vgsp: m137 - CE-hold liveness watchdog (bound deferred CE waits across passes; targets the guest cudaMalloc stall class)

1 files changed, 32 insertions, 0 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 86cee81..ac2e561 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -10398,6 +10398,38 @@ static void svc_pump_channels(struct nv_vgpu_device *vgpu)
 			bitmap_zero(ch->slot_done_epoch, SVC_GPFIFO_MAX_BITS);
 		ch->put_prev = put;
 
+		/*
+		 * m136 (SA-STALL M2): liveness watchdog for a held CE entry.
+		 * When this channel's entry loop is not re-entered (early
+		 * continue paths above, or no fresh pushes), a deferred entry
+		 * past the grace window holds the client's tracking-semaphore
+		 * release forever - the guest cudaMalloc stall class. Bound
+		 * every wait here, mirroring the in-loop expiry at :11283:
+		 * verify the exact slot, re-seed and author the fake release
+		 * when it still lags, mark the slot consumed and move on.
+		 */
+		if (ch->engine_type && ch->defer_since && ch->entry_want &&
+		    time_after_eq(jiffies, ch->defer_since +
+				  msecs_to_jiffies(gsp_sema_ce_grace_ms))) {
+			if (!svc_entry_want_reached(vgpu, ch)) {
+				pr_info_ratelimited("sema-chan: watchdog re-seed h=0x%x rank=%u want=%u expected=%u (slot lags)\n",
+					ch->h_object, ch->ce_index,
+					ch->entry_want,
+					ch->expected_sem_payload);
+				ch->expected_sem_payload = ch->entry_want;
+				if (!svc_fake_release(vgpu, ch))
+					svc.gpfifo_starved++;
+			}
+			ch->rescue_fail = 0;
+			ch->defer_since = 0;
+			ch->unread_va = 0;
+			ch->unread_until = 0;
+			svc_replay_note(ch, (u32)ch->last_get, put);
+			ch->last_get = (ch->last_get + 1) %
+				ch->gpfifo_entries;
+			svc.gpfifo_skipped++;
+		}
+
 		get_before = ch->last_get;
 		applied_before = svc.releases_applied;