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

[Andrew Morton <akpm@osdl.org>]

On Sat, 14 Oct 2006 21:20:01 -0600
Matthew Wilcox <matthew@wil.cx> wrote:

> On Sat, Oct 14, 2006 at 01:48:55PM -0700, Andrew Morton wrote:
> > On Sat, 14 Oct 2006 08:02:49 -0600
> > Matthew Wilcox <matthew@wil.cx> wrote:
> > 
> > > On Fri, Oct 13, 2006 at 09:41:35PM -0700, Andrew Morton wrote:
> > > > Bisection shows that this patch
> > > > (pci-check-that-mwi-bit-really-did-get-set.patch in Greg's PCI tree) breaks
> > > > suspend-to-disk on my Vaio.  It writes the suspend image and gets to the
> > > > point where it's supposed to power down, but doesn't.
> > > 
> > > How odd.  What driver is calling pci_set_mwi() on the suspend path?
> > 
> > ehci_pci_reinit().  I stuck a dump_stack() in there.  See
> > http://userweb.kernel.org/~akpm/s5000342.jpg
> 
> Thanks for the picture; that's really helpful.
> 
> I see.  We hibernate all the devices then wake them all back up again.
> No doubt there's a good reason for this.
> 
> Still doesn't make much sense, though.  As far as I can see, the only
> consequence of this particular patch is that 1) we do an additional read
> from PCI_COMMAND and 2) we can return -EINVAL in one additional case.
> But the only effect of returning EINVAL is a printk (for this particular
> driver):
> 
>         /* PCI Memory-Write-Invalidate cycle support is optional (uncommon) */
>         retval = pci_set_mwi(pdev);
>         if (!retval)
>                 ehci_dbg(ehci, "MWI active\n");
> 
>         ehci_port_power(ehci, 0);
> 
>         return 0;
> 
> So even if we return EINVAL ... big deal.
> 
> Is it possible reading PCI_COMMAND too quickly after writing it causes
> a foul-up?  That would be weird ...
> 
> so I suppose there's a few things I can ask you to try:

It seems to have stopped happening.  I don't know why.

gregkh-pci-pci-prevent-user-config-space-access-during-power-state-transitions.patch
still breaks suspend though: http://userweb.kernel.org/~akpm/s5000349.jpg