Re: [PATCH] Make Intel 8-way Xeons boot again

[Yinghai Lu <yinghai@kernel.org>]

On Sat, Jan 9, 2010 at 10:11 AM, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
>
> On Sat, 9 Jan 2010, Ananth N Mavinakayanahalli wrote:
>>
>> On an 8-way system with Intel Xeon X7350 CPUs, booting 2.6.32 or newer
>> kernels fails at:
>>
>> ...
>> CPU0: Intel(R) Xeon(R) CPU           X7350  @ 2.93GHz stepping 0b
>> Booting Node   0, Processors  #1 #2 #3 #4 #5 #6 #7 Ok.
>> Brought up 8 CPUs
>> Total of 8 processors activated (46906.05 BogoMIPS).
>>
>> Git bisect showed 2fbd07a5f as the offending commit.
>
> Ok, that commit definitely is buggy.
>
>> With the patch below, I am able to boot the latest Linus' git tree on
>> the machine. If this patch is correct, it needs to get into the stable
>> tree too.
>
> I don't think the patch is correct, though. The thing is, the AMD check
> seems to be the correct one: you can only use 'apic_flat' if all the APIC
> ID's are < 8.
>
> It doesn't matter _how_ many CPU's you have. If you have two CPU's, but
> one of them has an APIC ID >= 8, then you cannot use the flat APIC model,
> since it depends on a 8-bit bitfield.
>
> So your patch doesn't seem right either, because it still tests
> num_processors, which is bogus.
>
> In fact, I can't for the life of me understand why it treats different
> vendors differently. Why is that code not just a simple
>
>        /* Flat apic mode requires that all APIC ID's are in the range 0..7 */
>        if (apic == &apic_flat && max_physical_apicid >= 8)
>                apic = &apic_physflat;
>
> instead, with no crazy vendor tests.
>
> What am I missing?

according to Suresh, intel CPUs could use logical flat mode when total
num_cpus <=8 even some cpu's physical apicid > 0.

and init_apic_ldr should set ldr to the cpu to map cpu index to the cpu.

static void flat_init_apic_ldr(void)
{
        unsigned long val;
        unsigned long num, id;

        num = smp_processor_id();
        id = 1UL << num;
        apic_write(APIC_DFR, APIC_DFR_FLAT);
        val = apic_read(APIC_LDR) & ~APIC_LDR_MASK;
        val |= SET_APIC_LOGICAL_ID(id);
        apic_write(APIC_LDR, val);
}

in Ananth's case, APs are started, so LDR should be set correctly.

YH