Re: [OLE #76] Avoid infinite loop when doing a typelib marshalled IUnknown::QueryInterface

[Rob Shearman]

Mike Hearn wrote:
> Mike Hearn <mh(a)codeweavers.com>
> Avoid infinite loop when doing a typelib marshalled
> IUnknown::QueryInterface by only doing an extra QI if requested IID is
> not equal to marshalled IID.

Does this happen with only builtin oleaut32? If so, it should be fixed 
so that it doesn't cause an infinite loop. I don't think standard 
marshaled proxies should cause infinite loops (if I remember the test 
cases we have correctly).
However, this patch should go in as it is a good optimisation.

Rob

[OLE #76] Avoid infinite loop when doing a typelib marshalled IUnknown::QueryInterface

[Mike Hearn]

Rob Shearman wrote:
> Does this happen with only builtin oleaut32?

I'm not sure, I didn't try mixing native/builtins for this problem.
> If so, it should be fixed 
> so that it doesn't cause an infinite loop. I don't think standard 
> marshaled proxies should cause infinite loops (if I remember the test 
> cases we have correctly).

Yes, I also have a vague feeling that it may be covering up for an issue 
elsewhere, but can't quite place my finger on what. The problem was 
quite subtle and confusing (as everything in DCOM is <sigh>)

InstallShield gets a TLB proxy

xCall, IUnknown::QueryInterface
    -> rpc to stub -> IRemUnknown::RemQueryInterface marshals result

tmarshal.c:_unmarshal_interface called on the result of the QI

CoUnmarshalInterface( ... results of QI ... )

    IMarshal::UnmarshalInterface sets up the TLBProxy
    IUnknown::QueryInterface on the TLBProxy object
           xCall, IUnknown::QueryInterface
                 -> rpc to stub -> [infinite loop]

I have a feeling the QI in CoUnmarshalInterface should be a no-op 
because the typelib proxies should aggregate the proxy manager, which 
would optimize out a QI for an already unmarshalled interface.

So possibly the real issue is that we never made the tmarshaller 
aggregate the proxy manager. I don't recall if you did that or not, I 
know I didn't.
> However, this patch should go in as it is a good optimisation.

Right. It also unblocks InstallShield for the time being, letting us get 
back to the stage where it needs RPC re-entrancy.

thanks -mike