[PATCH] KNOWN_ISSUES: background threads + Logger monkey patch

Eric Wong normalperson at yhbt.net
Mon Jan 25 21:35:53 EST 2010

Thanks to Michael Guterl for informing us of the issues
and testing the monkey patch.
 Pushed out to the website and unicorn.git

 KNOWN_ISSUES               |    8 ++++++++
 examples/logger_mp_safe.rb |   25 +++++++++++++++++++++++++
 2 files changed, 33 insertions(+), 0 deletions(-)
 create mode 100644 examples/logger_mp_safe.rb

index 83b3584..e83e34e 100644
@@ -3,6 +3,14 @@
 Occasionally odd {issues}[link:ISSUES.html] arise without a transparent or
 acceptable solution.  Those issues are documented here.
+* When using "preload_app true", with apps using background threads
+  need to restart them in the after_fork hook because threads are never
+  shared with child processes.  Additionally, any synchronization
+  primitives (Mutexes, Monitors, ConditionVariables) should be
+  reinitialized in case they are held during fork time to avoid
+  deadlocks.  The core Ruby Logger class needlessly uses a MonitorMutex
+  which can be disabled with a {monkey patch}[link:examples/logger_mp_safe.rb]
 * Rails 2.3.2 bundles its own version of Rack.  This may cause subtle
   bugs when simultaneously loaded with the system-wide Rack Rubygem
   which Unicorn depends on.  Upgrading to Rails 2.3.4 (or later) is
diff --git a/examples/logger_mp_safe.rb b/examples/logger_mp_safe.rb
new file mode 100644
index 0000000..05ad3fa
--- /dev/null
+++ b/examples/logger_mp_safe.rb
@@ -0,0 +1,25 @@
+# Multi-Processing-safe monkey patch for Logger
+# This monkey patch fixes the case where "preload_app true" is used and
+# the application spawns a background thread upon being loaded.
+# This removes all lock from the Logger code and solely relies on the
+# underlying filesystem to handle write(2) system calls atomically when
+# O_APPEND is used.  This is safe in the presence of both multiple
+# threads (native or green) and multiple processes when writing to
+# a filesystem with POSIX O_APPEND semantics.
+# It should be noted that the original locking on Logger could _never_ be
+# considered reliable on non-POSIX filesystems with multiple processes,
+# either, so nothing is lost in that case.
+require 'logger'
+class Logger::LogDevice
+  def write(message)
+    @dev.syswrite(message)
+  end
+  def close
+    @dev.close
+  end
Eric Wong

More information about the mongrel-unicorn mailing list