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 static unsigned long raw_process_index = 0; /* index of the next process to run in the
33 * list of runnable processes */
34 static smx_ctx_raw_t raw_maestro_context;
36 extern raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
37 rawctx_entry_point_t entry_point, void* arg);
38 extern void raw_swapcontext(raw_stack_t* old, raw_stack_t new);
44 ".globl _raw_makecontext\n"
48 ".globl raw_makecontext\n"
52 ".globl raw_makecontext\n"
53 ".type raw_makecontext,@function\n"
54 "raw_makecontext:\n"/* Calling convention sets the arguments in rdi, rsi, rdx and rcx, respectively */
56 " mov %rdi,%rax\n" /* stack */
57 " add %rsi,%rax\n" /* size */
58 " andq $-16, %rax\n" /* align stack */
59 " movq $0, -8(%rax)\n" /* @return for func */
60 " mov %rdx,-16(%rax)\n" /* func */
61 " mov %rcx,-24(%rax)\n" /* arg/rdi */
62 " movq $0, -32(%rax)\n" /* rsi */
63 " movq $0, -40(%rax)\n" /* rdx */
64 " movq $0, -48(%rax)\n" /* rcx */
65 " movq $0, -56(%rax)\n" /* r8 */
66 " movq $0, -64(%rax)\n" /* r9 */
67 " movq $0, -72(%rax)\n" /* rbp */
68 " movq $0, -80(%rax)\n" /* rbx */
69 " movq $0, -88(%rax)\n" /* r12 */
70 " movq $0, -96(%rax)\n" /* r13 */
71 " movq $0, -104(%rax)\n" /* r14 */
72 " movq $0, -112(%rax)\n" /* r15 */
80 ".globl _raw_swapcontext\n"
84 ".globl raw_swapcontext\n"
88 ".globl raw_swapcontext\n"
89 ".type raw_swapcontext,@function\n"
90 "raw_swapcontext:\n" /* Calling convention sets the arguments in rdi and rsi, respectively */
104 " mov %rsp,(%rdi)\n" /* old */
105 " mov %rsi,%rsp\n" /* new */
122 #if defined(APPLE) || defined(_WIN32)
124 ".globl _raw_makecontext\n"
125 "_raw_makecontext:\n"
128 ".globl raw_makecontext\n"
129 ".type raw_makecontext,@function\n"
132 " movl 4(%esp),%eax\n" /* stack */
133 " addl 8(%esp),%eax\n" /* size */
134 " andl $-16, %eax\n" /* align stack */
135 " movl 12(%esp),%ecx\n" /* func */
136 " movl 16(%esp),%edx\n" /* arg */
137 " movl %edx, -4(%eax)\n"
138 " movl $0, -8(%eax)\n" /* @return for func */
139 " movl %ecx,-12(%eax)\n"
140 " movl $0, -16(%eax)\n" /* ebp */
141 " movl $0, -20(%eax)\n" /* ebx */
142 " movl $0, -24(%eax)\n" /* esi */
143 " movl $0, -28(%eax)\n" /* edi */
149 #if defined(APPLE) || defined(_WIN32)
151 ".globl _raw_swapcontext\n"
152 "_raw_swapcontext:\n"
155 ".globl raw_swapcontext\n"
156 ".type raw_swapcontext,@function\n"
159 " movl 4(%esp),%eax\n" /* old */
160 " movl 8(%esp),%edx\n" /* new */
165 " movl %esp,(%eax)\n"
176 /* If you implement raw contexts for other processors, don't forget to
177 update the definition of HAVE_RAWCTX in buildtools/Cmake/CompleteInFiles.cmake */
179 raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
180 rawctx_entry_point_t entry_point, void* arg) {
184 void raw_swapcontext(raw_stack_t* old, raw_stack_t new) {
190 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
192 #ifdef TIME_BENCH_PER_SR
193 #include "xbt/xbt_os_time.h"
194 #define NUM_THREADS 4
195 static xbt_os_timer_t timer;
196 static double time_thread_sr[NUM_THREADS];
197 static double time_thread_ssr[NUM_THREADS];
198 static double time_wasted_sr = 0;
199 static double time_wasted_ssr = 0;
200 static unsigned int sr_count = 0;
201 static unsigned int ssr_count = 0;
202 static char new_sr = 0;
205 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
206 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory);
207 static smx_context_t smx_ctx_raw_create_context(xbt_main_func_t code, int argc,
208 char **argv, void_pfn_smxprocess_t cleanup_func, smx_process_t process);
209 static void smx_ctx_raw_free(smx_context_t context);
210 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
211 static void smx_ctx_raw_stop(smx_context_t context);
212 static void smx_ctx_raw_suspend_serial(smx_context_t context);
213 static void smx_ctx_raw_resume_serial(smx_process_t first_process);
214 #ifdef TIME_BENCH_PER_SR
215 static void smx_ctx_raw_runall_serial(xbt_dynar_t processes);
216 void smx_ctx_raw_new_sr(void);
218 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_VERB("Total wasted time in %u SR: %f", sr_count, time_wasted_sr);
280 XBT_VERB("Total wasted time in %u SSR: %f", 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 process SIMIX process
302 smx_ctx_raw_create_context(xbt_main_func_t code, int argc, char **argv,
303 void_pfn_smxprocess_t cleanup_func,
304 smx_process_t process)
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 = SIMIX_context_stack_new();
321 raw_makecontext(context->malloced_stack,
322 smx_context_usable_stack_size,
323 (void_f_pvoid_t)smx_ctx_raw_wrapper, context);
326 if(process != NULL && raw_maestro_context==NULL)
327 raw_maestro_context = context;
330 MC_ignore_heap(&(raw_maestro_context->stack_top), sizeof(raw_maestro_context->stack_top));
334 return (smx_context_t) context;
338 * \brief Destroys a raw context.
339 * \param context a raw context
341 static void smx_ctx_raw_free(smx_context_t context)
344 SIMIX_context_stack_delete(((smx_ctx_raw_t) context)->malloced_stack);
346 smx_ctx_base_free(context);
350 * \brief Wrapper for the main function of a context.
351 * \param context a raw context
353 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context)
355 (context->super.code) (context->super.argc, context->super.argv);
357 smx_ctx_raw_stop((smx_context_t) context);
361 * \brief Stops a raw context.
363 * This function is called when the main function of the context if finished.
365 * \param context the current context
367 static void smx_ctx_raw_stop(smx_context_t context)
369 smx_ctx_base_stop(context);
370 simix_global->context_factory->suspend(context);
374 * \brief Suspends a running context and resumes another one or returns to
376 * \param context the current context
378 static void smx_ctx_raw_suspend_serial(smx_context_t context)
380 /* determine the next context */
381 smx_context_t next_context;
383 #ifdef TIME_BENCH_PER_SR
384 i = ++raw_process_index;
386 i = raw_process_index++;
388 if (i < xbt_dynar_length(simix_global->process_to_run)) {
389 /* execute the next process */
390 XBT_DEBUG("Run next process");
391 next_context = xbt_dynar_get_as(
392 simix_global->process_to_run, i, smx_process_t)->context;
395 /* all processes were run, return to maestro */
396 XBT_DEBUG("No more process to run");
397 next_context = (smx_context_t) raw_maestro_context;
399 SIMIX_context_set_current(next_context);
400 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top,
401 ((smx_ctx_raw_t) next_context)->stack_top);
405 * \brief Resumes sequentially all processes ready to run.
406 * \param first_process the first process to resume
408 static void smx_ctx_raw_resume_serial(smx_process_t first_process)
410 smx_ctx_raw_t context = (smx_ctx_raw_t) first_process->context;
411 SIMIX_context_set_current((smx_context_t) context);
412 raw_swapcontext(&raw_maestro_context->stack_top,
413 ((smx_ctx_raw_t) context)->stack_top);
416 #ifdef TIME_BENCH_PER_SR
417 static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
419 smx_process_t process;
423 unsigned long num_proc = xbt_dynar_length(simix_global->process_to_run);
425 unsigned int data_size = (num_proc / NUM_THREADS) + ((num_proc % NUM_THREADS) ? 1 : 0);
428 time_thread_ssr[0] = 0;
429 xbt_dynar_foreach(processes, cursor, process){
430 XBT_VERB("Schedule item %u of %lu",cursor,num_proc);
431 if(cursor >= t * data_size + data_size){
432 if(time_thread_ssr[t] > tmax)
433 tmax = time_thread_ssr[t];
435 time_thread_ssr[t] = 0;
439 ((smx_ctx_raw_t)process->context)->thread = t;
440 time_thread_sr[t] = 0;
443 xbt_os_cputimer_start(timer);
444 smx_ctx_raw_resume_serial(process);
445 xbt_os_cputimer_stop(timer);
446 elapsed = xbt_os_timer_elapsed(timer);
447 time_thread_ssr[t] += elapsed;
448 time_thread_sr[((smx_ctx_raw_t)process->context)->thread] += elapsed;
454 if(time_thread_ssr[t] > tmax)
455 tmax = time_thread_ssr[t];
457 for(cursor=0; cursor <= t; cursor++){
458 XBT_VERB("Time SSR thread %u = %f (max %f)", cursor, time_thread_ssr[cursor], tmax);
459 time_wasted_ssr += tmax - time_thread_ssr[cursor];
463 void smx_ctx_raw_new_sr(void)
469 for(i=0; i < NUM_THREADS; i++){
470 if(time_thread_sr[i] > tmax)
471 tmax = time_thread_sr[i];
474 for(i=0; i < NUM_THREADS; i++){
475 XBT_VERB("Time SR thread %u = %f (max %f)", i, time_thread_sr[i], tmax);
476 time_wasted_sr += tmax - time_thread_sr[i];
479 XBT_VERB("Total time SR %u = %f, %d", sr_count, tmax, xbt_dynar_length(simix_global->process_that_ran));
480 XBT_VERB("New scheduling round");
484 * \brief Resumes sequentially all processes ready to run.
486 static void smx_ctx_raw_runall_serial(void)
488 smx_process_t first_process =
489 xbt_dynar_get_as(simix_global->process_to_run, 0, smx_process_t);
490 raw_process_index = 1;
492 /* execute the first process */
493 smx_ctx_raw_resume_serial(first_process);
498 * \brief Suspends a running context and resumes another one or returns to
499 * the main function of the current worker thread.
500 * \param context the context of the current worker thread
502 static void smx_ctx_raw_suspend_parallel(smx_context_t context)
504 #ifdef CONTEXT_THREADS
505 /* determine the next context */
506 smx_process_t next_work = xbt_parmap_next(raw_parmap);
507 smx_context_t next_context;
508 raw_stack_t next_stack;
510 if (next_work != NULL) {
511 /* there is a next process to resume */
512 XBT_DEBUG("Run next process");
513 next_context = next_work->context;
514 next_stack = ((smx_ctx_raw_t) next_context)->stack_top;
517 /* all processes were run, go to the barrier */
518 XBT_DEBUG("No more processes to run");
520 unsigned long worker_id =
521 (unsigned long) xbt_os_thread_get_specific(raw_worker_id_key);
523 next_context = (smx_context_t)raw_workers_context[worker_id];
524 XBT_DEBUG("Restoring worker stack %lu (working threads = %lu)",
525 worker_id, raw_threads_working);
526 next_stack = ((smx_ctx_raw_t)next_context)->stack_top;
529 SIMIX_context_set_current(next_context);
530 raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top, next_stack);
535 * \brief Resumes sequentially in the current worker thread the processes ready
537 * \param first_process the first process to resume
539 static void smx_ctx_raw_resume_parallel(smx_process_t first_process)
541 #ifdef CONTEXT_THREADS
542 unsigned long worker_id = __sync_fetch_and_add(&raw_threads_working, 1);
543 xbt_os_thread_set_specific(raw_worker_id_key, (void*) worker_id);
544 smx_ctx_raw_t worker_context = (smx_ctx_raw_t)SIMIX_context_self();
545 raw_workers_context[worker_id] = worker_context;
546 XBT_DEBUG("Saving worker stack %lu", worker_id);
547 raw_stack_t* worker_stack = &(worker_context)->stack_top;
550 smx_context_t context = first_process->context;
551 SIMIX_context_set_current(context);
552 raw_swapcontext(worker_stack, ((smx_ctx_raw_t) context)->stack_top);
557 * \brief Resumes in parallel all processes ready to run.
559 static void smx_ctx_raw_runall_parallel(void)
561 #ifdef CONTEXT_THREADS
562 raw_threads_working = 0;
563 xbt_parmap_apply(raw_parmap, (void_f_pvoid_t) smx_ctx_raw_resume_parallel,
564 simix_global->process_to_run);
566 xbt_die("You asked for a parallel execution, but you don't have any threads.");
571 * \brief Resumes all processes ready to run.
573 static void smx_ctx_raw_runall(void)
575 unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
576 if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
577 XBT_DEBUG("Runall // %lu", nb_processes);
578 simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
579 smx_ctx_raw_runall_parallel();
581 XBT_DEBUG("Runall serial %lu", nb_processes);
582 simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
583 #ifdef TIME_BENCH_PER_SR
584 smx_ctx_raw_runall_serial(simix_global->process_to_run);
586 smx_ctx_raw_runall_serial();