[dpdk-dev] [PATCHv2 0/5] ethdev: add new API to retrieve RX/TX queue information

Walukiewicz, Miroslaw Miroslaw.Walukiewicz at intel.com
Thu Jun 18 16:37:15 CEST 2015



> -----Original Message-----
> From: Ananyev, Konstantin
> Sent: Thursday, June 18, 2015 4:17 PM
> To: Walukiewicz, Miroslaw; dev at dpdk.org
> Subject: RE: [dpdk-dev] [PATCHv2 0/5] ethdev: add new API to retrieve RX/TX
> queue information
> 
> Hi Mirek,
> 
> > -----Original Message-----
> > From: Walukiewicz, Miroslaw
> > Sent: Thursday, June 18, 2015 2:31 PM
> > To: Ananyev, Konstantin; dev at dpdk.org
> > Subject: RE: [dpdk-dev] [PATCHv2 0/5] ethdev: add new API to retrieve
> RX/TX queue information
> >
> > Konstantin,
> >
> > Is there a possibility to read information about available space in NIC queue
> for TX.
> 
> I suppose it is possible as some future addition.
> As I said in the commit message, I left some reserved space, so extra fields
> could be added to the structure without ABI breakage.
> For now, I just added some static/config information for the queue, but I think
> it is possible and plausible to have some runtime information too.
> 
> >
> > It is quite easy to compute (or even available directly)  and very useful
> especially for application sending multi-descriptor packets like
> > TCP TSO.
> >
> > Now there is no access to such information and the transmit packet function
> must be called to
> > be sure that there is available space.
> 
> Hmm, as I said I was thinking about adding some RT information in future:
> number of free descriptors (from SW point of view), index of next descriptor
> to process by SW, etc.
> But my thought it would be use by some watchdog thread to collect
> statistics/detect stall, etc.
> I didn't intend it to be used by IO thread.
> 
> I am not sure why do you need to call such function at RT?
> PMD wouldn't TX a packet, if there is not enough free TXDs for the whole
> packet.
> From other side upper layer, can't always calculate correctly how many TXD it
> would really need
> (context descriptors might be needed, etc).
>  Plus, even if nb_tx_free==X, in reality it could be there much more free
> TXDs, and SW just need
> to process them (and would do that at next  tx_burst() call.
> So, why just not:
> ...
> n = tx_burst(..., nb_tx);
> if (n < nb_tx) {requeue unsent of packets;}
> ?
> 
The case is TCP TSO operation as I said.

I really use the that method proposed by you but it very expensive way.

The TX path in TCP is very complex and expensive. 
It is desirable to create the TCP TSO segments that fits the space in queue to avoid re-queue  of packets 
making the TX path much more complex. 
For ixgbe the TSO segments are up to 40 descriptors (max 64K of data) and making the TX queue full is just simple.

Having the possibility for reading number of descriptors available to send can improve TCP TX process in the way 
that when there is no space in TX queue we can just not enter the complex TX path and spent time making more 
useful job than re-queue.
Also It is easy to fit the TSO segment to available space in TX queue instead of sending always a large 64K TCP segments 
and count on the result of tx_burst and re-queue in case of lack of space.

Mirek

> Konstantin
> 
> >
> > Mirek
> >
> > > -----Original Message-----
> > > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Konstantin
> Ananyev
> > > Sent: Thursday, June 18, 2015 3:19 PM
> > > To: dev at dpdk.org
> > > Subject: [dpdk-dev] [PATCHv2 0/5] ethdev: add new API to retrieve RX/TX
> > > queue information
> > >
> > > Add the ability for the upper layer to query RX/TX queue information.
> > > Right now supported for:
> > > ixgbe, i40e, e1000 PMDs.
> > >
> > > Konstantin Ananyev (5):
> > >   ethdev: add new API to retrieve RX/TX queue information
> > >   i40e: add support for eth_(rxq|txq)_info_get
> > >   ixgbe: add support for eth_(rxq|txq)_info_get
> > >   e1000: add support for eth_(rxq|txq)_info_get
> > >   testpmd: add new command to display RX/TX queue information
> > >
> > >  app/test-pmd/cmdline.c           | 48 +++++++++++++++++++++++++
> > >  app/test-pmd/config.c            | 67
> > > ++++++++++++++++++++++++++++++++++
> > >  app/test-pmd/testpmd.h           |  2 ++
> > >  drivers/net/e1000/e1000_ethdev.h | 12 +++++++
> > >  drivers/net/e1000/em_ethdev.c    |  2 ++
> > >  drivers/net/e1000/em_rxtx.c      | 38 ++++++++++++++++++++
> > >  drivers/net/e1000/igb_ethdev.c   |  4 +++
> > >  drivers/net/e1000/igb_rxtx.c     | 36 +++++++++++++++++++
> > >  drivers/net/i40e/i40e_ethdev.c   |  2 ++
> > >  drivers/net/i40e/i40e_ethdev.h   |  5 +++
> > >  drivers/net/i40e/i40e_rxtx.c     | 42 ++++++++++++++++++++++
> > >  drivers/net/ixgbe/ixgbe_ethdev.c |  4 +++
> > >  drivers/net/ixgbe/ixgbe_ethdev.h |  6 ++++
> > >  drivers/net/ixgbe/ixgbe_rxtx.c   | 42 ++++++++++++++++++++++
> > >  lib/librte_ether/rte_ethdev.c    | 54 ++++++++++++++++++++++++++++
> > >  lib/librte_ether/rte_ethdev.h    | 77
> > > +++++++++++++++++++++++++++++++++++++++-
> > >  16 files changed, 440 insertions(+), 1 deletion(-)
> > >
> > > --
> > > 1.8.5.3



More information about the dev mailing list