From: Martin Quinson <martin.quinson@ens-rennes.fr>
Date: Fri, 9 Aug 2019 22:24:53 +0000 (+0200)
Subject: introduce a simcall_blocking(), and improve the comments
X-Git-Tag: v3.24~204
X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/commitdiff_plain/afa62e573131c8a346d4703b58037ee1ff80e69e?hp=f2ae9283d896f8581020b2357fd2706d68ce9ede

introduce a simcall_blocking(), and improve the comments
---

diff --git a/include/simgrid/simix.hpp b/include/simgrid/simix.hpp
index 9662eb017d..2b979d735d 100644
--- a/include/simgrid/simix.hpp
+++ b/include/simgrid/simix.hpp
@@ -17,26 +17,29 @@
 #include <unordered_map>
 
 XBT_PUBLIC void simcall_run_kernel(std::function<void()> const& code);
-
-/** Execute some code in the kernel and block
- *
- * run_blocking() is a generic blocking simcall. It is given a callback
- * which is executed immediately in the SimGrid kernel. The callback is
- * responsible for setting the suitable logic for waking up the process
- * when needed.
- *
- * @ref simix::kernelSync() is a higher level wrapper for this.
- */
 XBT_PUBLIC void simcall_run_blocking(std::function<void()> const& code);
 
 namespace simgrid {
 namespace simix {
 
-/** Execute some code in the kernel/maestro
+/** Execute some code in kernel context on behalf of the user code.
+ *
+ * Every modification of the environment must be protected this way: every setter, constructor and similar.
+ * Getters don't have to be protected this way.
+ *
+ * This allows deterministic parallel simulation without any locking, even if almost nobody uses parallel simulation in
+ * SimGrid. More interestingly it makes every modification of the simulated world observable by the model-checker,
+ * allowing the whole MC business.
+ *
+ * It is highly inspired from the syscalls in a regular operating system, allowing the user code to get some specific
+ * code executed in the kernel context. But here, there is almost no security involved. Parameters get checked for
+ * finitness but that's all. The main goal remain to ensure reproductible ordering of uncomparable events (in [parallel]
+ * simulation) and observability of events (in model-checking).
  *
- *  This can be used to enforce mutual exclusion with other simcall.
- *  More importantly, this enforces a deterministic/reproducible ordering
- *  of the operation with respect to other simcalls.
+ * The code passed as argument is supposed to terminate at the exact same simulated timestamp.
+ * Do not use it if your code may block waiting for a subsequent event, e.g. if you lock a mutex,
+ * you may need to wait for that mutex to be unlocked by its current owner.
+ * Potentially blocking simcall must be issued using simcall_blocking(), right below in this file.
  */
 template <class F> typename std::result_of<F()>::type simcall(F&& code)
 {
@@ -54,6 +57,36 @@ template <class F> typename std::result_of<F()>::type simcall(F&& code)
   return result.get();
 }
 
+/** Execute some code (that does not return immediately) in kernel context
+ *
+ * This is very similar to simcall() right above, but the calling actor will not get rescheduled until
+ * actor->simcall_answer() is called explicitely.
+ *
+ * This is meant for blocking actions. For example, locking a mutex is a blocking simcall.
+ * First it's a simcall because that's obviously a modification of the world. Then, that's a blocking simcall because if
+ * the mutex happens not to be free, the actor is added to a queue of actors in the mutex. Every mutex->unlock() takes
+ * the first actor from the queue, mark it as current owner of the mutex and call actor->simcall_answer() to mark that
+ * this mutex is now unblocked and ready to run again. If the mutex is initially free, the calling actor is unblocked
+ * right away with actor->simcall_answer() once the mutex is marked as locked.
+ *
+ * If your code never calls actor->simcall_answer() itself, the actor will never return from its simcall.
+ */
+template <class F> typename std::result_of<F()>::type simcall_blocking(F&& code)
+{
+  // If we are in the maestro, we take the fast path and execute the
+  // code directly without simcall mashalling/unmarshalling/dispatch:
+  if (SIMIX_is_maestro())
+    return std::forward<F>(code)();
+
+  // If we are in the application, pass the code to the maestro which
+  // executes it for us and reports the result. We use a std::future which
+  // conveniently handles the success/failure value for us.
+  typedef typename std::result_of<F()>::type R;
+  simgrid::xbt::Result<R> result;
+  simcall_run_blocking([&result, &code] { simgrid::xbt::fulfill_promise(result, std::forward<F>(code)); });
+  return result.get();
+}
+
 XBT_ATTRIB_DEPRECATED_v325("Please manifest if you actually need this function")
     XBT_PUBLIC const std::vector<smx_actor_t>& process_get_runnable();
 
diff --git a/src/kernel/actor/ActorImpl.hpp b/src/kernel/actor/ActorImpl.hpp
index 2713a5659d..81a710a7a5 100644
--- a/src/kernel/actor/ActorImpl.hpp
+++ b/src/kernel/actor/ActorImpl.hpp
@@ -132,7 +132,8 @@ public:
 
   /** execute the pending simcall -- must be called from the maestro context */
   void simcall_handle(int value);
-  /** Terminates a simcall currently executed in maestro context */
+  /** Terminates a simcall currently executed in maestro context. The actor will be restarted in the next scheduling
+   * round */
   void simcall_answer();
 };