On Tue, Oct 13, 2020 at 05:03:58PM -0700, Jakub Kicinski wrote: > On Sun, 11 Oct 2020 11:09:41 +0200 Willy Tarreau wrote: > > while running some tests on my Breadbee board, I noticed poor network > > Tx performance. I had a look at the driver (macb, at91ether variant) > > and noticed that at91ether_start_xmit() immediately stops the queue > > after sending a frame and waits for the interrupt to restart the queue, > > causing a dead time after each packet is sent. > > > > The AT91RM9200 datasheet states that the controller supports two frames, > > one being sent and the other one being queued, so I performed minimal > > changes to support this. The transmit performance on my board has > > increased by 50% on medium-sized packets (HTTP traffic), and with large > > packets I can now reach line rate. > > > > Since this driver is shared by various platforms, I tried my best to > > isolate and limit the changes as much as possible and I think it's pretty > > reasonable as-is. I've run extensive tests and couldn't meet any > > unexpected situation (no stall, overflow nor lockup). > > > > There are 3 patches in this series. The first one adds the missing > > interrupt flag for RM9200 (TBRE, indicating the tx buffer is willing > > to take a new packet). The second one replaces the single skb with a > > 2-array and uses only index 0. It does no other change, this is just > > to prepare the code for the third one. The third one implements the > > queue. Packets are added at the tail of the queue, the queue is > > stopped at 2 packets and the interrupt releases 0, 1 or 2 depending > > on what the transmit status register reports. > > LGTM. There's always a chance that this will make other > designs explode, but short of someone from Cadence giving > us a timely review we have only one way to find that out.. :)
Not that much in fact, given that the at91ether_* functions are only used by AT91RM9200 (whose datasheet I used to do this) and Mstar which I used for the tests. I initially wanted to get my old SAM9G20 board to boot until I noticed that it doesn't even use the same set of functions, so the potential victims are extremely limited :-) > Applied, thanks! Thank you! Willy
