dev -> main

nodescraper/plugins/inband/dmesg/mce_utils.py

@@ -30,7 +30,7 @@
 
 _CORRECTABLE_SUMMARY_RE = re.compile(
     r"(?P<count>\d+)\s+correctable hardware errors detected in total in (?P<block>\w+) block"
-    r"(?:\s+on\s+(?P<cpu>CPU\d+))?",
+    r"(?:\s+on\s+(?P<cpu>CPU:?\d+))?",
     re.IGNORECASE,
 )
 
@@ -52,16 +52,20 @@
 )
 
 _MCE_CE_STATUS_RE = re.compile(
-    r"\[Hardware Error\]:.*?(?P<cpu>CPU\d+).*?MC\d+_STATUS\[[^\]]*\|CE\|[^\]]*\]",
+    r"\[Hardware Error\]:.*?(?P<cpu>CPU:?\d+).*?MC\d+_STATUS\[[^\]]*\|CE\|[^\]]*\]",
     re.IGNORECASE,
 )
 
 _MCE_UC_STATUS_RE = re.compile(
-    r"\[Hardware Error\]:.*?(?P<cpu>CPU\d+).*?MC\d+_STATUS\[[^\]]*\|UC\|[^\]]*\]",
+    r"\[Hardware Error\]:.*?(?P<cpu>CPU:?\d+).*?MC\d+_STATUS\[[^\]]*\|UC\|[^\]]*\]",
     re.IGNORECASE,
 )
 
 
+def _normalize_cpu_label(cpu: str) -> str:
+    return cpu.replace(":", "")
+
+
 def _add_count(counts: dict[str, int], part: str, amount: int) -> None:
     counts[part] = counts.get(part, 0) + amount
 
@@ -120,8 +124,9 @@ def parse_correctable_mce_counts(
 
         summary_match = _CORRECTABLE_SUMMARY_RE.search(line)
         if summary_match:
+            cpu = summary_match.group("cpu")
             part = _part_label(
-                cpu=summary_match.group("cpu"),
+                cpu=_normalize_cpu_label(cpu) if cpu else None,
                 block=summary_match.group("block"),
             )
             _add_count(counts, part, int(summary_match.group("count")))
@@ -132,7 +137,11 @@ def parse_correctable_mce_counts(
             bank = extract_mce_bank_from_line(line)
             if bank is not None and bank in ignored:
                 continue
-            part = status_match.group("cpu") if status_match.group("cpu") else "unknown"
+            part = (
+                _normalize_cpu_label(status_match.group("cpu"))
+                if status_match.group("cpu")
+                else "unknown"
+            )
             _add_count(counts, part, 1)
 
     return counts
@@ -170,7 +179,11 @@ def parse_uncorrectable_mce_counts(
             bank = extract_mce_bank_from_line(line)
             if bank is not None and bank in ignored:
                 continue
-            part = status_match.group("cpu") if status_match.group("cpu") else "unknown"
+            part = (
+                _normalize_cpu_label(status_match.group("cpu"))
+                if status_match.group("cpu")
+                else "unknown"
+            )
             _add_count(counts, part, 1)
 
     return counts

test/unit/plugin/test_dmesg_analyzer.py

@@ -1043,6 +1043,26 @@ def test_mce_threshold_raises_error_for_gpu(system_info):
     assert res.status == ExecutionStatus.ERROR
 
 
+def test_mce_threshold_raises_error_for_cpu_colon_status(system_info):
+    dmesg_content = (
+        "kern  :err   : 2038-01-19T00:00:00,000000+00:00 "
+        "[Hardware Error]: CPU:72 (00:00:0) MC60_STATUS[-|CE|Misc]: 0xabc\n"
+    )
+
+    analyzer = DmesgAnalyzer(system_info=system_info)
+    res = analyzer.analyze_data(
+        DmesgData(dmesg_content=dmesg_content),
+        args=DmesgAnalyzerArgs(check_unknown_dmesg_errors=False, mce_threshold=1),
+    )
+
+    threshold_events = [e for e in res.events if e.data.get("mce_threshold") == 1]
+    assert len(threshold_events) == 1
+    assert threshold_events[0].priority == EventPriority.ERROR
+    assert threshold_events[0].data["part"] == "CPU72"
+    assert threshold_events[0].data["correctable_mce_count"] == 1
+    assert res.status == ExecutionStatus.ERROR
+
+
 def test_mce_threshold_not_triggered_below_limit(system_info):
     dmesg_content = (
         "kern  :warn  : 2024-06-11T14:30:00,123456+00:00 "

test/unit/plugin/test_mce_utils.py

@@ -76,3 +76,39 @@ def test_parse_correctable_mce_counts_skips_ignored_banks():
     counts = parse_correctable_mce_counts(content, ignore_banks=frozenset({1, 2}))
 
     assert counts == {"CPU0": 1}
+
+
+def test_parse_correctable_mce_counts_cpu_colon_status():
+    content = (
+        "kern  :err   : 2038-01-19T00:00:00,000000+00:00 "
+        "[Hardware Error]: CPU:72 (00:00:0) MC60_STATUS[-|CE|Misc]: 0xabc\n"
+    )
+
+    counts = parse_correctable_mce_counts(content)
+
+    assert counts == {"CPU72": 1}
+
+
+def test_parse_correctable_mce_counts_both_cpu_formats():
+    content = (
+        "[Hardware Error]: CPU0 MC1_STATUS[0x0|CE|]: 0x1\n"
+        "kern  :err   : 2038-01-19T00:00:00,000000+00:00 "
+        "[Hardware Error]: CPU:72 (00:00:0) MC60_STATUS[-|CE|Misc]: 0xabc\n"
+        "kern  :warn  : 2024-06-11T14:30:00,123456+00:00 "
+        "mce: 2 correctable hardware errors detected in total in mc0 block on CPU:1\n"
+    )
+
+    counts = parse_correctable_mce_counts(content)
+
+    assert counts == {"CPU0": 1, "CPU72": 1, "CPU1/mc0": 2}
+
+
+def test_parse_uncorrectable_mce_counts_cpu_colon_status():
+    content = (
+        "kern  :err   : 2038-01-19T00:00:01,000000+00:00 "
+        "[Hardware Error]: CPU:72 (00:00:0) MC60_STATUS[-|UC|Misc]: 0xdef\n"
+    )
+
+    counts = parse_uncorrectable_mce_counts(content)
+
+    assert counts == {"CPU72": 1}