Hi Tony, I have checked the ACPI and MPSpec too. As you have mentioned, those specifications have no detailed definition for default trigger mode and polarity setting, only give some examples and state "refer to the bus specification". On the other hand, our platform uses Intel Boxboro and ICH10 chipsets, which are the same chipsets used by Intel Emarald-Ridge platform. The ICH10 has a PCI-ISA bridge, so I assume it's an ISA bus instead of EISA. Then I referred to x86 IOAPIC implementation and found that default ISA IRQ trigger/polarity setting is edge-rising on x86, but level-low on IA64. As the MPSpec 1.4 table 4-10 states, EL 2:2 2 Trigger mode of APIC I/O input signals: 00 = Conforms to specifications of bus (for example, ISA is edge triggered) 01 = Edge-triggered 10 = Reserved 11 = Level-triggered So at lease IA64 has incorrectly set the default trigger mode to level-triggered, which violates the specifications. The scenario to trigger the issue on our platform is as below: 1) System has plenty of IO cards, vector allocation mode is set to "global" by default. Because the system has limited interrupt vectors, it begins to sharing interrupt when running out of vectors. With current implementation, it will share interrupt vector among PCI IRQs(level-low) and ISA IRQs with default setting, then interrupt storm happens. 2) If we switch the vector allocation mode to "percpu", system willhave plenty of interrupt vectors, so no interrupt sharing and no interrupt storm.
So I think IA64 may have incorrectly set the default trigger/polarity mode. But I'm not sure whether it will cause any compatibility issues with existing platforms. Thanks! On 2012-3-15 5:12, Tony Luck wrote:
When handling Interrupt Source Override in MADT table, the default ISA IRQ trigger model and poliarity should be edge-rising. Current IA64 implmentation doesn't follow the specification and set default ISA IRQ trigger model as level-low. With that wrong configuration and when system runs out of interrupt vectors, it will cause vector sharing among edge triggered ISA IRQ and level triggered PCI IRQ, then interrupt storm. So change the code to follow the specification.I chased the reference in ACPI to the MPS spec: http://www.intel.com/design/pentium/datashts/24201606.pdf Which says in table 4-10 "for example. EISA is active-low for level triggered interrupt" and "for example, EISA is edge-triggered". It looks to me that if MADT says "high", we should set high, if it says "low" we should set low. Ditto for trigger vs. level. But if we see the value "0" - then we have more work to do to determine what the bus we are on uses for its default. Did you find some other specification that gives a better explanation? Or did I miss something - can we only get to this routine for ISA? You've tested this patch - and it works for you - but is there a risk that it will break someone else? -Tony
-- To unsubscribe from this list: send the line "unsubscribe linux-ia64" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html