Changeset 20115 in webkit for trunk/JavaScriptCore/kjs

Timestamp:

Mar 12, 2007, 8:09:37 AM (18 years ago)

Author:

ggaren

Message:

JavaScriptCore:

Reviewed by Oliver Hunt.

Fixed <rdar://problem/4681051> Installer crashes in KJS::Collector::
markOtherThreadConservatively(KJS::Collector::Thread*) trying to install
iLife 06 using Rosetta on an Intel Machine

The problem was that our thread-specific data destructor would modify the
list of active JavaScript threads without holding the JSLock, corrupting
the list. Corruption was especially likely if one JavaScript thread exited
while another was starting up.

• JavaScriptCore.exp:
• kjs/JSLock.cpp: Don't conflate locking the JSLock with registering a thread, since the thread-specific data destructor needs to lock without registering a thread. Instead, treat thread registration as a part of the convenience of the JSLock object, and whittle down JSLock::lock() to just the bits that actually do the locking. (KJS::JSLock::lock): (KJS::JSLock::registerThread):
• kjs/JSLock.h: Updated comments to mention the new behavior above, and other recent changes. (KJS::JSLock::JSLock):
• kjs/collector.cpp: (KJS::destroyRegisteredThread): Lock here. (KJS::Collector::registerThread): To match, assert that we're locked here.

JavaScriptGlue:

Reviewed by Oliver Hunt.

Updated in light of fix for <rdar://problem/4681051> Installer crashes
in KJS::Collector::markOtherThreadConservatively(KJS::Collector::Thread*)
trying to install iLife 06 using Rosetta on an Intel Machine

• JavaScriptGlue.cpp: (JSLockInterpreter): Ensure backwards compatibility by calling registerThread() when explicitly taking the JSLock. (This doesn't happen automatically anymore.) I doubt this actally matters, but in JavaScriptGlue territory, that kind of thinking will get you killed.

WebKitTools:

Reviewed by Oliver Hunt.

Beefed up --threaded mode in light of <rdar://problem/4681051> Installer
crashes in KJS::Collector::markOtherThreadConservatively(KJS::Collector::Thread*)
trying to install iLife 06 using Rosetta on an Intel Machine

--threaded mode now runs a bunch of different JavaScript threads, randomly
killing and respawning them. This was sufficient for reproducing the
bug on my MacBook Pro.

• DumpRenderTree/DumpRenderTree.m: (javaScriptThreads): (runJavaScriptThread): (startJavaScriptThreads): (stopJavaScriptThreads): (dumpRenderTree):

Location:

trunk/JavaScriptCore/kjs

Files:

• JSLock.cpp (modified) (3 diffs)
• JSLock.h (modified) (3 diffs)
• collector.cpp (modified) (3 diffs)

trunk/JavaScriptCore/kjs/JSLock.cpp

@@ -54 +54 @@
     }
     ++JSLockCount;
-
-    Collector::registerThread();
 }
 
@@ -68 +66 @@
         pthread_mutex_unlock(&JSMutex);
     }
+}
+
+void JSLock::registerThread()
+{
+    Collector::registerThread();
 }
 
@@ -101 +104 @@
 }
 
+void JSLock::registerThread()
+{
+}
+
 JSLock::DropAllLocks::DropAllLocks()
 {

trunk/JavaScriptCore/kjs/JSLock.h

@@ -31 +31 @@
     // To make it safe to use JavaScript on multiple threads, it is
     // important to lock before doing anything that allocates a
-    // garbage-collected object or which may affect other shared state
-    // such as the protect count hash table. The simplest way to do
-    // this is by having a local JSLock object for the scope
-    // where the lock must be held. The lock is recursive so nesting
-    // is ok.
+    // JavaScript data structure or that interacts with shared state
+    // such as the protect count hash table. The simplest way to lock
+    // is to create a local JSLock object in the scope where the lock 
+    // must be held. The lock is recursive so nesting is ok. The JSLock 
+    // object also acts as a convenience short-hand for running important
+    // initialization routines.
 
     // To avoid deadlock, sometimes it is necessary to temporarily
@@ -48 +49 @@
     class JSLock : Noncopyable {
     public:
-        JSLock() 
+        JSLock()
         {
             lock();
+            registerThread();
         }
 
@@ -61 +63 @@
         static void unlock();
         static int lockCount();
+
+        static void registerThread();
 
         class DropAllLocks : Noncopyable {

trunk/JavaScriptCore/kjs/collector.cpp

@@ -237 +237 @@
   Collector::Thread *thread = (Collector::Thread *)data;
 
+  // Can't use JSLock convenience object here because we don't want to re-register
+  // an exiting thread.
+  JSLock::lock();
+  
   if (registeredThreads == thread) {
     registeredThreads = registeredThreads->next;
   } else {
     Collector::Thread *last = registeredThreads;
-    for (Collector::Thread *t = registeredThreads->next; t != NULL; t = t->next) {
+    Collector::Thread *t;
+    for (t = registeredThreads->next; t != NULL; t = t->next) {
       if (t == thread) {
         last->next = t->next;
@@ -248 +253 @@
       last = t;
     }
-  }
+    ASSERT(t); // If t is NULL, we never found ourselves in the list.
+  }
+
+  JSLock::unlock();
 
   delete thread;
@@ -260 +268 @@
 void Collector::registerThread()
 {
+  ASSERT(JSLock::lockCount() > 0);
+  
   pthread_once(&registeredThreadKeyOnce, initializeRegisteredThreadKey);