From 8bdf8a640380fe3a010e319f6c3f7bca65fb9be4 Mon Sep 17 00:00:00 2001 From: Arthur Hanson Rasmusson Date: Wed, 30 Sep 2026 12:02:40 -0400 Subject: STATE-CURRENT: baseline regression (DtoH destination translation) and the targeted-fallback/retry work --- STATE-CURRENT.md | 23 +++++++++++++---------- 1 file changed, 13 insertions(+), 10 deletions(-) diff --git a/STATE-CURRENT.md b/STATE-CURRENT.md index 64824b3..fe2cd0f 100644 --- a/STATE-CURRENT.md +++ b/STATE-CURRENT.md @@ -1,6 +1,6 @@ # Open-IOV Current State -> **Snapshot: 2026-09-30 10:01.** This is a dated public status page, not live +> **Snapshot: 2026-09-30 11:43.** 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 known-knowns/unknowns document > is authoritative here: @@ -53,12 +53,15 @@ chain succeeds. class transfer bound on a compute channel whose work was denied by the engine-type gate; trusted host-injected pushes may now execute copy-engine work as well). -- **Runtime bring-up has started**: the harness now runs `nvidia-smi` and a - static-runtime `vectorAdd`-style client. The page-table walk resolves at - scale (thousands of successful walks per run), but the runtime's larger - buffers currently fail translation - host-side copies of 1 MB and smaller - report a missing-entry error on those virtual addresses. That translation - class is the new wall. +- **Runtime bring-up, with a baseline regression to clear first**: the + harness now runs `nvidia-smi` and a static-runtime `vectorAdd` client. The + runtime client currently fails at its first call - registration release + delivery lands seconds after an ~18 s deadline. Separately, the previously + verified copy baseline regressed: the device-to-host push is decoded and + executes, but its **destination** (a host staging address) has no + translation, so the copy fails with a missing-entry error. The operand + resolver now has a targeted allocation fallback and per-side miss tracing, + and a push whose operand is still under a scan is held and retried. - Recap of the fixes behind this point: all 20/20 compute-channel releases land (three decode paths routed to the exact counter slot); registration delivers the incrementing tracking release for every consumed submission; @@ -66,9 +69,9 @@ chain succeeds. the copy-engine decoder consults pinned translations before the scanned map, so every push-data operation executes (`memset=19473`, `ringops=10551` live). -- Active work: chase the runtime buffers' translations (the kernel-VA class), - with a CPU monitor now flagging kernel-VA translation failures; iterate on - the runtime path with one fresh boot per run. +- Active work: validate the targeted fallback resolves the device-to-host + destination (latest run still misses it), restore the verified 64/64 + baseline, then re-run the runtime client. - Observability: a size-keyed copy watch, a VRAM shadow dump/search, and an operation log are now module parameters, and the test's output survives guest-agent stalls via the serial console. -- cgit v1.2.3