Hi, David

From: David Laight <redacted> Data: Friday, May 30, 2014 11:35 PM

To: Duan Fugang-B38611; Eric Dumazet; ezequiel.garcia@free-electrons.com
Cc: Li Frank-B20596; davem@davemloft.net; netdev@vger.kernel.org;
shawn.guo@linaro.org; bhutchings@solarflare.com;
stephen@networkplumber.org
Subject: RE: [PATCH v1 4/6] net: fec: Increase buffer descriptor entry
number

From: fugang.duan@freescale.com

quoted

quoted

64KB TSO packet, with MSS=1460 -> 44 segments  (44*1460 = 64240) with
MSS=1448 (TCP timestamps) -> 45 segments  (45*1448 = 65160)

This software TSO emulation uses at least 2 descriptors per MSS

one descriptor to hold the headers (ethernet + ip + tcp) one
descriptor (or two) to hold the payload for this MSS

Thanks for Eric's detail explain.

If frag page data is not match the alignment for ethernet DMA
controller, there need three descriptor for one MSS:
One descriptor for headers, one for the first non-align bytes copied
from frag page, one for the rest of frag page.

So one frame may cost descriptor number is: 3 x 45

No - that is 45 frames, typically needing 3 ring entries each.

quoted

And I will add interrupt coalescing support for tx and rx, which also

cost some more descriptors.

quoted

So the descriptors slots set to 512 is not big, just is reasonable. Do

you think ?

Software TSO generates lots of separate ethernet frames, there is no
absolute requirement to be able to put all of them into the tx ring at
once.

The required size for the tx ring is much more likely to be related to any
interrupt mitigation that delays the refilling of ring entries.
512 sounds like a lot of tx ring entries.

The receive ring doesn't need to allow for fragments - your driver is in
control of allocating the buffers.
I didn't understand why you've based the number of rx ring entries on
PAGE_SIZE - IIRC that might be 64k, or even larger.
I'd have expected 128 or 256 rx ring entries to be typical, but how many
are needed is a separate issue.
If the system can keep up with the maximum ethernet data rate I'd expect
that a smaller number would be fine.
If it can't keep up you'll lose packets anyway.
Aggressive power saving (with wake on LAN) might need more.

David

Thanks for your information. I will do more test to weigh the smaller number for maximum data.
Summary you and Eric's thoughts, maybe rx set to 256, tx set to 512 are fine.

Thanks,
Andy