[Mono-devel-list] Problem in Marshaling Runtime Code?
vargaz at gmail.com
Wed Feb 23 10:10:35 EST 2005
Currently, we never unload assemblies even if all appdomains using them are
unloaded, so this isn't a big problem now. Later, when assemblies are unloaded,
we will fix this as well.
On Wed, 23 Feb 2005 06:11:35 -0800, Jesse Towner <townerj at gmail.com> wrote:
> I'm not sure if I have identified a potential problem or not, but here goes.
> I was examining the code in mono/mono/metadata/marshal.c and I noticed
> that there is a single method wrapper cache that is global to the entire
> process. However, it is not always the case that methods are also global
> to the process because sometimes they're local to the host AppDomain
> that their originating assembly has been loaded into. There also appears
> to be no mechanism for removing a particular wrapper from the cache, so
> it is quite possible that an internal call or some runtime code could
> make reference to a method that has been unloaded with its host AppDomain.
> I'm also not sure how this ties in with the GC implementation, that is,
> if the method wrapper cache induces a live reference to a particular
> method and the parent class, which somehow might stop an AppDomain from
> being unloaded. But if it doesn't cause the GC to improperly unload an
> AppDomain, then their is still the issue of not recollecting dead
> wrappers from the method wrapper cache. I'm not sure how XSP works, but
> if it is anything like IIS, then this could be a potential memory leak.
> Furthermore, since their is only a single method wrapper cache and
> guarding mutex, thread contention between threads (from different
> AppDomains) competing for access to the one and same cache could occur.
> This problem is exacerbated when one considers that the different
> threads may be trying to acquire method wrappers to methods that may be
> exclusive to a particular AppDomain.
> So unless I have missed something in the code, it seems this problem
> could rear its ugly head at some point in the future.
> If this is indeed an issue not handled elsewhere, the proposed solution
> would be to encapsulate the cache and other supporting variables and
> objects in a context structure per AppDomain. Then the code in marshal.c
> would then be modified to use the marshaling context for the current
> Jesse Towner, Lead Programmer, Hybrid Mobile Technologies, LLC
> Email Addresses: <jtowner at hybrid-mobile.com>, <townerj at gmail.com>
> PGP Public Key: http://www.virtuallyonline.net/personal/jesse/pgp.html
More information about the Mono-devel-list