off single-stepping). Only do this on arches (only x86 so far)
which classify single-step traps unambiguously.
This allows other parts of the kernel to be intentionally and
unintentionally sloppy about generating single-step traps. On
x86, at least the following places were unintentionally sloppy:
- all operations that context-switched [er]flags. Especially
spinlock_enter()/exit() and cpu_switch(). When single-stepped,
saving the flags leaves PSL_T set in the saved flags, so
restoring gives a trap that is spurious if it occurs after
single-step mode has been left. Switching contexts away from
a low priority thread gives especially long-lived saved copies.
- the vm86 emulation allows user mode to set PSL_T. This was
correct until vm86 bios call mode was unintentionally given
access to kdb handling its single-step traps.
Now these places are intentionally sloppy, but unexpected
debugger traps still cause panics if no debugger that handles
the trap is attached when the trap is delivered.
83d5e12de9