-/* context_raw - context switching with ucontextes from System V */
+/* context_raw - fast context switching inspired from System V ucontextes */
/* Copyright (c) 2009, 2010. The SimGrid Team.
* All rights reserved. */
#include "simix/private.h"
#include "xbt/parmap.h"
-
#ifdef HAVE_VALGRIND_VALGRIND_H
# include <valgrind/valgrind.h>
#endif /* HAVE_VALGRIND_VALGRIND_H */
-#ifdef _XBT_WIN32
-#include "win32_ucontext.h"
-#include "win32_ucontext.c"
-#else
-#include "ucontext.h"
-#endif
-
typedef char * raw_stack_t;
typedef void (*rawctx_entry_point_t)(void *);
typedef struct s_smx_ctx_raw {
- s_smx_ctx_base_t super; /* Fields of super implementation */
- char *malloced_stack; /* malloced area containing the stack */
- raw_stack_t stack_top; /* pointer to stack top (within previous area) */
- raw_stack_t old_stack_top; /* to whom I should return the control */
+ s_smx_ctx_base_t super; /* Fields of super implementation */
+ char *malloced_stack; /* malloced area containing the stack */
+ raw_stack_t stack_top; /* pointer to stack top (within previous area) */
#ifdef HAVE_VALGRIND_VALGRIND_H
- unsigned int valgrind_stack_id; /* the valgrind stack id */
+ unsigned int valgrind_stack_id; /* the valgrind stack id */
+#endif
+#ifdef TIME_BENCH
+ unsigned int thread; /* Just for measuring purposes */
#endif
} s_smx_ctx_raw_t, *smx_ctx_raw_t;
-smx_ctx_raw_t maestro_raw_context;
+#ifdef CONTEXT_THREADS
+static xbt_parmap_t raw_parmap;
+static raw_stack_t* raw_workers_stacks; /* space to save the worker stack in each thread */
+static unsigned long raw_threads_working; /* number of threads that have started their work */
+static xbt_os_thread_key_t raw_worker_id_key; /* thread-specific storage for the thread id */
+#endif
+
+static unsigned long raw_process_index = 0; /* index of the next process to run in the
+ * list of runnable processes */
+smx_ctx_raw_t raw_maestro_context;
extern raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
rawctx_entry_point_t entry_point, void* arg);
#ifdef PROCESSOR_i686
__asm__ (
+#if defined(APPLE) || defined(_WIN32)
+ ".text\n"
+ ".globl _raw_makecontext\n"
+ "_raw_makecontext:\n"
+#else
".text\n"
".globl raw_makecontext\n"
".type raw_makecontext,@function\n"
"raw_makecontext:\n"
+#endif
" movl 4(%esp),%eax\n" /* stack */
" addl 8(%esp),%eax\n" /* size */
" movl 12(%esp),%ecx\n" /* func */
);
__asm__ (
+#if defined(APPLE) || defined(_WIN32)
+ ".text\n"
+ ".globl _raw_swapcontext\n"
+ "_raw_swapcontext:\n"
+#else
".text\n"
".globl raw_swapcontext\n"
".type raw_swapcontext,@function\n"
"raw_swapcontext:\n"
+#endif
" movl 4(%esp),%eax\n" /* old */
" movl 8(%esp),%edx\n" /* new */
" pushl %ebp\n"
);
#elif PROCESSOR_x86_64
__asm__ (
+#if defined(APPLE) || defined(_WIN32)
+ ".text\n"
+ ".globl _raw_makecontext\n"
+ "_raw_makecontext:\n"
+#else
".text\n"
".globl raw_makecontext\n"
".type raw_makecontext,@function\n"
- "raw_makecontext:\n" /* Calling convention sets the arguments in rdi, rsi, rdx and rcx, respectively */
+ "raw_makecontext:\n"/* Calling convention sets the arguments in rdi, rsi, rdx and rcx, respectively */
+#endif
" movq %rdi,%rax\n" /* stack */
" addq %rsi,%rax\n" /* size */
" movq $0, -8(%rax)\n" /* @return for func */
);
__asm__ (
+#if defined(APPLE) || defined(_WIN32)
+ ".text\n"
+ ".globl _raw_swapcontext\n"
+ "_raw_swapcontext:\n"
+#else
".text\n"
".globl raw_swapcontext\n"
".type raw_swapcontext,@function\n"
"raw_swapcontext:\n" /* Calling convention sets the arguments in rdi and rsi, respectively */
+#endif
" pushq %rdi\n"
" pushq %rsi\n"
" pushq %rdx\n"
" retq\n"
);
#else
+
+/* If you implement raw contextes for other processors, don't forget to
+ update the definition of HAVE_RAWCTX in buildtools/Cmake/CompleteInFiles.cmake */
+
raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
rawctx_entry_point_t entry_point, void* arg) {
THROW_UNIMPLEMENTED;
XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
-#ifdef CONTEXT_THREADS
-static xbt_parmap_t parmap;
+#ifdef TIME_BENCH
+#include "xbt/xbt_os_time.h"
+#define NUM_THREADS 4
+static xbt_os_timer_t timer;
+static double time_thread_sr[NUM_THREADS];
+static double time_thread_ssr[NUM_THREADS];
+static double time_wasted_sr = 0;
+static double time_wasted_ssr = 0;
+static unsigned int sr_count = 0;
+static unsigned int ssr_count = 0;
+static char new_sr = 0;
#endif
-static smx_context_factory_t raw_factory;
-
static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
+static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory);
+static smx_context_t smx_ctx_raw_create_context(xbt_main_func_t code, int argc,
+ char **argv, void_pfn_smxprocess_t cleanup_func, void *data);
+static void smx_ctx_raw_free(smx_context_t context);
+static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
+static void smx_ctx_raw_stop(smx_context_t context);
+static void smx_ctx_raw_suspend_serial(smx_context_t context);
+static void smx_ctx_raw_resume_serial(smx_process_t first_process);
+static void smx_ctx_raw_runall_serial(void);
+static void smx_ctx_raw_suspend_parallel(smx_context_t context);
+static void smx_ctx_raw_resume_parallel(smx_process_t first_process);
+static void smx_ctx_raw_runall_parallel(void);
+static void smx_ctx_raw_runall(void);
+
+void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory)
+{
+ XBT_VERB("Using raw contexts. Because the glibc is just not good enough for us.");
+ smx_ctx_base_factory_init(factory);
+
+ (*factory)->finalize = smx_ctx_raw_factory_finalize;
+ (*factory)->create_context = smx_ctx_raw_create_context;
+ /* Do not overload that method (*factory)->finalize */
+ (*factory)->free = smx_ctx_raw_free;
+ (*factory)->stop = smx_ctx_raw_stop;
+ (*factory)->name = "smx_raw_context_factory";
+
+ if (SIMIX_context_is_parallel()) {
+#ifdef CONTEXT_THREADS
+ int nthreads = SIMIX_context_get_nthreads();
+ raw_parmap = xbt_parmap_new(nthreads);
+ raw_workers_stacks = xbt_new(raw_stack_t, nthreads);
+ xbt_os_thread_key_create(&raw_worker_id_key);
+#endif
+ if (SIMIX_context_get_parallel_threshold() > 1) {
+ /* choose dynamically */
+ (*factory)->runall = smx_ctx_raw_runall;
+ (*factory)->suspend = NULL;
+ }
+ else {
+ /* always parallel */
+ (*factory)->runall = smx_ctx_raw_runall_parallel;
+ (*factory)->suspend = smx_ctx_raw_suspend_parallel;
+ }
+ }
+ else {
+ /* always serial */
+ (*factory)->runall = smx_ctx_raw_runall_serial;
+ (*factory)->suspend = smx_ctx_raw_suspend_serial;
+ }
+#ifdef TIME_BENCH
+ timer = xbt_os_timer_new();
+#endif
+}
static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory)
-{
- if(parmap)
- xbt_parmap_destroy(parmap);
+{
+#ifdef TIME_BENCH
+ XBT_CRITICAL("Total wasted time in %u SR: %lf", sr_count, time_wasted_sr);
+ XBT_CRITICAL("Total wasted time in %u SSR: %lf", ssr_count, time_wasted_ssr);
+#endif
+
+#ifdef CONTEXT_THREADS
+ if (raw_parmap)
+ xbt_parmap_destroy(raw_parmap);
+ xbt_free(raw_workers_stacks);
+#endif
return smx_ctx_base_factory_finalize(factory);
}
-
static smx_context_t
smx_ctx_raw_create_context(xbt_main_func_t code, int argc, char **argv,
void_pfn_smxprocess_t cleanup_func,
cleanup_func,
data);
- /* If the user provided a function for the process then use it
- otherwise is the context for maestro */
+ /* if the user provided a function for the process then use it,
+ otherwise it is the context for maestro */
if (code) {
context->malloced_stack = xbt_malloc0(smx_context_stack_size);
context->stack_top =
context->malloced_stack + smx_context_stack_size);
#endif /* HAVE_VALGRIND_VALGRIND_H */
- }else{
- maestro_raw_context = context;
+ } else {
+ raw_maestro_context = context;
}
return (smx_context_t) context;
smx_ctx_base_free(context);
}
-static void smx_ctx_raw_suspend(smx_context_t context)
+static void smx_ctx_raw_wrapper(smx_ctx_raw_t context)
{
- smx_current_context = (smx_context_t)maestro_raw_context;
- raw_swapcontext(
- &((smx_ctx_raw_t) context)->stack_top,
- ((smx_ctx_raw_t) context)->old_stack_top);
+ (context->super.code) (context->super.argc, context->super.argv);
+
+ smx_ctx_raw_stop((smx_context_t) context);
}
static void smx_ctx_raw_stop(smx_context_t context)
{
smx_ctx_base_stop(context);
- smx_ctx_raw_suspend(context);
+ simix_global->context_factory->suspend(context);
}
-static void smx_ctx_raw_wrapper(smx_ctx_raw_t context)
-{
- (context->super.code) (context->super.argc, context->super.argv);
-
- smx_ctx_raw_stop((smx_context_t) context);
+static void smx_ctx_raw_suspend_serial(smx_context_t context)
+{
+ /* determine the next context */
+ smx_context_t next_context;
+ unsigned long int i = raw_process_index++;
+
+ if (i < xbt_dynar_length(simix_global->process_to_run)) {
+ /* execute the next process */
+ XBT_DEBUG("Run next process");
+ next_context = xbt_dynar_get_as(
+ simix_global->process_to_run,i, smx_process_t)->context;
+ }
+ else {
+ /* all processes were run, return to maestro */
+ XBT_DEBUG("No more process to run");
+ next_context = (smx_context_t) raw_maestro_context;
+ }
+ SIMIX_context_set_current(next_context);
+ raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top,
+ ((smx_ctx_raw_t) next_context)->stack_top);
}
-static void smx_ctx_raw_resume(smx_process_t process)
+static void smx_ctx_raw_resume_serial(smx_process_t first_process)
{
- smx_ctx_raw_t context = (smx_ctx_raw_t)process->context;
- smx_current_context = (smx_context_t)context;
- raw_swapcontext(
- &((smx_ctx_raw_t) context)->old_stack_top,
+ smx_ctx_raw_t context = (smx_ctx_raw_t) first_process->context;
+ SIMIX_context_set_current((smx_context_t) context);
+ raw_swapcontext(&raw_maestro_context->stack_top,
((smx_ctx_raw_t) context)->stack_top);
}
-
+#ifdef TIME_BENCH
static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
{
smx_process_t process;
unsigned int cursor;
+ double elapsed = 0;
+ double tmax = 0;
+ unsigned long num_proc = xbt_dynar_length(processes);
+ unsigned int t=0;
+ unsigned int data_size = (num_proc / NUM_THREADS) + ((num_proc % NUM_THREADS) ? 1 : 0);
+
+ ssr_count++;
+ time_thread_ssr[0] = 0;
xbt_dynar_foreach(processes, cursor, process) {
- DEBUG2("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
+ XBT_DEBUG("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
+ if(cursor >= t * data_size + data_size){
+ if(time_thread_ssr[t] > tmax)
+ tmax = time_thread_ssr[t];
+ t++;
+ time_thread_ssr[t] = 0;
+ }
+
+ if(new_sr){
+ ((smx_ctx_raw_t)process->context)->thread = t;
+ time_thread_sr[t] = 0;
+ }
+
+ xbt_os_timer_start(timer);
smx_ctx_raw_resume(process);
+ xbt_os_timer_stop(timer);
+ elapsed = xbt_os_timer_elapsed(timer);
+ time_thread_ssr[t] += elapsed;
+ time_thread_sr[((smx_ctx_raw_t)process->context)->thread] += elapsed;
+ }
+
+ if(new_sr)
+ new_sr = FALSE;
+
+ if(time_thread_ssr[t] > tmax)
+ tmax = time_thread_ssr[t];
+
+ for(cursor=0; cursor <= t; cursor++){
+ XBT_VERB("Time SSR thread %u = %lf (max %lf)", cursor, time_thread_ssr[cursor], tmax);
+ time_wasted_ssr += tmax - time_thread_ssr[cursor];
}
- xbt_dynar_reset(processes);
}
-static void smx_ctx_raw_runall_parallel(xbt_dynar_t processes)
+void smx_ctx_raw_new_sr(void);
+void smx_ctx_raw_new_sr(void)
{
- xbt_parmap_apply(parmap, (void_f_pvoid_t)smx_ctx_raw_resume, processes);
- xbt_dynar_reset(processes);
-}
+ int i;
+ double tmax = 0;
+ new_sr = TRUE;
+ sr_count++;
+ for(i=0; i < NUM_THREADS; i++){
+ if(time_thread_sr[i] > tmax)
+ tmax = time_thread_sr[i];
+ }
+
+ for(i=0; i < NUM_THREADS; i++){
+ XBT_VERB("Time SR thread %u = %lf (max %lf)", i, time_thread_sr[i], tmax);
+ time_wasted_sr += tmax - time_thread_sr[i];
+ }
-static smx_context_t smx_ctx_raw_self_parallel(void)
+ XBT_VERB("New scheduling round");
+}
+#else
+static void smx_ctx_raw_runall_serial(void)
{
- return smx_current_context;
+ if (!xbt_dynar_is_empty(simix_global->process_to_run)) {
+ smx_process_t first_process =
+ xbt_dynar_get_as(simix_global->process_to_run, 0, smx_process_t);
+ raw_process_index = 1;
+
+ /* execute the first process */
+ smx_ctx_raw_resume_serial(first_process);
+ }
}
+#endif
-static int smx_ctx_raw_get_thread_id(){
- return (int)(unsigned long)xbt_os_thread_get_extra_data();
+static void smx_ctx_raw_stop_parallel(smx_context_t context)
+{
+ smx_ctx_base_stop(context);
+ smx_ctx_raw_suspend_parallel(context);
}
-static void smx_ctx_raw_runall(xbt_dynar_t processes)
+static void smx_ctx_raw_suspend_parallel(smx_context_t context)
{
- if (SIMIX_context_is_parallel()
- && xbt_dynar_length(processes) >= SIMIX_context_get_parallel_threshold()) {
- DEBUG1("Runall // %lu", xbt_dynar_length(processes));
- raw_factory->self = smx_ctx_raw_self_parallel;
- raw_factory->get_thread_id = smx_ctx_raw_get_thread_id;
- smx_ctx_raw_runall_parallel(processes);
- } else {
- DEBUG1("Runall serial %lu", xbt_dynar_length(processes));
- raw_factory->self = smx_ctx_base_self;
- raw_factory->get_thread_id = smx_ctx_base_get_thread_id;
- smx_ctx_raw_runall_serial(processes);
+#ifdef CONTEXT_THREADS
+ /* determine the next context */
+ smx_process_t next_work = xbt_parmap_next(raw_parmap);
+ smx_context_t next_context;
+ raw_stack_t next_stack;
+
+ if (next_work != NULL) {
+ /* there is a next process to resume */
+ XBT_DEBUG("Run next process");
+ next_context = next_work->context;
+ next_stack = ((smx_ctx_raw_t) next_context)->stack_top;
+ }
+ else {
+ /* all processes were run, go to the barrier */
+ XBT_DEBUG("No more processes to run");
+ next_context = (smx_context_t) raw_maestro_context;
+ unsigned long worker_id =
+ (unsigned long) xbt_os_thread_get_specific(raw_worker_id_key);
+ XBT_DEBUG("Restoring worker stack %lu (working threads = %lu)",
+ worker_id, raw_threads_working);
+ next_stack = raw_workers_stacks[worker_id];
}
+
+ SIMIX_context_set_current(next_context);
+ raw_swapcontext(&((smx_ctx_raw_t) context)->stack_top, next_stack);
+#endif
}
-void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory)
+static void smx_ctx_raw_resume_parallel(smx_process_t first_process)
{
- VERB0("Using raw contexts. Because the glibc is just not good enough for us.");
- smx_ctx_base_factory_init(factory);
-
- (*factory)->finalize = smx_ctx_raw_factory_finalize;
- (*factory)->create_context = smx_ctx_raw_create_context;
- /* Do not overload that method (*factory)->finalize */
- (*factory)->free = smx_ctx_raw_free;
- (*factory)->stop = smx_ctx_raw_stop;
- (*factory)->suspend = smx_ctx_raw_suspend;
- (*factory)->name = "smx_raw_context_factory";
+#ifdef CONTEXT_THREADS
+ unsigned long worker_id = __sync_fetch_and_add(&raw_threads_working, 1);
+ xbt_os_thread_set_specific(raw_worker_id_key, (void*) worker_id);
+ XBT_DEBUG("Saving worker stack %lu", worker_id);
+ raw_stack_t* worker_stack = &raw_workers_stacks[worker_id];
+
+ smx_context_t context = first_process->context;
+ SIMIX_context_set_current(context);
+ raw_swapcontext(worker_stack, ((smx_ctx_raw_t) context)->stack_top);
+#endif
+}
- parmap = xbt_parmap_new(2);
- (*factory)->self = smx_ctx_raw_self_parallel;
- (*factory)->get_thread_id = smx_ctx_raw_get_thread_id;
- (*factory)->runall = smx_ctx_raw_runall;
+static void smx_ctx_raw_runall_parallel(void)
+{
+#ifdef CONTEXT_THREADS
+ raw_threads_working = 0;
+ xbt_parmap_apply(raw_parmap, (void_f_pvoid_t) smx_ctx_raw_resume_parallel,
+ simix_global->process_to_run);
+#endif
+}
- raw_factory = *factory;
+static void smx_ctx_raw_runall(void)
+{
+ unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
+ if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
+ XBT_DEBUG("Runall // %lu", nb_processes);
+ simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
+ smx_ctx_raw_runall_parallel();
+ } else {
+ XBT_DEBUG("Runall serial %lu", nb_processes);
+ simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
+ smx_ctx_raw_runall_serial();
+ }
}
