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

[Ananth N Mavinakayanahalli <ananth@in.ibm.com>]

On Tue, Jan 12, 2010 at 02:46:00PM -0800, Suresh Siddha wrote:
> On Sat, 2010-01-09 at 18:30 -0800, Ananth N Mavinakayanahalli wrote:
> > Linux version 2.6.33-rc3-bsect (ananth@llm69.in.ibm.com) (gcc version 4.1.2 20080704 (Red Hat 4.1.2-46)) #1 SMP Sun Jan 10 07:36:02 IST 2010
> > Command line: ro root=LABEL=/ rhgb console=tty0 console=ttyS0,9600n1 apic=debug
> > BIOS-provided physical RAM map:
> >  BIOS-e820: 0000000000000000 - 000000000009bc00 (usable)
> >  BIOS-e820: 000000000009bc00 - 00000000000a0000 (reserved)
> >  BIOS-e820: 00000000000e0000 - 0000000000100000 (reserved)
> >  BIOS-e820: 0000000000100000 - 00000000bff4b480 (usable)
> >  BIOS-e820: 00000000bff4b480 - 00000000bff57b40 (ACPI data)
> >  BIOS-e820: 00000000bff57b40 - 00000000c0000000 (reserved)
> >  BIOS-e820: 00000000d0000000 - 00000000e0000000 (reserved)
> >  BIOS-e820: 00000000fec00000 - 0000000100000000 (reserved)
> >  BIOS-e820: 0000000100000000 - 0000000840000000 (usable)
> > NX (Execute Disable) protection: active
> > DMI 2.4 present.
> > No AGP bridge found
> > last_pfn = 0x840000 max_arch_pfn = 0x400000000
> > x86 PAT enabled: cpu 0, old 0x7040600070406, new 0x7010600070106
> > last_pfn = 0xbff4b max_arch_pfn = 0x400000000
> > Scan SMP from ffff880000000000 for 1024 bytes.
> > Scan SMP from ffff88000009fc00 for 1024 bytes.
> > Scan SMP from ffff8800000f0000 for 65536 bytes.
> > Scan SMP from ffff88000009bc00 for 1024 bytes.
> > found SMP MP-table at [ffff88000009bd40] 9bd40
> >   mpc: 9d920-9dc84
> > init_memory_mapping: 0000000000000000-00000000bff4b000
> > init_memory_mapping: 0000000100000000-0000000840000000
> > RAMDISK: 37d4d000 - 37fef9e3
> > ACPI: RSDP 000000000009bde0 00014 (v00 M   IB)
> > ACPI: RSDT 00000000bff57ac0 00044 (v01 IBM    EXA01ZEU 00001000 IBM  45444F43)
> > ACPI: FACP 00000000bff57900 000F4 (v03 IBM    EXA01ZEU 00001000 IBM  45444F43)
> > ACPI: DSDT 00000000bff4b480 021B5 (v01 IBM    EXA01ZEU 00001000 INTL 20060707)
> > ACPI: FACS 00000000bff53780 00040
> > ACPI: APIC 00000000bff57800 000F4 (v01 IBM    EXA01ZEU 00001000 IBM  45444F43)
> 
> Ananth, This is an IBM EXA system using hurricane chipset.
> 
> And from http://www.redbooks.ibm.com/redbooks/pdfs/sg247630.pdf page
> 106, what is referred to as "special mode" in that page is the xapic's
> logical flat mode and what is referred as "logical mode" on that page is
> the xapic's logical cluster mode. And that page says that the logical
> mode is not used on x3850 M2 / x3950 M2 (it is not very clear to me that
> if the logical flat mode is fundamentally broken or not. but from your
> data and from this page, it looks like it).
> 
> In the past (long time back which predates x86_64 architecture)
> hurricane systems had issues with flat mode and hence 32bit architecture
> uses summit-specific code to handle this platform.
> 
> /* Hook from generic ACPI tables.c */
> static int summit_acpi_madt_oem_check(char *oem_id, char *oem_table_id)
> {
>         if (!strncmp(oem_id, "IBM", 3) &&
>             (!strncmp(oem_table_id, "SERVIGIL", 8)
>              || !strncmp(oem_table_id, "EXA", 3))){
>                 mark_tsc_unstable("Summit based system");
>                 use_cyclone = 1; /*enable cyclone-timer*/
>                 setup_summit();
>                 return 1;
>         }
>         return 0;
> }
> 
> So I think if you boot today's 32bit kernel on this platform, it will
> use this summit specific code which doesn't use logical flat mode.
> 
> I think we were just lucky why we haven't run into this problem before
> in 64-bit kernels.
> 
> Even with the existing code (after yesterday's revert by Linus), if all
> the apic id's happen to be less than or equal to 8, then 64-bit kernel
> will use flat mode again and that will cause an issue on your system
> again.

Thank you for the extensive analysis Suresh.

> Can you please check internally if the logical flat mode is broken on
> this platform and if so, either kernel need to have a summit specific
> check on 64bit too (just like 32bit) and not use flat mode or have the
> bios on your platform set the ACPI_FADT_APIC_PHYSICAL fadt flag set.
> Then the kernel will use physical mode irrespective of number of cpu's
> or their apic id's.

I have already sounded Chris McDermott, our System X platform person
about this issue. Chris should be able to give us more concrete answers
to your queries.

Ananth