19879 nvidia_smi: don't crash when no GPU is found

The bash agent plugin emits the section header before running
nvidia-smi -q -x, so any non-XML output (Linux "No devices were
found", Windows "ERROR: nvidia-smi.exe was not found in: ...")
was passed verbatim to ElementTree.fromstring and crashed
parse_nvidia_smi with ParseError. Treat ParseError as an empty
section.

Crash-Group-ID: 3616
Jira: SUP-28831
AI-Generated: true
Change-Id: I6bf22854809ab0ba4c1d5598a6162a53a8ac9fb0

Author: otAAAh

.werks/19879.md

@@ -0,0 +1,25 @@
+[//]: # (werk v3)
+# nvidia_smi: don't crash when no GPU is found
+
+key        | value
+---------- | ---
+date       | 2026-05-13T08:23:12.201670+00:00
+version    | 2.6.0b1
+class      | fix
+edition    | community
+component  | checks
+level      | 1
+compatible | yes
+
+When the agent host had `nvidia-smi` installed but no GPU was currently
+visible to the driver (the Linux plugin then forwarded `No devices were
+found`), or when `nvidia-smi.exe` was not installed at the path the
+Windows agent plugin searches (the plugin forwarded a `not found in:`
+error block), the agent emitted a plain-text message instead of the
+expected XML. The `nvidia_smi` section then crashed with `ParseError:
+syntax error: line 1, column 0` and produced a crash report on every
+check cycle.
+
+The check now treats a non-XML body as an empty section: no `nvidia_smi`
+services are reported for the affected host, and no crash report is
+generated.

cmk/plugins/nvidia/agent_based/nvidia_smi.py

@@ -146,7 +146,16 @@ def _let_pydantic_check_power_state(value: str | None) -> PowerState:
 
 
 def parse_nvidia_smi(string_table: StringTable) -> Section:
-    xml = ElementTree.fromstring("".join([element[0] for element in string_table]))
+    try:
+        xml = ElementTree.fromstring("".join([element[0] for element in string_table]))
+    except ElementTree.ParseError:
+        return Section(
+            timestamp=None,
+            driver_version=None,
+            cuda_version=None,
+            attached_gpus=None,
+            gpus={},
+        )
     # find the element name for power_readings
     power_readings_element = "gpu_power_readings"
     if xml.find(f"gpu/{power_readings_element}") is None:

tests/unit/cmk/plugins/nvidia/agent_based/test_nvidia_smi.py

@@ -273,47 +273,44 @@ def empty_value_store(monkeypatch: pytest.MonkeyPatch) -> None:
 )
 
 
-@pytest.mark.parametrize(
-    "string_table, expected_result",
-    [
-        (
-            STRING_TABLE,
-            SECTION,
-        ),
-    ],
+EMPTY_SECTION = nvidia_smi.Section(
+    timestamp=None,
+    driver_version=None,
+    cuda_version=None,
+    attached_gpus=None,
+    gpus={},
 )
-def test_parse_nvidia_smi(
-    string_table: StringTable,
-    expected_result: nvidia_smi.Section,
-) -> None:
-    assert nvidia_smi.parse_nvidia_smi(string_table) == expected_result
 
 
-# Inputs captured verbatim from real agents that emit non-XML on the nvidia_smi
-# section. They must be tolerated, not crash the parser with xml.etree.ParseError.
-#   * linux_no_devices: SUP-28831 — RHEL host with installed nvidia-smi but no
-#     visible GPU returns "No devices were found".
-#   * windows_nvidia_smi_missing: crash group 3616 — Windows agent plugin emits
-#     a "not found in:" error block when nvidia-smi.exe is absent.
-@pytest.mark.xfail(strict=True, reason="Crash group 3616: ParseError on non-XML nvidia-smi output")
+# linux_no_devices and windows_nvidia_smi_missing capture verbatim agent
+# outputs that today reach parse_nvidia_smi as non-XML and used to crash it
+# with xml.etree.ElementTree.ParseError:
+#   * linux_no_devices: SUP-28831 — RHEL host with installed nvidia-smi but
+#     no visible GPU returns "No devices were found".
+#   * windows_nvidia_smi_missing: crash group 3616 — Windows agent plugin
+#     emits a "not found in:" error block when nvidia-smi.exe is absent.
 @pytest.mark.parametrize(
-    "string_table",
+    "string_table, expected_result",
     [
-        pytest.param([["No devices were found"]], id="linux_no_devices"),
+        pytest.param(STRING_TABLE, SECTION, id="valid_xml"),
+        pytest.param([["No devices were found"]], EMPTY_SECTION, id="linux_no_devices"),
         pytest.param(
             [
                 ["ERROR: nvidia-smi.exe was not found in:"],
                 ["-  (configured path)"],
                 ["- C:\\Program Files\\NVIDIA Corporation\\NVSMI\\nvidia-smi.exe (default path)"],
                 ["- system PATH"],
             ],
+            EMPTY_SECTION,
             id="windows_nvidia_smi_missing",
         ),
     ],
 )
-def test_parse_nvidia_smi_non_xml_output(string_table: StringTable) -> None:
-    section = nvidia_smi.parse_nvidia_smi(string_table)
-    assert section.gpus == {}
+def test_parse_nvidia_smi(
+    string_table: StringTable,
+    expected_result: nvidia_smi.Section,
+) -> None:
+    assert nvidia_smi.parse_nvidia_smi(string_table) == expected_result
 
 
 @pytest.mark.parametrize(