PCI Error Recovery API Proposal (updated)

[Hidetoshi Seto]

Benjamin Herrenschmidt wrote:
>>>The definition of "pci_error_token" is not covered here. 
>>
>>What is the default type of pci_error_token in API 1)? You said "within
>>this function and after it returns, the driver shouldn't do any new
>>IOs." AER code is required to pass error severity (fatal or nonfatal) to
>>a driver when calling API 1). I refer this error token should be defined
>>as an integer type, which is passed with either PCIERR_FATAL_DETECTED or
>>PCIERR_NONFATAL_DETECTED. Please let me know what you think? 
> 
> The token should be an opaque type with accessors. You could define a
> pci_error_get_severity(token) to return the severity. The idea is to
> define accessors which return an error when the data requested isn't
> present in the error info. The actual content of the token is to be
> defined. I was thinking about a type plus a union. I was hoping Seto
> could provide something here ...

I agree that the token should be an opaque, implement-depends thing.
For example, it could be a bitmask like:

#define PCIERR_ERROR_DETECTED    0x00000001  /* fundamental, always ON */
#define PCIERR_VALID_INFO        0x00000010  /* optional */
#define PCIERR_SEVERITY_FATAL    0x00000100  /* optional */
#define PCIERR_SEVERITY_NONFATAL 0x00000200  /* optional */
  :

or to get more detail:

#define PCIERR_ERROR_DETECTED    0x00000001  /* fundamental, always ON */
#define PCIERR_VALID_PCIE_AER    0x00010000  /* optional */
  :
and define a function like:
  int pci_error_get_severity(pci_error_token *t)
  int pci_error_get_pcie_aer(pci_error_token *t, pcie_aer_bits *dat)

or (I think it would be better,) merge of them, etc.

My thought was, depends on the arch, some could have only a error bit,
some could have various info with well-defined struct/union/etc.

Anyway, the token should be a easy handle for basic use, and also
should be useful for advanced use.


Thanks,
H.Seto