0.89I vs 0.90f: performance drop on receive?

[Chris Siebenmann]

 I have a large collection of Digital PCs (the specific model is the
Digital PC 5510) with on-motherboard 21143's that have, with 0.90f,
finally both been recognized properly and have been able to lock onto
100baseTX from our 3com switches; previous driver versions (including at
least 0.90 and 0.89I) have had various problems and would only come up
at 10mbits if left to autonegociate.

 However, I have noticed an anomaly: receive speeds (but not transmit
speeds) to a system running 0.90f are about 10 to 15 mbits slower (as
measured with ttcp). If the sender is using 0.90f, I can get some
30-second bursts of the full 100mbits; if the sender is running 0.89I, I
haven't seen that yet. Systems running 0.90f can send at full wire speed
to systems running 0.89I.

 Has anyone seen this before? Does anyone have any ideas or comments?
Is this a natural consequence of properly handling whatever particular
21143 version these PCs have at last?

	- cks
More information: 0.89I identifies these chips as
	eth0: Digital DS21142/3 Tulip at 0xec00, 00 00 f8 70 e7 2c, IRQ 9.
whereas 0.90f chimes in with:
	eth0: Digital DS21143 Tulip at 0xec00, 00 00 f8 70 e7 2c, IRQ 9.
0.89I is forced to 100baseTX with 'options=3' in modprobe.