1 /* Copyright (c) 2009-2022. The SimGrid Team. All rights reserved. */
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. */
6 #include "simgrid/Exception.hpp"
7 #include "simgrid/modelchecker.h"
8 #include "src/internal_config.h"
9 #include "src/kernel/EngineImpl.hpp"
10 #include "src/kernel/actor/ActorImpl.hpp"
11 #include "xbt/parmap.hpp"
13 #include "src/kernel/context/ContextSwapped.hpp"
15 #include <boost/core/demangle.hpp>
27 #define _aligned_malloc __mingw_aligned_malloc
28 #define _aligned_free __mingw_aligned_free
32 #include <valgrind/valgrind.h>
34 #if HAVE_SANITIZER_ADDRESS_FIBER_SUPPORT
35 #include <sanitizer/asan_interface.h>
37 #if HAVE_SANITIZER_THREAD_FIBER_SUPPORT
38 #include <sanitizer/tsan_interface.h>
41 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(ker_context);
43 // The name of this function is currently hardcoded in MC (as string).
44 // Do not change it without fixing those references as well.
45 void smx_ctx_wrapper(simgrid::kernel::context::SwappedContext* context)
47 #if HAVE_SANITIZER_ADDRESS_FIBER_SUPPORT
48 __sanitizer_finish_switch_fiber(nullptr, &context->asan_ctx_->asan_stack_, &context->asan_ctx_->asan_stack_size_);
52 context->Context::stop();
53 } catch (simgrid::ForcefulKillException const&) {
54 XBT_DEBUG("Caught a ForcefulKillException");
55 } catch (simgrid::Exception const& e) {
56 XBT_INFO("Actor killed by an uncaught exception %s", boost::core::demangle(typeid(e).name()).c_str());
59 #if HAVE_SANITIZER_ADDRESS_FIBER_SUPPORT
60 context->asan_stop_ = true;
70 /* thread-specific storage for the worker's context */
71 thread_local SwappedContext* SwappedContext::worker_context_ = nullptr;
73 SwappedContext::SwappedContext(std::function<void()>&& code, smx_actor_t actor, SwappedContextFactory* factory)
74 : Context(std::move(code), actor, not code /* maestro if no code */), factory_(*factory)
76 // Save maestro (=first created context) in preparation for run_all
77 if (not is_parallel() && factory_.maestro_context_ == nullptr)
78 factory_.maestro_context_ = this;
81 xbt_assert((actor->get_stacksize() & 0xf) == 0, "Actor stack size should be multiple of 16");
82 if (guard_size > 0 && not MC_is_active()) {
83 #if PTH_STACKGROWTH != -1
85 "Stack overflow protection is known to be broken on your system: you stacks grow upwards (or detection is "
87 "Please disable stack guards with --cfg=contexts:guard-size:0");
88 /* Current code for stack overflow protection assumes that stacks are growing downward (PTH_STACKGROWTH == -1).
89 * Protected pages need to be put after the stack when PTH_STACKGROWTH == 1. */
92 size_t size = actor->get_stacksize() + guard_size;
94 /* Cannot use posix_memalign when SIMGRID_HAVE_MC. Align stack by hand, and save the
95 * pointer returned by xbt_malloc0. */
96 auto* alloc = static_cast<unsigned char*>(xbt_malloc0(size + xbt_pagesize));
97 stack_ = alloc - (reinterpret_cast<uintptr_t>(alloc) & (xbt_pagesize - 1)) + xbt_pagesize;
98 reinterpret_cast<unsigned char**>(stack_)[-1] = alloc;
99 #elif !defined(_WIN32)
101 xbt_assert(posix_memalign(&alloc, xbt_pagesize, size) == 0, "Failed to allocate stack.");
102 this->stack_ = static_cast<unsigned char*>(alloc);
104 this->stack_ = static_cast<unsigned char*>(_aligned_malloc(size, xbt_pagesize));
108 /* This is fatal. We are going to fail at some point when we try reusing this. */
110 mprotect(this->stack_, guard_size, PROT_NONE) != -1,
111 "Failed to protect stack: %s.\n"
112 "If you are running a lot of actors, you may be exceeding the amount of mappings allowed per process.\n"
113 "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
114 "Please see https://simgrid.org/doc/latest/Configuring_SimGrid.html#configuring-the-user-code-virtualization "
115 "for more information.",
118 this->stack_ = this->stack_ + guard_size;
120 this->stack_ = static_cast<unsigned char*>(xbt_malloc0(actor->get_stacksize()));
124 if (RUNNING_ON_VALGRIND)
125 this->valgrind_stack_id_ = VALGRIND_STACK_REGISTER(this->stack_, this->stack_ + actor->get_stacksize());
127 #if HAVE_SANITIZER_ADDRESS_FIBER_SUPPORT
128 this->asan_stack_ = get_stack_bottom();
130 #if HAVE_SANITIZER_THREAD_FIBER_SUPPORT
131 this->tsan_fiber_ = __tsan_create_fiber(0);
134 // not has_code(): in maestro context
135 #if HAVE_SANITIZER_THREAD_FIBER_SUPPORT
136 this->tsan_fiber_ = __tsan_get_current_fiber();
141 SwappedContext::~SwappedContext()
143 if (stack_ == nullptr) // maestro has no extra stack
146 #if HAVE_SANITIZER_THREAD_FIBER_SUPPORT
147 __tsan_destroy_fiber(tsan_fiber_);
150 if (valgrind_stack_id_ != 0)
151 VALGRIND_STACK_DEREGISTER(valgrind_stack_id_);
155 if (guard_size > 0 && not MC_is_active()) {
156 stack_ = stack_ - guard_size;
157 if (mprotect(stack_, guard_size, PROT_READ | PROT_WRITE) == -1) {
158 XBT_WARN("Failed to remove page protection: %s", strerror(errno));
159 /* try to pursue anyway */
162 /* Retrieve the saved pointer. See the initialization above. */
163 stack_ = reinterpret_cast<unsigned char**>(stack_)[-1];
166 #endif /* not windows */
171 unsigned char* SwappedContext::get_stack_bottom() const
173 // Depending on the stack direction, its bottom (that make_fcontext needs) may be the lower or higher end
174 #if PTH_STACKGROWTH == 1
177 return stack_ + get_actor()->get_stacksize();
181 void SwappedContext::stop()
184 /* We must cut the actor execution using an exception to properly free the C++ RAII variables */
185 throw ForcefulKillException();
188 void SwappedContext::swap_into(SwappedContext* to)
190 #if HAVE_SANITIZER_ADDRESS_FIBER_SUPPORT
191 void* fake_stack = nullptr;
192 to->asan_ctx_ = this;
193 __sanitizer_start_switch_fiber(this->asan_stop_ ? nullptr : &fake_stack, to->asan_stack_, to->asan_stack_size_);
195 #if HAVE_SANITIZER_THREAD_FIBER_SUPPORT
196 __tsan_switch_to_fiber(to->tsan_fiber_, 0);
199 swap_into_for_real(to);
201 #if HAVE_SANITIZER_ADDRESS_FIBER_SUPPORT
202 __sanitizer_finish_switch_fiber(fake_stack, &this->asan_ctx_->asan_stack_, &this->asan_ctx_->asan_stack_size_);
206 /** Maestro wants to run all ready actors */
207 void SwappedContextFactory::run_all()
209 const auto* engine = EngineImpl::get_instance();
210 /* This function is called by maestro at the beginning of a scheduling round to get all working threads executing some
211 * stuff It is much easier to understand what happens if you see the working threads as bodies that swap their soul
212 * for the ones of the simulated processes that must run.
215 // We lazily create the parmap so that all options are actually processed when doing so.
216 if (parmap_ == nullptr)
217 parmap_ = std::make_unique<simgrid::xbt::Parmap<smx_actor_t>>(get_nthreads(), get_parallel_mode());
219 // Usually, Parmap::apply() executes the provided function on all elements of the array.
220 // Here, the executed function does not return the control to the parmap before all the array is processed:
221 // - suspend() should switch back to the worker_context (either maestro or one of its minions) to return
222 // the control to the parmap. Instead, it uses parmap_->next() to steal another work, and does it directly.
223 // It only yields back to worker_context when the work array is exhausted.
224 // - So, resume() is only launched from the parmap for the first job of each minion.
226 [](const actor::ActorImpl* actor) {
227 auto* context = static_cast<SwappedContext*>(actor->context_.get());
230 engine->get_actors_to_run());
231 } else { // sequential execution
232 if (not engine->has_actors_to_run())
235 /* maestro is already saved in the first slot of workers_context_ */
236 const actor::ActorImpl* first_actor = engine->get_first_actor_to_run();
238 /* execute the first actor; it will chain to the others when using suspend() */
239 static_cast<SwappedContext*>(first_actor->context_.get())->resume();
243 /** Maestro wants to yield back to a given actor, so awake it on the current thread
245 * In parallel, it is only applied to the N first elements of the parmap array,
246 * where N is the amount of worker threads in the parmap.
247 * See SwappedContextFactory::run_all for details.
249 void SwappedContext::resume()
251 auto* old = static_cast<SwappedContext*>(self());
253 // Save my current soul (either maestro, or one of the minions) in a thread-specific area
254 worker_context_ = old;
256 // Switch my soul and the actor's one
257 Context::set_current(this);
258 old->swap_into(this);
259 // No body runs that soul anymore at this point, but it is stored in a safe place.
260 // When the executed actor will do a blocking action, ActorImpl::yield() will call suspend(), below.
263 /** The actor wants to yield back to maestro, because it is blocked in a simcall (i.e., in ActorImpl::yield())
265 * Actually, it does not really yield back to maestro, but directly into the next executable actor.
267 * This makes the parmap::apply awkward (see SwappedContextFactory::run_all()) because it only apply regularly
268 * on the few first elements of the array, but it saves a lot of context switches back to maestro,
269 * and directly forth to the next executable actor.
271 void SwappedContext::suspend()
273 SwappedContext* next_context;
275 // Get some more work to directly swap into the next executable actor instead of yielding back to the parmap
276 boost::optional<smx_actor_t> next_work = factory_.parmap_->next();
278 // There is a next soul to embody (ie, another executable actor)
279 XBT_DEBUG("Run next process");
280 next_context = static_cast<SwappedContext*>(next_work.get()->context_.get());
282 // All actors were run, go back to the parmap context
283 XBT_DEBUG("No more actors to run");
284 // worker_context_ is my own soul, stored in thread_local when starting the scheduling round
285 next_context = worker_context_;
286 // When given that soul, the body will wait for the next scheduling round
288 } else { // sequential execution
289 const auto* engine = EngineImpl::get_instance();
290 /* determine the next context */
291 unsigned long int i = factory_.process_index_;
292 factory_.process_index_++;
294 if (i < engine->get_actor_to_run_count()) {
295 /* Actually swap into the next actor directly without transiting to maestro */
296 XBT_DEBUG("Run next actor");
297 next_context = static_cast<SwappedContext*>(engine->get_actor_to_run_at(i)->context_.get());
299 /* all processes were run, actually return to maestro */
300 XBT_DEBUG("No more actors to run");
301 next_context = factory_.maestro_context_;
304 Context::set_current(next_context);
305 this->swap_into(next_context);
308 } // namespace context
309 } // namespace kernel
310 } // namespace simgrid