Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Make the plugin usable within a SMPI simulation
[simgrid.git] / src / bindings / python / simgrid_python.cpp
1 /* Copyright (c) 2018-2021. The SimGrid Team. All rights reserved.          */
2
3 /* This program is free software; you can redistribute it and/or modify it
4  * under the terms of the license (GNU LGPL) which comes with this package. */
5
6 #ifdef _WIN32
7 #warning Try to work around https://bugs.python.org/issue11566
8 #define _hypot hypot
9 #endif
10
11 #if defined(__GNUG__)
12 #pragma GCC diagnostic push
13 #pragma GCC diagnostic ignored "-Wunused-value"
14 #endif
15
16 #ifndef NDEBUG
17 /* Many tests are failing after pybind11 commit ad6bf5cd39ca64b4a9bf846b84b11c4c8df1c8e1 "Adding PyGILState_Check() in
18  *  object_api<>::operator(). (#2919)".
19  * See https://github.com/pybind/pybind11/commit/ad6bf5cd39ca64b4a9bf846b84b11c4c8df1c8e1
20  *
21  * The failing tests are mostly those with boost/raw/sysv contexts. As a workaround, define NDEBUG before pybind11
22  * includes.
23  */
24 #define NDEBUG
25 #define NDEBUG_LOCALLY_DEFINED
26 #endif
27
28 #include <pybind11/pybind11.h> // Must come before our own stuff
29
30 #include <pybind11/functional.h>
31 #include <pybind11/stl.h>
32
33 #ifdef NDEBUG_LOCALLY_DEFINED
34 #undef NDEBUG_LOCALLY_DEFINED
35 #undef NDEBUG
36 #endif
37
38 #if defined(__GNUG__)
39 #pragma GCC diagnostic pop
40 #endif
41
42 #include "src/kernel/context/Context.hpp"
43 #include <simgrid/Exception.hpp>
44 #include <simgrid/s4u/Actor.hpp>
45 #include <simgrid/s4u/Comm.hpp>
46 #include <simgrid/s4u/Engine.hpp>
47 #include <simgrid/s4u/Exec.hpp>
48 #include <simgrid/s4u/Host.hpp>
49 #include <simgrid/s4u/Mailbox.hpp>
50 #include <simgrid/version.h>
51
52 #include <algorithm>
53 #include <memory>
54 #include <string>
55 #include <vector>
56
57 namespace py = pybind11;
58 using simgrid::s4u::Actor;
59 using simgrid::s4u::ActorPtr;
60 using simgrid::s4u::Engine;
61 using simgrid::s4u::Host;
62 using simgrid::s4u::Mailbox;
63
64 XBT_LOG_NEW_DEFAULT_CATEGORY(python, "python");
65
66 namespace {
67
68 std::string get_simgrid_version()
69 {
70   int major;
71   int minor;
72   int patch;
73   sg_version_get(&major, &minor, &patch);
74   return simgrid::xbt::string_printf("%i.%i.%i", major, minor, patch);
75 }
76
77 /* Classes GilScopedAcquire and GilScopedRelease have the same purpose as pybind11::gil_scoped_acquire and
78  * pybind11::gil_scoped_release.  Refer to the manual of pybind11 for details:
79  * https://pybind11.readthedocs.io/en/stable/advanced/misc.html#global-interpreter-lock-gil
80  *
81  * The pybind11 versions are however too sophisticated (using TLS for example) and don't work well with all kinds of
82  * contexts.
83  * See also https://github.com/pybind/pybind11/issues/1276, which may be related.
84  *
85  * Briefly, GilScopedAcquire can be used on actor creation to acquire a new PyThreadState.  The PyThreadState has to be
86  * released for context switches (i.e. before simcalls). That's the purpose of GilScopedRelease.
87  *
88  * Like their pybind11 counterparts, both classes use a RAII pattern.
89  */
90 class XBT_PRIVATE GilScopedAcquire {
91   static PyThreadState* acquire()
92   {
93     PyThreadState* state = PyThreadState_New(PyInterpreterState_Head());
94     PyEval_AcquireThread(state);
95     return state;
96   }
97   static void release(PyThreadState* state)
98   {
99     PyEval_ReleaseThread(state);
100     PyThreadState_Clear(state);
101     PyThreadState_Delete(state);
102   }
103
104   std::unique_ptr<PyThreadState, decltype(&release)> thread_state{acquire(), &release};
105
106 public:
107   void reset() { thread_state.reset(); }
108 };
109
110 class XBT_PRIVATE GilScopedRelease {
111   std::unique_ptr<PyThreadState, decltype(&PyEval_RestoreThread)> thread_state{PyEval_SaveThread(),
112                                                                                &PyEval_RestoreThread};
113 };
114
115 } // namespace
116
117 PYBIND11_DECLARE_HOLDER_TYPE(T, boost::intrusive_ptr<T>)
118
119 PYBIND11_MODULE(simgrid, m)
120 {
121   m.doc() = "SimGrid userspace API";
122
123   m.attr("simgrid_version") = get_simgrid_version();
124
125   // Internal exception used to kill actors and sweep the RAII chimney (free objects living on the stack)
126   static py::object pyForcefulKillEx(py::register_exception<simgrid::ForcefulKillException>(m, "ActorKilled"));
127
128   /* this_actor namespace */
129   m.def_submodule("this_actor", "Bindings of the s4u::this_actor namespace.")
130       .def(
131           "info", [](const char* s) { XBT_INFO("%s", s); }, "Display a logging message of 'info' priority.")
132       .def(
133           "error", [](const char* s) { XBT_ERROR("%s", s); }, "Display a logging message of 'error' priority.")
134       .def("execute", py::overload_cast<double, double>(&simgrid::s4u::this_actor::execute),
135            py::call_guard<GilScopedRelease>(),
136            "Block the current actor, computing the given amount of flops at the given priority.", py::arg("flops"),
137            py::arg("priority") = 1)
138       .def("exec_init", py::overload_cast<double>(&simgrid::s4u::this_actor::exec_init),
139            py::call_guard<GilScopedRelease>())
140       .def("get_host", &simgrid::s4u::this_actor::get_host, "Retrieves host on which the current actor is located")
141       .def("set_host", &simgrid::s4u::this_actor::set_host, py::call_guard<GilScopedRelease>(),
142            "Moves the current actor to another host.", py::arg("dest"))
143       .def("sleep_for", static_cast<void (*)(double)>(&simgrid::s4u::this_actor::sleep_for),
144            py::call_guard<GilScopedRelease>(), "Block the actor sleeping for that amount of seconds.",
145            py::arg("duration"))
146       .def("sleep_until", static_cast<void (*)(double)>(&simgrid::s4u::this_actor::sleep_until),
147            py::call_guard<GilScopedRelease>(), "Block the actor sleeping until the specified timestamp.",
148            py::arg("duration"))
149       .def("suspend", &simgrid::s4u::this_actor::suspend, py::call_guard<GilScopedRelease>(),
150            "Suspend the current actor, that is blocked until resume()ed by another actor.")
151       .def("yield_", &simgrid::s4u::this_actor::yield, py::call_guard<GilScopedRelease>(), "Yield the actor")
152       .def("exit", &simgrid::s4u::this_actor::exit, py::call_guard<GilScopedRelease>(), "kill the current actor")
153       .def(
154           "on_exit",
155           [](py::object fun) {
156             fun.inc_ref(); // FIXME: why is this needed for tests like actor-kill and actor-lifetime?
157             simgrid::s4u::this_actor::on_exit([fun](bool /*failed*/) {
158               GilScopedAcquire py_context; // need a new context for callback
159               try {
160                 fun();
161               } catch (const py::error_already_set& e) {
162                 std::string what = e.what();
163                 py_context.reset();
164                 xbt_die("Error while executing the on_exit lambda: %s", what.c_str());
165               }
166             });
167           },
168           py::call_guard<GilScopedRelease>(), "");
169
170   /* Class Engine */
171   py::class_<Engine>(m, "Engine", "Simulation Engine")
172       .def(py::init([](std::vector<std::string> args) {
173         auto argc           = static_cast<int>(args.size());
174         std::vector<char*> argv(args.size() + 1); // argv[argc] is nullptr
175         std::transform(begin(args), end(args), begin(argv), [](std::string& s) { return &s.front(); });
176         // Currently this can be dangling, we should wrap this somehow.
177         return new simgrid::s4u::Engine(&argc, argv.data());
178       }))
179       .def_static("get_clock", &Engine::get_clock,
180                   "The simulation time, ie the amount of simulated seconds since the simulation start.")
181       .def("get_all_hosts", &Engine::get_all_hosts, "Returns the list of all hosts found in the platform")
182       .def("load_platform", &Engine::load_platform, "Load a platform file describing the environment")
183       .def("load_deployment", &Engine::load_deployment, "Load a deployment file and launch the actors that it contains")
184       .def("run", &Engine::run, py::call_guard<GilScopedRelease>(), "Run the simulation")
185       .def(
186           "register_actor",
187           [](Engine* e, const std::string& name, py::object fun_or_class) {
188             e->register_actor(name, [fun_or_class](std::vector<std::string> args) {
189               GilScopedAcquire py_context;
190               try {
191                 /* Convert the std::vector into a py::tuple */
192                 py::tuple params(args.size() - 1);
193                 for (size_t i = 1; i < args.size(); i++)
194                   params[i - 1] = py::cast(args[i]);
195
196                 py::object res = fun_or_class(*params);
197                 /* If I was passed a class, I just built an instance, so I need to call it now */
198                 if (py::isinstance<py::function>(res))
199                   res();
200               } catch (const py::error_already_set& ex) {
201                 bool ffk = ex.matches(pyForcefulKillEx);
202                 py_context.reset();
203                 if (ffk) {
204                   XBT_VERB("Actor killed");
205                   simgrid::ForcefulKillException::do_throw(); // Forward that ForcefulKill exception
206                 }
207                 throw;
208               }
209             });
210           },
211           "Registers the main function of an actor");
212
213   /* Class Host */
214   py::class_<simgrid::s4u::Host, std::unique_ptr<Host, py::nodelete>>(m, "Host", "Simulated host")
215       .def("by_name", &Host::by_name, "Retrieves a host from its name, or die")
216       .def("get_pstate_count", &Host::get_pstate_count, "Retrieve the count of defined pstate levels")
217       .def("get_pstate_speed", &Host::get_pstate_speed, "Retrieve the maximal speed at the given pstate")
218       .def_property(
219           "pstate", &Host::get_pstate,
220           [](Host* h, int i) {
221             GilScopedRelease gil_guard;
222             h->set_pstate(i);
223           },
224           "The current pstate")
225       .def("current", &Host::current, py::call_guard<GilScopedRelease>(),
226            "Retrieves the host on which the running actor is located.")
227       .def_property_readonly(
228           "name",
229           [](const Host* self) {
230             return std::string(self->get_name().c_str()); // Convert from xbt::string because of MC
231           },
232           "The name of this host")
233       .def_property_readonly(
234           "load", &Host::get_load,
235           "Returns the current computation load (in flops per second). This is the currently achieved speed.")
236       .def_property_readonly(
237           "speed", &Host::get_speed,
238           "The peak computing speed in flops/s at the current pstate, taking the external load into account. "
239           "This is the max potential speed.");
240
241   /* Class Mailbox */
242   py::class_<simgrid::s4u::Mailbox, std::unique_ptr<Mailbox, py::nodelete>>(m, "Mailbox", "Mailbox")
243       .def(
244           "__str__", [](const Mailbox* self) { return std::string("Mailbox(") + self->get_cname() + ")"; },
245           "Textual representation of the Mailbox`")
246       .def("by_name", &Mailbox::by_name, py::call_guard<GilScopedRelease>(), "Retrieve a Mailbox from its name")
247       .def_property_readonly(
248           "name",
249           [](const Mailbox* self) {
250             return std::string(self->get_name().c_str()); // Convert from xbt::string because of MC
251           },
252           "The name of that mailbox")
253       .def(
254           "put",
255           [](Mailbox* self, py::object data, int size) {
256             data.inc_ref();
257             self->put(data.ptr(), size);
258           },
259           py::call_guard<GilScopedRelease>(), "Blocking data transmission")
260       .def(
261           "put_async",
262           [](Mailbox* self, py::object data, int size) {
263             data.inc_ref();
264             return self->put_async(data.ptr(), size);
265           },
266           py::call_guard<GilScopedRelease>(), "Non-blocking data transmission")
267       .def(
268           "get",
269           [](Mailbox* self) {
270             py::object data = pybind11::reinterpret_steal<py::object>(self->get<PyObject>());
271             // data.dec_ref(); // FIXME: why does it break python-actor-create?
272             return data;
273           },
274           py::call_guard<GilScopedRelease>(), "Blocking data reception")
275       .def("set_receiver",
276          [](Mailbox* self, ActorPtr actor) {
277            self->set_receiver(actor);
278          },
279          py::call_guard<GilScopedRelease>(),
280          "Sets the actor as permanent receiver");
281
282   /* Class Comm */
283   py::class_<simgrid::s4u::Comm, simgrid::s4u::CommPtr>(m, "Comm", "Communication")
284       .def("test", &simgrid::s4u::Comm::test, py::call_guard<GilScopedRelease>(),
285            "Test whether the communication is terminated.")
286       .def("wait", &simgrid::s4u::Comm::wait, py::call_guard<GilScopedRelease>(),
287            "Block until the completion of that communication.")
288       // use py::overload_cast for wait_all/wait_any, until the overload marked XBT_ATTRIB_DEPRECATED_v332 is removed
289       .def_static("wait_all",
290                   py::overload_cast<const std::vector<simgrid::s4u::CommPtr>&>(&simgrid::s4u::Comm::wait_all),
291                   py::call_guard<GilScopedRelease>(), "Block until the completion of all communications in the list.")
292       .def_static(
293           "wait_any", py::overload_cast<const std::vector<simgrid::s4u::CommPtr>&>(&simgrid::s4u::Comm::wait_any),
294           py::call_guard<GilScopedRelease>(),
295           "Block until the completion of any communication in the list and return the index of the terminated one.");
296
297   /* Class Exec */
298   py::class_<simgrid::s4u::Exec, simgrid::s4u::ExecPtr>(m, "Exec", "Execution")
299       .def_property_readonly(
300           "remaining",
301           [](simgrid::s4u::ExecPtr self) {
302             GilScopedRelease gil_guard;
303             return self->get_remaining();
304           },
305           "Amount of flops that remain to be computed until completion.")
306       .def_property_readonly(
307           "remaining_ratio",
308           [](simgrid::s4u::ExecPtr self) {
309             GilScopedRelease gil_guard;
310             return self->get_remaining_ratio();
311           },
312           "Amount of work remaining until completion from 0 (completely done) to 1 (nothing done "
313           "yet).")
314       .def_property("host", &simgrid::s4u::Exec::get_host, &simgrid::s4u::Exec::set_host,
315                     "Host on which this execution runs. Only the first host is returned for parallel executions.")
316       .def("test", &simgrid::s4u::Exec::test, py::call_guard<GilScopedRelease>(),
317            "Test whether the execution is terminated.")
318       .def("cancel", &simgrid::s4u::Exec::cancel, py::call_guard<GilScopedRelease>(), "Cancel that execution.")
319       .def("start", &simgrid::s4u::Exec::start, py::call_guard<GilScopedRelease>(), "Start that execution.")
320       .def("wait", &simgrid::s4u::Exec::wait, py::call_guard<GilScopedRelease>(),
321            "Block until the completion of that execution.");
322
323   /* Class Actor */
324   py::class_<simgrid::s4u::Actor, ActorPtr>(m, "Actor",
325                                             "An actor is an independent stream of execution in your distributed "
326                                             "application")
327       .def(
328           "create",
329           [](py::str name, Host* host, py::object fun, py::args args) {
330             fun.inc_ref();  // FIXME: why is this needed for tests like exec-async, exec-dvfs and exec-remote?
331             args.inc_ref(); // FIXME: why is this needed for tests like actor-migrate?
332             return simgrid::s4u::Actor::create(name, host, [fun, args]() {
333               GilScopedAcquire py_context;
334               try {
335                 fun(*args);
336               } catch (const py::error_already_set& ex) {
337                 bool ffk = ex.matches(pyForcefulKillEx);
338                 py_context.reset();
339                 if (ffk) {
340                   XBT_VERB("Actor killed");
341                   simgrid::ForcefulKillException::do_throw(); // Forward that ForcefulKill exception
342                 }
343                 throw;
344               }
345             });
346           },
347           py::call_guard<GilScopedRelease>(), "Create an actor from a function or an object.")
348       .def_property(
349           "host", &Actor::get_host,
350           [](Actor* a, Host* h) {
351             GilScopedRelease gil_guard;
352             a->set_host(h);
353           },
354           "The host on which this actor is located")
355       .def_property_readonly("name", &Actor::get_cname, "The name of this actor.")
356       .def_property_readonly("pid", &Actor::get_pid, "The PID (unique identifier) of this actor.")
357       .def_property_readonly("ppid", &Actor::get_ppid,
358                              "The PID (unique identifier) of the actor that created this one.")
359       .def("by_pid", &Actor::by_pid, "Retrieve an actor by its PID")
360       .def("daemonize", &Actor::daemonize, py::call_guard<GilScopedRelease>(),
361            "This actor will be automatically terminated when the last non-daemon actor finishes (more info in the C++ "
362            "documentation).")
363       .def("is_daemon", &Actor::is_daemon,
364            "Returns True if that actor is a daemon and will be terminated automatically when the last non-daemon actor "
365            "terminates.")
366       .def("join", py::overload_cast<double>(&Actor::join, py::const_), py::call_guard<GilScopedRelease>(),
367            "Wait for the actor to finish (more info in the C++ documentation).", py::arg("timeout"))
368       .def("kill", &Actor::kill, py::call_guard<GilScopedRelease>(), "Kill that actor")
369       .def("kill_all", &Actor::kill_all, py::call_guard<GilScopedRelease>(), "Kill all actors but the caller.")
370       .def("self", &Actor::self, "Retrieves the current actor.")
371       .def("is_suspended", &Actor::is_suspended, "Returns True if that actor is currently suspended.")
372       .def("suspend", &Actor::suspend, py::call_guard<GilScopedRelease>(),
373            "Suspend that actor, that is blocked until resume()ed by another actor.")
374       .def("resume", &Actor::resume, py::call_guard<GilScopedRelease>(),
375            "Resume that actor, that was previously suspend()ed.");
376 }