow ID of TCP connections]
Hans Petter Selasky
hps at selasky.org
Wed Aug 20 19:25:13 UTC 2014
On 08/20/14 20:44, Navdeep Parhar wrote:
> On 08/20/14 00:34, Hans Petter Selasky wrote:
>> A month has passed since the last e-mail on this topic, and in the
>> meanwhile some new patches have been created and tested:
>> Basically the approach has been changed a little bit:
>> - The creation of hardware transmit rings has been made independent of
>> the TCP stack. This allows firewall applications to forward traffic into
>> hardware transmit rings aswell, and not only native TCP applications.
>> This should be one more reason to get the feature into the kernel.
>> - A hardware transmit ring basically can have two modes: FIXED-RATE or
>> AUTOMATIC-RATE. In the fixed rate mode all traffic is sent at a fixed
>> bytes per second rate. In the automatic mode you can configure a time
>> after which the TX queue must be empty. The hardware driver uses this to
>> configure the actual rate. In automatic mode you can also set an upper
>> and lower transmit rate limit.
>> - The MBUF has got a new field in the packet header: "txringid"
>> - IOCTLs for TCP v4 and v6 sockets has been updated to allow setting of
>> the "txringid" field in the mbuf.
>> The current patch [see attachment] should be much simpler and less
>> intrusive than the previous one.
>> Any comments ?
> Here are some thoughts. The first two bullets cover relatively
> minor issues, the rest are more important.
> - All of the mbuf pkthdr fields today have the same meaning no matter
> what the context. It is not clear what txringid's global meaning is.
> Is it even possible for driver foo to interpret it the same way as
> driver bar? What if the number of rings are different, or if the ring
> at the particular index for foo is setup differently than the ring at
> that same index for bar? You are attempting to influence the driver's
> txq selection and traditionally the mbuf's flowid has been used for
> this purpose. Have you considered allowing the user to set the flowid
> directly? And mark it as such via a new rsstype so the kernel will
> leave it alone.
At work so to speak, we have tried to make a simple approach that will
not break existing code, without trying to optimise the possibilities
and reduce memory footprint.
> - uint32_t -> m_flowid_t is plain gratuitous. Now we need to include
> mbuf.h in more places just to get this definition. What's the
> advantage of this? style(9) isn't too fond of typedefs either. Also,
> drivers *do* need to know the width of the flowid. At least lagg(4)
> looks at the high bits of the flowid (see flowid_shift in lagg). How
> high it can go depends on the width of the flowid.
The flowid should be typedef'ed. Else how can you know its type passing
flowid along function arguments and so on?
> - Interfaces can come and go, routes can change, and so the relationship
> between an inpcb and txringid is not stable at all. What happens when
> the outbound route for an inpcb changes?
This is managed separately by a daemon or such. The problem about using
the "inpcb" approach which you are suggesting, is that you limit the
rate control feature to traffic which is bound by sockets. Can your way
of doing rate control be useful to non-socket based firewall
applications, for example?
You also assume a 1:1 mapping between "inpcb" and the flowID, right.
What about M:N mappings, where multiple streams should share the same
flowID, because it makes more sense?
> - The in_ratectlreq structure that you propose is inadequate in its
> current form. For example, cxgbe's hardware can do rate limiting on a
> per-ring as well as per-connection basis, and it allows for pps,
> bandwidth, or min-max limits. I think this is the critical piece that
> we NIC maintainers must agree on before any code hits the core kernel:
> how to express a rate-limit policy in a standard way and allow for
> hardware assistance opportunistically. ipfw(4)'s dummynet is probably
> interested in this part too, so it's great that Luigi is paying
> attention to this thread.
My "in_ratectlreq" is a work in progress.
> - The RATECTL ioctls deal with in_ratectlreq so we need to standardize
> the ratectlreq structure before these ioctls can be considered generic
> ifnet ioctls. This is the reason cxgbetool (and not ifconfig) has a
> private ioctl to frob cxgbe's per-queue rate-limiters. I did not want
> to add ifnet ioctls that in reality were cxgbe only. Ditto for i2c
> ioctls. Now we have multiple drivers with i2c and melifaro@ is doing
> the right thing by promoting these private ioctls to a standard ifnet
> ioctl. Have you considered a private mlxtool as a stop gap measure?
It might end that we need to create our own tool for this, having vendor
specific IOCTLs, if we cannot agree how to do this in a general way.
> To summarize my take on all of this: we need a standard ratectlreq
> a standard way to associate an inpcb with one,
> and a standard
> way to pass on this info to if_transmit.
> After all this is in place we
> could even have a dummynet-ish software layer that implements rate
> limiters when the underlying hardware offers no assistance.
More information about the freebsd-current