Re: 8139cp: set ring address before enabling receiver

On Wed, 2012-11-21 at 15:40 -0500, Jeff Garzik wrote:

On 11/21/2012 03:18 PM, Ben Hutchings wrote:

quoted

On Wed, 2012-11-21 at 19:51 +0000, David Woodhouse wrote:

quoted

On Wed, 2012-11-21 at 13:12 -0500, Jeff Garzik wrote:

quoted

What sticks out at me from the commit message?

It was not tested on the famously quirky 8139 hardware at all.

While I have not looked at the 8139C+ data sheet in a while, sometimes
the hardware _did_ have a strange init order.

As this works in a simulator but fails on real hardware, it seems like
an obvious regression caused by an untested [on read hardware] patch.

The data sheet (v1.6, from http://realtek.info/pdf/rtl8139cp.pdf ) says
in §6.33 (C+ Command Register):
  "Enable C+ mode functions in C+CR register first,
  => Enable transmit/receive in Command register (offset 37h),
  => Configure other related registers (ex. Descriptor start address,
     TCR, RCR, ...)."

I understand the concern expressed in the offending commit message about
DMA happening to invalid addresses, and I'll look at the data sheet
harder to see when the DMA actually starts happening. But it definitely
seems that our current code isn't doing what the data sheet says.

I wonder if I can find one of these lying around and stick it in a
machine with an IOMMU...

You might be able to avoid disaster by doing:

1. Set MAC filter to drop everything
2. Enable RX DMA
3. Set RX DMA ring address
4. Set MAC filter according to current flags & multicast list

I'm assuming, knowing nothing about this particular hardware, that the
MAC filter register(s) will accept writes before RX DMA is enabled.

A larger point is that the commit was created to avoid imagined disaster 
on simulated hardware...

  ...and wound up creating behavior that is (a) contra to the data sheet 
and (b) breaks real hardware.

I wasn't suggesting anyone should change this again without testing on
real hardware.  But the 'imagined disaster' seems to be an obvious and
real race condition, which the driver is just more likely to win when
racing real hardware than when racing virtual hardware.

(It could be that the hardware pre-fetches DMA descriptors, in which
case this is a 'how did that ever work?' bug.  Alternately, there could
be a hidden enable bit that doesn't get set until the RX DMA ring
address is written, in which case the driver may need a quirk for
emulations that lack that.  An IOMMU should be able to answer these
questions.)

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.