> > Hi Eric,
> >
> > There is another problem with that patch - it broke ES7000, I kept
> > getting timer panics. It turned out that check_timer() runs
> before the
> > actual APIC destination is set up. The IO-APIC uses
> > cpu_to_logical_apicid to find the destination - which needs
> > cpu_2_logical_apicid[] to be filled - which only happens after
> > processors are booted. At the time when check_timer() runs, it will
> > always be BAD_APICID (0xFF - broadcast) as the IO-APIC rte
> destination
> > for the timer, but ES7000 hardware happened not to support
> 0xFF so it
> > panics. I used bios_cpu_apicid[] to bring it up, but
> > cpu_to_logical_apicid is the only one that is kept
> up-to-date in the
> > hotplug case, so I cannot replace it in the cpu_mask_to_apicid().
> >
> > There are probably some ways to fix this such as one below that I
> > tried (in mpparse.c):
> >
> > if (m->mpc_cpuflag & CPU_BOOTPROCESSOR) {
> > Dprintk(" Bootup CPU\n");
> > boot_cpu_physical_apicid = m->mpc_apicid;
> > + cpu_2_logical_apicid[num_processors] =
> m->mpc_apicid;
> > }
> > it worked, but looks more like a kludge of course. I think IO-APIC
> > setup has to happen after processors were brought online and so is
> > check_timer(), if timer is connected through the IO-APIC.
>
> The first cpu is brought online much earlier than the rest.
> So we just need to setup a table for boot cpu earlier. From
> the looks of it mach-es700 won't work if you compile a
> uniprocessor kernel for it right now.
Yea, I didn't even try this - but I think it will produce the same
result with regard to timer IOAPIC rte.
> We need to do this a little later than in mptable but this
> should be a fairly simple one or two line change.
Yes, it is maybe something like running map_cpu_to_logical_apicid() from
APIC_init() just before the setup_IO_APIC().
Thanks,
--Natalie
> If people keep breaking the subarchitectures by accident we
> might even inspire someone to make a comprehensible sub
> architecture implentation on x86 one of these days.
>
>
> Eric
>
"Protasevich, Natalie" <[email protected]> writes:
>> The first cpu is brought online much earlier than the rest.
>> So we just need to setup a table for boot cpu earlier. From
>> the looks of it mach-es700 won't work if you compile a
>> uniprocessor kernel for it right now.
>
> Yea, I didn't even try this - but I think it will produce the same
> result with regard to timer IOAPIC rte.
>
>> We need to do this a little later than in mptable but this
>> should be a fairly simple one or two line change.
>
> Yes, it is maybe something like running map_cpu_to_logical_apicid() from
> APIC_init() just before the setup_IO_APIC().
The core piece of the puzzle is cpu_mask_to_apicid(). At the
time we setup the io_apic TARGET_CPUS will just be the bootstrap
processor. So we might be able to get away with hard coding
the bootstrap processor in the non-SMP sections of the ioapic
startup code.
setup_ioapic_dest is going to fix things after we start the
cpus anyway so it should not be a problem.
Does that sound like a sane thing to do?
This code appears to affect all of the subarchitectures but the
default x86 one. So it is clearly not just an ES7000 problem.
Now to figure out why Linus's laptop hates this patch...
Eric