Fix stack overflow when logging system encounters an error (#8268)

2 files changed, 24 insertions, 2 deletions

diff --git a/changelog.d/8268.bugfix b/changelog.d/8268.bugfix
new file mode 100644
index 0000000000..4b15a60253
--- /dev/null
+++ b/changelog.d/8268.bugfix
@@ -0,0 +1 @@
+Fix stack overflow when stderr is redirected to the logging system, and the logging system encounters an error.
diff --git a/synapse/config/logger.py b/synapse/config/logger.py
index c96e6ef62a..13d6f6a3ea 100644
--- a/synapse/config/logger.py
+++ b/synapse/config/logger.py
@@ -17,6 +17,7 @@ import logging
 import logging.config
 import os
 import sys
+import threading
 from string import Template
 
 import yaml
@@ -25,6 +26,7 @@ from twisted.logger import (
     ILogObserver,
     LogBeginner,
     STDLibLogObserver,
+    eventAsText,
     globalLogBeginner,
 )
 
@@ -216,8 +218,9 @@ def _setup_stdlib_logging(config, log_config, logBeginner: LogBeginner):
     # system.
     observer = STDLibLogObserver()
 
-    def _log(event):
+    threadlocal = threading.local()
 
+    def _log(event):
         if "log_text" in event:
             if event["log_text"].startswith("DNSDatagramProtocol starting on "):
                 return
@@ -228,7 +231,25 @@ def _setup_stdlib_logging(config, log_config, logBeginner: LogBeginner):
             if event["log_text"].startswith("Timing out client"):
                 return
 
-        return observer(event)
+        # this is a workaround to make sure we don't get stack overflows when the
+        # logging system raises an error which is written to stderr which is redirected
+        # to the logging system, etc.
+        if getattr(threadlocal, "active", False):
+            # write the text of the event, if any, to the *real* stderr (which may
+            # be redirected to /dev/null, but there's not much we can do)
+            try:
+                event_text = eventAsText(event)
+                print("logging during logging: %s" % event_text, file=sys.__stderr__)
+            except Exception:
+                # gah.
+                pass
+            return
+
+        try:
+            threadlocal.active = True
+            return observer(event)
+        finally:
+            threadlocal.active = False
 
     logBeginner.beginLoggingTo([_log], redirectStandardIO=not config.no_redirect_stdio)
     if not config.no_redirect_stdio: