[dpdk-dev] [RFC] lib/librte_ether: consistent PMD batching behavior

Yang, Zhiyong zhiyong.yang at intel.com
Tue Feb 7 08:50:21 CET 2017


Hi, Adrien:

	Sorry for the late reply  due to Chinese new year.

> -----Original Message-----
> From: Adrien Mazarguil [mailto:adrien.mazarguil at 6wind.com]
> Sent: Tuesday, January 24, 2017 12:36 AM
> To: Richardson, Bruce <bruce.richardson at intel.com>
> Cc: Ananyev, Konstantin <konstantin.ananyev at intel.com>; Andrew
> Rybchenko <arybchenko at solarflare.com>; Yang, Zhiyong
> <zhiyong.yang at intel.com>; dev at dpdk.org; thomas.monjalon at 6wind.com
> Subject: Re: [dpdk-dev] [RFC] lib/librte_ether: consistent PMD batching
> behavior
> 
> On Fri, Jan 20, 2017 at 11:48:22AM +0000, Bruce Richardson wrote:
> > On Fri, Jan 20, 2017 at 11:24:40AM +0000, Ananyev, Konstantin wrote:
> > > >
> > > > From: Andrew Rybchenko [mailto:arybchenko at solarflare.com]
> > > > Sent: Friday, January 20, 2017 10:26 AM
> > > > To: Yang, Zhiyong <zhiyong.yang at intel.com>; dev at dpdk.org
> > > > Cc: thomas.monjalon at 6wind.com; Richardson, Bruce
> > > > <bruce.richardson at intel.com>; Ananyev, Konstantin
> > > > <konstantin.ananyev at intel.com>
> > > > Subject: Re: [dpdk-dev] [RFC] lib/librte_ether: consistent PMD
> > > > batching behavior
> > > >
> > > > On 01/20/2017 12:51 PM, Zhiyong Yang wrote:
> > > > The rte_eth_tx_burst() function in the file Rte_ethdev.h is
> > > > invoked to transmit output packets on the output queue for DPDK
> > > > applications as follows.
> > > >
> > > > static inline uint16_t
> > > > rte_eth_tx_burst(uint8_t port_id, uint16_t queue_id,
> > > >                  struct rte_mbuf **tx_pkts, uint16_t nb_pkts);
> > > >
> > > > Note: The fourth parameter nb_pkts: The number of packets to
> transmit.
> > > > The rte_eth_tx_burst() function returns the number of packets it
> > > > actually sent. The return value equal to *nb_pkts* means that all
> > > > packets have been sent, and this is likely to signify that other
> > > > output packets could be immediately transmitted again.
> > > > Applications that implement a "send as many packets to transmit as
> > > > possible" policy can check this specific case and keep invoking
> > > > the rte_eth_tx_burst() function until a value less than
> > > > *nb_pkts* is returned.
> > > >
> > > > When you call TX only once in rte_eth_tx_burst, you may get
> > > > different behaviors from different PMDs. One problem that every
> > > > DPDK user has to face is that they need to take the policy into
> > > > consideration at the app- lication level when using any specific
> > > > PMD to send the packets whether or not it is necessary, which
> > > > brings usage complexities and makes DPDK users easily confused
> > > > since they have to learn the details on TX function limit of
> > > > specific PMDs and have to handle the different return value: the
> > > > number of packets transmitted successfully for various PMDs. Some
> > > > PMDs Tx func- tions have a limit of sending at most 32 packets for
> > > > every invoking, some PMDs have another limit of at most 64 packets
> > > > once, another ones have imp- lemented to send as many packets to
> transmit as possible, etc. This will easily cause wrong usage for DPDK users.
> > > >
> > > > This patch proposes to implement the above policy in DPDK lib in
> > > > order to simplify the application implementation and avoid the
> > > > incorrect invoking as well. So, DPDK Users don't need to consider
> > > > the implementation policy and to write duplicated code at the
> > > > application level again when sending packets. In addition to it,
> > > > the users don't need to know the difference of specific PMD TX and
> > > > can transmit the arbitrary number of packets as they expect when
> > > > invoking TX API rte_eth_tx_burst, then check the return value to get
> the number of packets actually sent.
> > > >
> > > > How to implement the policy in DPDK lib? Two solutions are proposed
> below.
> > > >
> > > > Solution 1:
> > > > Implement the wrapper functions to remove some limits for each
> > > > specific PMDs as i40e_xmit_pkts_simple and ixgbe_xmit_pkts_simple
> do like that.
> > > >
> > > > > IMHO, the solution is a bit better since it:
> > > > > 1. Does not affect other PMDs at all
> > > > > 2. Could be a bit faster for the PMDs which require it since has
> > > > >no indirect
> > > > >    function call on each iteration
> > > > > 3. No ABI change
> > >
> > > I also would prefer solution number 1 for the reasons outlined by Andrew
> above.
> > > Also, IMO current limitation for number of packets to TX in some
> > > Intel PMD TX routines are sort of artificial:
> > > - they are not caused by any real HW limitations
> > > - avoiding them at PMD level shouldn't cause any performance or
> functional degradation.
> > > So I don't see any good reason why instead of fixing these
> > > limitations in our own PMDs we are trying to push them to the upper
> (rte_ethdev) layer.
> 
> For what it's worth, I agree with Konstantin. Wrappers should be as thin as
> possible on top of PMD functions, they are not helpers. We could define a
> set of higher level functions for this purpose though.
> 
> In the meantime, exposing and documenting PMD limitations seems safe
> enough.
> 
> We could assert that RX/TX burst requests larger than the size of the target
> queue are unlikely to be fully met (i.e. PMDs usually do not check for
> completions in the middle of a TX burst).

As a tool,  it is very important for its users to easily consume it and make it work
in a right way.  Sort of artificial limits will make things look like a little confused  and
make some users probably get into trouble when writing drivers. 
Why do we correct it and make it easier?  :)

Zhiyong


More information about the dev mailing list