* Fences for driver and IPC serialisation, scheduling and synchronisation.
*/
-#include <linux/sync_file.h>
-
+#include "drm-uapi/sync_file.h"
+#include "util/u_debug.h"
#include "util/u_inlines.h"
#include "intel/common/gen_gem.h"
struct pipe_fence_handle {
struct pipe_reference ref;
- struct iris_seqno *seqno[IRIS_BATCH_COUNT];
+
+ struct pipe_context *unflushed_ctx;
+
+ struct iris_fine_fence *fine[IRIS_BATCH_COUNT];
};
static void
{
struct iris_screen *screen = (struct iris_screen *)p_screen;
- for (unsigned i = 0; i < ARRAY_SIZE(fence->seqno); i++)
- iris_seqno_reference(screen, &fence->seqno[i], NULL);
+ for (unsigned i = 0; i < ARRAY_SIZE(fence->fine); i++)
+ iris_fine_fence_reference(screen, &fence->fine[i], NULL);
free(fence);
}
struct iris_screen *screen = (void *) ctx->screen;
struct iris_context *ice = (struct iris_context *)ctx;
+ /* We require DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT (kernel 5.2+) for
+ * deferred flushes. Just ignore the request to defer on older kernels.
+ */
+ if (!(screen->kernel_features & KERNEL_HAS_WAIT_FOR_SUBMIT))
+ flags &= ~PIPE_FLUSH_DEFERRED;
+
+ const bool deferred = flags & PIPE_FLUSH_DEFERRED;
+
if (flags & PIPE_FLUSH_END_OF_FRAME) {
ice->frame++;
}
}
- /* XXX PIPE_FLUSH_DEFERRED */
- for (unsigned i = 0; i < IRIS_BATCH_COUNT; i++)
- iris_batch_flush(&ice->batches[i]);
+ if (!deferred) {
+ for (unsigned i = 0; i < IRIS_BATCH_COUNT; i++)
+ iris_batch_flush(&ice->batches[i]);
+ }
if (!out_fence)
return;
pipe_reference_init(&fence->ref, 1);
+ if (deferred)
+ fence->unflushed_ctx = ctx;
+
for (unsigned b = 0; b < IRIS_BATCH_COUNT; b++) {
struct iris_batch *batch = &ice->batches[b];
- if (iris_seqno_signaled(batch->last_seqno))
- continue;
+ if (deferred && iris_batch_bytes_used(batch) > 0) {
+ struct iris_fine_fence *fine =
+ iris_fine_fence_new(batch, IRIS_FENCE_BOTTOM_OF_PIPE);
+ iris_fine_fence_reference(screen, &fence->fine[b], fine);
+ iris_fine_fence_reference(screen, &fine, NULL);
+ } else {
+ /* This batch has no commands queued up (perhaps we just flushed,
+ * or all the commands are on the other batch). Wait for the last
+ * syncobj on this engine - unless it's already finished by now.
+ */
+ if (iris_fine_fence_signaled(batch->last_fence))
+ continue;
- iris_seqno_reference(screen, &fence->seqno[b], batch->last_seqno);
+ iris_fine_fence_reference(screen, &fence->fine[b], batch->last_fence);
+ }
}
iris_fence_reference(ctx->screen, out_fence, NULL);
{
struct iris_context *ice = (struct iris_context *)ctx;
+ /* Unflushed fences from the same context are no-ops. */
+ if (ctx && ctx == fence->unflushed_ctx)
+ return;
+
+ /* XXX: We can't safely flush the other context, because it might be
+ * bound to another thread, and poking at its internals wouldn't
+ * be safe. In the future we should use MI_SEMAPHORE_WAIT and
+ * block until the other job has been submitted, relying on
+ * kernel timeslicing to preempt us until the other job is
+ * actually flushed and the seqno finally passes.
+ */
+ if (fence->unflushed_ctx) {
+ pipe_debug_message(&ice->dbg, CONFORMANCE, "%s",
+ "glWaitSync on unflushed fence from another context "
+ "is unlikely to work without kernel 5.8+\n");
+ }
+
+ /* Flush any current work in our context as it doesn't need to wait
+ * for this fence. Any future work in our context must wait.
+ */
for (unsigned b = 0; b < IRIS_BATCH_COUNT; b++) {
struct iris_batch *batch = &ice->batches[b];
- for (unsigned i = 0; i < ARRAY_SIZE(fence->seqno); i++) {
- struct iris_seqno *seqno = fence->seqno[i];
+ for (unsigned i = 0; i < ARRAY_SIZE(fence->fine); i++) {
+ struct iris_fine_fence *fine = fence->fine[i];
- if (iris_seqno_signaled(seqno))
+ if (iris_fine_fence_signaled(fine))
continue;
- iris_batch_add_syncobj(batch, seqno->syncobj, I915_EXEC_FENCE_WAIT);
+ iris_batch_flush(batch);
+ iris_batch_add_syncobj(batch, fine->syncobj, I915_EXEC_FENCE_WAIT);
}
}
}
struct pipe_fence_handle *fence,
uint64_t timeout)
{
+ struct iris_context *ice = (struct iris_context *)ctx;
struct iris_screen *screen = (struct iris_screen *)p_screen;
+ /* If we created the fence with PIPE_FLUSH_DEFERRED, we may not have
+ * flushed yet. Check if our syncobj is the current batch's signalling
+ * syncobj - if so, we haven't flushed and need to now.
+ *
+ * The Gallium docs mention that a flush will occur if \p ctx matches
+ * the context the fence was created with. It may be NULL, so we check
+ * that it matches first.
+ */
+ if (ctx && ctx == fence->unflushed_ctx) {
+ for (unsigned i = 0; i < IRIS_BATCH_COUNT; i++) {
+ struct iris_fine_fence *fine = fence->fine[i];
+
+ if (iris_fine_fence_signaled(fine))
+ continue;
+
+ if (fine->syncobj == iris_batch_get_signal_syncobj(&ice->batches[i]))
+ iris_batch_flush(&ice->batches[i]);
+ }
+
+ /* The fence is no longer deferred. */
+ fence->unflushed_ctx = NULL;
+ }
+
unsigned int handle_count = 0;
- uint32_t handles[ARRAY_SIZE(fence->seqno)];
- for (unsigned i = 0; i < ARRAY_SIZE(fence->seqno); i++) {
- struct iris_seqno *seqno = fence->seqno[i];
+ uint32_t handles[ARRAY_SIZE(fence->fine)];
+ for (unsigned i = 0; i < ARRAY_SIZE(fence->fine); i++) {
+ struct iris_fine_fence *fine = fence->fine[i];
- if (iris_seqno_signaled(seqno))
+ if (iris_fine_fence_signaled(fine))
continue;
- handles[handle_count++] = seqno->syncobj->handle;
+ handles[handle_count++] = fine->syncobj->handle;
}
if (handle_count == 0)
.timeout_nsec = rel2abs(timeout),
.flags = DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL
};
+
+ if (fence->unflushed_ctx) {
+ /* This fence had a deferred flush from another context. We can't
+ * safely flush it here, because the context might be bound to a
+ * different thread, and poking at its internals wouldn't be safe.
+ *
+ * Instead, use the WAIT_FOR_SUBMIT flag to block and hope that
+ * another thread submits the work.
+ */
+ args.flags |= DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT;
+ }
+
return gen_ioctl(screen->fd, DRM_IOCTL_SYNCOBJ_WAIT, &args) == 0;
}
struct iris_screen *screen = (struct iris_screen *)p_screen;
int fd = -1;
- for (unsigned i = 0; i < ARRAY_SIZE(fence->seqno); i++) {
- struct iris_seqno *seqno = fence->seqno[i];
+ /* Deferred fences aren't supported. */
+ if (fence->unflushed_ctx)
+ return -1;
+
+ for (unsigned i = 0; i < ARRAY_SIZE(fence->fine); i++) {
+ struct iris_fine_fence *fine = fence->fine[i];
- if (iris_seqno_signaled(seqno))
+ if (iris_fine_fence_signaled(fine))
continue;
struct drm_syncobj_handle args = {
- .handle = seqno->syncobj->handle,
+ .handle = fine->syncobj->handle,
.flags = DRM_SYNCOBJ_HANDLE_TO_FD_FLAGS_EXPORT_SYNC_FILE,
.fd = -1,
};
syncobj->handle = args.handle;
pipe_reference_init(&syncobj->ref, 1);
- struct iris_seqno *seqno = malloc(sizeof(*seqno));
- if (!seqno) {
+ struct iris_fine_fence *fine = calloc(1, sizeof(*fine));
+ if (!fine) {
free(syncobj);
*out = NULL;
return;
static const uint32_t zero = 0;
- /* Fences work in terms of iris_seqno, but we don't actually have a
+ /* Fences work in terms of iris_fine_fence, but we don't actually have a
* seqno for an imported fence. So, create a fake one which always
* returns as 'not signaled' so we fall back to using the sync object.
*/
- seqno->seqno = UINT32_MAX;
- seqno->map = &zero;
- seqno->syncobj = syncobj;
- seqno->flags = IRIS_SEQNO_END;
- pipe_reference_init(&seqno->reference, 1);
+ fine->seqno = UINT32_MAX;
+ fine->map = &zero;
+ fine->syncobj = syncobj;
+ fine->flags = IRIS_FENCE_END;
+ pipe_reference_init(&fine->reference, 1);
struct pipe_fence_handle *fence = calloc(1, sizeof(*fence));
if (!fence) {
- free(seqno);
+ free(fine);
free(syncobj);
*out = NULL;
return;
}
pipe_reference_init(&fence->ref, 1);
- fence->seqno[0] = seqno;
+ fence->fine[0] = fine;
*out = fence;
}