From c4069a4bff26f430a8a4614a5fab62abc24a8722 Mon Sep 17 00:00:00 2001
From: Serguei Mokhov <mokhov@cse.concordia.ca>
Date: Thu, 22 Feb 2024 09:09:04 -0500
Subject: [PATCH] feat(inputs.amd_rocm_smi): Add startup_error_behavior config
 option (#14872)

---
 plugins/inputs/amd_rocm_smi/README.md         |  6 +++
 plugins/inputs/amd_rocm_smi/amd_rocm_smi.go   | 33 ++++++++++++--
 .../inputs/amd_rocm_smi/amd_rocm_smi_test.go  | 45 +++++++++++++++++++
 plugins/inputs/amd_rocm_smi/sample.conf       |  6 +++
 4 files changed, 86 insertions(+), 4 deletions(-)

diff --git a/plugins/inputs/amd_rocm_smi/README.md b/plugins/inputs/amd_rocm_smi/README.md
index c8439bbc3..ae0cbd56e 100644
--- a/plugins/inputs/amd_rocm_smi/README.md
+++ b/plugins/inputs/amd_rocm_smi/README.md
@@ -22,6 +22,12 @@ See the [CONFIGURATION.md][CONFIGURATION.md] for more details.
   ## Optional: path to rocm-smi binary, defaults to $PATH via exec.LookPath
   # bin_path = "/opt/rocm/bin/rocm-smi"
 
+  ## Optional: specifies plugin behavior regarding missing rocm-smi binary
+  ## Available choices:
+  ##   - error: telegraf will return an error on startup
+  ##   - ignore: telegraf will ignore this plugin
+  # startup_error_behavior = "error"
+
   ## Optional: timeout for GPU polling
   # timeout = "5s"
 ```
diff --git a/plugins/inputs/amd_rocm_smi/amd_rocm_smi.go b/plugins/inputs/amd_rocm_smi/amd_rocm_smi.go
index fa9dddb40..eb890259e 100644
--- a/plugins/inputs/amd_rocm_smi/amd_rocm_smi.go
+++ b/plugins/inputs/amd_rocm_smi/amd_rocm_smi.go
@@ -23,8 +23,12 @@ var sampleConfig string
 const measurement = "amd_rocm_smi"
 
 type ROCmSMI struct {
-	BinPath string
-	Timeout config.Duration
+	BinPath              string          `toml:"bin_path"`
+	Timeout              config.Duration `toml:"timeout"`
+	StartupErrorBehavior string          `toml:"startup_error_behavior"`
+	Log                  telegraf.Logger `toml:"-"`
+
+	ignorePlugin bool
 }
 
 func (*ROCmSMI) SampleConfig() string {
@@ -33,8 +37,8 @@ func (*ROCmSMI) SampleConfig() string {
 
 // Gather implements the telegraf interface
 func (rsmi *ROCmSMI) Gather(acc telegraf.Accumulator) error {
-	if _, err := os.Stat(rsmi.BinPath); os.IsNotExist(err) {
-		return fmt.Errorf("rocm-smi binary not found in path %s, cannot query GPUs statistics", rsmi.BinPath)
+	if rsmi.ignorePlugin {
+		return nil
 	}
 
 	data := rsmi.pollROCmSMI()
@@ -46,6 +50,27 @@ func (rsmi *ROCmSMI) Gather(acc telegraf.Accumulator) error {
 	return nil
 }
 
+func (rsmi *ROCmSMI) Init() error {
+	if _, err := os.Stat(rsmi.BinPath); os.IsNotExist(err) {
+		binPath, err := exec.LookPath("rocm-smi")
+		if err != nil {
+			switch rsmi.StartupErrorBehavior {
+			case "ignore":
+				rsmi.ignorePlugin = true
+				rsmi.Log.Warnf("rocm-smi not found on the system, ignoring: %s", err)
+				return nil
+			case "", "error":
+				return fmt.Errorf("rocm-smi binary not found in path %s, cannot query GPUs statistics", rsmi.BinPath)
+			default:
+				return fmt.Errorf("unknown startup behavior setting: %s", rsmi.StartupErrorBehavior)
+			}
+		}
+		rsmi.BinPath = binPath
+	}
+
+	return nil
+}
+
 func init() {
 	inputs.Add("amd_rocm_smi", func() telegraf.Input {
 		return &ROCmSMI{
diff --git a/plugins/inputs/amd_rocm_smi/amd_rocm_smi_test.go b/plugins/inputs/amd_rocm_smi/amd_rocm_smi_test.go
index e38e0ff89..9d508cf25 100644
--- a/plugins/inputs/amd_rocm_smi/amd_rocm_smi_test.go
+++ b/plugins/inputs/amd_rocm_smi/amd_rocm_smi_test.go
@@ -11,6 +11,51 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+func TestErrorBehaviorError(t *testing.T) {
+	// make sure we can't find rocm-smi in $PATH somewhere
+	os.Unsetenv("PATH")
+	plugin := &ROCmSMI{
+		BinPath:              "/random/non-existent/path",
+		Log:                  &testutil.Logger{},
+		StartupErrorBehavior: "error",
+	}
+	require.Error(t, plugin.Init())
+}
+
+func TestErrorBehaviorDefault(t *testing.T) {
+	// make sure we can't find rocm-smi in $PATH somewhere
+	os.Unsetenv("PATH")
+	plugin := &ROCmSMI{
+		BinPath: "/random/non-existent/path",
+		Log:     &testutil.Logger{},
+	}
+	require.Error(t, plugin.Init())
+}
+
+func TestErrorBehaviorIgnore(t *testing.T) {
+	// make sure we can't find rocm-smi in $PATH somewhere
+	os.Unsetenv("PATH")
+	plugin := &ROCmSMI{
+		BinPath:              "/random/non-existent/path",
+		Log:                  &testutil.Logger{},
+		StartupErrorBehavior: "ignore",
+	}
+	require.NoError(t, plugin.Init())
+	acc := testutil.Accumulator{}
+	require.NoError(t, plugin.Gather(&acc))
+}
+
+func TestErrorBehaviorInvalidOption(t *testing.T) {
+	// make sure we can't find rocm-smi in $PATH somewhere
+	os.Unsetenv("PATH")
+	plugin := &ROCmSMI{
+		BinPath:              "/random/non-existent/path",
+		Log:                  &testutil.Logger{},
+		StartupErrorBehavior: "giveup",
+	}
+	require.Error(t, plugin.Init())
+}
+
 func TestGatherValidJSON(t *testing.T) {
 	tests := []struct {
 		name     string
diff --git a/plugins/inputs/amd_rocm_smi/sample.conf b/plugins/inputs/amd_rocm_smi/sample.conf
index aed15aae9..b4fb9ebb1 100644
--- a/plugins/inputs/amd_rocm_smi/sample.conf
+++ b/plugins/inputs/amd_rocm_smi/sample.conf
@@ -3,5 +3,11 @@
   ## Optional: path to rocm-smi binary, defaults to $PATH via exec.LookPath
   # bin_path = "/opt/rocm/bin/rocm-smi"
 
+  ## Optional: specifies plugin behavior regarding missing rocm-smi binary
+  ## Available choices:
+  ##   - error: telegraf will return an error on startup
+  ##   - ignore: telegraf will ignore this plugin
+  # startup_error_behavior = "error"
+
   ## Optional: timeout for GPU polling
   # timeout = "5s"