From 3d4777f614e31f3c3a4f2ce49b37054a10fd1fd3 Mon Sep 17 00:00:00 2001
From: Arthur Hanson Rasmusson <arthur@vgpu.io>
Date: Fri, 9 Oct 2026 22:11:03 -0400
Subject: vgsp-sema: rewind last_get on a fresh channel's second put drop
 (ISSUE-0002)

r73 c3 rank2: discovery read the previous session's residual put (67); the
pump consumed the live entries 0..1, the r48 resync handled the first drop,
but the guest's counter then settled at put=1 while last_get=2 and
userd_wait_fresh was already consumed. The loop condition last_get != put
then marched ~1023 slots the producer had not written, holding each stale
entry for the full CE grace (one fake release per 8 s) - the client looked
wedged and the exec timed out.

When a drop leaves the pump ahead of the producer on a channel with fewer
than half a ring of submissions, rewind last_get to put instead of marching.
A genuine ring wrap needs a full ring of submissions, so this cannot skip a
live entry.
---
 kernel/gsp_sema_service.c | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index fa76162..e972732 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -13130,6 +13130,33 @@ static void svc_pump_channels(struct nv_vgpu_device *vgpu)
 			if (ch->userd_wait_fresh) {
 				ch->last_get = 0;
 				ch->userd_wait_fresh = false;
+			} else if (ch->engine_type && ch->gpfifo_entries &&
+				   ch->last_get > put &&
+				   ch->submissions < ch->gpfifo_entries / 2) {
+				/*
+				 * ISSUE-0002 (r73): a SECOND (or later) drop
+				 * while the channel is still fresh. The
+				 * discovery-time put was residual (the previous
+				 * session's counter, e.g. 67) and the pump had
+				 * already consumed the live entries 0..k; the
+				 * guest's counter then settled at a value below
+				 * the pump's read position (r73 rank2: put=1
+				 * with last_get=2 after the r48 resync). The
+				 * loop condition last_get != put then marches
+				 * forward through ~1023 slots the producer has
+				 * not written in this epoch, holding each
+				 * unreadable stale entry for the full CE grace
+				 * (one fake release per 8 s) - the client
+				 * appears wedged. Never stay ahead of the
+				 * producer: rewind to its position. A genuine
+				 * ring wrap cannot be fresh (it needs a full
+				 * ring of submissions), so this cannot skip a
+				 * live entry.
+				 */
+				pr_info_ratelimited("sema-chan: fresh put drop h=0x%x rank=%u put=%u last=%u -> rewind\n",
+					ch->h_object, ch->ce_index, put,
+					(u32)ch->last_get);
+				ch->last_get = put;
 			}
 			bitmap_zero(ch->slot_done_epoch, SVC_GPFIFO_MAX_BITS);
 		}
-- 
cgit v1.2.3