looking at VG MLs, it seems that there's already some ongoing work (http://www.goop.org/~jeremy/valgrind/ get to #77) which would provide a better signal handling, so this might be a way to get through that (I didn't test it though)Apart from the duplication of the signal mechanism, it also introduces races, since we depend on the client to do part of the suspend code. It means the server state will not necessarily match the actual state of the client thread, which could cause trouble. I'm still not convinced we want that in the standard tree.
A+
-- Eric Pouech
