auto merge of #8234 : bblum/rust/assorted-fixes, r=brson

This fixes 4 bugs that prevented the extra::arc and extra::sync tests from passing on the new runtime.

* In ```Add SendDeferred trait``` I add a non-rescheduling ```send_deferred``` method to our various channel types. The ```extra::sync``` concurrency primitives need this guarantee so they can send while inside of an exclusive. (This fixes deterministic deadlocks seen with ```RUST_THREADS=1```.)
* In "Fix nasty double-free bug" I make sure that a ```ChanOne``` suppresses_finalize *before* rescheduling away to the receiver, so in case it gets a kill signal upon coming back, the destructor is inhibited as desired. (This is pretty uncommon on multiple CPUs but showed up always with ```RUST_THREADS=1```.)
* In ```Fix embarrassing bug where 'unkillable' would unwind improperly``` I make sure the task's unkillable counter stays consistent when a kill signal is received right at the start of an unkillable section. (This is a very uncommon race and can only occur with multiple CPUs.)
* In ```Don't fail from kill signals if already unwinding``` I do pretty much what it says on the tin. Surprising that it took the whole suite of sync/arc tests to expose this.

The other two commits are cleanup.

r @brson

1 files changed, 5 insertions, 4 deletions

diff --git a/src/libstd/rt/kill.rs b/src/libstd/rt/kill.rs
index 696f4a8c355..deec8dd37a6 100644
--- a/src/libstd/rt/kill.rs
+++ b/src/libstd/rt/kill.rs
@@ -530,13 +530,13 @@ impl Death {
 
     /// Fails if a kill signal was received.
     #[inline]
-    pub fn check_killed(&self) {
+    pub fn check_killed(&self, already_failing: bool) {
         match self.kill_handle {
             Some(ref kill_handle) =>
                 // The task may be both unkillable and killed if it does some
                 // synchronization during unwinding or cleanup (for example,
                 // sending on a notify port). In that case failing won't help.
-                if self.unkillable == 0 && kill_handle.killed() {
+                if self.unkillable == 0 && (!already_failing) && kill_handle.killed() {
                     fail!(KILLED_MSG);
                 },
             // This may happen during task death (see comments in collect_failure).
@@ -548,11 +548,12 @@ impl Death {
     /// All calls must be paired with a subsequent call to allow_kill.
     #[inline]
     pub fn inhibit_kill(&mut self, already_failing: bool) {
-        if self.unkillable == 0 {
+        self.unkillable += 1;
+        // May fail, hence must happen *after* incrementing the counter
+        if self.unkillable == 1 {
             rtassert!(self.kill_handle.is_some());
             self.kill_handle.get_mut_ref().inhibit_kill(already_failing);
         }
-        self.unkillable += 1;
     }
 
     /// Exit a possibly-nested unkillable section of code.