So you zero-out struct reg before calling into MD code. Then why would we need to explicitly set r_err and r_trapno ? Might be for procfs, but isn't it better to clear at top-level in procfs too ?

getcontext() seems to be immune to trapno/err leak for the same reason, top-level code clears the ucontext_t at start, and same is done for sigcontext.

I tried to explain the motivation in the description: I wanted to try and protect other (future) consumers of the interface. I did not catch procfs, for instance. I agree that the explicit zeroing is preferable there, though, and will fix it too. If you still prefer to limit the zeroing to sys_ptrace() itself and procfs, I don't insist on modifying the machdep code.