summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--accel-pppd/triton/triton.c183
-rw-r--r--accel-pppd/triton/triton.h2
-rw-r--r--accel-pppd/triton/triton_p.h4
3 files changed, 119 insertions, 70 deletions
diff --git a/accel-pppd/triton/triton.c b/accel-pppd/triton/triton.c
index 7a494a5a..ce054ce1 100644
--- a/accel-pppd/triton/triton.c
+++ b/accel-pppd/triton/triton.c
@@ -5,6 +5,9 @@
#include <string.h>
#include <unistd.h>
#include <limits.h>
+#include <assert.h>
+#include <ucontext.h>
+#include <setjmp.h>
#include <sys/resource.h>
#include "triton_p.h"
@@ -49,7 +52,9 @@ static struct triton_timer_t ru_timer = {
};
struct triton_context_t default_ctx;
-static struct triton_context_t __thread *this_ctx;
+static __thread struct triton_context_t *this_ctx;
+static __thread jmp_buf jmp_env;
+static __thread void *thread_frame;
#define log_debug2(fmt, ...)
@@ -77,6 +82,7 @@ static void __config_reload(void (*notify)(int))
}
static void ctx_thread(struct _triton_context_t *ctx);
+
static void* triton_thread(struct _triton_thread_t *thread)
{
sigset_t set;
@@ -92,21 +98,38 @@ static void* triton_thread(struct _triton_thread_t *thread)
sigaddset(&set, SIGUSR1);
sigaddset(&set, SIGQUIT);
+ thread_frame = __builtin_frame_address(0);
+
pthread_mutex_lock(&thread->sleep_lock);
pthread_mutex_unlock(&thread->sleep_lock);
while (1) {
spin_lock(&threads_lock);
- if (!list_empty(&ctx_queue) && !need_config_reload && triton_stat.thread_active <= thread_count) {
- thread->ctx = list_entry(ctx_queue.next, typeof(*thread->ctx), entry2);
- log_debug2("thread: %p: dequeued ctx %p\n", thread, thread->ctx);
- list_del(&thread->ctx->entry2);
- spin_unlock(&threads_lock);
- spin_lock(&thread->ctx->lock);
- thread->ctx->thread = thread;
- thread->ctx->queued = 0;
- spin_unlock(&thread->ctx->lock);
- __sync_sub_and_fetch(&triton_stat.context_pending, 1);
+ if ((!list_empty(&ctx_queue) || !list_empty(&thread->wakeup_list)) && !need_config_reload && triton_stat.thread_active <= thread_count) {
+ if (!list_empty(&thread->wakeup_list)) {
+ thread->ctx = list_entry(thread->wakeup_list.next, typeof(*thread->ctx), entry2);
+ log_debug2("thread: %p: wakeup ctx %p\n", thread, thread->ctx);
+ list_del(&thread->ctx->entry2);
+ spin_unlock(&threads_lock);
+
+ this_ctx = thread->ctx->ud;
+ if (thread->ctx->ud->before_switch)
+ thread->ctx->ud->before_switch(thread->ctx->ud, thread->ctx->bf_arg);
+
+ alloca(thread->ctx->uc->uc_stack.ss_size);
+ memcpy(thread_frame - thread->ctx->uc->uc_stack.ss_size, thread->ctx->uc->uc_stack.ss_sp, thread->ctx->uc->uc_stack.ss_size);
+ setcontext(thread->ctx->uc);
+ } else {
+ thread->ctx = list_entry(ctx_queue.next, typeof(*thread->ctx), entry2);
+ log_debug2("thread: %p: dequeued ctx %p\n", thread, thread->ctx);
+ list_del(&thread->ctx->entry2);
+ spin_unlock(&threads_lock);
+ spin_lock(&thread->ctx->lock);
+ thread->ctx->thread = thread;
+ thread->ctx->queued = 0;
+ spin_unlock(&thread->ctx->lock);
+ __sync_sub_and_fetch(&triton_stat.context_pending, 1);
+ }
} else {
if (triton_stat.thread_count > thread_count + triton_stat.context_sleeping) {
__sync_sub_and_fetch(&triton_stat.thread_active, 1);
@@ -118,7 +141,9 @@ static void* triton_thread(struct _triton_thread_t *thread)
_free(thread);
return NULL;
}
+
log_debug2("thread: %p: sleeping\n", thread);
+
if (!terminate)
list_add(&thread->entry2, &sleep_threads);
@@ -149,31 +174,33 @@ static void* triton_thread(struct _triton_thread_t *thread)
spin_unlock(&threads_lock);
}
- log_debug2("thread %p: ctx=%p %p\n", thread, thread->ctx, thread->ctx ? thread->ctx->thread : NULL);
- this_ctx = thread->ctx->ud;
- if (thread->ctx->ud->before_switch)
- thread->ctx->ud->before_switch(thread->ctx->ud, thread->ctx->bf_arg);
+ if (setjmp(jmp_env) == 0) {
+ log_debug2("thread %p: ctx=%p %p\n", thread, thread->ctx, thread->ctx ? thread->ctx->thread : NULL);
+ this_ctx = thread->ctx->ud;
+ if (thread->ctx->ud->before_switch)
+ thread->ctx->ud->before_switch(thread->ctx->ud, thread->ctx->bf_arg);
-cont:
- log_debug2("thread %p: switch to %p\n", thread, thread->ctx);
- ctx_thread(thread->ctx);
- log_debug2("thread %p: switch from %p %p\n", thread, thread->ctx, thread->ctx->thread);
+ while (1) {
+ log_debug2("thread %p: switch to %p\n", thread, thread->ctx);
+ ctx_thread(thread->ctx);
+ log_debug2("thread %p: switch from %p %p\n", thread, thread->ctx, thread->ctx->thread);
- spin_lock(&threads_lock);
- if (thread->ctx->pending && !thread->ctx->need_free) {
+ spin_lock(&threads_lock);
+ if (!thread->ctx->pending || thread->ctx->need_free)
+ break;
+ spin_unlock(&threads_lock);
+ }
+
+ thread->ctx->thread = NULL;
+ need_free = thread->ctx->need_free;
spin_unlock(&threads_lock);
- goto cont;
- }
- thread->ctx->thread = NULL;
- need_free = thread->ctx->need_free;
- spin_unlock(&threads_lock);
- if (need_free) {
- log_debug2("- context %p removed\n", thread->ctx);
- triton_context_release(thread->ctx);
+ if (need_free) {
+ log_debug2("- context %p removed\n", thread->ctx);
+ triton_context_release(thread->ctx);
+ }
+ thread->ctx = NULL;
}
-
- thread->ctx = NULL;
}
}
@@ -272,6 +299,7 @@ struct _triton_thread_t *create_thread()
pthread_attr_setstacksize(&attr, WORKER_STACK_SIZE);
memset(thread, 0, sizeof(*thread));
+ INIT_LIST_HEAD(&thread->wakeup_list);
pthread_mutex_init(&thread->sleep_lock, NULL);
pthread_cond_init(&thread->sleep_cond, NULL);
pthread_mutex_lock(&thread->sleep_lock);
@@ -419,36 +447,6 @@ void __export triton_context_set_priority(struct triton_context_t *ud, int prio)
ctx->priority = prio > 0;
}
-void __export triton_context_schedule()
-{
- struct _triton_context_t *ctx = (struct _triton_context_t *)this_ctx->tpd;
- struct _triton_thread_t *t = NULL;
-
- log_debug2("ctx %p: enter schedule\n", ctx);
- __sync_add_and_fetch(&triton_stat.context_sleeping, 1);
- __sync_sub_and_fetch(&triton_stat.thread_active, 1);
- pthread_mutex_lock(&ctx->thread->sleep_lock);
- while (1) {
- if (ctx->wakeup) {
- ctx->wakeup = 0;
- break;
- } else {
- if (!t && triton_stat.thread_count <= thread_count + triton_stat.context_sleeping) {
- t = create_thread();
- spin_lock(&threads_lock);
- list_add_tail(&t->entry, &threads);
- spin_unlock(&threads_lock);
- pthread_mutex_unlock(&t->sleep_lock);
- }
- pthread_cond_wait(&ctx->thread->sleep_cond, &ctx->thread->sleep_lock);
- }
- }
- pthread_mutex_unlock(&ctx->thread->sleep_lock);
- __sync_sub_and_fetch(&triton_stat.context_sleeping, 1);
- __sync_add_and_fetch(&triton_stat.thread_active, 1);
- log_debug2("ctx %p: exit schedule\n", ctx);
-}
-
struct triton_context_t __export *triton_context_self(void)
{
return this_ctx;
@@ -462,6 +460,50 @@ void triton_context_print(void)
printf("%p\n", ctx);
}
+static ucontext_t *alloc_context()
+{
+ ucontext_t *uc;
+ void *frame = __builtin_frame_address(0);
+ size_t stack_size = thread_frame - frame;
+
+ uc = malloc(sizeof(*uc) + stack_size);
+ uc->uc_stack.ss_sp = (void *)(uc + 1);
+ uc->uc_stack.ss_size = stack_size;
+ memcpy(uc->uc_stack.ss_sp, frame, stack_size);
+
+ return uc;
+}
+
+void __export triton_context_schedule()
+{
+ struct _triton_context_t *ctx = (struct _triton_context_t *)this_ctx->tpd;
+
+ log_debug2("ctx %p: enter schedule\n", ctx);
+ __sync_add_and_fetch(&triton_stat.context_sleeping, 1);
+
+ ctx->uc = alloc_context();
+
+ getcontext(ctx->uc);
+
+ barrier();
+
+ ctx = (struct _triton_context_t *)this_ctx->tpd;
+
+ spin_lock(&threads_lock);
+ if (ctx->wakeup) {
+ spin_unlock(&threads_lock);
+ ctx->wakeup = 0;
+ free(ctx->uc);
+ ctx->uc = NULL;
+ __sync_sub_and_fetch(&triton_stat.context_sleeping, 1);
+ log_debug2("ctx %p: exit schedule\n", ctx);
+ } else {
+ ctx->thread->ctx = NULL;
+ spin_unlock(&threads_lock);
+ longjmp(jmp_env, 1);
+ }
+}
+
void __export triton_context_wakeup(struct triton_context_t *ud)
{
struct _triton_context_t *ctx = (struct _triton_context_t *)ud->tpd;
@@ -481,26 +523,27 @@ void __export triton_context_wakeup(struct triton_context_t *ud)
return;
}
- pthread_mutex_lock(&ctx->thread->sleep_lock);
+ spin_lock(&threads_lock);
ctx->wakeup = 1;
- pthread_cond_signal(&ctx->thread->sleep_cond);
- pthread_mutex_unlock(&ctx->thread->sleep_lock);
+ if (ctx->thread->ctx != ctx) {
+ list_add_tail(&ctx->entry2, &ctx->thread->wakeup_list);
+ r = ctx->thread->ctx == NULL;
+ }
+ spin_unlock(&threads_lock);
+
+ if (r)
+ triton_thread_wakeup(ctx->thread);
}
int __export triton_context_call(struct triton_context_t *ud, void (*func)(void *), void *arg)
{
- struct _triton_context_t *ctx;
+ struct _triton_context_t *ctx = ud ? (struct _triton_context_t *)ud->tpd : (struct _triton_context_t *)default_ctx.tpd;
struct _triton_ctx_call_t *call = mempool_alloc(call_pool);
int r;
if (!call)
return -1;
- if (ud)
- ctx = (struct _triton_context_t *)ud->tpd;
- else
- ctx = (struct _triton_context_t *)default_ctx.tpd;
-
call->func = func;
call->arg = arg;
diff --git a/accel-pppd/triton/triton.h b/accel-pppd/triton/triton.h
index b66737cf..79bbee1b 100644
--- a/accel-pppd/triton/triton.h
+++ b/accel-pppd/triton/triton.h
@@ -138,6 +138,8 @@ void triton_terminate(void);
#define __export __attribute__((visibility("default")))
#define __unused __attribute__((unused))
+#define barrier() asm volatile ("" ::: "memory")
+
#undef offsetof
#ifdef __compiler_offsetof
#define offsetof(TYPE,MEMBER) __compiler_offsetof(TYPE,MEMBER)
diff --git a/accel-pppd/triton/triton_p.h b/accel-pppd/triton/triton_p.h
index 6c9f884d..a16ef98e 100644
--- a/accel-pppd/triton/triton_p.h
+++ b/accel-pppd/triton/triton_p.h
@@ -2,6 +2,7 @@
#define TRITON_P_H
#include <pthread.h>
+#include <ucontext.h>
#include <sys/epoll.h>
#include "triton.h"
@@ -18,6 +19,7 @@ struct _triton_thread_t
struct _triton_context_t *ctx;
pthread_mutex_t sleep_lock;
pthread_cond_t sleep_cond;
+ struct list_head wakeup_list;
};
struct _triton_context_t
@@ -43,6 +45,8 @@ struct _triton_context_t
int priority;
int refs;
+ ucontext_t *uc;
+
struct triton_context_t *ud;
void *bf_arg;
};
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-15 04:52:30 +0000