Bogus Uncorrectable Error reported on bit 32, syndrome 13

There is a problem in the Starfire's XDB algorithm that checks the syndrome bit to identify the bad bit and determine if it is a single or multiple error. The XDB is coded to expect a syndrome of 12 for bit 32. The syndrome for bit 32 really is 13. The result is that the XDB will request a Recordstop but instead of recording a single bit error (CE), it will record an multiple bit error (UE).

From the wfail output, we see something like the following:

redxl> wfail
(output omitted)
XDB   7.2   EccErrFlags[11:0] = 220
        EccFlg[5]: Uncorrectable error in ldat bus lo half, bits [71:0]
        EccFlg[11:8]: Error count = 2
ldat[ 71: 0]= D3 00FD0ECD 00000004 (xmux_par[5:0]= 02)  syn= 13: bit 32 [06]
(output omitted)

Bear in mind that the UE is misreported by the XDB only. Solaris detects and reports this error properly. As a result, only the Recordstop Dump File will reflect a UE with Bit 32 in error in the XDB output. The flip-side of this problem will be the XDB reporting a Syndrome 12 Correctable Error, but not identify which Bit was Corrected. In reality, Syndrome 12 maps to an Uncorrectable Error (UE), and cannot be mapped to a single bit.

 

• Only a recordstop is generated.
• Solaris properly detects, handles, and reports the error properly. Only the XDB output in the Recordstop file is in error.
• This error is in XDB code and has no relation to system board hardware.
• This problem will not be fixed.

 

Correctable ECC Error (CE) Processor X Dtags

From the wfail output, we see something like the following:

redxl> wfail
(output omitted)
CIC   7.2   ErrFlags[61:0] = 00000001 00000002   (after mask)
         ErrFlag[1]: Correctable ECC Error   (CE)  Processor 1 Dtags
         ErrFlag[32]: Repeated Error
     Proc 1 Dtag ECCSyn[13: 8] = 23:  CE: bit 00  Dtag SRAM 7.2.0
  FAIL Proc 7.1 in all  configs using CIC2: : Arbstop/Recordstop detected by cic
         (*** NOTE: Implicated FRU is sysboard 7)
(output omitted)

The above error should be analyzed in a way consistent with other Correctable Error recordstops. This means that the first instance or event for any given error against a particular DTag SRAM (in this example, CIC 7.2- DTag SRam 0) should be diagnosed as a soft error, and no action should be taken against it.

Swap the "Implicated FRU" (SB7 in example) when the third failure occurs on any one CIC.

NOTE: Blacklisting the affected CPU (proc 7.1 in example) could be used as a short term workaround.