vgsp-sema r616 diag: grfaultpoke - manually write a channel's error notifier (relayed-fault shape: info32=except, info16=engine, status=0xffff, notifier_err latch) and post its channel event, to test whether a synthetic engine-error edge ends a refused launch's cudaDeviceSynchronize spin (the authored sw release lands - measured 0x6 - but the kernel never completes for a never-executed launch)

1 files changed, 59 insertions, 0 deletions

diff --git a/kernel/main.c b/kernel/main.c
index e340b7b..9a5d912 100644
--- a/kernel/main.c
+++ b/kernel/main.c
@@ -3893,5 +3893,64 @@ static const struct kernel_param_ops nv_vgpu_rmpost_ops = {
 module_param_cb(rmpost, &nv_vgpu_rmpost_ops, NULL, 0200);
 MODULE_PARM_DESC(rmpost, "Post a GSP-RM event to the guest by event index (debug)");
 
+/*
+ * Debug: poke a channel's error notifier with an engine-exception shape and
+ * post its channel event (the relayed-fault delivery edge, manually):
+ *   echo 0x5c000019 > .../parameters/grfaultpoke
+ *   echo 0          > .../parameters/grfaultpoke   (every valid channel)
+ * Experiment: a refused/degenerate GR launch ends its client's
+ * cudaDeviceSynchronize spin if a synthetic engine error completes the wait
+ * (r616; see gsp_sema_channel_fault_poke).
+ */
+struct nv_vgpu_faultpoke_req {
+	u32 h_object;
+	int rc;
+	struct completion done;
+};
+
+static int nv_vgpu_faultpoke_thread(void *arg)
+{
+	struct nv_vgpu_faultpoke_req *req = arg;
+	struct nv_vgpu_device *vgpu;
+
+	mutex_lock(&nv_vgpu_driver.lock);
+	list_for_each_entry(vgpu, &nv_vgpu_driver.devices, list) {
+		if (atomic_read(&vgpu->vfio_dev_open_count) <= 0)
+			continue;
+		req->rc = gsp_sema_channel_fault_poke(vgpu, req->h_object);
+		pr_info("vgsp: grfaultpoke h=0x%x -> %d\n",
+			req->h_object, req->rc);
+	}
+	mutex_unlock(&nv_vgpu_driver.lock);
+
+	complete(&req->done);
+	return 0;
+}
+
+static int nv_vgpu_faultpoke_set(const char *val, const struct kernel_param *kp)
+{
+	static struct nv_vgpu_faultpoke_req req;
+	struct task_struct *task;
+
+	memset(&req, 0, sizeof(req));
+	req.rc = -ENODEV;
+	if (kstrtou32(val, 0, &req.h_object))
+		return -EINVAL;
+
+	init_completion(&req.done);
+	task = kthread_run(nv_vgpu_faultpoke_thread, &req, "vgpu-faultpoke");
+	if (IS_ERR(task))
+		return PTR_ERR(task);
+	wait_for_completion(&req.done);
+	return 0;
+}
+
+static const struct kernel_param_ops nv_vgpu_faultpoke_ops = {
+	.set = nv_vgpu_faultpoke_set,
+};
+
+module_param_cb(grfaultpoke, &nv_vgpu_faultpoke_ops, NULL, 0200);
+MODULE_PARM_DESC(grfaultpoke, "Write a channel error notifier + post its event (r616 experiment)");
+
 module_init(nv_vgpu_init);
 module_exit(nv_vgpu_exit);