Rewrite reap_dead_tasks to never grab the sched lock before a task lock

Doing so contradicts the locking order used everywhere else and causes
deadlocks.

Un-XFAIL task-perf-spawnalot

Closes #854

Author: brson

src/rt/rust_scheduler.cpp

@@ -100,25 +100,66 @@ rust_scheduler::number_of_live_tasks() {
 void
 rust_scheduler::reap_dead_tasks(int id) {
     I(this, lock.lock_held_by_current_thread());
-    for (size_t i = 0; i < dead_tasks.length(); ) {
+    if (dead_tasks.length() == 0) {
+        return;
+    }
+
+    // First make up copy of the dead_task list with the lock held
+    size_t dead_tasks_len = dead_tasks.length();
+    rust_task **dead_tasks_copy = (rust_task**)
+        srv->malloc(sizeof(rust_task*) * dead_tasks_len);
+    for (size_t i = 0; i < dead_tasks_len; ++i) {
         rust_task *task = dead_tasks[i];
+        dead_tasks_copy[i] = task;
+    }
+
+    // Now drop the lock and futz with the tasks. This avoids establishing
+    // a sched->lock then task->lock locking order, which would be devestating
+    // to performance.
+    lock.unlock();
+
+    for (size_t i = 0; i < dead_tasks_len; ++i) {
+        rust_task *task = dead_tasks_copy[i];
         task->lock.lock();
         // Make sure this task isn't still running somewhere else...
         if (task->can_schedule(id)) {
             I(this, task->tasks_waiting_to_join.is_empty());
-            dead_tasks.remove(task);
             DLOG(this, task,
                 "deleting unreferenced dead task %s @0x%" PRIxPTR,
                 task->name, task);
             task->lock.unlock();
+        } else {
+            task->lock.unlock();
+            dead_tasks_copy[i] = NULL;
+        }
+    }
+
+    // Now grab the lock again and remove the tasks that were truly dead
+    lock.lock();
+
+    for (size_t i = 0; i < dead_tasks_len; ++i) {
+        rust_task *task = dead_tasks_copy[i];
+        if (task) {
+            dead_tasks.remove(task);
+        }
+    }
+
+    // Now unlock again because we have to actually free the dead tasks,
+    // and that may end up wanting to do lock the task and sched locks
+    // again (via target->send)
+    lock.unlock();
+
+    for (size_t i = 0; i < dead_tasks_len; ++i) {
+        rust_task *task = dead_tasks_copy[i];
+        if (task) {
             task->deref();
             sync::decrement(kernel->live_tasks);
             kernel->wakeup_schedulers();
-            continue;
         }
-        task->lock.unlock();
-        ++i;
     }
+    srv->free(dead_tasks_copy);
+
+    lock.lock();
 }
 
 /**

src/rt/rust_task.cpp

@@ -90,6 +90,7 @@ rust_task::rust_task(rust_scheduler *sched, rust_task_list *state,
 
 rust_task::~rust_task()
 {
+    I(sched, !sched->lock.lock_held_by_current_thread());
     DLOG(sched, task, "~rust_task %s @0x%" PRIxPTR ", refcnt=%d",
          name, (uintptr_t)this, ref_count);
 

src/test/bench/task-perf-spawnalot.rs

@@ -1,7 +1,3 @@
-// xfail-stage1
-// xfail-stage2
-// xfail-stage3
-
 use std;
 import std::vec;
 import std::task;