i915/gem_exec_fence: Teach invalid-wait about invalid future fences

When we allow a wait on a future future fence, it must autoexpire if the
fence is never signaled by userspace. Also put future fences to work, as
the intention is to use them, along with WAIT_SUBMIT and semaphores, for
userspace to perform its own fine-grained scheduling. Or simply run
concurrent clients without having to flush batches between context
switches.

v2: Verify deadlock detection

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

1 files changed, 1098 insertions, 12 deletions

diff --git a/tests/i915/gem_exec_fence.c b/tests/i915/gem_exec_fence.c
index af08c385..e7e253c1 100644
--- a/tests/i915/gem_exec_fence.c
+++ b/tests/i915/gem_exec_fence.c
@@ -24,6 +24,7 @@
 #include <sys/ioctl.h>
 #include <sys/poll.h>
 #include <sys/signal.h>
+#include <sys/socket.h>
 
 #include "i915/gem.h"
 #include "i915/gem_ring.h"
@@ -31,6 +32,7 @@
 #include "igt_syncobj.h"
 #include "igt_sysfs.h"
 #include "igt_vgem.h"
+#include "ioctl_wrappers.h"
 #include "sw_sync.h"
 
 IGT_TEST_DESCRIPTION("Check that execbuf waits for explicit fences");
@@ -457,7 +459,7 @@ static void test_submit_fence(int i915, unsigned int engine)
 			.offset = offset,
 			.flags = EXEC_OBJECT_PINNED,
 		};
-		struct drm_i915_gem_execbuffer2 execbuf  = {
+		struct drm_i915_gem_execbuffer2 execbuf = {
 			.buffers_ptr = to_user_pointer(&obj),
 			.buffer_count = 1,
 		};
@@ -542,7 +544,7 @@ static void test_submitN(int i915, unsigned int engine, int count)
 		.offset = offset,
 		.flags = EXEC_OBJECT_PINNED,
 	};
-	struct drm_i915_gem_execbuffer2 execbuf  = {
+	struct drm_i915_gem_execbuffer2 execbuf = {
 		.buffers_ptr = to_user_pointer(&obj),
 		.buffer_count = 1,
 		.flags = engine | I915_EXEC_FENCE_OUT,
@@ -607,7 +609,7 @@ static void test_parallel(int i915, const struct intel_execution_engine2 *e)
 	int fence;
 	int x = 0;
 
-	fence = igt_cork_plug(&cork, i915),
+	fence = igt_cork_plug(&cork, i915);
 	spin = igt_spin_new(i915,
 			    .engine = e->flags,
 			    .fence = fence,
@@ -736,12 +738,12 @@ static void test_concurrent(int i915, const struct intel_execution_engine2 *e)
 	 * the master is ready.
 	 */
 
-	fence = igt_cork_plug(&cork, i915),
-	      spin = igt_spin_new(i915,
-				  .engine = e->flags,
-				  .fence = fence,
-				  .flags = (IGT_SPIN_FENCE_OUT |
-					    IGT_SPIN_FENCE_IN));
+	fence = igt_cork_plug(&cork, i915);
+	spin = igt_spin_new(i915,
+			    .engine = e->flags,
+			    .fence = fence,
+			    .flags = (IGT_SPIN_FENCE_OUT |
+				      IGT_SPIN_FENCE_IN));
 	close(fence);
 
 	i = 0;
@@ -1170,11 +1172,12 @@ static void test_syncobj_invalid_wait(int fd)
 	struct drm_i915_gem_exec_fence fence = {
 		.handle = syncobj_create(fd, 0),
 	};
+	int out;
 
 	memset(&execbuf, 0, sizeof(execbuf));
 	execbuf.buffers_ptr = to_user_pointer(&obj);
 	execbuf.buffer_count = 1;
-	execbuf.flags = I915_EXEC_FENCE_ARRAY;
+	execbuf.flags = I915_EXEC_FENCE_ARRAY | I915_EXEC_FENCE_OUT;
 	execbuf.cliprects_ptr = to_user_pointer(&fence);
 	execbuf.num_cliprects = 1;
 
@@ -1182,14 +1185,59 @@ static void test_syncobj_invalid_wait(int fd)
 	obj.handle = gem_create(fd, 4096);
 	gem_write(fd, obj.handle, 0, &bbe, sizeof(bbe));
 
-	/* waiting before the fence is set is invalid */
+	/* waiting before the fence is set is^W may be invalid */
 	fence.flags = I915_EXEC_FENCE_WAIT;
-	igt_assert_eq(__gem_execbuf(fd, &execbuf), -EINVAL);
+	if (__gem_execbuf_wr(fd, &execbuf)) {
+		igt_assert_eq(__gem_execbuf(fd, &execbuf), -EINVAL);
+		return;
+	}
+
+	/* If we do allow the wait on a future fence, it should autoexpire */
+	gem_sync(fd, obj.handle);
+	out = execbuf.rsvd2 >> 32;
+	igt_assert_eq(sync_fence_status(out), -ETIMEDOUT);
+	close(out);
 
 	gem_close(fd, obj.handle);
 	syncobj_destroy(fd, fence.handle);
 }
 
+static void test_syncobj_incomplete_wait_submit(int i915)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = batch_create(i915),
+	};
+	struct drm_i915_gem_exec_fence fence = {
+		.handle = syncobj_create(i915, 0),
+		.flags = I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_WAIT_SUBMIT,
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+
+		.flags = I915_EXEC_FENCE_ARRAY | I915_EXEC_FENCE_OUT,
+	};
+	int out;
+
+	/* waiting before the fence is set is^W may be invalid */
+	if (__gem_execbuf_wr(i915, &execbuf)) {
+		igt_assert_eq(__gem_execbuf(i915, &execbuf), -EINVAL);
+		return;
+	}
+
+	/* If we do allow the wait on a future fence, it should autoexpire */
+	gem_sync(i915, obj.handle);
+	out = execbuf.rsvd2 >> 32;
+	igt_assert_eq(sync_fence_status(out), -ETIMEDOUT);
+	close(out);
+
+	gem_close(i915, obj.handle);
+	syncobj_destroy(i915, fence.handle);
+}
+
 static void test_syncobj_invalid_flags(int fd)
 {
 	const uint32_t bbe = MI_BATCH_BUFFER_END;
@@ -2937,6 +2985,1038 @@ static void test_syncobj_backward_timeline_chain_engines(int fd, struct intel_en
 	teardown_timeline_chain_engines(&ctx);
 }
 
+static uint32_t future_batch(int i915, uint32_t offset)
+{
+	uint32_t handle = gem_create(i915, 4096);
+	const int gen = intel_gen(intel_get_drm_devid(i915));
+	uint32_t cs[16];
+	int i = 0;
+
+	cs[i] = MI_STORE_DWORD_IMM | (gen < 6 ? 1 << 22 : 0);
+	if (gen >= 8) {
+		cs[++i] = offset + 4000;
+		cs[++i] = 0;
+	} else if (gen >= 4) {
+		cs[++i] = 0;
+		cs[++i] = offset + 4000;
+	} else {
+		cs[i]--;
+		cs[++i] = offset + 4000;
+	}
+	cs[++i] = 1;
+	cs[i + 1] = MI_BATCH_BUFFER_END;
+	gem_write(i915, handle, 0, cs, sizeof(cs));
+
+	cs[i] = 2;
+	gem_write(i915, handle, 64, cs, sizeof(cs));
+
+	return handle;
+}
+
+static void test_syncobj_future(int i915, unsigned int engine)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence = {};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+	};
+	const struct intel_execution_engine2 *e;
+
+	/*
+	 * Client A is waiting on a future fence from B. So even though its
+	 * execbuf is called first, we need to hold it in a queue waiting on
+	 * B.
+	 */
+	igt_require(gem_scheduler_enabled(i915));
+
+	__for_each_physical_engine(i915, e) {
+		uint32_t result;
+
+		if (!gem_class_can_store_dword(i915, e->class))
+			continue;
+
+		igt_debug("waiting on future %s\n", e->name);
+		fence.handle = syncobj_create(i915, 0);
+
+		fence.flags = I915_EXEC_FENCE_WAIT;
+		execbuf.batch_start_offset = 0;
+		execbuf.flags = engine | I915_EXEC_FENCE_ARRAY;
+		execbuf.rsvd1 = 0;
+		gem_execbuf(i915, &execbuf); /* writes 1 */
+
+		fence.flags = I915_EXEC_FENCE_SIGNAL;
+		execbuf.batch_start_offset = 64;
+		execbuf.flags = e->flags | I915_EXEC_FENCE_ARRAY;
+		execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+		gem_execbuf(i915, &execbuf); /* writes 2 */
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		syncobj_destroy(i915, fence.handle);
+		gem_sync(i915, obj.handle); /* write hazard lies */
+		gem_read(i915, obj.handle, 4000, &result, sizeof(result));
+		igt_assert_eq(result, 1);
+	}
+
+	gem_close(i915, obj.handle);
+}
+
+static uint32_t future_submit_batch(int i915, uint32_t offset)
+{
+	uint32_t handle = gem_create(i915, 4096);
+	const int gen = intel_gen(intel_get_drm_devid(i915));
+	uint32_t cs[16];
+	int i = 0;
+
+	cs[i] = MI_STORE_DWORD_IMM | (gen < 6 ? 1 << 22 : 0);
+	if (gen >= 8) {
+		cs[++i] = offset + 4000;
+		cs[++i] = 0;
+	} else if (gen >= 4) {
+		cs[++i] = 0;
+		cs[++i] = offset + 4000;
+	} else {
+		cs[i]--;
+		cs[++i] = offset + 4000;
+	}
+	cs[++i] = 1;
+	cs[i + 1] = MI_BATCH_BUFFER_END;
+	igt_assert(i + 1 < ARRAY_SIZE(cs));
+	gem_write(i915, handle, 0, cs, sizeof(cs));
+
+	i = 0;
+	cs[i++] =
+		MI_SEMAPHORE_WAIT |
+		MI_SEMAPHORE_POLL |
+		MI_SEMAPHORE_SAD_EQ_SDD |
+		(4 - 2);
+	cs[i++] = 1;
+	cs[i++] = offset + 4000;
+	cs[i++] = 0;
+
+	cs[i] = MI_STORE_DWORD_IMM | (gen < 6 ? 1 << 22 : 0);
+	if (gen >= 8) {
+		cs[++i] = offset + 4000;
+		cs[++i] = 0;
+	} else if (gen >= 4) {
+		cs[++i] = 0;
+		cs[++i] = offset + 4000;
+	} else {
+		cs[i]--;
+		cs[++i] = offset + 4000;
+	}
+	cs[++i] = 2;
+	cs[++i] = MI_BATCH_BUFFER_END;
+	igt_assert(i < ARRAY_SIZE(cs));
+
+	gem_write(i915, handle, 64, cs, sizeof(cs));
+
+	return handle;
+}
+
+static void test_syncobj_future_submit(int i915, unsigned int engine)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_submit_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	const struct intel_execution_engine2 *e;
+
+	/*
+	 * Here we submit client A waiting on client B, but internally client
+	 * B has a semaphore that waits on client A. This relies on timeslicing
+	 * to reorder A before B, even though userspace has asked to submit
+	 * A & B simultaneously (and due to the sequence we will submit B
+	 * then A).
+	 */
+	igt_require(gem_scheduler_has_timeslicing(i915));
+
+	__for_each_physical_engine(i915, e) {
+		struct drm_i915_gem_exec_fence fence = { };
+		struct drm_i915_gem_execbuffer2 execbuf = {
+			.buffers_ptr = to_user_pointer(&obj),
+			.buffer_count = 1,
+			.cliprects_ptr = to_user_pointer(&fence),
+			.num_cliprects = 1,
+			.flags = engine | I915_EXEC_FENCE_ARRAY,
+		};
+		uint32_t result;
+		int out;
+
+		if (!gem_class_can_store_dword(i915, e->class))
+			continue;
+
+		igt_debug("waiting on future %s\n", e->name);
+		fence.handle = syncobj_create(i915, 0);
+
+		execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+		fence.flags = I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_WAIT_SUBMIT;
+		execbuf.batch_start_offset = 0;
+		execbuf.flags |= I915_EXEC_FENCE_OUT;
+		igt_require(__gem_execbuf_wr(i915, &execbuf) == 0); /* writes 1 */
+		execbuf.flags &= ~I915_EXEC_FENCE_OUT;
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+		fence.flags = I915_EXEC_FENCE_SIGNAL;
+		execbuf.batch_start_offset = 64;
+		execbuf.flags &= ~I915_EXEC_RING_MASK;
+		execbuf.flags |= e->flags;
+		gem_execbuf(i915, &execbuf); /* writes 2 */
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		syncobj_destroy(i915, fence.handle);
+		gem_sync(i915, obj.handle); /* write hazard lies */
+		gem_read(i915, obj.handle, 4000, &result, sizeof(result));
+		igt_assert_eq(result, 2);
+
+		/* check we didn't autotimeout */
+		out = execbuf.rsvd2 >> 32;
+		igt_assert_eq(sync_fence_status(out), 1);
+		close(out);
+	}
+
+	gem_close(i915, obj.handle);
+}
+
+static void test_syncobj_future_past(int i915, unsigned int engine, int flags)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence = {
+		.handle = syncobj_create(i915, 0),
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+		.flags = engine | I915_EXEC_FENCE_ARRAY,
+	};
+	uint32_t result;
+	int out;
+
+	fence.flags = I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_SIGNAL | flags;
+	execbuf.batch_start_offset = 0;
+	execbuf.flags |= I915_EXEC_FENCE_OUT;
+	igt_require(__gem_execbuf_wr(i915, &execbuf) == 0); /* writes 1 */
+	execbuf.flags &= ~I915_EXEC_FENCE_OUT;
+
+	gem_sync(i915, obj.handle); /* write hazard lies */
+	gem_read(i915, obj.handle, 4000, &result, sizeof(result));
+	igt_assert_eq(result, 1);
+
+	/* check we didn't autotimeout */
+	out = execbuf.rsvd2 >> 32;
+	igt_assert_eq(sync_fence_status(out), 1);
+	close(out);
+
+	gem_close(i915, obj.handle);
+	syncobj_destroy(i915, fence.handle);
+}
+
+static void test_syncobj_future_self(int i915, unsigned int engine, int flags)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence = {
+		.handle = syncobj_create(i915, 0),
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+		.flags = engine | I915_EXEC_FENCE_ARRAY,
+	};
+	int out;
+
+	fence.flags = I915_EXEC_FENCE_WAIT | flags;
+	execbuf.batch_start_offset = 0;
+	execbuf.flags |= I915_EXEC_FENCE_OUT;
+	igt_require(__gem_execbuf_wr(i915, &execbuf) == 0); /* writes 1 */
+	execbuf.flags &= ~I915_EXEC_FENCE_OUT;
+
+	fence.flags = I915_EXEC_FENCE_SIGNAL;
+	execbuf.batch_start_offset = 64;
+	gem_execbuf(i915, &execbuf); /* writes 2 */
+
+	gem_sync(i915, obj.handle); /* write hazard lies */
+
+	gem_close(i915, obj.handle);
+	syncobj_destroy(i915, fence.handle);
+
+	out = execbuf.rsvd2 >> 32;
+	igt_assert_eq(sync_fence_status(out), -EDEADLK);
+	close(out);
+}
+
+static void test_syncobj_future_pair(int i915, unsigned int engine)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = submitN_batches(i915, 24 << 20, 2),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence[2] = {
+		{ .handle = syncobj_create(i915, 0) },
+		{ .handle = syncobj_create(i915, 0) }
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(fence),
+		.num_cliprects = 2,
+	};
+	const struct intel_execution_engine2 *e;
+
+	__for_each_physical_engine(i915, e) {
+		int out = 0;
+
+		gem_write(i915, obj.handle, 0, &out, sizeof(out));
+		fence[0].handle = syncobj_create(i915, 0);
+		fence[1].handle = syncobj_create(i915, 0);
+
+		fence[0].flags = I915_EXEC_FENCE_SIGNAL;
+		fence[1].flags =
+			I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_WAIT_SUBMIT;
+		execbuf.batch_start_offset = 1024;
+		execbuf.flags =
+			engine | I915_EXEC_FENCE_ARRAY | I915_EXEC_FENCE_OUT;
+		execbuf.rsvd1 = gem_context_create(i915);
+		igt_require(__gem_execbuf_wr(i915, &execbuf) == 0);
+		gem_context_destroy(i915, execbuf.rsvd1);
+		execbuf.rsvd2 >>= 32;
+
+		fence[0].flags =
+			I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_WAIT_SUBMIT;
+		fence[1].flags = I915_EXEC_FENCE_SIGNAL;
+		execbuf.batch_start_offset = 2048;
+		execbuf.flags =
+			e->flags | I915_EXEC_FENCE_ARRAY | I915_EXEC_FENCE_OUT;
+		execbuf.rsvd1 = gem_context_create(i915);
+		gem_execbuf_wr(i915, &execbuf);
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		syncobj_destroy(i915, fence[0].handle);
+		syncobj_destroy(i915, fence[1].handle);
+
+		gem_sync(i915, obj.handle); /* write hazard lies */
+
+		out = execbuf.rsvd2;
+		igt_assert_eq(sync_fence_status(out), 1);
+		close(out);
+
+		out = execbuf.rsvd2 >> 32;
+		igt_assert_eq(sync_fence_status(out), 1);
+		close(out);
+
+		gem_read(i915, obj.handle, 0, &out, sizeof(out));
+		igt_assert_eq(out, 16);
+	}
+
+	gem_close(i915, obj.handle);
+}
+
+static void test_syncobj_future_group(int i915, unsigned int engine, int count)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = submitN_batches(i915, 24 << 20, count),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence[count];
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(fence),
+		.num_cliprects = count,
+		.flags = engine | I915_EXEC_FENCE_ARRAY | I915_EXEC_FENCE_OUT,
+	};
+	int out[count];
+	uint32_t result;
+
+	for (int i = 0; i < count; i++) {
+		fence[i].handle = syncobj_create(i915, 0);
+		fence[i].flags =
+			I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_WAIT_SUBMIT;
+	}
+
+	for (int i = 0; i < count; i++) {
+		fence[i].flags = I915_EXEC_FENCE_SIGNAL;
+
+		execbuf.batch_start_offset = 1024 * (i + 1);
+		execbuf.rsvd1 = gem_context_create(i915);
+		gem_execbuf_wr(i915, &execbuf);
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		out[i] = execbuf.rsvd2 >> 32;
+		fence[i].flags =
+			I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_WAIT_SUBMIT;
+	}
+	gem_sync(i915, obj.handle); /* write hazard lies */
+
+	/* As both batches were waiting for the other to start -- deadlock? */
+	for (int i = 0; i < count; i++)  {
+		syncobj_destroy(i915, fence[i].handle);
+		igt_assert_eq(sync_fence_status(out[i]), 1);
+		close(out[i]);
+	}
+
+	/* Nevertheless, we ignored^Wresolved the deadlock and let them run */
+	gem_read(i915, obj.handle, 0, &result, sizeof(result));
+	igt_assert_eq(result, 8 * count);
+	gem_close(i915, obj.handle);
+}
+
+static void
+test_syncobj_future_deadlock(int i915, unsigned int engine, int flags)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence = {};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+	};
+	const struct intel_execution_engine2 *e;
+
+	__for_each_physical_engine(i915, e) {
+		int out;
+
+		if (!gem_class_can_store_dword(i915, e->class))
+			continue;
+
+		fence.handle = syncobj_create(i915, 0),
+
+		fence.flags = I915_EXEC_FENCE_WAIT | flags;
+		execbuf.batch_start_offset = 0;
+		execbuf.flags = engine | I915_EXEC_FENCE_ARRAY,
+		execbuf.flags |= I915_EXEC_FENCE_OUT;
+		execbuf.rsvd1 = 0;
+		gem_execbuf_wr(i915, &execbuf); /* writes 1 */
+
+		fence.flags = I915_EXEC_FENCE_SIGNAL;
+		execbuf.batch_start_offset = 64;
+		execbuf.flags = e->flags | I915_EXEC_FENCE_ARRAY,
+		execbuf.flags |= I915_EXEC_FENCE_OUT | I915_EXEC_FENCE_IN;
+		execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+		execbuf.rsvd2 >>= 32;
+		gem_execbuf_wr(i915, &execbuf); /* writes 2 */
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		syncobj_destroy(i915, fence.handle);
+		gem_sync(i915, obj.handle);
+
+		/* How should this deadlock be resolved? */
+		out = execbuf.rsvd2;
+		igt_assert_eq(sync_fence_status(out), -EDEADLK);
+		close(out);
+
+		out = execbuf.rsvd2 >> 32;
+		igt_assert_eq(sync_fence_status(out), -EDEADLK);
+		close(out);
+	}
+
+	gem_close(i915, obj.handle);
+}
+
+static bool fd_wait(int dmabuf, int timeout)
+{
+	struct pollfd pfd = { .fd = dmabuf, .events = POLLOUT };
+	return poll(&pfd, 1, timeout) == 1;
+}
+
+static void
+test_syncobj_future_dmabuf_deadlock(int i915, unsigned int engine)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence = {};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+	};
+	const struct intel_execution_engine2 *e;
+
+	__for_each_physical_engine(i915, e) {
+		int dmabuf;
+
+		if (!gem_class_can_store_dword(i915, e->class))
+			continue;
+
+		fence.handle = syncobj_create(i915, 0),
+
+		fence.flags = I915_EXEC_FENCE_WAIT;
+		execbuf.batch_start_offset = 0;
+		execbuf.flags = engine | I915_EXEC_FENCE_ARRAY,
+		execbuf.flags |= I915_EXEC_FENCE_OUT;
+		execbuf.rsvd1 = 0;
+		gem_execbuf_wr(i915, &execbuf); /* writes 1 */
+
+		dmabuf = prime_handle_to_fd(i915, obj.handle);
+		igt_assert_eq(fd_wait(dmabuf, 50), 0);
+		close(dmabuf);
+
+		fence.flags = I915_EXEC_FENCE_SIGNAL;
+		execbuf.batch_start_offset = 64;
+		execbuf.flags = e->flags | I915_EXEC_FENCE_ARRAY,
+		execbuf.flags |= I915_EXEC_FENCE_IN;
+		execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+		execbuf.rsvd2 >>= 32;
+		gem_execbuf(i915, &execbuf); /* writes 2 */
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		dmabuf = prime_handle_to_fd(i915, obj.handle);
+		igt_assert_eq(fd_wait(dmabuf, -1), 1);
+		close(dmabuf);
+
+		syncobj_destroy(i915, fence.handle);
+		close(execbuf.rsvd2);
+	}
+
+	gem_close(i915, obj.handle);
+}
+
+static void
+test_syncobj_future_export_deadlock(int i915, unsigned int engine)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence = {};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+	};
+	const struct intel_execution_engine2 *e;
+
+	__for_each_physical_engine(i915, e) {
+		int fd;
+
+		if (!gem_class_can_store_dword(i915, e->class))
+			continue;
+
+		fence.handle = syncobj_create(i915, 0),
+
+		fence.flags = I915_EXEC_FENCE_WAIT;
+		execbuf.batch_start_offset = 0;
+		execbuf.flags = engine | I915_EXEC_FENCE_ARRAY,
+		execbuf.flags |= I915_EXEC_FENCE_OUT;
+		execbuf.rsvd1 = 0;
+		gem_execbuf_wr(i915, &execbuf); /* writes 1 */
+
+		fd = syncobj_handle_to_fd(i915, fence.handle, 0);
+		igt_assert_eq(fd_wait(fd, 50), 0);
+		close(fd);
+
+		fence.flags = I915_EXEC_FENCE_SIGNAL;
+		execbuf.batch_start_offset = 64;
+		execbuf.flags = e->flags | I915_EXEC_FENCE_ARRAY,
+		execbuf.flags |= I915_EXEC_FENCE_IN;
+		execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+		execbuf.rsvd2 >>= 32;
+		gem_execbuf(i915, &execbuf); /* writes 2 */
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		fd = syncobj_handle_to_fd(i915, fence.handle, 0);
+		igt_assert_eq(fd_wait(fd, -1), 1);
+		close(fd);
+
+		syncobj_destroy(i915, fence.handle);
+		close(execbuf.rsvd2);
+	}
+
+	gem_close(i915, obj.handle);
+}
+
+static void
+test_syncobj_future_import(int i915, unsigned int engine, bool signal)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = batch_create(i915),
+	};
+	struct drm_i915_gem_exec_fence fence = {
+		.handle = syncobj_create(i915, 0),
+		.flags = I915_EXEC_FENCE_WAIT,
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+
+		.flags = engine | I915_EXEC_FENCE_ARRAY | I915_EXEC_FENCE_OUT,
+	};
+	int timeline = sw_sync_timeline_create();
+	int syncpt, out;
+
+	gem_execbuf_wr(i915, &execbuf);
+	out = execbuf.rsvd2 >> 32;
+
+	syncpt = sw_sync_timeline_create_fence(timeline, 1);
+	syncobj_import_sync_file(i915, fence.handle, syncpt);
+	close(syncpt);
+
+	igt_assert_eq(sync_fence_wait(out, 50), -ETIME);
+	igt_assert_eq(sync_fence_status(out), 0);
+
+	if (signal)
+		sw_sync_timeline_inc(timeline, 1);
+
+	igt_assert_eq(sync_fence_wait(out, -1), 0);
+	igt_assert_eq(sync_fence_status(out), signal ? 1 : -ETIMEDOUT);
+
+	close(out);
+	close(timeline);
+	gem_close(i915, obj.handle);
+	syncobj_destroy(i915, fence.handle);
+}
+
+static void
+test_syncobj_future_cycle(int i915, unsigned int engine, int flags)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	const struct intel_execution_engine2 *e1, *e2;
+
+	__for_each_physical_engine(i915, e1) {
+		if (!gem_class_can_store_dword(i915, e1->class))
+			continue;
+
+	__for_each_physical_engine(i915, e2) {
+		struct drm_i915_gem_exec_fence fence = {};
+		struct drm_i915_gem_execbuffer2 execbuf = {
+			.buffers_ptr = to_user_pointer(&obj),
+			.buffer_count = 1,
+			.cliprects_ptr = to_user_pointer(&fence),
+			.num_cliprects = 1,
+			.flags = engine | I915_EXEC_FENCE_ARRAY,
+		};
+		int out;
+
+		if (!gem_class_can_store_dword(i915, e1->class))
+			continue;
+
+		fence.handle = syncobj_create(i915, 0);
+		fence.flags = I915_EXEC_FENCE_WAIT | flags;
+		execbuf.batch_start_offset = 0;
+		execbuf.flags |= I915_EXEC_FENCE_OUT;
+		igt_require(__gem_execbuf_wr(i915, &execbuf) == 0);
+
+		fence.flags = 0;
+		execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+		execbuf.rsvd2 >>= 32;
+		execbuf.flags &= ~I915_EXEC_RING_MASK;
+		execbuf.flags |= e1->flags | I915_EXEC_FENCE_IN;
+		gem_execbuf_wr(i915, &execbuf);
+		gem_context_destroy(i915, execbuf.rsvd1);
+		close(execbuf.rsvd2);
+
+		fence.flags = I915_EXEC_FENCE_SIGNAL;
+		execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+		execbuf.rsvd2 >>= 32;
+		execbuf.flags &= ~I915_EXEC_RING_MASK;
+		execbuf.flags |= e2->flags;
+		execbuf.batch_start_offset = 64;
+		gem_execbuf_wr(i915, &execbuf); /* writes 2 */
+		gem_context_destroy(i915, execbuf.rsvd1);
+
+		syncobj_destroy(i915, fence.handle);
+		gem_sync(i915, obj.handle);
+
+
+		/* How should this deadlock be resolved? */
+		out = execbuf.rsvd2 >> 32;
+		igt_assert_eq(sync_fence_status(out), -EDEADLK);
+		close(out);
+
+		out = execbuf.rsvd2;
+		igt_assert_eq(sync_fence_status(out), -EDEADLK);
+		close(out);
+	}}
+
+	gem_close(i915, obj.handle);
+}
+
+static void test_syncobj_future_robust_fd(int i915, unsigned int engine)
+{
+	int child = gem_reopen_driver(i915);
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = batch_create(child),
+	};
+	struct drm_i915_gem_exec_fence fence = {
+		.flags = I915_EXEC_FENCE_WAIT,
+		.handle = syncobj_create(child, 0),
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+		.flags = engine | I915_EXEC_FENCE_ARRAY | I915_EXEC_FENCE_OUT,
+	};
+	int out;
+
+	gem_execbuf_wr(child, &execbuf);
+	close(child);
+
+	out = execbuf.rsvd2 >> 32;
+	sync_fence_wait(out, -1);
+	igt_assert_eq(sync_fence_status(out), 1);
+	close(out);
+}
+
+static void sendfd(int socket, int fd)
+{
+	char buf[CMSG_SPACE(sizeof(fd))];
+	struct iovec io = { .iov_base = (char *)"ABC", .iov_len = 3 };
+	struct msghdr msg = {
+		.msg_iov = &io,
+		.msg_iovlen = 1,
+		.msg_control = buf,
+		.msg_controllen = CMSG_LEN(sizeof(fd)),
+	};
+	struct cmsghdr *cmsg = CMSG_FIRSTHDR(&msg);
+
+	cmsg->cmsg_level = SOL_SOCKET;
+	cmsg->cmsg_type = SCM_RIGHTS;
+	cmsg->cmsg_len = msg.msg_controllen;
+	memcpy(CMSG_DATA(cmsg), &fd, sizeof(fd));
+
+	igt_assert(sendmsg(socket, &msg, 0) != -1);
+}
+
+static int recvfd(int socket)
+{
+	char m_buffer[256], c_buffer[256];
+	struct iovec io = {
+		.iov_base = m_buffer,
+		.iov_len = sizeof(m_buffer),
+	};
+	struct msghdr msg = {
+		.msg_iov = &io,
+		.msg_iovlen = 1,
+		.msg_control = c_buffer,
+		.msg_controllen = sizeof(c_buffer),
+	};
+
+	igt_assert(recvmsg(socket, &msg, 0) != -1);
+	return *(int *)CMSG_DATA(CMSG_FIRSTHDR(&msg));
+}
+
+static void test_syncobj_future_robust_process(int i915, unsigned int engine)
+{
+	int out, sv[2];
+
+	igt_require(socketpair(AF_UNIX, SOCK_DGRAM, 0, sv) == 0);
+
+	igt_fork(child, 1) {
+		int fd = gem_reopen_driver(i915);
+		struct drm_i915_gem_exec_object2 obj = {
+			.handle = batch_create(fd),
+		};
+		struct drm_i915_gem_exec_fence fence = {
+			.flags = I915_EXEC_FENCE_WAIT,
+			.handle = syncobj_create(fd, 0),
+		};
+		struct drm_i915_gem_execbuffer2 execbuf = {
+			.buffers_ptr = to_user_pointer(&obj),
+			.buffer_count = 1,
+			.cliprects_ptr = to_user_pointer(&fence),
+			.num_cliprects = 1,
+			.flags = engine | I915_EXEC_FENCE_ARRAY | I915_EXEC_FENCE_OUT,
+		};
+
+		gem_execbuf_wr(fd, &execbuf);
+		sendfd(sv[0], execbuf.rsvd2 >> 32);
+	}
+	close(sv[0]);
+	igt_waitchildren();
+
+	out = recvfd(sv[1]);
+	close(sv[1]);
+
+	sync_fence_wait(out, -1);
+	igt_assert_eq(sync_fence_status(out), 1);
+	close(out);
+}
+
+static void
+test_syncobj_future_suspend(int i915, unsigned int engine, int target)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.offset = 24 << 20,
+		.handle = future_submit_batch(i915, 24 << 20),
+		.flags = EXEC_OBJECT_PINNED,
+	};
+	struct drm_i915_gem_exec_fence fence = {
+		.handle = syncobj_create(i915, 0)
+       	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+		.flags = engine | I915_EXEC_FENCE_ARRAY,
+	};
+	uint32_t result;
+	int out;
+
+	igt_require(gem_scheduler_has_timeslicing(i915));
+
+	execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+	fence.flags = I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_WAIT_SUBMIT;
+	execbuf.batch_start_offset = 0;
+	execbuf.flags |= I915_EXEC_FENCE_OUT;
+	igt_require(__gem_execbuf_wr(i915, &execbuf) == 0); /* writes 1 */
+	execbuf.flags &= ~I915_EXEC_FENCE_OUT;
+	gem_context_destroy(i915, execbuf.rsvd1);
+
+	igt_system_suspend_autoresume(target, SUSPEND_TEST_NONE);
+
+	execbuf.rsvd1 = gem_context_clone_with_engines(i915, 0);
+	fence.flags = I915_EXEC_FENCE_SIGNAL;
+	execbuf.batch_start_offset = 64;
+	gem_execbuf(i915, &execbuf); /* writes 2 */
+	gem_context_destroy(i915, execbuf.rsvd1);
+
+	gem_sync(i915, obj.handle); /* write hazard lies */
+	gem_read(i915, obj.handle, 4000, &result, sizeof(result));
+	igt_assert_eq(result, 2);
+
+	/* check we didn't autotimeout */
+	out = execbuf.rsvd2 >> 32;
+	igt_assert_eq(sync_fence_status(out), 1);
+	close(out);
+
+	syncobj_destroy(i915, fence.handle);
+	gem_close(i915, obj.handle);
+}
+
+static bool has_future_syncobj(int i915)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = batch_create(i915),
+	};
+	struct drm_i915_gem_exec_fence fence = {
+		.handle = syncobj_create(i915, 0),
+		.flags = I915_EXEC_FENCE_WAIT | I915_EXEC_FENCE_SIGNAL,
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.cliprects_ptr = to_user_pointer(&fence),
+		.num_cliprects = 1,
+		.flags = I915_EXEC_FENCE_ARRAY,
+	};
+	bool result;
+
+	result = __gem_execbuf(i915, &execbuf) == 0;
+	gem_close(i915, obj.handle);
+
+	return result;
+}
+
+static void syncobj_futures(int i915)
+{
+	const struct intel_execution_engine2 *e;
+
+	igt_fixture {
+		igt_require(gem_scheduler_enabled(i915));
+		igt_require(has_future_syncobj(i915));
+	}
+
+	igt_subtest_with_dynamic("syncobj-future") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future(i915, e->flags);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-past") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_past(i915, e->flags, 0);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-submit") {
+		igt_require(intel_gen(intel_get_drm_devid(i915)) >= 8);
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_submit(i915, e->flags);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-submit-past") {
+		igt_require(intel_gen(intel_get_drm_devid(i915)) >= 8);
+		__for_each_physical_engine(i915, e) {
+			if (!gem_class_can_store_dword(i915, e->class))
+				continue;
+
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_past(i915, e->flags,
+						I915_EXEC_FENCE_WAIT_SUBMIT);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-self") {
+		__for_each_physical_engine(i915, e) {
+			if (!gem_class_can_store_dword(i915, e->class))
+				continue;
+
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_self(i915, e->flags, 0);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-self-submit") {
+		__for_each_physical_engine(i915, e) {
+			if (!gem_class_can_store_dword(i915, e->class))
+				continue;
+
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_self(i915, e->flags,
+						I915_EXEC_FENCE_WAIT_SUBMIT);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-pair") {
+		igt_require(intel_gen(intel_get_drm_devid(i915)) >= 8);
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_pair(i915, e->flags);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-group") {
+		igt_require(intel_gen(intel_get_drm_devid(i915)) >= 8);
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_group(i915, e->flags, 67);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-deadlock") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_deadlock(i915, e->flags, 0);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-submit-deadlock") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_deadlock(i915, e->flags,
+							     I915_EXEC_FENCE_WAIT_SUBMIT);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-cycle") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_cycle(i915, e->flags, 0);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-submit-cycle") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_cycle(i915, e->flags,
+							  I915_EXEC_FENCE_WAIT_SUBMIT);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-dmabuf-deadlock") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_dmabuf_deadlock(i915, e->flags);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-export-deadlock") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_export_deadlock(i915, e->flags);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-import") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_import(i915, e->flags, true);
+		}
+	}
+	igt_subtest_with_dynamic("syncobj-future-import-incomplete") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_import(i915, e->flags, false);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-robust-fd") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_robust_fd(i915, e->flags);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-robust-process") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_robust_process(i915, e->flags);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-S0") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_suspend(i915, e->flags,
+							    SUSPEND_STATE_FREEZE);
+		}
+	}
+
+	igt_subtest_with_dynamic("syncobj-future-S3") {
+		__for_each_physical_engine(i915, e) {
+			igt_dynamic_f("%s", e->name)
+				test_syncobj_future_suspend(i915, e->flags,
+							    SUSPEND_STATE_MEM);
+		}
+	}
+}
+
 igt_main
 {
 	const struct intel_execution_engine2 *e;
@@ -3157,6 +4237,9 @@ igt_main
 		igt_subtest("syncobj-invalid-flags")
 			test_syncobj_invalid_flags(i915);
 
+		igt_subtest("syncobj-incomplete-wait-submit")
+			test_syncobj_incomplete_wait_submit(i915);
+
 		igt_subtest("syncobj-signal")
 			test_syncobj_signal(i915);
 
@@ -3175,6 +4258,9 @@ igt_main
 		igt_subtest("syncobj-channel")
 			test_syncobj_channel(i915);
 
+		igt_subtest_group
+			syncobj_futures(i915);
+
 		igt_fixture {
 			igt_stop_hang_detector();
 		}