Re: [PATCH net-next v4 4/4] net: ocelot: add FDMA support

[Vladimir Oltean]

On Mon, Dec 06, 2021 at 10:28:46AM +0100, Clément Léger wrote:
> Le Sat, 4 Dec 2021 13:43:43 +0000,
> Vladimir Oltean <vladimir.oltean@xxxxxxx> a écrit :
> 
> > On Fri, Dec 03, 2021 at 06:19:16PM +0100, Clément Léger wrote:
> > > Ethernet frames can be extracted or injected autonomously to or from
> > > the device’s DDR3/DDR3L memory and/or PCIe memory space. Linked list
> > > data structures in memory are used for injecting or extracting Ethernet
> > > frames. The FDMA generates interrupts when frame extraction or
> > > injection is done and when the linked lists need updating.
> > > 
> > > The FDMA is shared between all the ethernet ports of the switch and
> > > uses a linked list of descriptors (DCB) to inject and extract packets.
> > > Before adding descriptors, the FDMA channels must be stopped. It would
> > > be inefficient to do that each time a descriptor would be added so the
> > > channels are restarted only once they stopped.
> > > 
> > > Both channels uses ring-like structure to feed the DCBs to the FDMA.
> > > head and tail are never touched by hardware and are completely handled
> > > by the driver. On top of that, page recycling has been added and is
> > > mostly taken from gianfar driver.
> > > 
> > > Co-developed-by: Alexandre Belloni <alexandre.belloni@xxxxxxxxxxx>
> > > Signed-off-by: Alexandre Belloni <alexandre.belloni@xxxxxxxxxxx>
> > > Signed-off-by: Clément Léger <clement.leger@xxxxxxxxxxx>
> > > ---  
> > 
> > Doesn't look too bad. Did the page reuse make any difference to the
> > throughput, or is the interaction with the FDMA extraction channel where
> > the bottleneck is?
> 
> With a standard MTU, the results did not improved a lot... TCP RX add a
> small improvement (~4MBit/s) but that is the only one.
> Here are the new results with the FDMA:
> 
> TCP TX: 48.2 Mbits/sec
> TCP RX: 60.9 Mbits/sec
> UDP TX: 28.8 Mbits/sec
> UDP RX: 18.8 Mbits/sec
> 
> In jumbo mode (9000 bytes frames), there is improvements:
> 
> TCP TX: 74.4 Mbits/sec
> TCP RX: 109 Mbits/sec
> UDP TX: 105 Mbits/sec
> UDP RX: 51.6 Mbits/sec

Yeah, I don't know what else to tell you. Sorry.

> > > diff --git a/include/soc/mscc/ocelot.h b/include/soc/mscc/ocelot.h
> > > index 11c99fcfd341..2667a203e10f 100644
> > > --- a/include/soc/mscc/ocelot.h
> > > +++ b/include/soc/mscc/ocelot.h
> > > @@ -692,6 +692,12 @@ struct ocelot {
> > >  	/* Protects the PTP clock */
> > >  	spinlock_t			ptp_clock_lock;
> > >  	struct ptp_pin_desc		ptp_pins[OCELOT_PTP_PINS_NUM];
> > > +
> > > +	struct ocelot_fdma		*fdma;
> > > +	/* Napi context used by FDMA. Needs to be in ocelot to avoid using a
> > > +	 * backpointer in ocelot_fdma
> > > +	 */
> > > +	struct napi_struct		napi;  
> > 
> > Can it at least be dynamically allocated, and kept as a pointer here?
> 
> If it is dynamically allocated, then container_of can't be used anymore
> in the napi poll function. I could move it back in struct fdma but
> then, I would need a backpointer to ocelot in the fdma struct.
> Or I could use napi->dev and access the ocelot_port_private to then get
> the ocelot pointer but I have not seen much driver using the napi->dev
> field. Tell me what you would like.

If you want to move it back to struct ocelot_fdma, you can do that, I'm
fine with that now :) Sorry for the trouble.