Messages in this thread | | | Date | Tue, 10 Nov 2009 00:15:36 +0000 (GMT) | From | "Maciej W. Rozycki" <> | Subject | Re: [PATCH -tip] x86,apic: Use PAGE_SIZE instead of numbers |
| |
On Mon, 9 Nov 2009, H. Peter Anvin wrote:
> In theory we could have more than one ioapic packed into a single page, > and it is also entirely plausible we'll support other page sizes in x86 > at some point. However, it's probably easier to flag something as > PAGE_SIZE and have to fix it up later than have magic constants, so I > think it's probably the right thing to do.
Hmm, the MPS said in Chapter 3.6.5 "APIC Memory Mapping":
"Non-default APIC base addresses can be used if the MP configuration table is provided. (Refer to Chapter 4.) However, the local APIC base address must be aligned on a 4K boundary, and the I/O APIC base address must be aligned on a 1K boundary."
This probably still stands; I suppose it would be safer to define IOAPIC_SLOT_SIZE to 1024 and use it by default, expanding all reservations as needed where less than PAGE_SIZE / IOAPIC_SLOT_SIZE I/O APICs would be mapped in a page. This is relatively recent a piece of code -- how much has it been tested?
Well, actually not much as a quick search has revealed this message:
http://marc.info/?l=linux-kernel&m=118114792006520
which shows page alignment of I/O APICs clearly does not stand, and moreover there are two pairs of I/O APIC in the system reported which share a page each. In this case the ranges requested do not make sense and some resource insertions will silently fail. And also while page aliases created in fixmaps here should not harm, they make me feel a little bit chilly...
Overall this piece of code needs an overhaul, fixing resource allocation and reusing fixmaps where possible.
Maciej
| |