* [patch non-x86] PCI-PCI bridges fix
@ 2001-11-22 12:04 Ivan Kokshaysky
2001-11-22 21:49 ` Paul Mackerras
0 siblings, 1 reply; 3+ messages in thread
From: Ivan Kokshaysky @ 2001-11-22 12:04 UTC (permalink / raw)
To: Linus Torvalds; +Cc: Richard Henderson, linux-kernel
Bridge resources weren't added to the resource tree. This was
harmless for most real-life configurations (especially on Alpha),
but certainly it will be a problem with the hotplug stuff
(thinking of ES45).
Ivan.
--- 2.4.15p9/drivers/pci/setup-bus.c Fri Oct 5 05:47:08 2001
+++ linux/drivers/pci/setup-bus.c Wed Nov 21 20:16:24 2001
@@ -201,6 +201,16 @@ pbus_assign_resources(struct pci_bus *bu
b->resource[0]->end = ranges->io_end - 1;
b->resource[1]->end = ranges->mem_end - 1;
+ /* Add bridge resources to the resource tree. */
+ if (b->resource[0]->end > b->resource[0]->start &&
+ request_resource(bus->resource[0], b->resource[0]) < 0)
+ printk(KERN_ERR "PCI: failed to reserve IO "
+ "for bus %d\n", b->number);
+ if (b->resource[1]->end > b->resource[1]->start &&
+ request_resource(bus->resource[1], b->resource[1]) < 0)
+ printk(KERN_ERR "PCI: failed to reserve MEM "
+ "for bus %d\n", b->number);
+
pci_setup_bridge(b);
}
}
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [patch non-x86] PCI-PCI bridges fix
2001-11-22 12:04 [patch non-x86] PCI-PCI bridges fix Ivan Kokshaysky
@ 2001-11-22 21:49 ` Paul Mackerras
2001-11-23 12:36 ` Ivan Kokshaysky
0 siblings, 1 reply; 3+ messages in thread
From: Paul Mackerras @ 2001-11-22 21:49 UTC (permalink / raw)
To: Ivan Kokshaysky; +Cc: Linus Torvalds, Richard Henderson, linux-kernel
Ivan Kokshaysky writes:
> Bridge resources weren't added to the resource tree. This was
> harmless for most real-life configurations (especially on Alpha),
> but certainly it will be a problem with the hotplug stuff
> (thinking of ES45).
>
> Ivan.
>
> --- 2.4.15p9/drivers/pci/setup-bus.c Fri Oct 5 05:47:08 2001
> +++ linux/drivers/pci/setup-bus.c Wed Nov 21 20:16:24 2001
> @@ -201,6 +201,16 @@ pbus_assign_resources(struct pci_bus *bu
> b->resource[0]->end = ranges->io_end - 1;
> b->resource[1]->end = ranges->mem_end - 1;
>
> + /* Add bridge resources to the resource tree. */
> + if (b->resource[0]->end > b->resource[0]->start &&
> + request_resource(bus->resource[0], b->resource[0]) < 0)
> + printk(KERN_ERR "PCI: failed to reserve IO "
> + "for bus %d\n", b->number);
There is code in pci_read_bridge_bases which will decide under some
circumstances that a PCI-PCI bridge is transparent and set
child->resource[i] = child->parent->resource[i];
for i = 0, 1 or 2. What will happen with your request_resource in
this case is that the resource will end up being its own parent and
its own child and all allocations against it will fail. :)
I hit this just yesterday on PPC. On our RS/6000 boxes, if you have a
PCI-PCI bridge with nothing behind it, the firmware will configure it
with all the apertures closed, i.e. with base set larger than limit.
Then pci_read_bridge_bases goes and decides that the bridge is
transparent and consequently everything stops working. :(
Does anyone really need that "assuming transparent" stuff in
pci_read_bridge_bases? If so we need to add a check for the case
where base == limit + 0x1000 (for I/O) or base == limit + 0x100000
(for memory) and conclude that the aperture is closed in that case.
Paul.
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [patch non-x86] PCI-PCI bridges fix
2001-11-22 21:49 ` Paul Mackerras
@ 2001-11-23 12:36 ` Ivan Kokshaysky
0 siblings, 0 replies; 3+ messages in thread
From: Ivan Kokshaysky @ 2001-11-23 12:36 UTC (permalink / raw)
To: Paul Mackerras; +Cc: Linus Torvalds, Richard Henderson, linux-kernel
On Fri, Nov 23, 2001 at 08:49:30AM +1100, Paul Mackerras wrote:
> There is code in pci_read_bridge_bases which will decide under some
> circumstances that a PCI-PCI bridge is transparent and set
>
> child->resource[i] = child->parent->resource[i];
>
> for i = 0, 1 or 2. What will happen with your request_resource in
> this case is that the resource will end up being its own parent and
> its own child and all allocations against it will fail. :)
These code paths shouldn't be mixed in the first place, i.e. if
you use stuff in setup-bus.c, don't use pci_read_bridge_bases.
Basically, these are two different approaches to PCI setup:
"We trust our firmware" (i386, ia64) and "We are the firmware"
(alpha, arm, partially parisc).
> I hit this just yesterday on PPC. On our RS/6000 boxes, if you have a
> PCI-PCI bridge with nothing behind it, the firmware will configure it
> with all the apertures closed, i.e. with base set larger than limit.
Right, as required by PCI-PCI bridge specs.
> Then pci_read_bridge_bases goes and decides that the bridge is
> transparent and consequently everything stops working. :(
I think that "assuming transparent" code is not correct. Yes, PCI-PCI
bridge specs allow "transparent" (i.e. subtractive decoding) bridges,
but such bridges must have a class code 0x60401. IMO, we need something
like this instead:
if (!dev) /* It's a host bus, nothing to read */
return;
+ if (dev->class & 1) {
+ printk("Subtractive decoding bridge %s: assuming transparent\n",
+ dev->name);
+ for(i=0; i<3; i++)
+ child->resource[i] = child->parent->resource[i];
+ return;
+ }
Of course, this requires testing on problematic machines.
BTW, currently almost every PC user sees those "Unknown bridge resource"
messages just because a graphic card behind the AGP bridge doesn't use
IO or prefetchable memory...
Ivan.
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2001-11-23 12:37 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2001-11-22 12:04 [patch non-x86] PCI-PCI bridges fix Ivan Kokshaysky
2001-11-22 21:49 ` Paul Mackerras
2001-11-23 12:36 ` Ivan Kokshaysky
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®