1 /* Copyright (c) 2009-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 /** \file RawContext.cpp
8 * Fast context switching inspired from SystemV ucontexts.
10 * In contrast to System V context, it does not touch the signal mask
11 * which avoids making a system call (at least on Linux).
20 #include <xbt/parmap.h>
21 #include <xbt/dynar.h>
23 #include "smx_private.h"
24 #include "smx_private.hpp"
27 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
29 // ***** Class definitions
35 class RawContextFactory;
37 class RawContext : public Context {
39 void* stack_ = nullptr;
40 /** pointer to top the stack stack */
41 void* stack_top_ = nullptr;
43 friend class RawContextFactory;
44 RawContext(std::function<void()> code,
45 void_pfn_smxprocess_t cleanup_func,
46 smx_process_t process);
49 static void wrapper(void* arg);
51 void suspend() override;
54 void suspend_serial();
55 void suspend_parallel();
57 void resume_parallel();
60 class RawContextFactory : public ContextFactory {
64 RawContext* create_context(std::function<void()> code,
65 void_pfn_smxprocess_t, smx_process_t process) override;
66 void run_all() override;
68 void run_all_adaptative();
69 void run_all_serial();
70 void run_all_parallel();
73 ContextFactory* raw_factory()
75 XBT_VERB("Using raw contexts. Because the glibc is just not good enough for us.");
76 return new RawContextFactory();
82 // ***** Loads of static stuff
84 #ifdef HAVE_THREAD_CONTEXTS
85 static xbt_parmap_t raw_parmap;
86 static simgrid::simix::RawContext** raw_workers_context; /* space to save the worker context in each thread */
87 static uintptr_t raw_threads_working; /* number of threads that have started their work */
88 static xbt_os_thread_key_t raw_worker_id_key; /* thread-specific storage for the thread id */
90 #ifdef ADAPTIVE_THRESHOLD
91 #define SCHED_ROUND_LIMIT 5
92 static xbt_os_timer_t round_time;
93 static double par_time,seq_time;
94 static double par_ratio,seq_ratio;
95 static int reached_seq_limit, reached_par_limit;
96 static unsigned int par_proc_that_ran = 0,seq_proc_that_ran = 0; /* Counters of processes that have run in SCHED_ROUND_LIMIT scheduling rounds */
97 static unsigned int seq_sched_round=0, par_sched_round=0; /* Amount of SR that ran serial/parallel*/
98 /*Varables used to calculate running variance and mean*/
99 static double prev_avg_par_proc=0,prev_avg_seq_proc=0;
100 static double delta=0;
101 static double s_par_proc=0,s_seq_proc=0; /*Standard deviation of number of processes computed in par/seq during the current simulation*/
102 static double avg_par_proc=0,sd_par_proc=0;
103 static double avg_seq_proc=0,sd_seq_proc=0;
104 static long long par_window=(long long)HUGE_VAL,seq_window=0;
106 static unsigned long raw_process_index = 0; /* index of the next process to run in the
107 * list of runnable processes */
108 static simgrid::simix::RawContext* raw_maestro_context;
110 static bool raw_context_parallel = false;
111 #ifdef ADAPTIVE_THRESHOLD
112 static bool raw_context_adaptative = false;
115 // ***** Raw context routines
117 typedef void (*rawctx_entry_point_t)(void *);
119 typedef void* raw_stack_t;
120 extern "C" raw_stack_t raw_makecontext(void* malloced_stack, int stack_size,
121 rawctx_entry_point_t entry_point, void* arg);
122 extern "C" void raw_swapcontext(raw_stack_t* old, raw_stack_t new_context);
126 #if defined(__APPLE__)
128 ".globl _raw_makecontext\n"
129 "_raw_makecontext:\n"
130 #elif defined(_WIN32)
132 ".globl raw_makecontext\n"
136 ".globl raw_makecontext\n"
137 ".type raw_makecontext,@function\n"
138 "raw_makecontext:\n"/* Calling convention sets the arguments in rdi, rsi, rdx and rcx, respectively */
140 " mov %rdi,%rax\n" /* stack */
141 " add %rsi,%rax\n" /* size */
142 " andq $-16, %rax\n" /* align stack */
143 " movq $0, -8(%rax)\n" /* @return for func */
144 " mov %rdx,-16(%rax)\n" /* func */
145 " mov %rcx,-24(%rax)\n" /* arg/rdi */
146 " movq $0, -32(%rax)\n" /* rsi */
147 " movq $0, -40(%rax)\n" /* rdx */
148 " movq $0, -48(%rax)\n" /* rcx */
149 " movq $0, -56(%rax)\n" /* r8 */
150 " movq $0, -64(%rax)\n" /* r9 */
151 " movq $0, -72(%rax)\n" /* rbp */
152 " movq $0, -80(%rax)\n" /* rbx */
153 " movq $0, -88(%rax)\n" /* r12 */
154 " movq $0, -96(%rax)\n" /* r13 */
155 " movq $0, -104(%rax)\n" /* r14 */
156 " movq $0, -112(%rax)\n" /* r15 */
162 #if defined(__APPLE__)
164 ".globl _raw_swapcontext\n"
165 "_raw_swapcontext:\n"
166 #elif defined(_WIN32)
168 ".globl raw_swapcontext\n"
172 ".globl raw_swapcontext\n"
173 ".type raw_swapcontext,@function\n"
174 "raw_swapcontext:\n" /* Calling convention sets the arguments in rdi and rsi, respectively */
188 " mov %rsp,(%rdi)\n" /* old */
189 " mov %rsi,%rsp\n" /* new */
206 #if defined(__APPLE__) || defined(_WIN32)
208 ".globl _raw_makecontext\n"
209 "_raw_makecontext:\n"
212 ".globl raw_makecontext\n"
213 ".type raw_makecontext,@function\n"
216 " movl 4(%esp),%eax\n" /* stack */
217 " addl 8(%esp),%eax\n" /* size */
218 " andl $-16, %eax\n" /* align stack */
219 " movl 12(%esp),%ecx\n" /* func */
220 " movl 16(%esp),%edx\n" /* arg */
221 " movl %edx, -4(%eax)\n"
222 " movl $0, -8(%eax)\n" /* @return for func */
223 " movl %ecx,-12(%eax)\n"
224 " movl $0, -16(%eax)\n" /* ebp */
225 " movl $0, -20(%eax)\n" /* ebx */
226 " movl $0, -24(%eax)\n" /* esi */
227 " movl $0, -28(%eax)\n" /* edi */
233 #if defined(__APPLE__) || defined(_WIN32)
235 ".globl _raw_swapcontext\n"
236 "_raw_swapcontext:\n"
239 ".globl raw_swapcontext\n"
240 ".type raw_swapcontext,@function\n"
243 " movl 4(%esp),%eax\n" /* old */
244 " movl 8(%esp),%edx\n" /* new */
249 " movl %esp,(%eax)\n"
260 /* If you implement raw contexts for other processors, don't forget to
261 update the definition of HAVE_RAW_CONTEXTS in tools/cmake/CompleteInFiles.cmake */
263 raw_stack_t raw_makecontext(void* malloced_stack, int stack_size,
264 rawctx_entry_point_t entry_point, void* arg) {
268 void raw_swapcontext(raw_stack_t* old, raw_stack_t new_context) {
274 // ***** Method definitions
279 RawContextFactory::RawContextFactory()
280 : ContextFactory("RawContextFactory")
282 #ifdef ADAPTIVE_THRESHOLD
283 raw_context_adaptative = (SIMIX_context_get_parallel_threshold() > 1);
285 raw_context_parallel = SIMIX_context_is_parallel();
286 if (raw_context_parallel) {
287 #ifdef HAVE_THREAD_CONTEXTS
288 int nthreads = SIMIX_context_get_nthreads();
289 xbt_os_thread_key_create(&raw_worker_id_key);
291 raw_parmap = nullptr;
292 raw_workers_context = xbt_new(RawContext*, nthreads);
293 raw_maestro_context = nullptr;
295 // TODO, if(SIMIX_context_get_parallel_threshold() > 1) => choose dynamically
297 #ifdef ADAPTIVE_THRESHOLD
298 round_time = xbt_os_timer_new();
299 reached_seq_limit = 0;
300 reached_par_limit = 0;
304 RawContextFactory::~RawContextFactory()
306 #ifdef HAVE_THREAD_CONTEXTS
308 xbt_parmap_destroy(raw_parmap);
309 xbt_free(raw_workers_context);
313 RawContext* RawContextFactory::create_context(std::function<void()> code,
314 void_pfn_smxprocess_t cleanup, smx_process_t process)
316 return this->new_context<RawContext>(std::move(code),
320 void RawContext::wrapper(void* arg)
322 RawContext* context = (RawContext*) arg;
327 RawContext::RawContext(std::function<void()> code,
328 void_pfn_smxprocess_t cleanup, smx_process_t process)
329 : Context(std::move(code), cleanup, process)
332 this->stack_ = SIMIX_context_stack_new();
333 this->stack_top_ = raw_makecontext(this->stack_,
334 smx_context_usable_stack_size,
338 if(process != nullptr && raw_maestro_context == nullptr)
339 raw_maestro_context = this;
342 &raw_maestro_context->stack_top_,
343 sizeof(raw_maestro_context->stack_top_));
347 RawContext::~RawContext()
349 SIMIX_context_stack_delete(this->stack_);
352 void RawContext::stop()
358 void RawContextFactory::run_all()
360 #ifdef ADAPTIVE_THRESHOLD
361 if (raw_context_adaptative)
362 run_all_adaptative();
365 if (raw_context_parallel)
371 void RawContextFactory::run_all_serial()
373 smx_process_t first_process =
374 xbt_dynar_get_as(simix_global->process_to_run, 0, smx_process_t);
375 raw_process_index = 1;
376 static_cast<RawContext*>(first_process->context)->resume_serial();
379 void RawContextFactory::run_all_parallel()
381 #ifdef HAVE_THREAD_CONTEXTS
382 raw_threads_working = 0;
383 if (raw_parmap == nullptr)
384 raw_parmap = xbt_parmap_new(
385 SIMIX_context_get_nthreads(), SIMIX_context_get_parallel_mode());
386 xbt_parmap_apply(raw_parmap,
388 smx_process_t process = static_cast<smx_process_t>(arg);
389 RawContext* context = static_cast<RawContext*>(process->context);
390 context->resume_parallel();
392 simix_global->process_to_run);
394 xbt_die("You asked for a parallel execution, but you don't have any threads.");
398 void RawContext::suspend()
400 if (raw_context_parallel)
401 RawContext::suspend_parallel();
403 RawContext::suspend_serial();
406 void RawContext::suspend_serial()
408 /* determine the next context */
409 RawContext* next_context = nullptr;
411 i = raw_process_index++;
412 if (i < xbt_dynar_length(simix_global->process_to_run)) {
413 /* execute the next process */
414 XBT_DEBUG("Run next process");
415 next_context = (RawContext*) xbt_dynar_get_as(
416 simix_global->process_to_run, i, smx_process_t)->context;
419 /* all processes were run, return to maestro */
420 XBT_DEBUG("No more process to run");
421 next_context = (RawContext*) raw_maestro_context;
423 SIMIX_context_set_current(next_context);
424 raw_swapcontext(&this->stack_top_, next_context->stack_top_);
427 void RawContext::suspend_parallel()
429 #ifdef HAVE_THREAD_CONTEXTS
430 /* determine the next context */
431 smx_process_t next_work = (smx_process_t) xbt_parmap_next(raw_parmap);
432 RawContext* next_context = nullptr;
434 if (next_work != NULL) {
435 /* there is a next process to resume */
436 XBT_DEBUG("Run next process");
437 next_context = (RawContext*) next_work->context;
440 /* all processes were run, go to the barrier */
441 XBT_DEBUG("No more processes to run");
442 uintptr_t worker_id = (uintptr_t)
443 xbt_os_thread_get_specific(raw_worker_id_key);
444 next_context = (RawContext*) raw_workers_context[worker_id];
445 XBT_DEBUG("Restoring worker stack %zu (working threads = %zu)",
446 worker_id, raw_threads_working);
449 SIMIX_context_set_current(next_context);
450 raw_swapcontext(&this->stack_top_, next_context->stack_top_);
454 void RawContext::resume()
456 if (raw_context_parallel)
462 void RawContext::resume_serial()
464 SIMIX_context_set_current(this);
465 raw_swapcontext(&raw_maestro_context->stack_top_, this->stack_top_);
468 void RawContext::resume_parallel()
470 #ifdef HAVE_THREAD_CONTEXTS
471 uintptr_t worker_id = __sync_fetch_and_add(&raw_threads_working, 1);
472 xbt_os_thread_set_specific(raw_worker_id_key, (void*) worker_id);
473 RawContext* worker_context = (RawContext*) SIMIX_context_self();
474 raw_workers_context[worker_id] = worker_context;
475 XBT_DEBUG("Saving worker stack %zu", worker_id);
476 SIMIX_context_set_current(this);
477 raw_swapcontext(&worker_context->stack_top_, this->stack_top_);
479 xbt_die("Parallel execution disabled");
484 * \brief Resumes all processes ready to run.
486 #ifdef ADAPTIVE_THRESHOLD
487 void RawContectFactory::run_all_adaptative()
489 unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
490 unsigned long threshold = SIMIX_context_get_parallel_threshold();
491 reached_seq_limit = (seq_sched_round % SCHED_ROUND_LIMIT == 0);
492 reached_par_limit = (par_sched_round % SCHED_ROUND_LIMIT == 0);
494 if(reached_seq_limit && reached_par_limit){
495 par_ratio = (par_proc_that_ran != 0) ? (par_time / (double)par_proc_that_ran) : 0;
496 seq_ratio = (seq_proc_that_ran != 0) ? (seq_time / (double)seq_proc_that_ran) : 0;
497 if(seq_ratio > par_ratio){
498 if(nb_processes < avg_par_proc) {
499 threshold = (threshold>2) ? threshold - 1 : threshold ;
500 SIMIX_context_set_parallel_threshold(threshold);
503 if(nb_processes > avg_seq_proc){
504 SIMIX_context_set_parallel_threshold(threshold+1);
509 if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
510 simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
511 if (nb_processes < par_window){
513 xbt_os_walltimer_start(round_time);
514 smx_ctx_raw_runall_parallel();
515 xbt_os_walltimer_stop(round_time);
516 par_time += xbt_os_timer_elapsed(round_time);
518 prev_avg_par_proc = avg_par_proc;
519 delta = nb_processes - avg_par_proc;
520 avg_par_proc = (par_sched_round==1) ? nb_processes : avg_par_proc + delta / (double) par_sched_round;
522 if(par_sched_round>=2){
523 s_par_proc = s_par_proc + (nb_processes - prev_avg_par_proc) * delta;
524 sd_par_proc = sqrt(s_par_proc / (par_sched_round-1));
525 par_window = (int) (avg_par_proc + sd_par_proc);
530 par_proc_that_ran += nb_processes;
532 smx_ctx_raw_runall_parallel();
535 simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
536 if(nb_processes > seq_window){
538 xbt_os_walltimer_start(round_time);
539 smx_ctx_raw_runall_serial();
540 xbt_os_walltimer_stop(round_time);
541 seq_time += xbt_os_timer_elapsed(round_time);
543 prev_avg_seq_proc = avg_seq_proc;
544 delta = (nb_processes-avg_seq_proc);
545 avg_seq_proc = (seq_sched_round==1) ? nb_processes : avg_seq_proc + delta / (double) seq_sched_round;
547 if(seq_sched_round>=2){
548 s_seq_proc = s_seq_proc + (nb_processes - prev_avg_seq_proc)*delta;
549 sd_seq_proc = sqrt(s_seq_proc / (seq_sched_round-1));
550 seq_window = (int) (avg_seq_proc - sd_seq_proc);
555 seq_proc_that_ran += nb_processes;
557 smx_ctx_raw_runall_serial();
565 void RawContextFactory::run_all_adaptative()
567 unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
568 if (SIMIX_context_is_parallel()
569 && (unsigned long) SIMIX_context_get_parallel_threshold() < nb_processes) {
570 raw_context_parallel = true;
571 XBT_DEBUG("Runall // %lu", nb_processes);
572 this->run_all_parallel();
574 XBT_DEBUG("Runall serial %lu", nb_processes);
575 raw_context_parallel = false;
576 this->run_all_serial();