Improve the ktrace locking somewhat to reduce overhead:

- Depessimize userret() in kernels where KTRACE is enabled by doing an
  unlocked check of the per-process queue of pending events before
  acquiring any locks.  Previously ktr_userret() unconditionally acquired
  the global ktrace_sx lock on every return to userland for every thread,
  even if ktrace wasn't enabled for the thread.
- Optimize the locking in exit() to first perform an unlocked read of
  p_traceflag to see if ktrace is enabled and only acquire locks and
  teardown ktrace if the test succeeds.  Also, explicitly disable tracing
  before draining any pending events so the pending events actually get
  written out.  The unlocked read is safe because proc lock is acquired
  earlier after single-threading so p_traceflag can't change between then
  and this check (well, it can currently due to a bug in ktrace I will fix
  next, but that race existed prior to this change as well).

Reviewed by:	rwatson

1 files changed, 2 insertions, 0 deletions

diff --git a/sys/kern/kern_ktrace.c b/sys/kern/kern_ktrace.c
index 2b7ee33..0ca4a30 100644
--- a/sys/kern/kern_ktrace.c
+++ b/sys/kern/kern_ktrace.c
@@ -444,6 +444,8 @@ void
 ktruserret(struct thread *td)
 {
 
+	if (STAILQ_EMPTY(&td->td_proc->p_ktr))
+		return;
 	ktrace_enter(td);
 	sx_xlock(&ktrace_sx);
 	ktr_drain(td);