[Intel-gfx] [PATCH i-g-t 2/2] 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
---
tests/i915/gem_exec_fence.c | 680 +++-
1 file changed, 677 insertions(+), 3 deletions(-)
diff --git a/tests/i915/gem_exec_fence.c b/tests/i915/gem_exec_fence.c
index 4140bff24..eb1165080 100644
--- a/tests/i915/gem_exec_fence.c
+++ b/tests/i915/gem_exec_fence.c
@@ -1123,11 +1123,12 @@ static void test_syncobj_invalid_wait(int fd)
struct drm_i915_gem_exec_fence fence = {
.handle = syncobj_create(fd, 0),
};
+ int out;
memset(, 0, sizeof(execbuf));
execbuf.buffers_ptr = to_user_pointer();
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();
execbuf.num_cliprects = 1;
@@ -1135,14 +1136,59 @@ static void test_syncobj_invalid_wait(int fd)
obj.handle = gem_create(fd, 4096);
gem_write(fd, obj.handle, 0, , 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, ), -EINVAL);
+ if (__gem_execbuf_wr(fd, )) {
+ igt_assert_eq(__gem_execbuf(fd, ), -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(),
+ .buffer_count = 1,
+
+ .cliprects_ptr = to_user_pointer(),
+ .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, )) {
+ igt_assert_eq(__gem_execbuf(i915, ), -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;
@@ -1289,6 +1335,517 @@ static void test_syncobj_wait(int fd)
}
}
+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(),
+ .buffer_count = 1,
+ .cliprects_ptr = to_user_pointer(),
+ .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
[Intel-gfx] [PATCH i-g-t 2/2] 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
---
tests/i915/gem_exec_fence.c | 558 +++-
1 file changed, 555 insertions(+), 3 deletions(-)
diff --git a/tests/i915/gem_exec_fence.c b/tests/i915/gem_exec_fence.c
index 4b0d87e4d..e51b7452e 100644
--- a/tests/i915/gem_exec_fence.c
+++ b/tests/i915/gem_exec_fence.c
@@ -46,6 +46,15 @@ struct sync_merge_data {
#define SYNC_IOC_MERGE _IOWR(SYNC_IOC_MAGIC, 3, struct sync_merge_data)
#endif
+#define MI_SEMAPHORE_WAIT (0x1c << 23)
+#define MI_SEMAPHORE_POLL (1 << 15)
+#define MI_SEMAPHORE_SAD_GT_SDD (0 << 12)
+#define MI_SEMAPHORE_SAD_GTE_SDD (1 << 12)
+#define MI_SEMAPHORE_SAD_LT_SDD (2 << 12)
+#define MI_SEMAPHORE_SAD_LTE_SDD (3 << 12)
+#define MI_SEMAPHORE_SAD_EQ_SDD (4 << 12)
+#define MI_SEMAPHORE_SAD_NEQ_SDD (5 << 12)
+
static void store(int fd, const struct intel_execution_engine2 *e,
int fence, uint32_t target, unsigned offset_value)
{
@@ -907,11 +916,12 @@ static void test_syncobj_invalid_wait(int fd)
struct drm_i915_gem_exec_fence fence = {
.handle = syncobj_create(fd, 0),
};
+ int out;
memset(, 0, sizeof(execbuf));
execbuf.buffers_ptr = to_user_pointer();
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();
execbuf.num_cliprects = 1;
@@ -919,14 +929,59 @@ static void test_syncobj_invalid_wait(int fd)
obj.handle = gem_create(fd, 4096);
gem_write(fd, obj.handle, 0, , 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, ), -EINVAL);
+ if (__gem_execbuf_wr(fd, )) {
+ igt_assert_eq(__gem_execbuf(fd, ), -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(),
+ .buffer_count = 1,
+
+ .cliprects_ptr = to_user_pointer(),
+ .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, )) {
+ igt_assert_eq(__gem_execbuf(i915, ), -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;
@@ -1073,6 +1128,398 @@ static void test_syncobj_wait(int fd)
}
}
+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)
+{
+
[Intel-gfx] [PATCH i-g-t 2/2] 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
---
tests/i915/gem_exec_fence.c | 430 +++-
1 file changed, 427 insertions(+), 3 deletions(-)
diff --git a/tests/i915/gem_exec_fence.c b/tests/i915/gem_exec_fence.c
index 17fdaebd5..374b273e4 100644
--- a/tests/i915/gem_exec_fence.c
+++ b/tests/i915/gem_exec_fence.c
@@ -47,6 +47,15 @@ struct sync_merge_data {
#define SYNC_IOC_MERGE _IOWR(SYNC_IOC_MAGIC, 3, struct sync_merge_data)
#endif
+#define MI_SEMAPHORE_WAIT (0x1c << 23)
+#define MI_SEMAPHORE_POLL (1 << 15)
+#define MI_SEMAPHORE_SAD_GT_SDD (0 << 12)
+#define MI_SEMAPHORE_SAD_GTE_SDD (1 << 12)
+#define MI_SEMAPHORE_SAD_LT_SDD (2 << 12)
+#define MI_SEMAPHORE_SAD_LTE_SDD (3 << 12)
+#define MI_SEMAPHORE_SAD_EQ_SDD (4 << 12)
+#define MI_SEMAPHORE_SAD_NEQ_SDD (5 << 12)
+
static void store(int fd, const struct intel_execution_engine2 *e,
int fence, uint32_t target, unsigned offset_value)
{
@@ -913,11 +922,12 @@ static void test_syncobj_invalid_wait(int fd)
struct drm_i915_gem_exec_fence fence = {
.handle = syncobj_create(fd, 0),
};
+ int out;
memset(, 0, sizeof(execbuf));
execbuf.buffers_ptr = to_user_pointer();
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();
execbuf.num_cliprects = 1;
@@ -925,14 +935,59 @@ static void test_syncobj_invalid_wait(int fd)
obj.handle = gem_create(fd, 4096);
gem_write(fd, obj.handle, 0, , 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, ), -EINVAL);
+ if (__gem_execbuf_wr(fd, )) {
+ igt_assert_eq(__gem_execbuf(fd, ), -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(),
+ .buffer_count = 1,
+
+ .cliprects_ptr = to_user_pointer(),
+ .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, )) {
+ igt_assert_eq(__gem_execbuf(i915, ), -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;
@@ -1079,6 +1134,319 @@ static void test_syncobj_wait(int fd)
}
}
+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)
+{
+
