Patch: Defer bus_config_intr() until bus_alloc_resource()..
John Baldwin
jhb at FreeBSD.org
Wed Jun 2 08:04:27 PDT 2004
On Tuesday 01 June 2004 06:33 pm, M. Warner Losh wrote:
> In message: <20040601221109.GA21063 at ns1.xcllnt.net>
>
> Marcel Moolenaar <marcel at xcllnt.net> writes:
> : On Tue, Jun 01, 2004 at 05:42:57PM -0400, John Baldwin wrote:
> : > > > > > I need the patch below in order to turn on bus_config_intr()
> : > > > > > when using the I/O APICs.
> :
> : *snip*
> :
> : > > > > I appreciate what you're trying to do but I don't like this
> : > > > > approach.
> :
> : *snip*
> :
> : > > > Well, arguably it exposes an improper layering violation when
> : > > > bus_config_intr() was added.
> :
> : *snip*
> :
> : > > I still don't like this.
> :
> : *snip*
> :
> : > FreeBSD has no format for this and I don't really have time to add
> : > multiple resource types (which is what ACPI does, and probably what we
> : > will need to do eventually) to struct resource.
> :
> : Ok, let's relax for a minute. Clearly, the addition of bus_config_intr()
> : was done in a biggest-bang-for-the-buck approach and I think we can
> : assume for now that it's there in a way that isn't easily extensible.
> :
> : So, the root problem now is that we need to consume the ACPI info in
> : a way that makes it available for later use, i.e., when we actually
> : allocate the IRQ resource. The solution needs to be simple so that
> : jhb@ can close the immediate problem and it should also be a step in
> : the right direction, or at least not a step in the opposite direction,
> : so that njl@ doesn't get cornered by it at some later time.
> :
> : Let me give the numbers:
> : o We need 2 bits for the polarity,
> : o We need 2 bits for the trigger mode,
> : o r_flags is 32 bits of which only 16 are in use.
> :
> : Can we not stick the IRQ properties in r_flags, just like we
> : stick the alignment properties there, remove bus_config_intr()
> : and make sure the MD backend queries the flags for polarity and
> : trigger (0 values will mean "default" and thus preserve backward
> : compatibility)?
> :
> : This change should only affect 2 places: the place where we set
> : the resource and the place where we actually install the interrupt
> : and its handler. It's roughly in line with what jhb@ wants to achieve
> : and I think it removes the code that njl@ is objecting to the most,
> : right?
> :
> : Oh, and it's not intended to be pretty.
>
> I like it. Or should I say "Given that we're looking for an
> evolutionary change to shoe horn this in, I think it is better than
> any of the other ugly kludges."
The only PITA is that when we route PCI interrupts we have to push the
trigger/polarity back to the alloc_resource() that called route_interrupt().
I guess we could just ignore ACPIs settings and hardcode level/low for PCI
interrupts. What I was thinking about on the way home, btw, was to have
struct resource be a base class sort of for resoures, and to have an
irq_resource, mem_resource, dma_resource, etc. Each class would have an
alloc() function that returns a template that is handed to
resource_list_alloc() so that you would do something like:
struct resource *res;
res = irq_resource_alloc(TRIGGER_LEVEL, POLARITY_LOW);
bus_alloc_resource(..., res);
And bus_alloc_resource() just fills out the common struct resource fields.
But really alloc_resource() should be overridden per type as well so that you
could make 'prefetchable' and 'alignment' properties of just memory
resources, but both of those factor into the alloc() algorithm. I guess we
just need a much better resource abstraction. *sigh*
--
John Baldwin <jhb at FreeBSD.org> <>< http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve" = http://www.FreeBSD.org
More information about the freebsd-acpi
mailing list