Bloaty: Refcounting and Memory Bloat
Rem Total Mean StdDev Rem
Total Mean StdDev Per-Class Rem
78 nsCodebasePrincipal -3 258 ( -1.33 +/- 1.90) 1
5 ( 0.56 +/- 0.83) 36 36
I suspect that the refcount is going negative on this class because the mRefCnt
field isn't really being used, and the mJSPrincipals.refcount field is instead.
Maybe somewhere along the way, the nsCodebasePrincipal is getting QI'd to
nsISupports and that's incrementing the other mRefCnt(?). I couldn't catch it
in the act.
We really should fix this so that it doesn't get in the way of tracking down
other refcounting problems.
Do you have any test case that reproduces this?
There's only one vtable pointer in the nsCodebasePrincipal object, and it points
to the overridden forms of AddRef and Release. So a QI should never result in an
object whose AddRef and Release affect nsISupports::mRefCnt.
Ok, well I guess that theory's wrong. But I got this by recompiling with BLOATY
defined in nsISupportsUtil.h and visiting www.mozilla.org, www.yahoo.com and
then drilling down into the yahoo site for several pages. I'll try to reproduce
it again later today.
I just checked in changes to assert if Release or AddRef of a
nsCodebasePrincipal is called with a mRefcnt != 0. We'll see if anyone else can
reproduce this problem.
Well, that's not really what I had in mind when filing this bug. I'd like to
see you fix the bloat statistics to reflect reality so that people won't think
that nsCodebasePrincipals are leaking. The statistics can get off if
the NS_IMPL_ADDREF/RELEASE macros aren't used, or if the
NS_LOG_ADDREF/NS_LOG_RELEASE macros aren't used consistently in hand-written
AddRef/Release implementations. The statistics can also get off if there's a
Also, I'm not sure your assertion is useful for the case where the
nsCodebasePrincipal isn't getting released/destroyed.
I tried turning on the about:bloat code to look at the nsCodebasePrincipal
refcounts. I tried loading www.mozilla.org, then www.yahoo.com, and then going
down several layers deep into yahoo. I never saw any negative refcounts in the
I did see leaks of nsDocuments, which are holding refs to nsCodebasePrincipals
and preventing them from being released. I've filed bug 16466 on that. When that
bug is resolved, I'll ensure again that nsCodebasePrincipals are indeed released
after the associated page goes away.
Ok, I can't reproduce this now either. Maybe something else got fixed and
cleared it up. Maybe we should mark it WORKSFORME.
Done. Thanks for retrying.
Bulk moving all Browser Security bugs to new Security: General component. The
previous Security component for Browser will be deleted.