Since the code is inside the loop for function data, it should be 
freeing the function's custom data, not the interface's.

It isn't tracked per connection, unlike NTLM authentication, and so the 
server will return a 401 error and try to get us to authenticate again. 
However, this doesn't work as the authentication information is assumed 
by the code to be valid for the whole connection.

It is different to the other base types as it has a different size on 
the wire to in memory, so it can't just be set to the buffer when 
unmarshalling.

This fixes a regression introduced with the MAD patch.

wined3d: Fixed potential reference of freed backBuffer array in IWineD3DDeviceImpl_SetFrontBackBuffers.

advapi32: Add support for using the Mac Keychain services as a backend for the credential functions instead of the registry.

Add stubs for these so that the tests link.

ole32: Clean up properly in ClientRpcChannelBuffer_SendReceive in the case where PostMessageW fails.

ole32: Store the dispatch parameters in the message state structure to allow them to be initialised earlier in the sequence of IRpcChannelBuffer calls.

wined3d: Rewrite condition in vshader_program_add_param() to actually distinguish between two cases.

Adjust type of loop variable in INT21_Ioctl_Char().