vgsp: add gsp_phys_rmcall_channel to force one GR channel variant

Lets a bisection pair any promote shape with any channel shape, in
particular V6 (SKIP_CTXBUFFER_ALLOC, KMD-managed) with the server-managed
VA-only promote: with skip-ctxbuffer the server's kgrctxMapCtxBuffers does
not pre-book the engine context vaList, so the VA branch's
vaListSetManaged(NV_FALSE) gate should pass and the per-buffer commit can
program the engine context VA. Default 0 keeps the mode-selected walk
byte-identical. Not hardware-tested (GPU wedged; see Handoff ADDENDUM 51).

3 files changed, 60 insertions, 0 deletions

diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index fae657c..3a65361 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -7249,6 +7249,40 @@ static int gsp_phys_rmcall_alloc_channel(void __iomem *bar0)
 		vars = &gsp_phys_rmcall_channel_variants[0];
 		attempts = 1;
 	}
+	/*
+	 * Optional explicit force (gsp_phys_rmcall_channel=N): pin the walk to
+	 * one variant after the mode-based selection so a bisection can pair
+	 * any promote shape with any channel shape (e.g. V6 skip-ctxbuffer +
+	 * the srvctx VA-only promote).
+	 */
+	if (gsp_phys_rmcall_engine_is_gr() &&
+	    gsp_phys_rmcall_channel_value() != 0) {
+		u32 force = (u32)gsp_phys_rmcall_channel_value();
+		u32 j;
+		bool found = false;
+
+		for (j = 0; j < ARRAY_SIZE(gsp_phys_rmcall_channel_variants);
+		     j++) {
+			if (gsp_phys_rmcall_channel_variants[j].id != force)
+				continue;
+			vars = &gsp_phys_rmcall_channel_variants[j];
+			found = true;
+			break;
+		}
+		if (!found &&
+		    gsp_phys_rmcall_channel_variant_skipctx.id == force) {
+			vars = &gsp_phys_rmcall_channel_variant_skipctx;
+			found = true;
+		}
+		if (!found) {
+			pr_err("vgsp-boot: rm-call: gsp_phys_rmcall_channel=%u is not a known variant id (1/3/4/5/6); aborting the ladder\n",
+			       force);
+			return -EINVAL;
+		}
+		attempts = 1;
+		pr_info("vgsp-boot: rm-call: gsp_phys_rmcall_channel=%u forces channel variant V%u (%s) alone\n",
+			force, vars[0].id, vars[0].name);
+	}
 
 	ret = -EIO;
 	for (i = 0; i < attempts; i++) {
diff --git a/kernel/gsp_phys_gpu.c b/kernel/gsp_phys_gpu.c
index 94d6652..54fa84a 100644
--- a/kernel/gsp_phys_gpu.c
+++ b/kernel/gsp_phys_gpu.c
@@ -201,6 +201,25 @@ bool gsp_phys_srv_ctxcommit_value(void)
 EXPORT_SYMBOL_GPL(gsp_phys_srv_ctxcommit_value);
 
 /*
+ * Force one channel-alloc variant id on the GR ladder (default 0 = the
+ * promote-mode-selected default). 5 = V5 (client memdescs, no skip-ctxbuffer;
+ * the server books its managed engine context), 6 = V6 (SKIP_CTXBUFFER_ALLOC,
+ * KMD-managed: the server skips the ctx-buffer map/book, so a client VA-only
+ * promote can add and commit the engine context VA).
+ */
+static int gsp_phys_rmcall_channel_param;
+module_param_named(gsp_phys_rmcall_channel, gsp_phys_rmcall_channel_param,
+		   int, 0644);
+MODULE_PARM_DESC(gsp_phys_rmcall_channel,
+		 "Force one GR channel-alloc variant id (1/3/4/5/6); 0 = mode-selected default walk");
+
+int gsp_phys_rmcall_channel_value(void)
+{
+	return gsp_phys_rmcall_channel_param;
+}
+EXPORT_SYMBOL_GPL(gsp_phys_rmcall_channel_value);
+
+/*
  * P5 wave-1 shadow work forwarding (default 0). When set, the sema service
  * requests one real CE copy on the debugfs ladder's channel for each decoded
  * guest CE push, using only the ladder's module-owned buffers; the software
diff --git a/kernel/gsp_phys_gpu.h b/kernel/gsp_phys_gpu.h
index a270df8..925bfb2 100644
--- a/kernel/gsp_phys_gpu.h
+++ b/kernel/gsp_phys_gpu.h
@@ -90,6 +90,13 @@ bool gsp_phys_srv_ctxbuf_value(void);
 bool gsp_phys_srv_ctxcommit_value(void);
 
 /*
+ * gsp_phys_rmcall_channel: force one channel-alloc variant id on the GR
+ * ladder (0 = the promote-mode-selected default). 5 = V5 (client memdescs,
+ * no skip-ctxbuffer), 6 = V6 (SKIP_CTXBUFFER_ALLOC, KMD-managed ctx).
+ */
+int gsp_phys_rmcall_channel_value(void);
+
+/*
  * True when the gsp_phys_work_forward module parameter is set (default 0: the
  * guest's CE pushes stay software-emulated exactly as before). When set, the
  * sema service shadows each decoded CE push as one copy on the debugfs ladder's