Compaq EV67 Network Card User Manual


 
Alpha 21264/EV67 Hardware Reference Manual
Error Detection and Error Handling 8–3
Dcache Data Single-Bit Correctable ECC Error
3. The virtual address associated with the error is available in the VA register.
4. The PALcode flushes the error block by temporarily disabling
DC_CTL[DCTAG_PAR_EN] and evicting the block using two HW_LD instruc-
tions. The onchip duplicate tag provides the correct victim address and cache
coherence state.
If a retried load instruction detects the Dcache tag parity error, the memory reference
may have already been retired, so the EXC_ADDR is not available. In this case, the
error is uncorrectable and the Mbox performs the following actions:
Either DC_STAT[TPERR_P0] or DC_STAT[TPERR_P1] is set, indicating the
source of the error.
When enabled, a machine check (MCHK) is posted. The MCHK is taken when not
in PALmode.
8.4 Dcache Data Single-Bit Correctable ECC Error
The following operations may cause Dcache data ECC errors:
Load instructions
Stores of less than quadword length
Dcache victim read transactions
The hardware flow used for Dcache data ECC errors depends on the event that
caused the error.
8.4.1 Load Instruction
Loads that read data from the Dcache may do so either in the same cycle as the Dcache
tag probe (typical case) or in some subsequent cycle (load-queue retry). The hardware
functional flows for these two error cases differ slightly.
When a load instruction reads the Dcache data array in the same cycle as the tag array,
if an ECC error occurs on the LSD ECC error detectors, then the Ibox stops retiring
instructions and does not resume retiring until after hardware recovers from the error.
If an ECC error occurs on the LSD ECC error detectors, when a load instruction reads
the Dcache tag array before it reads the Dcache data array, then the load instruction may
have already been retired. In either case:
The incorrect data is written into the load instruction’s destination register;
however, the load queue retains the state associated with the load instruction.
A consumer of the load instruction’s data may be issued before the error is
recognized; however, the Ibox will invoke a replay trap at an instruction that is
older than (or equal to) any instruction that consumes the load instruction’s data,
and then stalls the replayed Istream in the map stage of the pipeline until the error is
corrected.
Given a READ_ERR read-type from the Mbox for the error load instruction, the
Cbox scrubs the block in the Dcache by evicting the block into the victim buffer
(thereby scrubbing it) and writing it back into the Dcache as follows:
C_STAT[DSTREAM_DC_ERR] is set.