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. */
8 #include "smx_private.h"
9 #include "xbt/parmap.h"
10 #include "xbt/dynar.h"
13 typedef char * raw_stack_t;
14 typedef void (*rawctx_entry_point_t)(void *);
16 typedef struct s_smx_ctx_raw {
17 s_smx_ctx_base_t super; /* Fields of super implementation */
18 char *malloced_stack; /* malloced area containing the stack */
19 raw_stack_t stack_top; /* pointer to stack top (within previous area) */
20 #ifdef TIME_BENCH_PER_SR
21 unsigned int thread; /* Just for measuring purposes */
23 } s_smx_ctx_raw_t, *smx_ctx_raw_t;
25 #ifdef CONTEXT_THREADS
26 static xbt_parmap_t raw_parmap;
27 static smx_ctx_raw_t* raw_workers_context; /* space to save the worker context in each thread */
28 static unsigned long raw_threads_working; /* number of threads that have started their work */
29 static xbt_os_thread_key_t raw_worker_id_key; /* thread-specific storage for the thread id */
32 #ifdef ADAPTIVE_THRESHOLD
33 #define SCHED_ROUND_LIMIT 5
34 xbt_os_timer_t round_time;
35 double par_time,seq_time;
36 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 */
37 static unsigned int seq_sched_round, par_sched_round; /* Amount of SR that ran serial/parallel*/
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 " andq $-16, %rax\n" /* align stack */
67 " movq $0, -8(%rax)\n" /* @return for func */
68 " mov %rdx,-16(%rax)\n" /* func */
69 " mov %rcx,-24(%rax)\n" /* arg/rdi */
70 " movq $0, -32(%rax)\n" /* rsi */
71 " movq $0, -40(%rax)\n" /* rdx */
72 " movq $0, -48(%rax)\n" /* rcx */
73 " movq $0, -56(%rax)\n" /* r8 */
74 " movq $0, -64(%rax)\n" /* r9 */
75 " movq $0, -72(%rax)\n" /* rbp */
76 " movq $0, -80(%rax)\n" /* rbx */
77 " movq $0, -88(%rax)\n" /* r12 */
78 " movq $0, -96(%rax)\n" /* r13 */
79 " movq $0, -104(%rax)\n" /* r14 */
80 " movq $0, -112(%rax)\n" /* r15 */
88 ".globl _raw_swapcontext\n"
92 ".globl raw_swapcontext\n"
96 ".globl raw_swapcontext\n"
97 ".type raw_swapcontext,@function\n"
98 "raw_swapcontext:\n" /* Calling convention sets the arguments in rdi and rsi, respectively */
112 " mov %rsp,(%rdi)\n" /* old */
113 " mov %rsi,%rsp\n" /* new */
130 #if defined(APPLE) || defined(_WIN32)
132 ".globl _raw_makecontext\n"
133 "_raw_makecontext:\n"
136 ".globl raw_makecontext\n"
137 ".type raw_makecontext,@function\n"
140 " movl 4(%esp),%eax\n" /* stack */
141 " addl 8(%esp),%eax\n" /* size */
142 " andl $-16, %eax\n" /* align stack */
143 " movl 12(%esp),%ecx\n" /* func */
144 " movl 16(%esp),%edx\n" /* arg */
145 " movl %edx, -4(%eax)\n"
146 " movl $0, -8(%eax)\n" /* @return for func */
147 " movl %ecx,-12(%eax)\n"
148 " movl $0, -16(%eax)\n" /* ebp */
149 " movl $0, -20(%eax)\n" /* ebx */
150 " movl $0, -24(%eax)\n" /* esi */
151 " movl $0, -28(%eax)\n" /* edi */
157 #if defined(APPLE) || defined(_WIN32)
159 ".globl _raw_swapcontext\n"
160 "_raw_swapcontext:\n"
163 ".globl raw_swapcontext\n"
164 ".type raw_swapcontext,@function\n"
167 " movl 4(%esp),%eax\n" /* old */
168 " movl 8(%esp),%edx\n" /* new */
173 " movl %esp,(%eax)\n"
184 /* If you implement raw contexts for other processors, don't forget to
185 update the definition of HAVE_RAWCTX in buildtools/Cmake/CompleteInFiles.cmake */
187 raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
188 rawctx_entry_point_t entry_point, void* arg) {
192 void raw_swapcontext(raw_stack_t* old, raw_stack_t new) {
198 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
200 #ifdef TIME_BENCH_PER_SR
201 #include "xbt/xbt_os_time.h"
202 #define NUM_THREADS 4
203 static xbt_os_timer_t timer;
204 static double time_thread_sr[NUM_THREADS];
205 static double time_thread_ssr[NUM_THREADS];
206 static double time_wasted_sr = 0;
207 static double time_wasted_ssr = 0;
208 static unsigned int sr_count = 0;
209 static unsigned int ssr_count = 0;
210 static char new_sr = 0;
213 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
214 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory);
215 static smx_context_t smx_ctx_raw_create_context(xbt_main_func_t code, int argc,
216 char **argv, void_pfn_smxprocess_t cleanup_func, smx_process_t process);
217 static void smx_ctx_raw_free(smx_context_t context);
218 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
219 static void smx_ctx_raw_stop(smx_context_t context);
220 static void smx_ctx_raw_suspend_serial(smx_context_t context);
221 static void smx_ctx_raw_resume_serial(smx_process_t first_process);
222 #ifdef TIME_BENCH_PER_SR
223 static void smx_ctx_raw_runall_serial(xbt_dynar_t processes);
224 void smx_ctx_raw_new_sr(void);
226 static void smx_ctx_raw_runall_serial(void);
228 static void smx_ctx_raw_suspend_parallel(smx_context_t context);
229 static void smx_ctx_raw_resume_parallel(smx_process_t first_process);
230 static void smx_ctx_raw_runall_parallel(void);
231 static void smx_ctx_raw_runall(void);
234 * \brief Initializes the raw context factory.
235 * \param factory where to initialize the factory
237 void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory)
240 XBT_VERB("Using raw contexts. Because the glibc is just not good enough for us.");
241 smx_ctx_base_factory_init(factory);
243 (*factory)->finalize = smx_ctx_raw_factory_finalize;
244 (*factory)->create_context = smx_ctx_raw_create_context;
245 /* Do not overload that method (*factory)->finalize */
246 (*factory)->free = smx_ctx_raw_free;
247 (*factory)->stop = smx_ctx_raw_stop;
248 (*factory)->name = "smx_raw_context_factory";
250 if (SIMIX_context_is_parallel()) {
251 #ifdef CONTEXT_THREADS
252 int nthreads = SIMIX_context_get_nthreads();
253 xbt_os_thread_key_create(&raw_worker_id_key);
254 raw_parmap = xbt_parmap_new(nthreads, SIMIX_context_get_parallel_mode());
255 raw_workers_context = xbt_new(smx_ctx_raw_t, nthreads);
256 raw_maestro_context=NULL;
259 if (SIMIX_context_get_parallel_threshold() > 1) {
260 /* choose dynamically */
261 (*factory)->runall = smx_ctx_raw_runall;
262 (*factory)->suspend = NULL;
265 /* always parallel */
266 (*factory)->runall = smx_ctx_raw_runall_parallel;
267 (*factory)->suspend = smx_ctx_raw_suspend_parallel;
272 (*factory)->runall = smx_ctx_raw_runall_serial;
273 (*factory)->suspend = smx_ctx_raw_suspend_serial;
275 #ifdef TIME_BENCH_PER_SR
276 timer = xbt_os_timer_new();
281 * \brief Finalizes the raw context factory.
282 * \param factory the raw context factory
284 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory)
286 #ifdef TIME_BENCH_PER_SR
287 XBT_VERB("Total wasted time in %u SR: %f", sr_count, time_wasted_sr);
288 XBT_VERB("Total wasted time in %u SSR: %f", ssr_count, time_wasted_ssr);
291 #ifdef CONTEXT_THREADS
293 xbt_parmap_destroy(raw_parmap);
294 xbt_free(raw_workers_context);
296 return smx_ctx_base_factory_finalize(factory);
300 * \brief Creates a new raw context.
301 * \param code main function of this context or NULL to create the maestro
303 * \param argc argument number
304 * \param argv arguments to pass to the main function
305 * \param cleanup_func a function to call to free the user data when the
307 * \param process SIMIX process
310 smx_ctx_raw_create_context(xbt_main_func_t code, int argc, char **argv,
311 void_pfn_smxprocess_t cleanup_func,
312 smx_process_t process)
315 smx_ctx_raw_t context =
316 (smx_ctx_raw_t) smx_ctx_base_factory_create_context_sized(
317 sizeof(s_smx_ctx_raw_t),
324 /* if the user provided a function for the process then use it,
325 otherwise it is the context for maestro */
327 context->malloced_stack = SIMIX_context_stack_new();
329 raw_makecontext(context->malloced_stack,
330 smx_context_usable_stack_size,
331 (void_f_pvoid_t)smx_ctx_raw_wrapper, context);
334 if(process != NULL && raw_maestro_context==NULL)
335 raw_maestro_context = context;
338 MC_ignore_heap(&(raw_maestro_context->stack_top), sizeof(raw_maestro_context->stack_top));
342 return (smx_context_t) context;
346 * \brief Destroys a raw context.
347 * \param context a raw context
349 static void smx_ctx_raw_free(smx_context_t context)
352 SIMIX_context_stack_delete(((smx_ctx_raw_t) context)->malloced_stack);
354 smx_ctx_base_free(context);
358 * \brief Wrapper for the main function of a context.
359 * \param context a raw context
361 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context)
363 (context->super.code) (context->super.argc, context->super.argv);
365 smx_ctx_raw_stop((smx_context_t) context);
369 * \brief Stops a raw context.
371 * This function is called when the main function of the context if finished.
373 * \param context the current context
375 static void smx_ctx_raw_stop(smx_context_t context)
377 smx_ctx_base_stop(context);
378 simix_global->context_factory->suspend(context);
382 * \brief Suspends a running context and resumes another one or returns to
384 * \param context the current context
386 static void smx_ctx_raw_suspend_serial(smx_context_t context)
388 /* determine the next context */
389 smx_context_t next_context;
391 #ifdef TIME_BENCH_PER_SR
392 i = ++raw_process_index;
394 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;
431 unsigned long num_proc = xbt_dynar_length(simix_global->process_to_run);
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_VERB("Schedule item %u of %lu",cursor,num_proc);
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_serial(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 = %f (max %f)", cursor, time_thread_ssr[cursor], tmax);
467 time_wasted_ssr += tmax - time_thread_ssr[cursor];
471 void smx_ctx_raw_new_sr(void)
477 for(i=0; i < NUM_THREADS; i++){
478 if(time_thread_sr[i] > tmax)
479 tmax = time_thread_sr[i];
482 for(i=0; i < NUM_THREADS; i++){
483 XBT_CRITICAL("Time SR thread %u = %f (max %f)", i, time_thread_sr[i], tmax);
484 time_wasted_sr += tmax - time_thread_sr[i];
487 XBT_CRITICAL("Total time SR %u = %f, %d", sr_count, tmax, xbt_dynar_length(simix_global->process_that_ran));
488 XBT_CRITICAL("New scheduling round");
492 * \brief Resumes sequentially all processes ready to run.
494 static void smx_ctx_raw_runall_serial(void)
496 smx_process_t first_process =
497 xbt_dynar_get_as(simix_global->process_to_run, 0, smx_process_t);
498 raw_process_index = 1;
500 /* execute the first process */
501 smx_ctx_raw_resume_serial(first_process);
506 * \brief Suspends a running context and resumes another one or returns to
507 * the main function of the current worker thread.
508 * \param context the context of the current worker thread
510 static void smx_ctx_raw_suspend_parallel(smx_context_t context)
512 #ifdef CONTEXT_THREADS
513 /* determine the next context */
514 smx_process_t next_work = xbt_parmap_next(raw_parmap);
515 smx_context_t next_context;
516 raw_stack_t next_stack;
518 if (next_work != NULL) {
519 /* there is a next process to resume */
520 XBT_DEBUG("Run next process");
521 next_context = next_work->context;
522 next_stack = ((smx_ctx_raw_t) next_context)->stack_top;
525 /* all processes were run, go to the barrier */
526 XBT_DEBUG("No more processes to run");
528 unsigned long worker_id =
529 (unsigned long) xbt_os_thread_get_specific(raw_worker_id_key);
531 next_context = (smx_context_t)raw_workers_context[worker_id];
532 XBT_DEBUG("Restoring worker stack %lu (working threads = %lu)",
533 worker_id, raw_threads_working);
534 next_stack = ((smx_ctx_raw_t)next_context)->stack_top;
537 SIMIX_context_set_current(next_context);
538 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top, next_stack);
543 * \brief Resumes sequentially in the current worker thread the processes ready
545 * \param first_process the first process to resume
547 static void smx_ctx_raw_resume_parallel(smx_process_t first_process)
549 #ifdef CONTEXT_THREADS
550 unsigned long worker_id = __sync_fetch_and_add(&raw_threads_working, 1);
551 xbt_os_thread_set_specific(raw_worker_id_key, (void*) worker_id);
552 smx_ctx_raw_t worker_context = (smx_ctx_raw_t)SIMIX_context_self();
553 raw_workers_context[worker_id] = worker_context;
554 XBT_DEBUG("Saving worker stack %lu", worker_id);
555 raw_stack_t* worker_stack = &(worker_context)->stack_top;
558 smx_context_t context = first_process->context;
559 SIMIX_context_set_current(context);
560 raw_swapcontext(worker_stack, ((smx_ctx_raw_t) context)->stack_top);
565 * \brief Resumes in parallel all processes ready to run.
567 static void smx_ctx_raw_runall_parallel(void)
569 #ifdef CONTEXT_THREADS
570 raw_threads_working = 0;
571 xbt_parmap_apply(raw_parmap, (void_f_pvoid_t) smx_ctx_raw_resume_parallel,
572 simix_global->process_to_run);
574 xbt_die("You asked for a parallel execution, but you don't have any threads.");
579 * \brief Resumes all processes ready to run.
581 #ifdef ADAPTIVE_THRESHOLD
582 static void smx_ctx_raw_runall(void)
584 unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
586 if(seq_sched_round % SCHED_ROUND_LIMIT == 0 && par_sched_round % SCHED_ROUND_LIMIT == 0){
589 if((seq_time / (double)seq_proc_that_ran) > (par_time / (double)par_proc_that_ran)){
590 SIMIX_context_set_parallel_threshold(SIMIX_context_get_parallel_threshold() - 1);
592 SIMIX_context_set_parallel_threshold(SIMIX_context_get_parallel_threshold() + 1);
594 par_time = 0; par_proc_that_ran = 0;
595 seq_time = 0; seq_proc_that_ran = 0;
597 round_time = xbt_os_timer_new();
598 if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
599 XBT_DEBUG("Runall // %lu", nb_processes);
600 simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
601 xbt_os_cputimer_start(round_time);
602 smx_ctx_raw_runall_parallel();
603 xbt_os_cputimer_stop(round_time);
604 par_time += xbt_os_timer_elapsed(round_time);
605 par_proc_that_ran += xbt_dynar_length(simix_global->process_that_ran);
606 par_sched_round += 1;
608 XBT_DEBUG("Runall serial %lu", nb_processes);
609 simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
610 xbt_os_cputimer_start(round_time);
611 #ifdef TIME_BENCH_PER_SR
612 smx_ctx_raw_runall_serial(simix_global->process_to_run);
614 smx_ctx_raw_runall_serial();
616 xbt_os_cputimer_stop(round_time);
617 seq_time += xbt_os_timer_elapsed(round_time);
618 seq_proc_that_ran += xbt_dynar_length(simix_global->process_that_ran);
619 seq_sched_round += 1;
625 static void smx_ctx_raw_runall(void)
627 unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
628 if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
629 XBT_DEBUG("Runall // %lu", nb_processes);
630 simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
631 smx_ctx_raw_runall_parallel();
633 XBT_DEBUG("Runall serial %lu", nb_processes);
634 simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
635 #ifdef TIME_BENCH_PER_SR
636 smx_ctx_raw_runall_serial(simix_global->process_to_run);
638 smx_ctx_raw_runall_serial();