1 /* context_raw - fast context switching inspired from System V ucontexts */
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 unsigned int parmap_index;
25 #ifdef HAVE_VALGRIND_VALGRIND_H
26 unsigned int valgrind_stack_id; /* the valgrind stack id */
28 #ifdef TIME_BENCH_PER_SR
29 unsigned int thread; /* Just for measuring purposes */
31 } s_smx_ctx_raw_t, *smx_ctx_raw_t;
33 #ifdef CONTEXT_THREADS
34 static xbt_parmap_t raw_parmap;
35 static smx_ctx_raw_t* raw_workers_context; /* space to save the worker context in each thread */
36 static unsigned long raw_threads_working; /* number of threads that have started their work */
37 static xbt_os_thread_key_t raw_worker_id_key; /* thread-specific storage for the thread id */
40 static unsigned long raw_process_index = 0; /* index of the next process to run in the
41 * list of runnable processes */
42 static smx_ctx_raw_t raw_maestro_context;
44 extern raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
45 rawctx_entry_point_t entry_point, void* arg);
46 extern void raw_swapcontext(raw_stack_t* old, raw_stack_t new);
52 ".globl _raw_makecontext\n"
56 ".globl raw_makecontext\n"
60 ".globl raw_makecontext\n"
61 ".type raw_makecontext,@function\n"
62 "raw_makecontext:\n"/* Calling convention sets the arguments in rdi, rsi, rdx and rcx, respectively */
64 " mov %rdi,%rax\n" /* stack */
65 " add %rsi,%rax\n" /* size */
66 " movq $0, -8(%rax)\n" /* @return for func */
67 " mov %rdx,-16(%rax)\n" /* func */
68 " mov %rcx,-24(%rax)\n" /* arg/rdi */
69 " movq $0, -32(%rax)\n" /* rsi */
70 " movq $0, -40(%rax)\n" /* rdx */
71 " movq $0, -48(%rax)\n" /* rcx */
72 " movq $0, -56(%rax)\n" /* r8 */
73 " movq $0, -64(%rax)\n" /* r9 */
74 " movq $0, -72(%rax)\n" /* rbp */
75 " movq $0, -80(%rax)\n" /* rbx */
76 " movq $0, -88(%rax)\n" /* r12 */
77 " movq $0, -96(%rax)\n" /* r13 */
78 " movq $0, -104(%rax)\n" /* r14 */
79 " movq $0, -112(%rax)\n" /* r15 */
87 ".globl _raw_swapcontext\n"
91 ".globl raw_swapcontext\n"
95 ".globl raw_swapcontext\n"
96 ".type raw_swapcontext,@function\n"
97 "raw_swapcontext:\n" /* Calling convention sets the arguments in rdi and rsi, respectively */
111 " mov %rsp,(%rdi)\n" /* old */
112 " mov %rsi,%rsp\n" /* new */
129 #if defined(APPLE) || defined(_WIN32)
131 ".globl _raw_makecontext\n"
132 "_raw_makecontext:\n"
135 ".globl raw_makecontext\n"
136 ".type raw_makecontext,@function\n"
139 " movl 4(%esp),%eax\n" /* stack */
140 " addl 8(%esp),%eax\n" /* size */
141 " movl 12(%esp),%ecx\n" /* func */
142 " movl 16(%esp),%edx\n" /* arg */
143 " movl %edx, -4(%eax)\n"
144 " movl $0, -8(%eax)\n" /* @return for func */
145 " movl %ecx,-12(%eax)\n"
146 " movl $0, -16(%eax)\n" /* ebp */
147 " movl $0, -20(%eax)\n" /* ebx */
148 " movl $0, -24(%eax)\n" /* esi */
149 " movl $0, -28(%eax)\n" /* edi */
155 #if defined(APPLE) || defined(_WIN32)
157 ".globl _raw_swapcontext\n"
158 "_raw_swapcontext:\n"
161 ".globl raw_swapcontext\n"
162 ".type raw_swapcontext,@function\n"
165 " movl 4(%esp),%eax\n" /* old */
166 " movl 8(%esp),%edx\n" /* new */
171 " movl %esp,(%eax)\n"
182 /* If you implement raw contexts for other processors, don't forget to
183 update the definition of HAVE_RAWCTX in buildtools/Cmake/CompleteInFiles.cmake */
185 raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
186 rawctx_entry_point_t entry_point, void* arg) {
190 void raw_swapcontext(raw_stack_t* old, raw_stack_t new) {
196 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
198 #ifdef TIME_BENCH_PER_SR
199 #include "xbt/xbt_os_time.h"
200 #define NUM_THREADS 4
201 static xbt_os_timer_t timer;
202 static double time_thread_sr[NUM_THREADS];
203 static double time_thread_ssr[NUM_THREADS];
204 static double time_wasted_sr = 0;
205 static double time_wasted_ssr = 0;
206 static unsigned int sr_count = 0;
207 static unsigned int ssr_count = 0;
208 static char new_sr = 0;
211 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
212 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory);
213 static smx_context_t smx_ctx_raw_create_context(xbt_main_func_t code, int argc,
214 char **argv, void_pfn_smxprocess_t cleanup_func, void *data);
215 static void smx_ctx_raw_free(smx_context_t context);
216 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
217 static void smx_ctx_raw_stop(smx_context_t context);
218 static void smx_ctx_raw_suspend_serial(smx_context_t context);
219 static void smx_ctx_raw_resume_serial(smx_process_t first_process);
220 static void smx_ctx_raw_runall_serial(void);
221 static void smx_ctx_raw_suspend_parallel(smx_context_t context);
222 static void smx_ctx_raw_resume_parallel(smx_process_t first_process);
223 static void smx_ctx_raw_runall_parallel(void);
224 static void smx_ctx_raw_runall(void);
227 * \brief Initializes the raw context factory.
228 * \param factory where to initialize the factory
230 void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory)
233 XBT_VERB("Using raw contexts. Because the glibc is just not good enough for us.");
234 smx_ctx_base_factory_init(factory);
236 (*factory)->finalize = smx_ctx_raw_factory_finalize;
237 (*factory)->create_context = smx_ctx_raw_create_context;
238 /* Do not overload that method (*factory)->finalize */
239 (*factory)->free = smx_ctx_raw_free;
240 (*factory)->stop = smx_ctx_raw_stop;
241 (*factory)->name = "smx_raw_context_factory";
243 if (SIMIX_context_is_parallel()) {
244 #ifdef CONTEXT_THREADS
245 int nthreads = SIMIX_context_get_nthreads();
246 xbt_os_thread_key_create(&raw_worker_id_key);
247 raw_parmap = xbt_parmap_new(nthreads, SIMIX_context_get_parallel_mode());
248 raw_workers_context = xbt_new(smx_ctx_raw_t, nthreads);
249 raw_maestro_context=NULL;
252 if (SIMIX_context_get_parallel_threshold() > 1) {
253 /* choose dynamically */
254 (*factory)->runall = smx_ctx_raw_runall;
255 (*factory)->suspend = NULL;
258 /* always parallel */
259 (*factory)->runall = smx_ctx_raw_runall_parallel;
260 (*factory)->suspend = smx_ctx_raw_suspend_parallel;
265 (*factory)->runall = smx_ctx_raw_runall_serial;
266 (*factory)->suspend = smx_ctx_raw_suspend_serial;
268 #ifdef TIME_BENCH_PER_SR
269 timer = xbt_os_timer_new();
274 * \brief Finalizes the raw context factory.
275 * \param factory the raw context factory
277 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory)
279 #ifdef TIME_BENCH_PER_SR
280 XBT_CRITICAL("Total wasted time in %u SR: %f", sr_count, time_wasted_sr);
281 XBT_CRITICAL("Total wasted time in %u SSR: %f", ssr_count, time_wasted_ssr);
284 #ifdef CONTEXT_THREADS
286 xbt_parmap_destroy(raw_parmap);
287 xbt_free(raw_workers_context);
289 return smx_ctx_base_factory_finalize(factory);
293 * \brief Creates a new raw context.
294 * \param code main function of this context or NULL to create the maestro
296 * \param argc argument number
297 * \param argv arguments to pass to the main function
298 * \param cleanup_func a function to call to free the user data when the
300 * \param data user data
303 smx_ctx_raw_create_context(xbt_main_func_t code, int argc, char **argv,
304 void_pfn_smxprocess_t cleanup_func,
308 smx_ctx_raw_t context =
309 (smx_ctx_raw_t) smx_ctx_base_factory_create_context_sized(
310 sizeof(s_smx_ctx_raw_t),
317 /* if the user provided a function for the process then use it,
318 otherwise it is the context for maestro */
320 context->malloced_stack = xbt_malloc0(smx_context_stack_size);
322 raw_makecontext(context->malloced_stack, smx_context_stack_size,
323 (void_f_pvoid_t) smx_ctx_raw_wrapper, context);
325 #ifdef HAVE_VALGRIND_VALGRIND_H
326 context->valgrind_stack_id =
327 VALGRIND_STACK_REGISTER(context->malloced_stack,
328 context->malloced_stack + smx_context_stack_size);
329 #endif /* HAVE_VALGRIND_VALGRIND_H */
332 if(data != NULL && raw_maestro_context==NULL)
333 raw_maestro_context = context;
336 MC_ignore_heap(&(raw_maestro_context->stack_top), sizeof(raw_maestro_context->stack_top));
340 return (smx_context_t) context;
344 * \brief Destroys a raw context.
345 * \param context a raw context
347 static void smx_ctx_raw_free(smx_context_t context)
351 #ifdef HAVE_VALGRIND_VALGRIND_H
352 VALGRIND_STACK_DEREGISTER(((smx_ctx_raw_t)
353 context)->valgrind_stack_id);
354 #endif /* HAVE_VALGRIND_VALGRIND_H */
356 free(((smx_ctx_raw_t) context)->malloced_stack);
358 smx_ctx_base_free(context);
362 * \brief Wrapper for the main function of a context.
363 * \param context a raw context
365 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context)
367 (context->super.code) (context->super.argc, context->super.argv);
369 smx_ctx_raw_stop((smx_context_t) context);
373 * \brief Stops a raw context.
375 * This function is called when the main function of the context if finished.
377 * \param context the current context
379 static void smx_ctx_raw_stop(smx_context_t context)
381 smx_ctx_base_stop(context);
382 simix_global->context_factory->suspend(context);
386 * \brief Suspends a running context and resumes another one or returns to
388 * \param context the current context
390 static void smx_ctx_raw_suspend_serial(smx_context_t context)
392 /* determine the next context */
393 smx_context_t next_context;
394 unsigned long int i = raw_process_index++;
396 if (i < xbt_dynar_length(simix_global->process_to_run)) {
397 /* execute the next process */
398 XBT_DEBUG("Run next process");
399 next_context = xbt_dynar_get_as(
400 simix_global->process_to_run, i, smx_process_t)->context;
403 /* all processes were run, return to maestro */
404 XBT_DEBUG("No more process to run");
405 next_context = (smx_context_t) raw_maestro_context;
407 SIMIX_context_set_current(next_context);
408 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top,
409 ((smx_ctx_raw_t) next_context)->stack_top);
413 * \brief Resumes sequentially all processes ready to run.
414 * \param first_process the first process to resume
416 static void smx_ctx_raw_resume_serial(smx_process_t first_process)
418 smx_ctx_raw_t context = (smx_ctx_raw_t) first_process->context;
419 SIMIX_context_set_current((smx_context_t) context);
420 raw_swapcontext(&raw_maestro_context->stack_top,
421 ((smx_ctx_raw_t) context)->stack_top);
424 #ifdef TIME_BENCH_PER_SR
425 static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
427 smx_process_t process;
432 unsigned long num_proc = xbt_dynar_length(processes);
434 unsigned int data_size = (num_proc / NUM_THREADS) + ((num_proc % NUM_THREADS) ? 1 : 0);
437 time_thread_ssr[0] = 0;
438 xbt_dynar_foreach(processes, cursor, process) {
439 XBT_DEBUG("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
440 if(cursor >= t * data_size + data_size){
441 if(time_thread_ssr[t] > tmax)
442 tmax = time_thread_ssr[t];
444 time_thread_ssr[t] = 0;
448 ((smx_ctx_raw_t)process->context)->thread = t;
449 time_thread_sr[t] = 0;
452 xbt_os_cputimer_start(timer);
453 smx_ctx_raw_resume(process);
454 xbt_os_cputimer_stop(timer);
455 elapsed = xbt_os_timer_elapsed(timer);
456 time_thread_ssr[t] += elapsed;
457 time_thread_sr[((smx_ctx_raw_t)process->context)->thread] += elapsed;
463 if(time_thread_ssr[t] > tmax)
464 tmax = time_thread_ssr[t];
466 for(cursor=0; cursor <= t; cursor++){
467 XBT_VERB("Time SSR thread %u = %f (max %f)", cursor, time_thread_ssr[cursor], tmax);
468 time_wasted_ssr += tmax - time_thread_ssr[cursor];
472 void smx_ctx_raw_new_sr(void);
473 void smx_ctx_raw_new_sr(void)
479 for(i=0; i < NUM_THREADS; i++){
480 if(time_thread_sr[i] > tmax)
481 tmax = time_thread_sr[i];
484 for(i=0; i < NUM_THREADS; i++){
485 XBT_VERB("Time SR thread %u = %f (max %f)", i, time_thread_sr[i], tmax);
486 time_wasted_sr += tmax - time_thread_sr[i];
489 XBT_VERB("New scheduling round");
494 * \brief Resumes sequentially all processes ready to run.
496 static void smx_ctx_raw_runall_serial(void)
498 smx_process_t first_process =
499 xbt_dynar_get_as(simix_global->process_to_run, 0, smx_process_t);
500 raw_process_index = 1;
502 /* execute the first process */
503 smx_ctx_raw_resume_serial(first_process);
508 * \brief Suspends a running context and resumes another one or returns to
509 * the main function of the current worker thread.
510 * \param context the context of the current worker thread
512 static void smx_ctx_raw_suspend_parallel(smx_context_t context)
514 #ifdef CONTEXT_THREADS
515 /* determine the next context */
516 smx_process_t next_work;
517 smx_context_t next_context;
518 raw_stack_t next_stack;
520 next_work = xbt_parmap_next(raw_parmap,
521 &((smx_ctx_raw_t) context)->parmap_index);
522 if (next_work != NULL) {
523 /* there is a next process to resume */
524 XBT_DEBUG("Run next process");
525 next_context = next_work->context;
526 next_stack = ((smx_ctx_raw_t) next_context)->stack_top;
527 ((smx_ctx_raw_t) next_context)->parmap_index =
528 ((smx_ctx_raw_t) context)->parmap_index;
531 /* all processes were run, go to the barrier */
532 XBT_DEBUG("No more processes to run");
534 unsigned long worker_id =
535 (unsigned long) xbt_os_thread_get_specific(raw_worker_id_key);
537 next_context = (smx_context_t)raw_workers_context[worker_id];
538 XBT_DEBUG("Restoring worker stack %lu (working threads = %lu)",
539 worker_id, raw_threads_working);
540 next_stack = ((smx_ctx_raw_t)next_context)->stack_top;
543 SIMIX_context_set_current(next_context);
544 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top, next_stack);
549 * \brief Resumes sequentially in the current worker thread the processes ready
551 * \param first_process the first process to resume
553 static void smx_ctx_raw_resume_parallel(smx_process_t first_process)
555 #ifdef CONTEXT_THREADS
556 unsigned long worker_id = __sync_fetch_and_add(&raw_threads_working, 1);
557 xbt_os_thread_set_specific(raw_worker_id_key, (void*) worker_id);
558 smx_ctx_raw_t worker_context = (smx_ctx_raw_t)SIMIX_context_self();
559 raw_workers_context[worker_id] = worker_context;
560 XBT_DEBUG("Saving worker stack %lu", worker_id);
561 raw_stack_t* worker_stack = &(worker_context)->stack_top;
562 unsigned int index = 0;
564 first_process = xbt_parmap_next(raw_parmap, &index);
565 if (first_process != NULL) {
566 smx_context_t context = first_process->context;
567 ((smx_ctx_raw_t) context)->parmap_index = index;
568 SIMIX_context_set_current(context);
569 raw_swapcontext(worker_stack, ((smx_ctx_raw_t) context)->stack_top);
575 * \brief Resumes in parallel all processes ready to run.
577 static void smx_ctx_raw_runall_parallel(void)
579 #ifdef CONTEXT_THREADS
580 raw_threads_working = 0;
581 xbt_parmap_apply(raw_parmap, (void_f_pvoid_t) smx_ctx_raw_resume_parallel,
582 simix_global->process_to_run,
583 SIMIX_context_get_parallel_threshold(), 1);
584 // 1 + SIMIX_context_get_parallel_threshold() / SIMIX_context_get_nthreads(), 1);
586 xbt_die("You asked for a parallel execution, but you don't have any threads.");
591 * \brief Resumes all processes ready to run.
593 static void smx_ctx_raw_runall(void)
595 unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
596 if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
597 XBT_DEBUG("Runall // %lu", nb_processes);
598 simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
599 smx_ctx_raw_runall_parallel();
601 XBT_DEBUG("Runall serial %lu", nb_processes);
602 simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
603 smx_ctx_raw_runall_serial();