-/* Copyright (c) 2009-2018. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2009-2019. The SimGrid Team. All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
+#include "simgrid/Exception.hpp"
#include "simgrid/modelchecker.h"
#include "src/internal_config.h"
#include "src/kernel/context/context_private.hpp"
namespace kernel {
namespace context {
-/* Sequential execution */
-unsigned long SwappedContext::process_index_;
-
-/* Parallel execution */
-simgrid::xbt::Parmap<smx_actor_t>* SwappedContext::parmap_;
-std::atomic<uintptr_t> SwappedContext::threads_working_; /* number of threads that have started their work */
+/* rank of the execution thread */
thread_local uintptr_t SwappedContext::worker_id_; /* thread-specific storage for the thread id */
-std::vector<SwappedContext*> SwappedContext::workers_context_; /* space to save the worker's context in each thread */
-void SwappedContext::initialize()
+SwappedContextFactory::SwappedContextFactory() : ContextFactory(), parallel_(SIMIX_context_is_parallel())
{
- parmap_ = nullptr;
- workers_context_.clear();
- workers_context_.resize(SIMIX_context_get_nthreads(), nullptr);
+ parmap_ = nullptr; // will be created lazily with the right parameters if needed (ie, in parallel)
+ workers_context_.resize(parallel_ ? SIMIX_context_get_nthreads() : 1, nullptr);
}
-void SwappedContext::finalize()
+SwappedContext::SwappedContext(std::function<void()>&& code, smx_actor_t actor, SwappedContextFactory* factory)
+ : Context(std::move(code), actor), factory_(factory)
{
- delete parmap_;
- parmap_ = nullptr;
- workers_context_.clear();
-}
+ // Save maestro (=context created first) in preparation for run_all
+ if (not factory->parallel_ && factory_->workers_context_[0] == nullptr)
+ factory_->workers_context_[0] = this;
-SwappedContext* SwappedContext::maestro_context_ = nullptr;
-
-SwappedContext::SwappedContext(std::function<void()> code, void_pfn_smxprocess_t cleanup_func, smx_actor_t process)
- : Context(std::move(code), cleanup_func, process)
-{
if (has_code()) {
if (smx_context_guard_size > 0 && not MC_is_active()) {
#if SIMGRID_HAVE_MC
/* Cannot use posix_memalign when SIMGRID_HAVE_MC. Align stack by hand, and save the
* pointer returned by xbt_malloc0. */
- char* alloc = (char*)xbt_malloc0(size + xbt_pagesize);
- stack_ = alloc - ((uintptr_t)alloc & (xbt_pagesize - 1)) + xbt_pagesize;
- *((void**)stack_ - 1) = alloc;
+ unsigned char* alloc = static_cast<unsigned char*>(xbt_malloc0(size + xbt_pagesize));
+ stack_ = alloc - (reinterpret_cast<uintptr_t>(alloc) & (xbt_pagesize - 1)) + xbt_pagesize;
+ reinterpret_cast<unsigned char**>(stack_)[-1] = alloc;
#elif !defined(_WIN32)
- if (posix_memalign(&this->stack_, xbt_pagesize, size) != 0)
+ void* alloc;
+ if (posix_memalign(&alloc, xbt_pagesize, size) != 0)
xbt_die("Failed to allocate stack.");
+ this->stack_ = static_cast<unsigned char*>(alloc);
#else
- this->stack_ = _aligned_malloc(size, xbt_pagesize);
+ this->stack_ = static_cast<unsigned char*>(_aligned_malloc(size, xbt_pagesize));
#endif
#ifndef _WIN32
/* This is fatal. We are going to fail at some point when we try reusing this. */
}
#endif
- this->stack_ = (char*)this->stack_ + smx_context_guard_size;
+ this->stack_ = this->stack_ + smx_context_guard_size;
} else {
- this->stack_ = xbt_malloc0(smx_context_stack_size);
+ this->stack_ = static_cast<unsigned char*>(xbt_malloc0(smx_context_stack_size));
}
+#if PTH_STACKGROWTH == -1
+ ASAN_ONLY(this->asan_stack_ = this->stack_ + smx_context_usable_stack_size);
+#else
+ ASAN_ONLY(this->asan_stack_ = this->stack_);
+#endif
#if HAVE_VALGRIND_H
- unsigned int valgrind_stack_id =
- VALGRIND_STACK_REGISTER(this->stack_, (char*)this->stack_ + smx_context_stack_size);
- memcpy((char*)this->stack_ + smx_context_usable_stack_size, &valgrind_stack_id, sizeof valgrind_stack_id);
+ unsigned int valgrind_stack_id = VALGRIND_STACK_REGISTER(this->stack_, this->stack_ + smx_context_stack_size);
+ memcpy(this->stack_ + smx_context_usable_stack_size, &valgrind_stack_id, sizeof valgrind_stack_id);
#endif
}
}
SwappedContext::~SwappedContext()
{
- if (stack_ == nullptr)
+ if (stack_ == nullptr) // maestro has no extra stack
return;
#if HAVE_VALGRIND_H
unsigned int valgrind_stack_id;
- memcpy(&valgrind_stack_id, (char*)stack_ + smx_context_usable_stack_size, sizeof valgrind_stack_id);
+ memcpy(&valgrind_stack_id, stack_ + smx_context_usable_stack_size, sizeof valgrind_stack_id);
VALGRIND_STACK_DEREGISTER(valgrind_stack_id);
#endif
#ifndef _WIN32
if (smx_context_guard_size > 0 && not MC_is_active()) {
- stack_ = (char*)stack_ - smx_context_guard_size;
+ stack_ = stack_ - smx_context_guard_size;
if (mprotect(stack_, smx_context_guard_size, PROT_READ | PROT_WRITE) == -1) {
XBT_WARN("Failed to remove page protection: %s", strerror(errno));
/* try to pursue anyway */
}
#if SIMGRID_HAVE_MC
/* Retrieve the saved pointer. See SIMIX_context_stack_new above. */
- stack_ = *((void**)stack_ - 1);
+ stack_ = reinterpret_cast<unsigned char**>(stack_)[-1];
#endif
}
#endif /* not windows */
xbt_free(stack_);
}
-/** Maestro wants to run all read_to_run actors */
-void SwappedContext::run_all()
+unsigned char* SwappedContext::get_stack()
{
- if (simix_global->process_to_run.empty())
- return;
- smx_actor_t first_process = simix_global->process_to_run.front();
- process_index_ = 1;
- /* execute the first process */
- static_cast<SwappedContext*>(first_process->context_)->resume();
+ return stack_;
}
-/** Maestro wants to yield back to a given actor */
+void SwappedContext::stop()
+{
+ Context::stop();
+ /* We must cut the actor execution using an exception to properly free the C++ RAII variables */
+ throw ForcefulKillException();
+}
+
+/** Maestro wants to run all ready actors */
+void SwappedContextFactory::run_all()
+{
+ /* This function is called by maestro at the beginning of a scheduling round to get all working threads executing some
+ * stuff It is much easier to understand what happens if you see the working threads as bodies that swap their soul
+ * for the ones of the simulated processes that must run.
+ */
+ if (parallel_) {
+ threads_working_ = 0;
+
+ // We lazily create the parmap so that all options are actually processed when doing so.
+ if (parmap_ == nullptr)
+ parmap_.reset(
+ new simgrid::xbt::Parmap<smx_actor_t>(SIMIX_context_get_nthreads(), SIMIX_context_get_parallel_mode()));
+
+ // Usually, Parmap::apply() executes the provided function on all elements of the array.
+ // Here, the executed function does not return the control to the parmap before all the array is processed:
+ // - suspend() should switch back to the worker_context (either maestro or one of its minions) to return
+ // the control to the parmap. Instead, it uses parmap_->next() to steal another work, and does it directly.
+ // It only yields back to worker_context when the work array is exhausted.
+ // - So, resume() is only launched from the parmap for the first job of each minion.
+ parmap_->apply(
+ [](smx_actor_t process) {
+ SwappedContext* context = static_cast<SwappedContext*>(process->context_.get());
+ context->resume();
+ },
+ simix_global->actors_to_run);
+ } else { // sequential execution
+ if (simix_global->actors_to_run.empty())
+ return;
+
+ /* maestro is already saved in the first slot of workers_context_ */
+ smx_actor_t first_actor = simix_global->actors_to_run.front();
+ process_index_ = 1;
+ /* execute the first actor; it will chain to the others when using suspend() */
+ static_cast<SwappedContext*>(first_actor->context_.get())->resume();
+ }
+}
+
+/** Maestro wants to yield back to a given actor, so awake it on the current thread
+ *
+ * In parallel, it is only applied to the N first elements of the parmap array,
+ * where N is the amount of worker threads in the parmap.
+ * See SwappedContextFactory::run_all for details.
+ */
void SwappedContext::resume()
{
- // Maestro is always the calling thread of this function (ie, self() == maestro)
- SwappedContext* old = static_cast<SwappedContext*>(self());
- Context::set_current(this);
- old->swap_into(this);
+ if (factory_->parallel_) {
+ // Save the thread number (my body) in an os-thread-specific area
+ worker_id_ = factory_->threads_working_.fetch_add(1, std::memory_order_relaxed);
+ // Save my current soul (either maestro, or one of the minions) in a permanent area
+ SwappedContext* worker_context = static_cast<SwappedContext*>(self());
+ factory_->workers_context_[worker_id_] = worker_context;
+ // Switch my soul and the actor's one
+ Context::set_current(this);
+ worker_context->swap_into(this);
+ // No body runs that soul anymore at this point, but it is stored in a safe place.
+ // When the executed actor will do a blocking action, ActorImpl::yield() will call suspend(), below.
+ } else { // sequential execution
+ SwappedContext* old = static_cast<SwappedContext*>(self());
+ Context::set_current(this);
+ old->swap_into(this);
+ }
}
-/** The actor wants to yield back to maestro */
+/** The actor wants to yield back to maestro, because it is blocked in a simcall (i.e., in ActorImpl::yield())
+ *
+ * Actually, it does not really yield back to maestro, but directly into the next executable actor.
+ *
+ * This makes the parmap::apply awkward (see ParallelUContext::run_all()) because it only apply regularly
+ * on the few first elements of the array, but it saves a lot of context switches back to maestro,
+ * and directly forth to the next executable actor.
+ */
void SwappedContext::suspend()
{
- /* determine the next context */
- SwappedContext* next_context;
- unsigned long int i = process_index_;
- process_index_++;
-
- if (i < simix_global->process_to_run.size()) {
- /* Actually swap into the next actor directly without transiting to maestro */
- XBT_DEBUG("Run next process");
- next_context = static_cast<SwappedContext*>(simix_global->process_to_run[i]->context_);
- } else {
- /* all processes were run, actually return to maestro */
- XBT_DEBUG("No more process to run");
- next_context = static_cast<SwappedContext*>(get_maestro());
+ if (factory_->parallel_) {
+ // Get some more work to directly swap into the next executable actor instead of yielding back to the parmap
+ boost::optional<smx_actor_t> next_work = factory_->parmap_->next();
+ SwappedContext* next_context;
+ if (next_work) {
+ // There is a next soul to embody (ie, another executable actor)
+ XBT_DEBUG("Run next process");
+ next_context = static_cast<SwappedContext*>(next_work.get()->context_.get());
+ } else {
+ // All actors were run, go back to the parmap context
+ XBT_DEBUG("No more actors to run");
+ // worker_id_ is the identity of my body, stored in thread_local when starting the scheduling round
+ next_context = factory_->workers_context_[worker_id_];
+ // When given that soul, the body will wait for the next scheduling round
+ }
+
+ // Get the next soul to run, either from another actor or the initial minion's one
+ Context::set_current(next_context);
+ this->swap_into(next_context);
+
+ } else { // sequential execution
+ /* determine the next context */
+ SwappedContext* next_context;
+ unsigned long int i = factory_->process_index_;
+ factory_->process_index_++;
+
+ if (i < simix_global->actors_to_run.size()) {
+ /* Actually swap into the next actor directly without transiting to maestro */
+ XBT_DEBUG("Run next actor");
+ next_context = static_cast<SwappedContext*>(simix_global->actors_to_run[i]->context_.get());
+ } else {
+ /* all processes were run, actually return to maestro */
+ XBT_DEBUG("No more actors to run");
+ next_context = factory_->workers_context_[0];
+ }
+ Context::set_current(next_context);
+ this->swap_into(next_context);
}
- Context::set_current(next_context);
- this->swap_into(next_context);
}
} // namespace context
