Hygon RAS -Data Poisoning Feature

一、设计思想：

系统致命问题本地化处理

Poison Data粒度：Cache line粒度 64Byte

Poison Data传播：propagates througn with poison status in Memor、Cache、Link wtih

Deferred Errors：不可纠正尚未消费 Uncorrected and have not be consumed

故障缓解：比如硬件重配、Memory page retired

二、Log：

1）专用，不可覆盖

MCA_DESTAT

MCA_DEADDR

2） 非专用，可覆盖

if MCA_CONFIG[LogDeferredErrorsInMcaStat]==1, it can be loged by the MCA_STATUS and MCA_ADDR registers.

A deferred error logged in these registers may be overwritten by a later uncorrected or fatal error.

三、通告：

可配置产生中断 if MCA_CONFIG[DeferredIntTypeSupported]==1, MCA_CONFIG[DeferredIntType]

四、潜在源

Clear Poison

CLZERO is published in Core::X86::Cpuid::FeatureExtIdEbx[CLZERO]. CLZERO

is the only mechanism that ensures removal of the poison indicator from a cache line

Enabled

DRAM ECC must be enabled