So I accidentally noticed this morning that my previous batch of
e1000 link negotiation patches (specifically, commit 39bb8ee737...)
only fix the problem on q35, but (at least on OS X 10.8 and 10.9)
not also on piix.
After digging around a bit, I came up with the following two issues:
- 1/2 set_phy_ctrl() seems to act on the value about to be written
to the phy ctrl. register, but not actually write the value
to the register! Assuming that's not intentional, once I fix
it and write the register, I can use the factored-out check
for autoneg availability from earlier ( have_autoneg() ) if
I move it before set_phy_ctrl() to avoid a forward declaration.
- 2/2 the timer delay requested from set_phy_ctrl() is currently 500,
which, at least on piix + OS X + the built-in e1000 driver, fires
too early (on piix, OS X takes longer to start paying attention
to LSC interrupts than it does on q35, for some reason I'm not
yet clear about). Increasing the timer delay to 5500 (by trial
and error, possibly specific to my dev machine -- a macpro1,1)
has it fire, bring the link up, and inject the LSC late enough
that the OS X driver will be ready for it.
I'm curious what you all think of this. If e.g. patch 1/2 makes sense,
I'd be OK with having it go in without 2/2, which may feel a bit hacky :)
But mainly I'm looking to be educated about what I might have missed :)
Thanks much,
Gabriel
Gabriel L. Somlo (2):
e1000: clean up set_phy_ctrl function
e1000: adjust initial autoneg timing (for piix/osx)
hw/net/e1000.c | 27 ++++++++++++++-------------
1 file changed, 14 insertions(+), 13 deletions(-)
--
1.9.3
