Re: gem driver for sun4m

[Julian Coleman <jdc%coris.org.uk@localhost>]

Hi,

> Downloading seems to be consistently faster than uploading. Without TLS 
> overhead, both download and upload speeds scale pretty linearly with 
> frequency. Except when MP gets involved, and then at least on iperf3, it's 
> simply worse. Which I do find hilarious.

On sparc, interrupts are sent to one CPU, so I would expect at least the
same speed with MP as SP.  Just guessing, but I wonder if there is something
else in the stack that is causing this.

In the PR, you wrote:

>  For the gem driver itself, GEM_SBUS_CFG_BSIZE64 was picked
> experimentally. I had difficulty tracking down the appropriate
> documentation on the appropriate size, but it tolerated this value fine
> in the subsequent testing.

When I added the SBus GEM support, I also couldn't find any documentation
for the best values - the only information I have is the:

  Specification for Sbus GEM Gigabit Ethernet ASIC (100-5544-03)

On sparc64, it worked with GEM_SBUS_CFG_BSIZE128, so I left it with that.  
I notice that you have:

+#if defined(__sparc64__)
+       sbus_cfg = GEM_SBUS_CFG_BSIZE128 | GEM_SBUS_CFG_PARITY |
+           GEM_SBUS_CFG_BMODE64;
+#else
+       sbus_cfg = GEM_SBUS_CFG_BSIZE64;
+#endif

I assumed that enabling GEM_SBUS_CFG_PARITY made sense, so I wonder if it
would also work on sparc.  However, this is really just nit-picking.

Thanks for your work on this!

Regards,

Julian

--