We've seen a couple reports where extensions that run JS code that deals with XPConnect off the main thread end up calling into nsScriptSecurityManager and then crashing (http://crash-stats.mozilla.com/report/index/2dcc69c6-ff24-4d30-9594-ba34f2100517 is one example).
We need to figure out why XPConnect is getting a hold of the main thread's security manager on these other threads. Without this, any extension that runs JS off the main thread (even if it's careful about what objects it uses off the main thread) risks crashing.
All of the magic wrapper code uses this which accesses a global
1048 XPCConvert::NativeInterface2JSObject(XPCLazyCallContext& lccx,
1244 if(NS_SUCCEEDED(rv) && wrapper)
1255 if (allowNativeWrapper &&
1256 !xpc_SameScope(wrapper->GetScope(), xpcscope, &sameOrigin))
1309 if(!JS_IsSystemObject(ccx, flat))
1329 nsIScriptSecurityManager *ssm =
Surprisingly, that's not actually the problem (I have no clue why, I suspect it might be because it's impossible for !xpc_SameScope to be true off the main thread). The problem shown in the crash report in comment 0 is that the "default" security manager for a not-main thread is not null, which is definitely a bug.
For my reference, these are my notes from talking with mrbkap about how to reproduce this issue:
From chrome, using nsIThreadManager, dispatch a runnable to some other code I own.
Touch xpconnect. Create an object using Components.classes.(some threadsafe interface), then touch it in a way that isn't quickstubbed.
Maybe crashes in nsScriptSecurityManager (somewhere), due to bug in GetDefaultSecurityManager.
* Could put NS_ABORT_IF_FALSE(IsMainThread) in nsScriptSecurityManager::CanAccess
Created attachment 474904 [details] [diff] [review]
Testcase. The new browser test triggers an assertion I added to GetDefaultSecurityManager and kills the browser.
Created attachment 474934 [details] [diff] [review]
Returns null when we're not the main thread.
Created attachment 474954 [details] [diff] [review]
Ack, sorry. qref'ed.
Blake, since this is crashing plugins, should this block?
Yes please. Firefox crashes every few minutes and that really gets my goat.
Yeah, I think so.
Could you please manage fixing this already in the next release? Current Firefox is nearly unusable because of so many crashes (mainly invoked by Yoono which I don't want to miss) and I don't want to switch back to 3.6.4 (should be Xulrunner 184.108.40.206, right?) where the problem didn't exist.
*** Bug 597287 has been marked as a duplicate of this bug. ***
We want this on the branch but wouldn't block a release on it, as it looks like there are only ~1000 crashes total per release. Please ask for approval on the patch when ready and we'll get it in the current in-flight update.
Is there a possibility of our own code calling nsScriptSecurityManager off the main thread and thus this fix would cause issues?
Whoops; didn't mean to resolve this, since we still need to land on branches.
Does this same fix apply to 1.9.1? We have it marked needed there as well. Approving for 220.127.116.11. Please ask for approval for 18.104.22.168 if it is needed there as well.
We use the status1.9.x fields to track branch status
IOW change status1.9.2 from "wanted" to ".11-fixed" when you land, and similarly for the 1.9.1 branch
Fixed on branches.
Verified for 1.9.1 and 1.9.2 based on passing unit test.