diff options
| -rw-r--r-- | STATE-CURRENT.md | 32 |
1 files changed, 28 insertions, 4 deletions
diff --git a/STATE-CURRENT.md b/STATE-CURRENT.md index ce10d48..aacb166 100644 --- a/STATE-CURRENT.md +++ b/STATE-CURRENT.md @@ -1,7 +1,7 @@ # Open-IOV Current State > **Snapshot: 2026-10-05 (reorganized; simulator status current 2026-10-04, -> physical/guest campaign data as of 2026-10-06 ~10:55; debugging-options +> physical/guest campaign data as of 2026-10-06 ~18:15; debugging-options > catalogue added 2026-10-05).** This is a dated public > status page, not live state; it is refreshed on every progress check. Live > working notes are maintained outside this repository; the public @@ -96,9 +96,33 @@ isolation, performance, portability). front-end (PBDMA), not by the GR engine, so they do not prove engine execution. A full individual acceptance-sweep cycle was live at snapshot time. The intermittent guest `cudaMalloc` stall class now has concrete - (unvalidated) fix candidates: pool-VA resolution ordering before fallback, - and verification of an unreadable entry before it is consumed. Host - clean; pstore empty. + (unvalidated) fix candidates: pool-VA resolution ordering before fallback, + and verification of an unreadable entry before it is consumed. Host + clean; pstore empty. **Update 2026-10-06 ~18:15:** the forwarded-launch + campaign closed the context and deposit gaps and exhausted the + method-form space. (a) The TSG-sibling commit lane (mode 31) gives the + execution channel a committed engine context (`ENG_CTX_VALID` set, + stage 3/4 green, no wedge) - the KMD skip-ctx channel alone cannot; + FBMEM context buffers, a free ChID for the sibling, and dropping the + shared-global FECS entry from the sibling's promote were each required + and accepted. (b) The inline-QMD copy was proven to execute: when the + destination is a module page inside the engine's proven ring-lookup + window the front-end deposits the QMD (the guest-VA destination never + received one despite an accepted PDE install) - a correction to the + earlier "destination stays zero" note. (c) With a committed context and + a working deposit the schedule is still absent across every method form + (stock window-only window, explicit copy PCAS, the QMD-virtualization + enable, and the two-step require-scheduling flow); the program-pointer + page offset is now preserved and the runner gained a launch-failure + fast-exit plus a flake-only retry wrapper. (d) A wired golden sibling + cannot carry a class object (constructs refused before/after its + promote), closing the golden-execution lane without a shared + subcontext. Ranked next: the context-share (VEID subcontext) route; an + init-fidelity sweep (shared-memory window and valid-span overflow area); + resolving the privileged register aperture for a real front-end / + scheduler readback; and a supervised working-channel capture for a + function-level diff. Host clean, GPU green, same boot, no reboots + (in-session function-level reset recovery only). - **Guest frontier: FIRST GREEN END-TO-END GUEST CUDA, NOW UNDER HARDENING (2026-10-04 ~16:20; hardening 2026-10-05 overnight).** The emulation-only ("no-claim") campaign runs an unmodified guest CUDA program to completion |