diff options
| -rw-r--r-- | kernel/gsp_sema_service.c | 20 |
1 files changed, 20 insertions, 0 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index 6a1595e..4a86235 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -11498,6 +11498,26 @@ static void svc_counter_publish(struct nv_vgpu_device *vgpu, if (!pa) return; + /* + * COMPLETION GATE (r205 mail/matmul regression): GP_PUT is the + * client's SUBMISSION position; the pump's last_get is its next + * unconsumed entry. While any submitted entry is outstanding + * (last_get != put) the pushes up to put have NOT all completed, so + * the put must never be published into the completion word: doing so + * short-circuits every client wait (r205 mail: rank-0's word jumped + * to the submitted 108 at last_get=104 - `pub-adv h=0x5c000019 + * pa=...ff0 val=108 cur=104` - ~1 s before the engine ran + * `output retirement wait: 1/968 ... settled`, so the host's + * sub-put release payloads were then forward-only-skipped and the + * client read a stale buffer and failed; matmul's stage ordering + * broke the same way). Only the release applies deliver completions; + * this publish is the catch-up for entries the pump consumed WITHOUT + * a decoded release, so it may run solely when nothing is + * outstanding - the same equality the entry loop consumes to + * (wrap-safe for the 1024-entry rings). + */ + if (ch->last_get != put) + return; svc.counter_pub_calls++; if (gsp_sema_dbg && !ch->engine_type && svc.counter_pub_dbg < 96) { |