1 /* context_raw - fast context switching inspired from System V ucontexts */
3 /* Copyright (c) 2009-2014. 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);
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 " andq $-16, %rax\n" /* align stack */
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 " andl $-16, %eax\n" /* align stack */
142 " movl 12(%esp),%ecx\n" /* func */
143 " movl 16(%esp),%edx\n" /* arg */
144 " movl %edx, -4(%eax)\n"
145 " movl $0, -8(%eax)\n" /* @return for func */
146 " movl %ecx,-12(%eax)\n"
147 " movl $0, -16(%eax)\n" /* ebp */
148 " movl $0, -20(%eax)\n" /* ebx */
149 " movl $0, -24(%eax)\n" /* esi */
150 " movl $0, -28(%eax)\n" /* edi */
156 #if defined(APPLE) || defined(_WIN32)
158 ".globl _raw_swapcontext\n"
159 "_raw_swapcontext:\n"
162 ".globl raw_swapcontext\n"
163 ".type raw_swapcontext,@function\n"
166 " movl 4(%esp),%eax\n" /* old */
167 " movl 8(%esp),%edx\n" /* new */
172 " movl %esp,(%eax)\n"
183 /* If you implement raw contexts for other processors, don't forget to
184 update the definition of HAVE_RAWCTX in buildtools/Cmake/CompleteInFiles.cmake */
186 raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
187 rawctx_entry_point_t entry_point, void* arg) {
191 void raw_swapcontext(raw_stack_t* old, raw_stack_t new) {
197 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
199 #ifdef TIME_BENCH_PER_SR
200 #include "xbt/xbt_os_time.h"
201 #define NUM_THREADS 4
202 static xbt_os_timer_t timer;
203 static double time_thread_sr[NUM_THREADS];
204 static double time_thread_ssr[NUM_THREADS];
205 static double time_wasted_sr = 0;
206 static double time_wasted_ssr = 0;
207 static unsigned int sr_count = 0;
208 static unsigned int ssr_count = 0;
209 static char new_sr = 0;
212 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
213 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory);
214 static smx_context_t smx_ctx_raw_create_context(xbt_main_func_t code, int argc,
215 char **argv, void_pfn_smxprocess_t cleanup_func, smx_process_t process);
216 static void smx_ctx_raw_free(smx_context_t context);
217 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
218 static void smx_ctx_raw_stop(smx_context_t context);
219 static void smx_ctx_raw_suspend_serial(smx_context_t context);
220 static void smx_ctx_raw_resume_serial(smx_process_t first_process);
221 static void smx_ctx_raw_runall_serial(void);
222 static void smx_ctx_raw_suspend_parallel(smx_context_t context);
223 static void smx_ctx_raw_resume_parallel(smx_process_t first_process);
224 static void smx_ctx_raw_runall_parallel(void);
225 static void smx_ctx_raw_runall(void);
228 * \brief Initializes the raw context factory.
229 * \param factory where to initialize the factory
231 void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory)
234 XBT_VERB("Using raw contexts. Because the glibc is just not good enough for us.");
235 smx_ctx_base_factory_init(factory);
237 (*factory)->finalize = smx_ctx_raw_factory_finalize;
238 (*factory)->create_context = smx_ctx_raw_create_context;
239 /* Do not overload that method (*factory)->finalize */
240 (*factory)->free = smx_ctx_raw_free;
241 (*factory)->stop = smx_ctx_raw_stop;
242 (*factory)->name = "smx_raw_context_factory";
244 if (SIMIX_context_is_parallel()) {
245 #ifdef CONTEXT_THREADS
246 int nthreads = SIMIX_context_get_nthreads();
247 xbt_os_thread_key_create(&raw_worker_id_key);
248 raw_parmap = xbt_parmap_new(nthreads, SIMIX_context_get_parallel_mode());
249 raw_workers_context = xbt_new(smx_ctx_raw_t, nthreads);
250 raw_maestro_context=NULL;
253 if (SIMIX_context_get_parallel_threshold() > 1) {
254 /* choose dynamically */
255 (*factory)->runall = smx_ctx_raw_runall;
256 (*factory)->suspend = NULL;
259 /* always parallel */
260 (*factory)->runall = smx_ctx_raw_runall_parallel;
261 (*factory)->suspend = smx_ctx_raw_suspend_parallel;
266 (*factory)->runall = smx_ctx_raw_runall_serial;
267 (*factory)->suspend = smx_ctx_raw_suspend_serial;
269 #ifdef TIME_BENCH_PER_SR
270 timer = xbt_os_timer_new();
275 * \brief Finalizes the raw context factory.
276 * \param factory the raw context factory
278 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory)
280 #ifdef TIME_BENCH_PER_SR
281 XBT_CRITICAL("Total wasted time in %u SR: %f", sr_count, time_wasted_sr);
282 XBT_CRITICAL("Total wasted time in %u SSR: %f", ssr_count, time_wasted_ssr);
285 #ifdef CONTEXT_THREADS
287 xbt_parmap_destroy(raw_parmap);
288 xbt_free(raw_workers_context);
290 return smx_ctx_base_factory_finalize(factory);
294 * \brief Creates a new raw context.
295 * \param code main function of this context or NULL to create the maestro
297 * \param argc argument number
298 * \param argv arguments to pass to the main function
299 * \param cleanup_func a function to call to free the user data when the
301 * \param process SIMIX process
304 smx_ctx_raw_create_context(xbt_main_func_t code, int argc, char **argv,
305 void_pfn_smxprocess_t cleanup_func,
306 smx_process_t process)
309 smx_ctx_raw_t context =
310 (smx_ctx_raw_t) smx_ctx_base_factory_create_context_sized(
311 sizeof(s_smx_ctx_raw_t),
318 /* if the user provided a function for the process then use it,
319 otherwise it is the context for maestro */
321 context->malloced_stack = SIMIX_context_stack_new();
323 raw_makecontext(context->malloced_stack, smx_context_stack_size,
324 (void_f_pvoid_t) smx_ctx_raw_wrapper, context);
326 #ifdef HAVE_VALGRIND_VALGRIND_H
327 context->valgrind_stack_id =
328 VALGRIND_STACK_REGISTER(context->malloced_stack,
329 context->malloced_stack + smx_context_stack_size);
330 #endif /* HAVE_VALGRIND_VALGRIND_H */
333 if(process != NULL && raw_maestro_context==NULL)
334 raw_maestro_context = context;
337 MC_ignore_heap(&(raw_maestro_context->stack_top), sizeof(raw_maestro_context->stack_top));
341 return (smx_context_t) context;
345 * \brief Destroys a raw context.
346 * \param context a raw context
348 static void smx_ctx_raw_free(smx_context_t context)
352 #ifdef HAVE_VALGRIND_VALGRIND_H
353 VALGRIND_STACK_DEREGISTER(((smx_ctx_raw_t)
354 context)->valgrind_stack_id);
355 #endif /* HAVE_VALGRIND_VALGRIND_H */
357 SIMIX_context_stack_delete(((smx_ctx_raw_t) context)->malloced_stack);
360 smx_ctx_base_free(context);
364 * \brief Wrapper for the main function of a context.
365 * \param context a raw context
367 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context)
369 (context->super.code) (context->super.argc, context->super.argv);
371 smx_ctx_raw_stop((smx_context_t) context);
375 * \brief Stops a raw context.
377 * This function is called when the main function of the context if finished.
379 * \param context the current context
381 static void smx_ctx_raw_stop(smx_context_t context)
383 smx_ctx_base_stop(context);
384 simix_global->context_factory->suspend(context);
388 * \brief Suspends a running context and resumes another one or returns to
390 * \param context the current context
392 static void smx_ctx_raw_suspend_serial(smx_context_t context)
394 /* determine the next context */
395 smx_context_t next_context;
396 unsigned long int i = raw_process_index++;
398 if (i < xbt_dynar_length(simix_global->process_to_run)) {
399 /* execute the next process */
400 XBT_DEBUG("Run next process");
401 next_context = xbt_dynar_get_as(
402 simix_global->process_to_run, i, smx_process_t)->context;
405 /* all processes were run, return to maestro */
406 XBT_DEBUG("No more process to run");
407 next_context = (smx_context_t) raw_maestro_context;
409 SIMIX_context_set_current(next_context);
410 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top,
411 ((smx_ctx_raw_t) next_context)->stack_top);
415 * \brief Resumes sequentially all processes ready to run.
416 * \param first_process the first process to resume
418 static void smx_ctx_raw_resume_serial(smx_process_t first_process)
420 smx_ctx_raw_t context = (smx_ctx_raw_t) first_process->context;
421 SIMIX_context_set_current((smx_context_t) context);
422 raw_swapcontext(&raw_maestro_context->stack_top,
423 ((smx_ctx_raw_t) context)->stack_top);
426 #ifdef TIME_BENCH_PER_SR
427 static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
429 smx_process_t process;
434 unsigned long num_proc = xbt_dynar_length(processes);
436 unsigned int data_size = (num_proc / NUM_THREADS) + ((num_proc % NUM_THREADS) ? 1 : 0);
439 time_thread_ssr[0] = 0;
440 xbt_dynar_foreach(processes, cursor, process) {
441 XBT_DEBUG("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
442 if(cursor >= t * data_size + data_size){
443 if(time_thread_ssr[t] > tmax)
444 tmax = time_thread_ssr[t];
446 time_thread_ssr[t] = 0;
450 ((smx_ctx_raw_t)process->context)->thread = t;
451 time_thread_sr[t] = 0;
454 xbt_os_cputimer_start(timer);
455 smx_ctx_raw_resume(process);
456 xbt_os_cputimer_stop(timer);
457 elapsed = xbt_os_timer_elapsed(timer);
458 time_thread_ssr[t] += elapsed;
459 time_thread_sr[((smx_ctx_raw_t)process->context)->thread] += elapsed;
465 if(time_thread_ssr[t] > tmax)
466 tmax = time_thread_ssr[t];
468 for(cursor=0; cursor <= t; cursor++){
469 XBT_VERB("Time SSR thread %u = %f (max %f)", cursor, time_thread_ssr[cursor], tmax);
470 time_wasted_ssr += tmax - time_thread_ssr[cursor];
474 void smx_ctx_raw_new_sr(void);
475 void smx_ctx_raw_new_sr(void)
481 for(i=0; i < NUM_THREADS; i++){
482 if(time_thread_sr[i] > tmax)
483 tmax = time_thread_sr[i];
486 for(i=0; i < NUM_THREADS; i++){
487 XBT_VERB("Time SR thread %u = %f (max %f)", i, time_thread_sr[i], tmax);
488 time_wasted_sr += tmax - time_thread_sr[i];
491 XBT_VERB("New scheduling round");
496 * \brief Resumes sequentially all processes ready to run.
498 static void smx_ctx_raw_runall_serial(void)
500 smx_process_t first_process =
501 xbt_dynar_get_as(simix_global->process_to_run, 0, smx_process_t);
502 raw_process_index = 1;
504 /* execute the first process */
505 smx_ctx_raw_resume_serial(first_process);
510 * \brief Suspends a running context and resumes another one or returns to
511 * the main function of the current worker thread.
512 * \param context the context of the current worker thread
514 static void smx_ctx_raw_suspend_parallel(smx_context_t context)
516 #ifdef CONTEXT_THREADS
517 /* determine the next context */
518 smx_process_t next_work = xbt_parmap_next(raw_parmap);
519 smx_context_t next_context;
520 raw_stack_t next_stack;
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;
529 /* all processes were run, go to the barrier */
530 XBT_DEBUG("No more processes to run");
532 unsigned long worker_id =
533 (unsigned long) xbt_os_thread_get_specific(raw_worker_id_key);
535 next_context = (smx_context_t)raw_workers_context[worker_id];
536 XBT_DEBUG("Restoring worker stack %lu (working threads = %lu)",
537 worker_id, raw_threads_working);
538 next_stack = ((smx_ctx_raw_t)next_context)->stack_top;
541 SIMIX_context_set_current(next_context);
542 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top, next_stack);
547 * \brief Resumes sequentially in the current worker thread the processes ready
549 * \param first_process the first process to resume
551 static void smx_ctx_raw_resume_parallel(smx_process_t first_process)
553 #ifdef CONTEXT_THREADS
554 unsigned long worker_id = __sync_fetch_and_add(&raw_threads_working, 1);
555 xbt_os_thread_set_specific(raw_worker_id_key, (void*) worker_id);
556 smx_ctx_raw_t worker_context = (smx_ctx_raw_t)SIMIX_context_self();
557 raw_workers_context[worker_id] = worker_context;
558 XBT_DEBUG("Saving worker stack %lu", worker_id);
559 raw_stack_t* worker_stack = &(worker_context)->stack_top;
562 smx_context_t context = first_process->context;
563 SIMIX_context_set_current(context);
564 raw_swapcontext(worker_stack, ((smx_ctx_raw_t) context)->stack_top);
569 * \brief Resumes in parallel all processes ready to run.
571 static void smx_ctx_raw_runall_parallel(void)
573 #ifdef CONTEXT_THREADS
574 raw_threads_working = 0;
575 xbt_parmap_apply(raw_parmap, (void_f_pvoid_t) smx_ctx_raw_resume_parallel,
576 simix_global->process_to_run);
578 xbt_die("You asked for a parallel execution, but you don't have any threads.");
583 * \brief Resumes all processes ready to run.
585 static void smx_ctx_raw_runall(void)
587 unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
588 if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
589 XBT_DEBUG("Runall // %lu", nb_processes);
590 simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
591 smx_ctx_raw_runall_parallel();
593 XBT_DEBUG("Runall serial %lu", nb_processes);
594 simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
595 smx_ctx_raw_runall_serial();