Re: [PATCH 1/2] [PCI] Check that MWI bit really did get set

[Andrew Morton <akpm@osdl.org>]

On Mon, 16 Oct 2006 10:44:30 +1000
Paul Mackerras <paulus@samba.org> wrote:

> Andrew Morton writes:
> 
> > Let me restore the words from my earlier email which you removed so that
> > you could say that:
> > 
> >   For you the driver author to make assumptions about what's happening
> >   inside pci_set_mwi() is a layering violation.  Maybe the bridge got
> >   hot-unplugged.  Maybe the attempt to set MWI caused some synchronous PCI
> >   error.  For example, take a look at the various implementations of
> >   pci_ops.read() around the place - various of them can fail for various
> >   reasons.  
> 
> Maybe aliens are firing a ray-gun at the card.  I think it's
> fundamentally wrong for the driver to deny service completely because
> of a maybe.
> 
> Either there was a transient error that only affected the attempt to
> set MWI, in which case a printk (inside pci_set_mwi!) is appropriate,
> and we carry on.  Or there is a persistent error condition, in which
> case the driver will see something else fail soon enough - something
> that the driver actually needs to have working in order to operate -
> and fail at that point.
> 
> For the driver to stop and refuse to go any further because of an
> error in pci_set_mwi has far more disadvantages than advantages.
> 

Sure.

So I think what we're needing in this case is:

- A modified version of Willy's patch which returns 0 if MWI was enabled,
  1 if MWI isn't available.

- A printk if something went bad

  It appears that the various functions which try to match the line sizes
  already have printks if something went wrong, but they're using
  KERN_DEBUG facility level and that warning would dupliate the new one in
  pci_set_mwi().

  And although the various implementations of pci_read_config_foo() and
  pci_write_config_foo() can return error codes, we have so many instances
  where we're not checking it, I don't think it's practical to try to start
  doing that everywhere.

- drop the __must_check.

Question is, should pci_set_mwi() ever return -EFOO?  I guess it should, in
the case where setting the line size didn't work out.