Robert Mustacchi <firstname.lastname@example.org>
RFD 137 CPU Autoreplacement and ID Synthesis
Today, the illumos fault management architecture (FMA) watches for issues with hardware and notes when they are faulty. One of the things that it looks after are issues with CPUs. For example, if a CPU has a series of errors occur on it, FMA may end up making the decision to offline the CPU.
In such cases, as part of offlining the CPU, FMA opens a case that records that the part is faulty. The case can subsequently be closed either because a component is replaced or the operator explicitly acquits the component. The way that FMA knows that a component has been replaced is based upon serial number or similar information that is part of the FRU (field replaceable unit) and contained in the FMRI (fault management resource indicator).
Consequently, if a component lacks such information, FMA will not be able to automatically close the case and it will fall to the operator to ensure that the proper component was replaced and inform the system.
As it happens, this is a major problem. This problem is magnified due to the labeling issues outlined in the current draft of our RAS Specifications. To summarize, the CPU socket and the silkscreen often do not align.
Traditionally CPUs have not exposed much identifying information. During the Pentium III era, there was a short-lived serial number exposed in cpuid. However, this was deemed to have privacy concerns, especially because all software could read it.
In recent server SKUs (multi-socket Xeons), Intel has offered a means of identifying a processor through what they call PPIN, which is intended for inventory management. See the Public E5 v4 Product Family Documentation for more information. Based on this, it is possible for us to construct something that we can use for RAS purposes. This information will not be available to a non-global zone. It will require cooperation from the systems firmware.
Based on available information, will we synthesize an illumos-specific ID string for a processor. This scheme will vary based on the vendor and will allow itself to be versioned in subsequent generations of systems.
The basic scheme is the following:
iv0 portion of the name is used to indicate that this is illumos
version zero. It is expected that for a given processor, the scheme
should not change so as not to cause the system to think that the CPUs
in the system were replaced. However, this allows a newer generation to
use a different schema.
The next part of the string is the name of the vendor. So, in this case we would use INTC. If we had an AMD scheme for this, we would use 'AMD' in its place. Finally, the latter part is specific to the actually processor family.
Intel Xeon Scheme
The initial scheme will have the following form:
Here both the CPUID signature and the PPIN will be printed as hexadecimal values. These will not be zero padded values. There will be no leading 0x used for either entry.
Use of versioning
The core use of versioning here is to deal with a subsequent CPU generation changing the required information for uniqueness. The version is specific to the vendor. Let's look at an example of how such versioning might occur.
Let's call the vendor of such processors LINK and assume that their initial scheme looked like Intel's. However, in their new generation of processors, let's imagine what would happen if they changed to providing a classic RFC 4112 UUID.
For all previous, supported generations of LINK based processors,
nothing would change. They would use the existing
However, the new generation and onwards would have a system that looked
To emphasize, the version number here really represents a generation of processors that have their own unique way of numbering. If multiple generations of a vendor's processor are guaranteed to be unique and use the same schema then we should not change this. However, incrementing the version number when we have these changes will allow us to more easily tell that these things cannot be directly comparable, especially if for some reason there would be overlap.
If this changed again in another generation, then we would move to an
Exposure via FM
To consume this information, a new routine will be added to the Intel
cpu_module.h header which will allow it to obtain a
const char *
value that has the string, assuming that one exists. If it does, it will
be added to the
nvlist_t payload that is consumed by
FM_IOC_PHYSCPU_INFO ioctl. This will then be consumed
by the corresponding FM topology modules in user land to allow for
autoreplace to work.
One of the biggest challenges that we have with this scheme is that the FRU of the processor can change in this scheme. This can happen for a couple of reasons:
FM detected an error on an older system and we rebooted to a system that now has support for the synthetic serial number.
FM detected an error on a new system that supports the synthetic serial numbers and rebooted to a system prior to this support.
FM detected an error on a new system that supports the synthetic serial numbers and a BIOS change locked the ability to synthesize the serial number.
FM detected an error on a new system that does not support the synthetic serial number, but was rebooted and a BIOS change unlocked the ability to synthesize the serial number.
All of these cases represent instances where we cannot know for certain whether or not a replacement occurred. In fact, the odds are high that the replacement did not occur. As such, we need to make sure that FM is enhanced such that in all of these cases a fault is not considered to have been autoreplaced and that instead the operator must manually clear such faults.
This means that the only time the hc schema CPU module would consider a replacement as having occured is if the original faulted FMRI had a synthesized serial number and the replacement part did as well. Othewrise, we will always consider the fault as persisting.
The only caveat wiht this approach is that if we have one of these changes and the part has not been replaced, we can end up with two FRUs that both have faults (one with and one without). However, because these FRUs will refer to the same logical CPUs, we should still see them being retired.