Read first: Read the mechanism and worked trace first
A read times out after occasional recovery events. Retain the request identity and reset/link epoch rather than correlating only its address; another request may later reuse that address.
Keep an error path in the operation record — decision record
Apply the stated local assumptions; record actual platform values before using this as an implementation checklist.
Decision / observation | Required evidence | Consequence |
|---|---|---|
Request outstanding | Requester/Tag, byte range and issue epoch | Retain unfinished byte accounting |
Recovery or reset observed | Protocol/platform cancellation and recovery policy | Do not guess that old tracking may immediately be reused |
Later Completion arrives | Identity and applicable epoch policy | Prevent matching an old result to unrelated new work |
Multiple operations may use the same address, and identifiers can later be reused. Retain the request identity and apply the protocol/platform’s recovery policy before recycling tracking.
Collect once, use across reviews: LTSSM history, per-lane margin, AER counters, TLP trace for failing transaction, config space snapshot, firmware enumeration log.
Protocol analyzer at slot level — correlates with driver timeline.
Interpret AER severity/status using the selected revision, mask/configuration and platform recovery policy.
A high replay rate is link-layer evidence; correlate physical margin, clocks, reset/recovery events, configuration and observation interval before assigning SI.