Fix sff panic deadlock, share the evdev probe, assorted review cleanups

sff.go

@@ -15,10 +15,15 @@ type sff struct {
 
 	// Every transport touch happens on the loop goroutine: requests execute
 	// one at a time, each admitted by the module type's own gate first.
-	reqs  chan func()
+	reqs  chan sffReq
 	admit func()
 }
 
+type sffReq struct {
+	fn   func()
+	done chan struct{}
+}
+
 func openSFF(ifname string) (*sff, error) {
 	devLink, err := os.Readlink("/sys/class/net/" + ifname + "/device")
 	if err != nil {
@@ -34,7 +39,7 @@ func openSFF(ifname string) (*sff, error) {
 	s := &sff{
 		ifname: ifname,
 		path:   "/sys/kernel/debug/ixgbe/" + filepath.Base(devLink) + "/sff_i2c",
-		reqs:   make(chan func()),
+		reqs:   make(chan sffReq),
 		admit:  func() {},
 	}
 	if _, err := os.Stat(s.path); err != nil {
@@ -45,16 +50,23 @@ func openSFF(ifname string) (*sff, error) {
 }
 
 func (s *sff) loop() {
-	defer holdPanic()
-	for fn := range s.reqs {
-		s.admit()
-		fn()
+	for r := range s.reqs {
+		s.serve(r)
 	}
 }
 
+// done closes even when the request panics, and the recover keeps the loop
+// serving, so no exec caller is ever stranded waiting on a dead loop.
+func (s *sff) serve(r sffReq) {
+	defer close(r.done)
+	defer holdPanic()
+	s.admit()
+	r.fn()
+}
+
 func (s *sff) exec(fn func()) {
 	done := make(chan struct{})
-	s.reqs <- func() { fn(); close(done) }
+	s.reqs <- sffReq{fn: fn, done: done}
 	<-done
 }