|
|
Chasing the sequence-dependent stall produced what looked like a major finding and then
refuted it. Replaying the fuzz stalls deterministically at invoke one thousand and eight,
plain identical zero-payload invokes stall at exactly the same index so it is count-driven
rather than content-driven, and a second application stalls at the same index, which
looked like a global resource leak letting any host wedge the root of trust with about a
thousand ordinary calls.
The check that killed it: that index is suspiciously near a power of two, so before
reporting a platform defect the host driver's ring submit was read. The write pointer
grows monotonically and is never wrapped against the ring size, so with the observed ring
size and pointer step the ring holds exactly one thousand and twenty-four submits, and
with the submits already used at boot the overflow is predicted one invoke later than the
observed stall.
The stall is therefore our own harness writing past its ring allocation, a host-side bug
rather than a platform one, retracted before it entered the findings. Two consequences are
kept: any observation requiring more than about a thousand invokes in one session is
suspect and must be re-run with a wrapping ring, and the immediate wedge found earlier is
unaffected since it fires within a handful of invokes.
|