#include "main/mtypes.h"
#include "main/glthread.h"
-#include "main/marshal.h"
-#include "main/marshal_generated.h"
+#include "main/glthread_marshal.h"
+#include "main/hash.h"
+#include "util/u_atomic.h"
#include "util/u_thread.h"
-#ifdef HAVE_PTHREAD
static void
-glthread_allocate_batch(struct gl_context *ctx)
+glthread_unmarshal_batch(void *job, int thread_index)
{
- 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)
-{
- size_t pos = 0;
+ struct glthread_batch *batch = (struct glthread_batch*)job;
+ struct gl_context *ctx = batch->ctx;
+ int pos = 0;
+ int used = batch->used;
+ uint8_t *buffer = batch->buffer;
_glapi_set_dispatch(ctx->CurrentServerDispatch);
- while (pos < batch->used)
- pos += _mesa_unmarshal_dispatch_cmd(ctx, &batch->buffer[pos]);
+ while (pos < used) {
+ const struct marshal_cmd_base *cmd =
+ (const struct marshal_cmd_base *)&buffer[pos];
- assert(pos == batch->used);
+ _mesa_unmarshal_dispatch[cmd->cmd_id](ctx, cmd);
+ pos += cmd->cmd_size;
+ }
- free(batch);
+ assert(pos == used);
+ 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);
-
- u_thread_setname("mesa_glthread");
-
- 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
_mesa_glthread_init(struct gl_context *ctx)
{
- struct glthread_state *glthread = calloc(1, sizeof(*glthread));
+ struct glthread_state *glthread = &ctx->GLThread;
+
+ assert(!glthread->enabled);
- if (!glthread)
+ if (!util_queue_init(&glthread->queue, "gl", MARSHAL_MAX_BATCHES - 2,
+ 1, 0)) {
return;
+ }
+
+ glthread->VAOs = _mesa_NewHashTable();
+ if (!glthread->VAOs) {
+ util_queue_destroy(&glthread->queue);
+ return;
+ }
+ glthread->CurrentVAO = &glthread->DefaultVAO;
ctx->MarshalExec = _mesa_create_marshal_table(ctx);
if (!ctx->MarshalExec) {
- free(glthread);
+ _mesa_DeleteHashTable(glthread->VAOs);
+ util_queue_destroy(&glthread->queue);
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;
- ctx->GLThread = glthread;
+ glthread->enabled = true;
+ glthread->stats.queue = &glthread->queue;
+ ctx->CurrentClientDispatch = ctx->MarshalExec;
- glthread_allocate_batch(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, 0);
+ util_queue_fence_wait(&fence);
+ util_queue_fence_destroy(&fence);
+}
- pthread_create(&glthread->thread, NULL, glthread_worker, ctx);
+static void
+free_vao(GLuint key, void *data, void *userData)
+{
+ free(data);
}
void
_mesa_glthread_destroy(struct gl_context *ctx)
{
- struct glthread_state *glthread = ctx->GLThread;
+ struct glthread_state *glthread = &ctx->GLThread;
- if (!glthread)
+ if (!glthread->enabled)
return;
- _mesa_glthread_flush_batch(ctx);
-
- pthread_mutex_lock(&glthread->mutex);
- glthread->shutdown = true;
- pthread_cond_broadcast(&glthread->new_work);
- pthread_mutex_unlock(&glthread->mutex);
+ _mesa_glthread_finish(ctx);
+ util_queue_destroy(&glthread->queue);
- /* 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);
+ for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++)
+ util_queue_fence_destroy(&glthread->batches[i].fence);
- /* 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);
+ _mesa_HashDeleteAll(glthread->VAOs, free_vao, NULL);
+ _mesa_DeleteHashTable(glthread->VAOs);
- free(glthread);
- ctx->GLThread = NULL;
+ ctx->GLThread.enabled = false;
- _mesa_glthread_restore_dispatch(ctx);
+ _mesa_glthread_restore_dispatch(ctx, "destroy");
}
void
-_mesa_glthread_restore_dispatch(struct gl_context *ctx)
+_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.
if (_glapi_get_dispatch() == ctx->MarshalExec) {
ctx->CurrentClientDispatch = ctx->CurrentServerDispatch;
_glapi_set_dispatch(ctx->CurrentClientDispatch);
+#if 0
+ printf("glthread disabled: %s\n", func);
+#endif
}
}
-static void
-_mesa_glthread_flush_batch_locked(struct gl_context *ctx)
+void
+_mesa_glthread_disable(struct gl_context *ctx, const char *func)
{
- struct glthread_state *glthread = ctx->GLThread;
- struct glthread_batch *batch = glthread->batch;
+ _mesa_glthread_finish_before(ctx, func);
+ _mesa_glthread_restore_dispatch(ctx, func);
+}
- if (!batch->used)
+void
+_mesa_glthread_flush_batch(struct gl_context *ctx)
+{
+ struct glthread_state *glthread = &ctx->GLThread;
+ if (!glthread->enabled)
return;
- /* Immediately reallocate a new batch, since the next marshalled call would
- * just do it.
- */
- glthread_allocate_batch(ctx);
+ struct glthread_batch *next = &glthread->batches[glthread->next];
+ if (!next->used)
+ return;
/* Debug: execute the batch immediately from this thread.
*
* 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;
}
- *glthread->batch_queue_tail = batch;
- glthread->batch_queue_tail = &batch->next;
- pthread_cond_broadcast(&glthread->new_work);
-}
-
-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)
- return;
+ p_atomic_add(&glthread->stats.num_offloaded_items, next->used);
- pthread_mutex_lock(&glthread->mutex);
- _mesa_glthread_flush_batch_locked(ctx);
- pthread_mutex_unlock(&glthread->mutex);
+ util_queue_add_job(&glthread->queue, next, &next->fence,
+ glthread_unmarshal_batch, NULL, 0);
+ glthread->last = glthread->next;
+ glthread->next = (glthread->next + 1) % MARSHAL_MAX_BATCHES;
}
/**
void
_mesa_glthread_finish(struct gl_context *ctx)
{
- struct glthread_state *glthread = ctx->GLThread;
-
- if (!glthread)
+ struct glthread_state *glthread = &ctx->GLThread;
+ if (!glthread->enabled)
return;
/* If this is called from the worker thread, then we've hit a path that
* 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;
- pthread_mutex_lock(&glthread->mutex);
-
- if (!(glthread->batch_queue || glthread->busy)) {
- if (glthread->batch && glthread->batch->used) {
- struct _glapi_table *dispatch = _glapi_get_dispatch();
- glthread_unmarshal_batch(ctx, glthread->batch);
- _glapi_set_dispatch(dispatch);
- glthread_allocate_batch(ctx);
- }
- } else {
- _mesa_glthread_flush_batch_locked(ctx);
- while (glthread->batch_queue || glthread->busy)
- pthread_cond_wait(&glthread->work_done, &glthread->mutex);
+ struct glthread_batch *last = &glthread->batches[glthread->last];
+ struct glthread_batch *next = &glthread->batches[glthread->next];
+ bool synced = false;
+
+ if (!util_queue_fence_is_signalled(&last->fence)) {
+ util_queue_fence_wait(&last->fence);
+ synced = true;
+ }
+
+ if (next->used) {
+ p_atomic_add(&glthread->stats.num_direct_items, next->used);
+
+ /* 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);
+
+ /* 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;
}
- pthread_mutex_unlock(&glthread->mutex);
+ if (synced)
+ p_atomic_inc(&glthread->stats.num_syncs);
}
-#endif
+void
+_mesa_glthread_finish_before(struct gl_context *ctx, const char *func)
+{
+ _mesa_glthread_finish(ctx);
+
+ /* Uncomment this if you want to know where glthread syncs. */
+ /*printf("fallback to sync: %s\n", func);*/
+}