Re: [PATCH v1] x86/platform/intel-mid: Make IRQ allocation a bit more flexible

[Bjorn Helgaas]

On Fri, Aug 04, 2017 at 04:43:15PM +0300, Andy Shevchenko wrote:
> On Fri, 2017-08-04 at 08:24 -0500, Bjorn Helgaas wrote:
> > On Fri, Aug 04, 2017 at 03:09:39PM +0300, Andy Shevchenko wrote:
> > > On Thu, 2017-08-03 at 17:08 -0500, Bjorn Helgaas wrote:
> > > > On Mon, Jul 24, 2017 at 08:34:02PM +0300, Andy Shevchenko wrote:
> > > > > In the future we would use dynamic allocation for IRQ which
> > > > > brings
> > > > > non-1:1 mapping for IOAPIC domain. Thus, we need to respect
> > > > > return
> > > > > value
> > > > > of mp_map_gsi_to_irq() and assign it back to the device
> > > > > structure.
> > > > > 
> > > > > Besides that we need to read GSI from interrupt pin register to
> > > > > avoid
> > > > > cases when some drivers will try to initialize PCI device twice
> > > > > in a
> > > > > row
> > > > > which will call pcibios_enable_irq() twice as well.
> > > > 
> > > > This seems sort of suspect.  It doesn't seem robust to rely on the
> > > > value of PCI_INTERRUPT_LINE being zero before
> > > > pcibios_enable_irq().
> > > 
> > > So, you are telling that it's possible to get pcibios_enable_irq()
> > > called with PCI_INTERRUPT_LINE == 0?
> > 
> > I don't know about that.  I'm just saying that it sounds like you're
> > relying on some condition that is hard to verify.  PCI_INTERRUPT_LINE
> > is read/write and it's hard to know what, if anything, BIOS did with
> > it.
> 
> _On this platform_ BIOS provides a static mapping for IRQ lines (there
> is no separate PIR, just IOAPIC). So, that register is initialized by
> firmware. We are not suppose to touch it _on this platform_.
> 
> > > > Can't we make pcibios_enable_irq() idempotent?  I guess I don't
> > > > understand the real problem here.
> > > 
> > > Currently there is no problem here.
> > > 
> > > Firmware assigns IRQ line and wrote this to the configuration space.
> > > Intel MID uses 1:1 mapping (IRQ domain is STRICT), so, it just
> > > allocates
> > > a vector inside kernel with the very same number.
> > > 
> > > But...
> > > 
> > > If we switch to dynamic allocation (it's a default for ACPI case),
> > > _on
> > > this platform_ we will get wrong allocation in the IOAPIC since
> > > mapping
> > > is not 1:1 anymore.
> > > 
> > > That's what I'm trying to avoid (dev->irq after this patch points
> > > correctly to dynamically allocated number).
> > > 
> > > For old SFI enabled platforms it will not make any change.
> > > 
> > > Note, there is no legacy PIC on this platform (as same case as for
> > > ACPI
> > > HW reduced platforms), just IOAPIC.
> > > 
> > > > Is this really two separate patches that could be separated?
> > > 
> > > I didn't get this, what separation you see might be applied here?
> > 
> > The "Besides that" in your changelog suggested that "here is another
> > change lumped into this patch."  I was wondering if the "read GSI from
> > PCI_INTERRUPT_LINE" part could be separated out.  That makes me wonder
> > whether this is really safe, since GSIs are 32 bits, but
> > PCI_INTERRUPT_LINE is only 8 bits.  I don't know enough to know why
> > this is safe in this case.
> 
> "Besides that" part means "we have a second circumstance we have to
> follow to avoid breaking working platforms".
> 
> The old code, if you look into patch, uses dev->irq as a predefined
> constant, which is not in case of dynamic allocation, so, if we respect
> return value from mp_map_gsi_to_irq() we may not use dev->irq as
> _source_ of IRQ mapping anymore.
> 
> I didn't check if we can do things in reversed order than I have done,
> i.e. apply part 2 as separate change and then respect the return value.
> It looks like it may be done in this way.
> 
> However, the patch is applied and I don't see any strong reason
> to revert it and bring again as a set of two patches.

I agree, thanks for your explanations.