1 files changed, 25 insertions, 6 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 8e6401d..7bd1b26 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -47161,7 +47161,13 @@ static void svc_scrub_one(struct nv_vgpu_device *vgpu, struct svc_scrub *s)
 		s->put_last = put;
 		s->consumed = put ? put - 1 : s->entries - 1;
 	} else if (s->submitted == s->authored &&
-		   s->authored_gen == s->pa_gen) {
+		   s->authored_gen == s->pa_gen &&
+		   !(s->registered && !s->submitted)) {
+		/*
+		 * CONT.414: re-author the release pair every pass while a
+		 * scanner-registered buffer has no observed submissions, so a
+		 * guest rewrite of the polled word cannot re-block the wait.
+		 */
 		return;		/* nothing new; no guest writes */
 	}
 
@@ -47181,15 +47187,28 @@ static void svc_scrub_one(struct nv_vgpu_device *vgpu, struct svc_scrub *s)
 	 */
 	svc_write_ap(vgpu, s->userd_pa + SVC_USERD_GP_GET, s->userd_as,
 		     &put, sizeof(put));
+	/*
+	 * CONT.414: a scanner-registered channel buffer with no observed
+	 * USERD GP_PUT submissions is the X DDX's vidheap scrub wait: the
+	 * guest RM polls the finish payload at channel buffer + 0x6c004
+	 * (channelWaitForFinishPayload) while the host still reads GP_PUT=0,
+	 * so the put-based forward authoring would write sema=entries-1 and
+	 * fin=0 - the exact values that leave channelWaitForFreeEntry and
+	 * channelWaitForFinishPayload blocked (r1729: Xorg holds the NVRM
+	 * RM API write lock in that wait, glxgears and nvidia-modeset D on
+	 * os_acquire_rwlock_write). Author the proven all-ones release pair
+	 * (CONT.408/409) until a real put advance gives us progress to
+	 * mirror; 0xffffffff can never equal a putIndex below the ring size
+	 * and the folded finish counter is at its maximum.
+	 */
 	{
-		u32 sema = s->consumed;
+		bool release = s->registered && !s->submitted;
+		u32 idle = 0xffffffffu;
+		u32 sema = release ? idle : s->consumed;
+		u32 fin = release ? idle : (u32)s->submitted;
 
 		svc_write_ap(vgpu, s->hsem_pa, SVC_APERTURE_SYS, &sema,
 			     sizeof(sema));
-	}
-	{
-		u32 fin = (u32)s->submitted;
-
 		svc_write_ap(vgpu, s->fin_pa, SVC_APERTURE_SYS, &fin,
 			     sizeof(fin));
 	}