Error Classifications

Errors are categorized into three classes that specify the severity of an error as listed below. Note also the specification defines the entity that should handle the error based on its severity:

• Correctable errors — handled by hardware

• Uncorrectable errors-nonfatal — handled by device-specific software

• Uncorrectable errors-fatal — handled by system software

By defining errors into these classes, error handling software can be partitioned into separate handlers to perform the actions required of a given platform. The actions taken based on severity of an error might range from monitoring the effects of correctable errors on system performance to simply resetting the system in the event of a fatal error.

Note that regardless of the severity of a given error, software can establish a policy whereby the system is notified of all errors for the purpose of tracking and logging each category.

Correctable Errors

The specification defines correctable errors as those errors that are corrected solely by hardware. Such errors may have an impact on performance (i.e., latency and bandwidth), but no information is lost as a result of the error. These types of error can be reported to software, which can take a variety of actions, Including:

• log the error

• update calculations of PCI Express performance

• track errors to project possible weaknesses within the fabric. This can suggest areas where greater potential exists for fatal errors in the future.

Uncorrectable Non-Fatal Errors

Non-fatal errors mean that the integrity of the PCI Express fabric is not affected, but that information has been lost. Non-fatal errors typically mean that a transaction has been corrupted, and the PCI Express hardware cannot correct the error. However, the PCI Express fabric continues to function correctly and other transactions are unaffected. Recovery from a non-fatal error may or may not be possible. The possibility of recovery rests in the hands of the device-specific software associated with the requester that initiated the transaction.

Uncorrectable Fatal Errors

Fatal errors indicate that a link in the PCI Express fabric is no longer reliable. Data has been lost and every attempt to recover data under software control will likely fail also. Such conditions affect all transactions that traverse a given link. In some cases, the error condition leading to the fatal errors may be resolved by resetting the link. Alternatively, the specification invites implementation-specific approaches, in which software may attempt to limit the effects of the failure. The specification does not define any particular actions that should or could be taken by software.