Um, I believe you are confused.  When we get the trace trap and diddle
frame.tf_eflags inside trap(), we are changing the *trap frame*, not the
*syscall frame*, and in fact, this is exactly what we wanted to do.  We
get a trace trap on the first instruction in the kernel, turn off the
bit, and return.

Now, the comment about it not actually getting turned back on on exit
is correct, but the rest of what you said is wrong.


------------------------------------------------------------------------------