#include "main/glthread.h"
#include "main/marshal.h"
#include "main/marshal_generated.h"
+#include "util/u_atomic.h"
+#include "util/u_thread.h"
-#ifdef HAVE_PTHREAD
static void
-glthread_allocate_batch(struct gl_context *ctx)
-{
- struct glthread_state *glthread = ctx->GLThread;
-
- /* TODO: handle memory allocation failure. */
- glthread->batch = malloc(sizeof(*glthread->batch));
- if (!glthread->batch)
- return;
- memset(glthread->batch, 0, offsetof(struct glthread_batch, buffer));
-}
-
-static void
-glthread_unmarshal_batch(struct gl_context *ctx, struct glthread_batch *batch)
+glthread_unmarshal_batch(void *job, int thread_index)
{
+ struct glthread_batch *batch = (struct glthread_batch*)job;
+ struct gl_context *ctx = batch->ctx;
size_t pos = 0;
_glapi_set_dispatch(ctx->CurrentServerDispatch);
pos += _mesa_unmarshal_dispatch_cmd(ctx, &batch->buffer[pos]);
assert(pos == batch->used);
-
- free(batch);
+ batch->used = 0;
}
-static void *
-glthread_worker(void *data)
+static void
+glthread_thread_initialization(void *job, int thread_index)
{
- struct gl_context *ctx = data;
- struct glthread_state *glthread = ctx->GLThread;
+ struct gl_context *ctx = (struct gl_context*)job;
- ctx->Driver.SetBackgroundContext(ctx);
+ ctx->Driver.SetBackgroundContext(ctx, &ctx->GLThread->stats);
_glapi_set_context(ctx);
-
- pthread_mutex_lock(&glthread->mutex);
-
- while (true) {
- struct glthread_batch *batch;
-
- /* Block (dropping the lock) until new work arrives for us. */
- while (!glthread->batch_queue && !glthread->shutdown) {
- pthread_cond_broadcast(&glthread->work_done);
- pthread_cond_wait(&glthread->new_work, &glthread->mutex);
- }
-
- batch = glthread->batch_queue;
-
- if (glthread->shutdown && !batch) {
- pthread_cond_broadcast(&glthread->work_done);
- pthread_mutex_unlock(&glthread->mutex);
- return NULL;
- }
- glthread->batch_queue = batch->next;
- if (glthread->batch_queue_tail == &batch->next)
- glthread->batch_queue_tail = &glthread->batch_queue;
-
- glthread->busy = true;
- pthread_mutex_unlock(&glthread->mutex);
-
- glthread_unmarshal_batch(ctx, batch);
-
- pthread_mutex_lock(&glthread->mutex);
- glthread->busy = false;
- }
-
- /* UNREACHED */
- return NULL;
}
void
if (!glthread)
return;
+ if (!util_queue_init(&glthread->queue, "gl", MARSHAL_MAX_BATCHES - 2,
+ 1, 0)) {
+ free(glthread);
+ return;
+ }
+
ctx->MarshalExec = _mesa_create_marshal_table(ctx);
if (!ctx->MarshalExec) {
+ util_queue_destroy(&glthread->queue);
free(glthread);
return;
}
- ctx->CurrentClientDispatch = ctx->MarshalExec;
-
- pthread_mutex_init(&glthread->mutex, NULL);
- pthread_cond_init(&glthread->new_work, NULL);
- pthread_cond_init(&glthread->work_done, NULL);
+ for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++) {
+ glthread->batches[i].ctx = ctx;
+ util_queue_fence_init(&glthread->batches[i].fence);
+ }
- glthread->batch_queue_tail = &glthread->batch_queue;
+ glthread->stats.queue = &glthread->queue;
+ ctx->CurrentClientDispatch = ctx->MarshalExec;
ctx->GLThread = glthread;
- glthread_allocate_batch(ctx);
-
- pthread_create(&glthread->thread, NULL, glthread_worker, ctx);
+ /* Execute the thread initialization function in the thread. */
+ struct util_queue_fence fence;
+ util_queue_fence_init(&fence);
+ util_queue_add_job(&glthread->queue, ctx, &fence,
+ glthread_thread_initialization, NULL);
+ util_queue_fence_wait(&fence);
+ util_queue_fence_destroy(&fence);
}
void
if (!glthread)
return;
- _mesa_glthread_flush_batch(ctx);
+ _mesa_glthread_finish(ctx);
+ util_queue_destroy(&glthread->queue);
- pthread_mutex_lock(&glthread->mutex);
- glthread->shutdown = true;
- pthread_cond_broadcast(&glthread->new_work);
- pthread_mutex_unlock(&glthread->mutex);
-
- /* Since this waits for the thread to exit, it means that all queued work
- * will have been completed.
- */
- pthread_join(glthread->thread, NULL);
-
- pthread_cond_destroy(&glthread->new_work);
- pthread_cond_destroy(&glthread->work_done);
- pthread_mutex_destroy(&glthread->mutex);
-
- /* Due to the join above, there should be one empty batch allocated at this
- * point, and no batches queued.
- */
- assert(!glthread->batch->used);
- assert(!glthread->batch->next);
- free(glthread->batch);
- assert(!glthread->batch_queue);
+ for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++)
+ util_queue_fence_destroy(&glthread->batches[i].fence);
free(glthread);
ctx->GLThread = NULL;
- /* Remove ourselves from the dispatch table. */
- ctx->CurrentClientDispatch = ctx->CurrentServerDispatch;
- _glapi_set_dispatch(ctx->CurrentClientDispatch);
+ _mesa_glthread_restore_dispatch(ctx, "destroy");
+}
+
+void
+_mesa_glthread_restore_dispatch(struct gl_context *ctx, const char *func)
+{
+ /* Remove ourselves from the dispatch table except if another ctx/thread
+ * already installed a new dispatch table.
+ *
+ * Typically glxMakeCurrent will bind a new context (install new table) then
+ * old context might be deleted.
+ */
+ if (_glapi_get_dispatch() == ctx->MarshalExec) {
+ ctx->CurrentClientDispatch = ctx->CurrentServerDispatch;
+ _glapi_set_dispatch(ctx->CurrentClientDispatch);
+#if 0
+ printf("glthread disabled: %s\n", func);
+#endif
+ }
}
void
_mesa_glthread_flush_batch(struct gl_context *ctx)
{
struct glthread_state *glthread = ctx->GLThread;
- struct glthread_batch *batch;
-
if (!glthread)
return;
- batch = glthread->batch;
- if (!batch->used)
+ struct glthread_batch *next = &glthread->batches[glthread->next];
+ if (!next->used)
return;
- /* Immediately reallocate a new batch, since the next marshalled call would
- * just do it.
- */
- glthread_allocate_batch(ctx);
-
/* Debug: execute the batch immediately from this thread.
*
* Note that glthread_unmarshal_batch() changes the dispatch table so we'll
* need to restore it when it returns.
*/
if (false) {
- glthread_unmarshal_batch(ctx, batch);
+ glthread_unmarshal_batch(next, 0);
_glapi_set_dispatch(ctx->CurrentClientDispatch);
return;
}
- pthread_mutex_lock(&glthread->mutex);
- *glthread->batch_queue_tail = batch;
- glthread->batch_queue_tail = &batch->next;
- pthread_cond_broadcast(&glthread->new_work);
- pthread_mutex_unlock(&glthread->mutex);
+ p_atomic_add(&glthread->stats.num_offloaded_items, next->used);
+
+ util_queue_add_job(&glthread->queue, next, &next->fence,
+ glthread_unmarshal_batch, NULL);
+ glthread->last = glthread->next;
+ glthread->next = (glthread->next + 1) % MARSHAL_MAX_BATCHES;
}
/**
_mesa_glthread_finish(struct gl_context *ctx)
{
struct glthread_state *glthread = ctx->GLThread;
-
if (!glthread)
return;
* dri interface entrypoints), in which case we don't need to actually
* synchronize against ourself.
*/
- if (pthread_self() == glthread->thread)
+ if (u_thread_is_self(glthread->queue.threads[0]))
return;
- _mesa_glthread_flush_batch(ctx);
+ struct glthread_batch *last = &glthread->batches[glthread->last];
+ struct glthread_batch *next = &glthread->batches[glthread->next];
+ bool synced = false;
- pthread_mutex_lock(&glthread->mutex);
+ if (!util_queue_fence_is_signalled(&last->fence)) {
+ util_queue_fence_wait(&last->fence);
+ synced = true;
+ }
- while (glthread->batch_queue || glthread->busy)
- pthread_cond_wait(&glthread->work_done, &glthread->mutex);
+ if (next->used) {
+ p_atomic_add(&glthread->stats.num_direct_items, next->used);
- pthread_mutex_unlock(&glthread->mutex);
-}
+ /* Since glthread_unmarshal_batch changes the dispatch to direct,
+ * restore it after it's done.
+ */
+ struct _glapi_table *dispatch = _glapi_get_dispatch();
+ glthread_unmarshal_batch(next, 0);
+ _glapi_set_dispatch(dispatch);
-#endif
+ /* It's not a sync because we don't enqueue partial batches, but
+ * it would be a sync if we did. So count it anyway.
+ */
+ synced = true;
+ }
+
+ if (synced)
+ p_atomic_inc(&glthread->stats.num_syncs);
+}