sema-svc: sema_skip_ops parameter (apply releases without running pushed memsets)

3 files changed, 34 insertions, 7 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 0c0e85b..e21c997 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -2258,7 +2258,7 @@ static void svc_decode_stream(struct nv_vgpu_device *vgpu, const u8 *buf,
 				svc_rel_va_seen = va;
 				/* The push's work always runs; see the
 				 * REPORT_SEM comment below. */
-				if (allow_ops && n_ops) {
+				if (allow_ops && n_ops && !gsp_sema_skip_ops) {
 					svc_run_ops(vgpu, ops, n_ops);
 					svc.ring_ops++;
 					ops_runs++;
@@ -2341,11 +2341,11 @@ static void svc_decode_stream(struct nv_vgpu_device *vgpu, const u8 *buf,
 					 * through a trusted ring translation, so this
 					 * cannot run swept-RAM false positives.
 					 */
-					if (allow_ops && n_ops) {
-						svc_run_ops(vgpu, ops, n_ops);
-						svc.ring_ops++;
-						ops_runs++;
-					}
+				if (allow_ops && n_ops && !gsp_sema_skip_ops) {
+					svc_run_ops(vgpu, ops, n_ops);
+					svc.ring_ops++;
+					ops_runs++;
+				}
 					n_ops = 0;
 					if (svc_sem_read(vgpu, va, &cur))
 						break;
@@ -2542,7 +2542,7 @@ static void svc_decode_stream(struct nv_vgpu_device *vgpu, const u8 *buf,
 				 * Ops are only ever collected from pushes read
 				 * through a trusted ring translation.
 				 */
-				if (allow_ops && n_ops) {
+				if (allow_ops && n_ops && !gsp_sema_skip_ops) {
 					svc_run_ops(vgpu, ops, n_ops);
 					svc.ring_ops++;
 					ops_runs++;
@@ -3362,6 +3362,18 @@ unsigned gsp_sema_gr_grace_ms = 8000;
  */
 bool gsp_sema_fake_rel;
 
+/*
+ * Skip executing a push's collected work (CE memsets/copies into the vidmem
+ * shadow) while still applying its semaphore release. The client only ever
+ * observes the release; the shadow writes exist so the emulation can walk
+ * page tables it built, and the critical translations are injected from guest
+ * truth instead. Running thousands of memsets per context creation is what
+ * pushes the client past its context-init deadline; skipping them trades
+ * emulation-internal fidelity for the latency the client deadline requires.
+ * Off by default; enabled per run through the module parameter.
+ */
+bool gsp_sema_skip_ops;
+
 static unsigned long svc_gr_ring_tried;	/* ranks that got a diagnostic line */
 
 static void svc_gr_ring_autopin(struct nv_vgpu_device *vgpu,
diff --git a/kernel/gsp_sema_service.h b/kernel/gsp_sema_service.h
index 8cd9e61..615d092 100644
--- a/kernel/gsp_sema_service.h
+++ b/kernel/gsp_sema_service.h
@@ -78,6 +78,12 @@ extern bool gsp_sema_gr_rings;
 extern bool gsp_sema_fake_rel;
 
 /*
+ * Skip executing a push's collected work (CE memsets/copies) while still
+ * applying its release; see the definition for the trade-off.
+ */
+extern bool gsp_sema_skip_ops;
+
+/*
  * Grace window (ms) an untranslatable GR GPFIFO entry is left pending before
  * the pump consumes it as stale. The GR push pages are pinned from guest
  * truth a moment after the channels appear; the window lets those pins land
diff --git a/kernel/main.c b/kernel/main.c
index 952daf1..590d956 100644
--- a/kernel/main.c
+++ b/kernel/main.c
@@ -2441,6 +2441,15 @@ module_param_named(sema_fake_rel, gsp_sema_fake_rel, bool, 0644);
 MODULE_PARM_DESC(sema_fake_rel, "Acknowledge unreadable CE releases directly (default 0)");
 
 /*
+ * Skip executing the pushed work (CE memsets/copies) while still applying the
+ * semaphore releases. The client only observes releases; skipping the shadow
+ * writes removes the per-push execution cost that dominates the context-init
+ * latency budget.
+ */
+module_param_named(sema_skip_ops, gsp_sema_skip_ops, bool, 0644);
+MODULE_PARM_DESC(sema_skip_ops, "Apply releases without executing pushed memsets (default 0)");
+
+/*
  * Grace window (ms) an untranslatable GR entry stays pending before the pump
  * consumes it as stale, so a late push-buffer pin can still make it readable.
  */