1 /* context_raw - fast context switching inspired from System V ucontextes */
3 /* Copyright (c) 2009-2013. The SimGrid Team.
4 * All rights reserved. */
6 /* This program is free software; you can redistribute it and/or modify it
7 * under the terms of the license (GNU LGPL) which comes with this package. */
9 #include "smx_private.h"
10 #include "xbt/parmap.h"
13 #ifdef HAVE_VALGRIND_VALGRIND_H
14 # include <valgrind/valgrind.h>
15 #endif /* HAVE_VALGRIND_VALGRIND_H */
17 typedef char * raw_stack_t;
18 typedef void (*rawctx_entry_point_t)(void *);
20 typedef struct s_smx_ctx_raw {
21 s_smx_ctx_base_t super; /* Fields of super implementation */
22 char *malloced_stack; /* malloced area containing the stack */
23 raw_stack_t stack_top; /* pointer to stack top (within previous area) */
24 #ifdef HAVE_VALGRIND_VALGRIND_H
25 unsigned int valgrind_stack_id; /* the valgrind stack id */
27 #ifdef TIME_BENCH_PER_SR
28 unsigned int thread; /* Just for measuring purposes */
30 } s_smx_ctx_raw_t, *smx_ctx_raw_t;
32 #ifdef CONTEXT_THREADS
33 static xbt_parmap_t raw_parmap;
34 static smx_ctx_raw_t* raw_workers_context; /* space to save the worker context in each thread */
35 static unsigned long raw_threads_working; /* number of threads that have started their work */
36 static xbt_os_thread_key_t raw_worker_id_key; /* thread-specific storage for the thread id */
39 static unsigned long raw_process_index = 0; /* index of the next process to run in the
40 * list of runnable processes */
41 static smx_ctx_raw_t raw_maestro_context;
43 extern raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
44 rawctx_entry_point_t entry_point, void* arg);
45 extern void raw_swapcontext(raw_stack_t* old, raw_stack_t new);
47 #ifdef PROCESSOR_x86_64
51 ".globl _raw_makecontext\n"
55 ".globl raw_makecontext\n"
59 ".globl raw_makecontext\n"
60 ".type raw_makecontext,@function\n"
61 "raw_makecontext:\n"/* Calling convention sets the arguments in rdi, rsi, rdx and rcx, respectively */
63 " mov %rdi,%rax\n" /* stack */
64 " add %rsi,%rax\n" /* size */
65 " movq $0, -8(%rax)\n" /* @return for func */
66 " mov %rdx,-16(%rax)\n" /* func */
67 " mov %rcx,-24(%rax)\n" /* arg/rdi */
68 " movq $0, -32(%rax)\n" /* rsi */
69 " movq $0, -40(%rax)\n" /* rdx */
70 " movq $0, -48(%rax)\n" /* rcx */
71 " movq $0, -56(%rax)\n" /* r8 */
72 " movq $0, -64(%rax)\n" /* r9 */
73 " movq $0, -72(%rax)\n" /* rbp */
74 " movq $0, -80(%rax)\n" /* rbx */
75 " movq $0, -88(%rax)\n" /* r12 */
76 " movq $0, -96(%rax)\n" /* r13 */
77 " movq $0, -104(%rax)\n" /* r14 */
78 " movq $0, -112(%rax)\n" /* r15 */
86 ".globl _raw_swapcontext\n"
90 ".globl raw_swapcontext\n"
94 ".globl raw_swapcontext\n"
95 ".type raw_swapcontext,@function\n"
96 "raw_swapcontext:\n" /* Calling convention sets the arguments in rdi and rsi, respectively */
110 " mov %rsp,(%rdi)\n" /* old */
111 " mov %rsi,%rsp\n" /* new */
128 #if defined(APPLE) || defined(_WIN32)
130 ".globl _raw_makecontext\n"
131 "_raw_makecontext:\n"
134 ".globl raw_makecontext\n"
135 ".type raw_makecontext,@function\n"
138 " movl 4(%esp),%eax\n" /* stack */
139 " addl 8(%esp),%eax\n" /* size */
140 " movl 12(%esp),%ecx\n" /* func */
141 " movl 16(%esp),%edx\n" /* arg */
142 " movl %edx, -4(%eax)\n"
143 " movl $0, -8(%eax)\n" /* @return for func */
144 " movl %ecx,-12(%eax)\n"
145 " movl $0, -16(%eax)\n" /* ebp */
146 " movl $0, -20(%eax)\n" /* ebx */
147 " movl $0, -24(%eax)\n" /* esi */
148 " movl $0, -28(%eax)\n" /* edi */
154 #if defined(APPLE) || defined(_WIN32)
156 ".globl _raw_swapcontext\n"
157 "_raw_swapcontext:\n"
160 ".globl raw_swapcontext\n"
161 ".type raw_swapcontext,@function\n"
164 " movl 4(%esp),%eax\n" /* old */
165 " movl 8(%esp),%edx\n" /* new */
170 " movl %esp,(%eax)\n"
181 /* If you implement raw contexts for other processors, don't forget to
182 update the definition of HAVE_RAWCTX in buildtools/Cmake/CompleteInFiles.cmake */
184 raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
185 rawctx_entry_point_t entry_point, void* arg) {
189 void raw_swapcontext(raw_stack_t* old, raw_stack_t new) {
195 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
197 #ifdef TIME_BENCH_PER_SR
198 #include "xbt/xbt_os_time.h"
199 #define NUM_THREADS 4
200 static xbt_os_timer_t timer;
201 static double time_thread_sr[NUM_THREADS];
202 static double time_thread_ssr[NUM_THREADS];
203 static double time_wasted_sr = 0;
204 static double time_wasted_ssr = 0;
205 static unsigned int sr_count = 0;
206 static unsigned int ssr_count = 0;
207 static char new_sr = 0;
210 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
211 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory);
212 static smx_context_t smx_ctx_raw_create_context(xbt_main_func_t code, int argc,
213 char **argv, void_pfn_smxprocess_t cleanup_func, void *data);
214 static void smx_ctx_raw_free(smx_context_t context);
215 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
216 static void smx_ctx_raw_stop(smx_context_t context);
217 static void smx_ctx_raw_suspend_serial(smx_context_t context);
218 static void smx_ctx_raw_resume_serial(smx_process_t first_process);
219 static void smx_ctx_raw_runall_serial(void);
220 static void smx_ctx_raw_suspend_parallel(smx_context_t context);
221 static void smx_ctx_raw_resume_parallel(smx_process_t first_process);
222 static void smx_ctx_raw_runall_parallel(void);
223 static void smx_ctx_raw_runall(void);
226 * \brief Initializes the raw context factory.
227 * \param factory where to initialize the factory
229 void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory)
232 XBT_VERB("Using raw contexts. Because the glibc is just not good enough for us.");
233 smx_ctx_base_factory_init(factory);
235 (*factory)->finalize = smx_ctx_raw_factory_finalize;
236 (*factory)->create_context = smx_ctx_raw_create_context;
237 /* Do not overload that method (*factory)->finalize */
238 (*factory)->free = smx_ctx_raw_free;
239 (*factory)->stop = smx_ctx_raw_stop;
240 (*factory)->name = "smx_raw_context_factory";
242 if (SIMIX_context_is_parallel()) {
243 #ifdef CONTEXT_THREADS
244 int nthreads = SIMIX_context_get_nthreads();
245 xbt_os_thread_key_create(&raw_worker_id_key);
246 raw_parmap = xbt_parmap_new(nthreads, SIMIX_context_get_parallel_mode());
247 raw_workers_context = xbt_new(smx_ctx_raw_t, nthreads);
248 raw_maestro_context=NULL;
251 if (SIMIX_context_get_parallel_threshold() > 1) {
252 /* choose dynamically */
253 (*factory)->runall = smx_ctx_raw_runall;
254 (*factory)->suspend = NULL;
257 /* always parallel */
258 (*factory)->runall = smx_ctx_raw_runall_parallel;
259 (*factory)->suspend = smx_ctx_raw_suspend_parallel;
264 (*factory)->runall = smx_ctx_raw_runall_serial;
265 (*factory)->suspend = smx_ctx_raw_suspend_serial;
267 #ifdef TIME_BENCH_PER_SR
268 timer = xbt_os_timer_new();
273 * \brief Finalizes the raw context factory.
274 * \param factory the raw context factory
276 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory)
278 #ifdef TIME_BENCH_PER_SR
279 XBT_CRITICAL("Total wasted time in %u SR: %lf", sr_count, time_wasted_sr);
280 XBT_CRITICAL("Total wasted time in %u SSR: %lf", ssr_count, time_wasted_ssr);
283 #ifdef CONTEXT_THREADS
285 xbt_parmap_destroy(raw_parmap);
286 xbt_free(raw_workers_context);
288 return smx_ctx_base_factory_finalize(factory);
292 * \brief Creates a new raw context.
293 * \param code main function of this context or NULL to create the maestro
295 * \param argc argument number
296 * \param argv arguments to pass to the main function
297 * \param cleanup_func a function to call to free the user data when the
299 * \param data user data
302 smx_ctx_raw_create_context(xbt_main_func_t code, int argc, char **argv,
303 void_pfn_smxprocess_t cleanup_func,
307 smx_ctx_raw_t context =
308 (smx_ctx_raw_t) smx_ctx_base_factory_create_context_sized(
309 sizeof(s_smx_ctx_raw_t),
316 /* if the user provided a function for the process then use it,
317 otherwise it is the context for maestro */
319 context->malloced_stack = xbt_malloc0(smx_context_stack_size);
321 raw_makecontext(context->malloced_stack, smx_context_stack_size,
322 (void_f_pvoid_t) smx_ctx_raw_wrapper, context);
324 #ifdef HAVE_VALGRIND_VALGRIND_H
325 context->valgrind_stack_id =
326 VALGRIND_STACK_REGISTER(context->malloced_stack,
327 context->malloced_stack + smx_context_stack_size);
328 #endif /* HAVE_VALGRIND_VALGRIND_H */
331 if(data != NULL && raw_maestro_context==NULL)
332 raw_maestro_context = context;
335 MC_ignore_heap(&(raw_maestro_context->stack_top), sizeof(raw_maestro_context->stack_top));
339 return (smx_context_t) context;
343 * \brief Destroys a raw context.
344 * \param context a raw context
346 static void smx_ctx_raw_free(smx_context_t context)
350 #ifdef HAVE_VALGRIND_VALGRIND_H
351 VALGRIND_STACK_DEREGISTER(((smx_ctx_raw_t)
352 context)->valgrind_stack_id);
353 #endif /* HAVE_VALGRIND_VALGRIND_H */
355 free(((smx_ctx_raw_t) context)->malloced_stack);
357 smx_ctx_base_free(context);
361 * \brief Wrapper for the main function of a context.
362 * \param context a raw context
364 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context)
366 (context->super.code) (context->super.argc, context->super.argv);
368 smx_ctx_raw_stop((smx_context_t) context);
372 * \brief Stops a raw context.
374 * This function is called when the main function of the context if finished.
376 * \param context the current context
378 static void smx_ctx_raw_stop(smx_context_t context)
380 smx_ctx_base_stop(context);
381 simix_global->context_factory->suspend(context);
385 * \brief Suspends a running context and resumes another one or returns to
387 * \param context the current context
389 static void smx_ctx_raw_suspend_serial(smx_context_t context)
391 /* determine the next context */
392 smx_context_t next_context;
393 unsigned long int i = raw_process_index++;
395 if (i < xbt_dynar_length(simix_global->process_to_run)) {
396 /* execute the next process */
397 XBT_DEBUG("Run next process");
398 next_context = xbt_dynar_get_as(
399 simix_global->process_to_run, i, smx_process_t)->context;
402 /* all processes were run, return to maestro */
403 XBT_DEBUG("No more process to run");
404 next_context = (smx_context_t) raw_maestro_context;
406 SIMIX_context_set_current(next_context);
407 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top,
408 ((smx_ctx_raw_t) next_context)->stack_top);
412 * \brief Resumes sequentially all processes ready to run.
413 * \param first_process the first process to resume
415 static void smx_ctx_raw_resume_serial(smx_process_t first_process)
417 smx_ctx_raw_t context = (smx_ctx_raw_t) first_process->context;
418 SIMIX_context_set_current((smx_context_t) context);
419 raw_swapcontext(&raw_maestro_context->stack_top,
420 ((smx_ctx_raw_t) context)->stack_top);
423 #ifdef TIME_BENCH_PER_SR
424 static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
426 smx_process_t process;
431 unsigned long num_proc = xbt_dynar_length(processes);
433 unsigned int data_size = (num_proc / NUM_THREADS) + ((num_proc % NUM_THREADS) ? 1 : 0);
436 time_thread_ssr[0] = 0;
437 xbt_dynar_foreach(processes, cursor, process) {
438 XBT_DEBUG("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
439 if(cursor >= t * data_size + data_size){
440 if(time_thread_ssr[t] > tmax)
441 tmax = time_thread_ssr[t];
443 time_thread_ssr[t] = 0;
447 ((smx_ctx_raw_t)process->context)->thread = t;
448 time_thread_sr[t] = 0;
451 xbt_os_cputimer_start(timer);
452 smx_ctx_raw_resume(process);
453 xbt_os_cputimer_stop(timer);
454 elapsed = xbt_os_timer_elapsed(timer);
455 time_thread_ssr[t] += elapsed;
456 time_thread_sr[((smx_ctx_raw_t)process->context)->thread] += elapsed;
462 if(time_thread_ssr[t] > tmax)
463 tmax = time_thread_ssr[t];
465 for(cursor=0; cursor <= t; cursor++){
466 XBT_VERB("Time SSR thread %u = %lf (max %lf)", cursor, time_thread_ssr[cursor], tmax);
467 time_wasted_ssr += tmax - time_thread_ssr[cursor];
471 void smx_ctx_raw_new_sr(void);
472 void smx_ctx_raw_new_sr(void)
478 for(i=0; i < NUM_THREADS; i++){
479 if(time_thread_sr[i] > tmax)
480 tmax = time_thread_sr[i];
483 for(i=0; i < NUM_THREADS; i++){
484 XBT_VERB("Time SR thread %u = %lf (max %lf)", i, time_thread_sr[i], tmax);
485 time_wasted_sr += tmax - time_thread_sr[i];
488 XBT_VERB("New scheduling round");
493 * \brief Resumes sequentially all processes ready to run.
495 static void smx_ctx_raw_runall_serial(void)
497 smx_process_t first_process =
498 xbt_dynar_get_as(simix_global->process_to_run, 0, smx_process_t);
499 raw_process_index = 1;
501 /* execute the first process */
502 smx_ctx_raw_resume_serial(first_process);
507 * \brief Suspends a running context and resumes another one or returns to
508 * the main function of the current worker thread.
509 * \param context the context of the current worker thread
511 static void smx_ctx_raw_suspend_parallel(smx_context_t context)
513 #ifdef CONTEXT_THREADS
514 /* determine the next context */
515 smx_process_t next_work = xbt_parmap_next(raw_parmap);
516 smx_context_t next_context;
517 raw_stack_t next_stack;
519 if (next_work != NULL) {
520 /* there is a next process to resume */
521 XBT_DEBUG("Run next process");
522 next_context = next_work->context;
523 next_stack = ((smx_ctx_raw_t) next_context)->stack_top;
526 /* all processes were run, go to the barrier */
527 XBT_DEBUG("No more processes to run");
529 unsigned long worker_id =
530 (unsigned long) xbt_os_thread_get_specific(raw_worker_id_key);
532 next_context = (smx_context_t)raw_workers_context[worker_id];
533 XBT_DEBUG("Restoring worker stack %lu (working threads = %lu)",
534 worker_id, raw_threads_working);
535 next_stack = ((smx_ctx_raw_t)next_context)->stack_top;
538 SIMIX_context_set_current(next_context);
539 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top, next_stack);
544 * \brief Resumes sequentially in the current worker thread the processes ready
546 * \param first_process the first process to resume
548 static void smx_ctx_raw_resume_parallel(smx_process_t first_process)
550 #ifdef CONTEXT_THREADS
551 unsigned long worker_id = __sync_fetch_and_add(&raw_threads_working, 1);
552 xbt_os_thread_set_specific(raw_worker_id_key, (void*) worker_id);
553 smx_ctx_raw_t worker_context = (smx_ctx_raw_t)SIMIX_context_self();
554 raw_workers_context[worker_id] = worker_context;
555 XBT_DEBUG("Saving worker stack %lu", worker_id);
556 raw_stack_t* worker_stack = &(worker_context)->stack_top;
559 smx_context_t context = first_process->context;
560 SIMIX_context_set_current(context);
561 raw_swapcontext(worker_stack, ((smx_ctx_raw_t) context)->stack_top);
566 * \brief Resumes in parallel all processes ready to run.
568 static void smx_ctx_raw_runall_parallel(void)
570 #ifdef CONTEXT_THREADS
571 raw_threads_working = 0;
572 xbt_parmap_apply(raw_parmap, (void_f_pvoid_t) smx_ctx_raw_resume_parallel,
573 simix_global->process_to_run);
575 xbt_die("You asked for a parallel execution, but you don't have any threads.")
580 * \brief Resumes all processes ready to run.
582 static void smx_ctx_raw_runall(void)
584 unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
585 if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
586 XBT_DEBUG("Runall // %lu", nb_processes);
587 simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
588 smx_ctx_raw_runall_parallel();
590 XBT_DEBUG("Runall serial %lu", nb_processes);
591 simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
592 smx_ctx_raw_runall_serial();