1 /* context_raw - context switching with ucontextes from System V */
3 /* Copyright (c) 2009, 2010. 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 "simix/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 raw_stack_t old_stack_top; /* to whom I should return the control */
25 #ifdef HAVE_VALGRIND_VALGRIND_H
26 unsigned int valgrind_stack_id; /* the valgrind stack id */
29 unsigned int thread; /* Just for measuring purposes */
31 } s_smx_ctx_raw_t, *smx_ctx_raw_t;
33 smx_ctx_raw_t maestro_raw_context;
35 extern raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
36 rawctx_entry_point_t entry_point, void* arg);
37 extern void raw_swapcontext(raw_stack_t* old, raw_stack_t new);
41 #if defined(APPLE) || defined(_WIN32)
43 ".globl _raw_makecontext\n"
47 ".globl raw_makecontext\n"
48 ".type raw_makecontext,@function\n"
51 " movl 4(%esp),%eax\n" /* stack */
52 " addl 8(%esp),%eax\n" /* size */
53 " movl 12(%esp),%ecx\n" /* func */
54 " movl 16(%esp),%edx\n" /* arg */
55 " movl %edx, -4(%eax)\n"
56 " movl $0, -8(%eax)\n" /* @return for func */
57 " movl %ecx,-12(%eax)\n"
58 " movl $0, -16(%eax)\n" /* ebp */
59 " movl $0, -20(%eax)\n" /* ebx */
60 " movl $0, -24(%eax)\n" /* esi */
61 " movl $0, -28(%eax)\n" /* edi */
67 #if defined(APPLE) || defined(_WIN32)
69 ".globl _raw_swapcontext\n"
73 ".globl raw_swapcontext\n"
74 ".type raw_swapcontext,@function\n"
77 " movl 4(%esp),%eax\n" /* old */
78 " movl 8(%esp),%edx\n" /* new */
91 #elif PROCESSOR_x86_64
93 #if defined(APPLE) || defined(_WIN32)
95 ".globl _raw_makecontext\n"
99 ".globl raw_makecontext\n"
100 ".type raw_makecontext,@function\n"
101 "raw_makecontext:\n"/* Calling convention sets the arguments in rdi, rsi, rdx and rcx, respectively */
103 " movq %rdi,%rax\n" /* stack */
104 " addq %rsi,%rax\n" /* size */
105 " movq $0, -8(%rax)\n" /* @return for func */
106 " movq %rdx,-16(%rax)\n" /* func */
107 " movq %rcx,-24(%rax)\n" /* arg/rdi */
108 " movq $0, -32(%rax)\n" /* rsi */
109 " movq $0, -40(%rax)\n" /* rdx */
110 " movq $0, -48(%rax)\n" /* rcx */
111 " movq $0, -56(%rax)\n" /* r8 */
112 " movq $0, -64(%rax)\n" /* r9 */
113 " movq $0, -72(%rax)\n" /* rbp */
114 " movq $0, -80(%rax)\n" /* rbx */
115 " movq $0, -88(%rax)\n" /* r12 */
116 " movq $0, -96(%rax)\n" /* r13 */
117 " movq $0, -104(%rax)\n" /* r14 */
118 " movq $0, -112(%rax)\n" /* r15 */
124 #if defined(APPLE) || defined(_WIN32)
126 ".globl _raw_swapcontext\n"
127 "_raw_swapcontext:\n"
130 ".globl raw_swapcontext\n"
131 ".type raw_swapcontext,@function\n"
132 "raw_swapcontext:\n" /* Calling convention sets the arguments in rdi and rsi, respectively */
146 " movq %rsp,(%rdi)\n" /* old */
147 " movq %rsi,%rsp\n" /* new */
164 /* If you implement raw contextes for other processors, don't forget to
165 update the definition of HAVE_RAWCTX in buildtools/Cmake/AddTests.cmake */
167 raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
168 rawctx_entry_point_t entry_point, void* arg) {
172 void raw_swapcontext(raw_stack_t* old, raw_stack_t new) {
178 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
180 #ifdef CONTEXT_THREADS
181 static xbt_parmap_t parmap;
185 #include "xbt/xbt_os_time.h"
186 #define NUM_THREADS 4
187 static xbt_os_timer_t timer;
188 static double time_thread_sr[NUM_THREADS];
189 static double time_thread_ssr[NUM_THREADS];
190 static double time_wasted_sr = 0;
191 static double time_wasted_ssr = 0;
192 static unsigned int sr_count = 0;
193 static unsigned int ssr_count = 0;
194 static char new_sr = 0;
197 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
199 static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory)
202 XBT_CRITICAL("Total wasted time in %u SR: %lf", sr_count, time_wasted_sr);
203 XBT_CRITICAL("Total wasted time in %u SSR: %lf", ssr_count, time_wasted_ssr);
206 #ifdef CONTEXT_THREADS
208 xbt_parmap_destroy(parmap);
210 return smx_ctx_base_factory_finalize(factory);
215 smx_ctx_raw_create_context(xbt_main_func_t code, int argc, char **argv,
216 void_pfn_smxprocess_t cleanup_func,
220 smx_ctx_raw_t context =
221 (smx_ctx_raw_t) smx_ctx_base_factory_create_context_sized(
222 sizeof(s_smx_ctx_raw_t),
229 /* If the user provided a function for the process then use it
230 otherwise is the context for maestro */
232 context->malloced_stack = xbt_malloc0(smx_context_stack_size);
234 raw_makecontext(context->malloced_stack, smx_context_stack_size,
235 (void(*)(void*))smx_ctx_raw_wrapper,context);
237 #ifdef HAVE_VALGRIND_VALGRIND_H
238 context->valgrind_stack_id =
239 VALGRIND_STACK_REGISTER(context->malloced_stack,
240 context->malloced_stack + smx_context_stack_size);
241 #endif /* HAVE_VALGRIND_VALGRIND_H */
244 maestro_raw_context = context;
247 return (smx_context_t) context;
250 static void smx_ctx_raw_free(smx_context_t context)
255 #ifdef HAVE_VALGRIND_VALGRIND_H
256 VALGRIND_STACK_DEREGISTER(((smx_ctx_raw_t)
257 context)->valgrind_stack_id);
258 #endif /* HAVE_VALGRIND_VALGRIND_H */
260 free(((smx_ctx_raw_t)context)->malloced_stack);
262 smx_ctx_base_free(context);
265 static void smx_ctx_raw_suspend(smx_context_t context)
267 SIMIX_context_set_current((smx_context_t) maestro_raw_context);
269 &((smx_ctx_raw_t) context)->stack_top,
270 ((smx_ctx_raw_t) context)->old_stack_top);
273 static void smx_ctx_raw_stop(smx_context_t context)
275 smx_ctx_base_stop(context);
276 smx_ctx_raw_suspend(context);
279 static void smx_ctx_raw_wrapper(smx_ctx_raw_t context)
281 (context->super.code) (context->super.argc, context->super.argv);
283 smx_ctx_raw_stop((smx_context_t) context);
286 static void smx_ctx_raw_resume(smx_process_t process)
288 smx_ctx_raw_t context = (smx_ctx_raw_t)process->context;
289 SIMIX_context_set_current((smx_context_t) context);
291 &((smx_ctx_raw_t) context)->old_stack_top,
292 ((smx_ctx_raw_t) context)->stack_top);
296 static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
298 smx_process_t process;
303 unsigned long num_proc = xbt_dynar_length(processes);
305 unsigned int data_size = (num_proc / NUM_THREADS) + ((num_proc % NUM_THREADS) ? 1 : 0);
308 time_thread_ssr[0] = 0;
309 xbt_dynar_foreach(processes, cursor, process) {
310 XBT_DEBUG("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
311 if(cursor >= t * data_size + data_size){
312 if(time_thread_ssr[t] > tmax)
313 tmax = time_thread_ssr[t];
315 time_thread_ssr[t] = 0;
319 ((smx_ctx_raw_t)process->context)->thread = t;
320 time_thread_sr[t] = 0;
323 xbt_os_timer_start(timer);
324 smx_ctx_raw_resume(process);
325 xbt_os_timer_stop(timer);
326 elapsed = xbt_os_timer_elapsed(timer);
327 time_thread_ssr[t] += elapsed;
328 time_thread_sr[((smx_ctx_raw_t)process->context)->thread] += elapsed;
334 if(time_thread_ssr[t] > tmax)
335 tmax = time_thread_ssr[t];
337 for(cursor=0; cursor <= t; cursor++){
338 XBT_VERB("Time SSR thread %u = %lf (max %lf)", cursor, time_thread_ssr[cursor], tmax);
339 time_wasted_ssr += tmax - time_thread_ssr[cursor];
343 void smx_ctx_raw_new_sr(void);
344 void smx_ctx_raw_new_sr(void)
350 for(i=0; i < NUM_THREADS; i++){
351 if(time_thread_sr[i] > tmax)
352 tmax = time_thread_sr[i];
355 for(i=0; i < NUM_THREADS; i++){
356 XBT_VERB("Time SR thread %u = %lf (max %lf)", i, time_thread_sr[i], tmax);
357 time_wasted_sr += tmax - time_thread_sr[i];
360 XBT_VERB("New scheduling round");
363 static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
365 smx_process_t process;
368 xbt_dynar_foreach(processes, cursor, process) {
369 XBT_DEBUG("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
370 smx_ctx_raw_resume(process);
375 static void smx_ctx_raw_runall_parallel(xbt_dynar_t processes)
377 #ifdef CONTEXT_THREADS
378 xbt_parmap_apply(parmap, (void_f_pvoid_t)smx_ctx_raw_resume, processes);
382 static void smx_ctx_raw_runall(xbt_dynar_t processes)
384 if (xbt_dynar_length(processes) >= SIMIX_context_get_parallel_threshold()) {
385 XBT_DEBUG("Runall // %lu", xbt_dynar_length(processes));
386 smx_ctx_raw_runall_parallel(processes);
388 XBT_DEBUG("Runall serial %lu", xbt_dynar_length(processes));
389 smx_ctx_raw_runall_serial(processes);
393 void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory)
395 XBT_VERB("Using raw contexts. Because the glibc is just not good enough for us.");
396 smx_ctx_base_factory_init(factory);
398 (*factory)->finalize = smx_ctx_raw_factory_finalize;
399 (*factory)->create_context = smx_ctx_raw_create_context;
400 /* Do not overload that method (*factory)->finalize */
401 (*factory)->free = smx_ctx_raw_free;
402 (*factory)->stop = smx_ctx_raw_stop;
403 (*factory)->suspend = smx_ctx_raw_suspend;
404 (*factory)->name = "smx_raw_context_factory";
406 if (SIMIX_context_is_parallel()) {
407 #ifdef CONTEXT_THREADS
408 parmap = xbt_parmap_new(SIMIX_context_get_nthreads());
410 if (SIMIX_context_get_parallel_threshold() > 1) {
411 /* choose dynamically */
412 (*factory)->runall = smx_ctx_raw_runall;
415 /* always parallel */
416 (*factory)->runall = smx_ctx_raw_runall_parallel;
421 (*factory)->runall = smx_ctx_raw_runall_serial;
424 timer = xbt_os_timer_new();