A bunch of focus related stuff leaks per-window via a cycle that includes the global object, and, if a key is pressed in that window, drags in an XBL document for each window where a key is pressed (which is 30K of logged (via nsTraceRefcnt) objects per window, probably a good bit more (50% more? 500% more?) of unlogged objects). The following diagram explains what I'm seeing (where |A| mFoo--> |B| represents A having a member mFoo that has an owning pointer to B): Another GlobalWindowImpl, in one case but not the others ^ | mOpener mChromeEventHandler-- ------------------ \ -------------- |GlobalWindowImpl| -> |nsWindowRoot| ------------------ -------------- ^ -----mFocusController mListenerManager / / / / / / / / ------------------------ / / / |nsEventListenerManager| <-- mFocusedWindow v ------------------------ ------------------- m???Listeners |nsFocusController| / \ ------------------- <--------- v ----------------------- |nsXBLWindowKeyHandler| ----------------------- | | (if a key is pressed) | v ------------------- |nsXBLDocumentInfo| ------------------- | v ------------ |nsDocument|----> Content nodes ------------

This seems related to hyatt's fix for bug 54023. He changed: nsIChromeEventHandler* mChromeEventHandler; // Weak Reference to nsCOMPtr<nsIChromeEventHandler> mChromeEventHandler; // [Strong] We break it when we get torn down. but I don't see any code to break the cycle when the global is torn down. Such code would be easy to write, though....

Well, the above patch works and it doesn't seem to do anything harmful after about a minute of using the browser. Are there any problems that it could cause?

sr=hyatt

looks right to me, r=saari

Fix checked in 2001-06-26 14:29 PDT.

Marking [PDT+] per conversation with chofmann.

Fix checked into branch 2001-06-30 12:01 PDT.

marking verifed as per developers comments.