[Intel-gfx] [PATCH i-g-t 2/2] lib: Spin fast, retire early

[Chris Wilson]

When using the pollable spinner, we often want to use it as a means of
ensuring the task is running on the GPU before switching to something
else. In which case we don't want to add extra delay inside the spinner,
but the current 1000 NOPs add on order of 5us, which is often larger
than the target latency.

v2: Don't change perf_pmu as that is sensitive to the extra CPU latency
from a tight GPU spinner.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Reviewed-by: Antonio Argenziano <antonio.argenziano at intel.com> #v1
Reviewed-by: Joonas Lahtinen <joonas.lahtinen at linux.intel.com> #v1
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
 lib/igt_dummyload.c       | 3 ++-
 lib/igt_dummyload.h       | 1 +
 tests/gem_ctx_isolation.c | 1 +
 tests/gem_eio.c           | 1 +
 tests/gem_exec_latency.c  | 4 ++--
 5 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/lib/igt_dummyload.c b/lib/igt_dummyload.c
index 94efdf745..7beb66244 100644
--- a/lib/igt_dummyload.c
+++ b/lib/igt_dummyload.c
@@ -199,7 +199,8 @@ emit_recursive_batch(igt_spin_t *spin,
 	 * between function calls, that appears enough to keep SNB out of
 	 * trouble. See https://bugs.freedesktop.org/show_bug.cgi?id=102262
 	 */
-	batch += 1000;
+	if (!(opts->flags & IGT_SPIN_FAST))
+		batch += 1000;
 
 	/* recurse */
 	r = &relocs[obj[BATCH].relocation_count++];
diff --git a/lib/igt_dummyload.h b/lib/igt_dummyload.h
index c794f2544..e80a12451 100644
--- a/lib/igt_dummyload.h
+++ b/lib/igt_dummyload.h
@@ -52,6 +52,7 @@ struct igt_spin_factory {
 
 #define IGT_SPIN_FENCE_OUT (1 << 0)
 #define IGT_SPIN_POLL_RUN  (1 << 1)
+#define IGT_SPIN_FAST      (1 << 2)
 
 igt_spin_t *
 __igt_spin_batch_factory(int fd, const struct igt_spin_factory *opts);
diff --git a/tests/gem_ctx_isolation.c b/tests/gem_ctx_isolation.c
index 2e19e8c03..4325e1c28 100644
--- a/tests/gem_ctx_isolation.c
+++ b/tests/gem_ctx_isolation.c
@@ -560,6 +560,7 @@ static void inject_reset_context(int fd, unsigned int engine)
 	struct igt_spin_factory opts = {
 		.ctx = gem_context_create(fd),
 		.engine = engine,
+		.flags = IGT_SPIN_FAST,
 	};
 	igt_spin_t *spin;
 
diff --git a/tests/gem_eio.c b/tests/gem_eio.c
index 0ec1aaec9..3162a3170 100644
--- a/tests/gem_eio.c
+++ b/tests/gem_eio.c
@@ -160,6 +160,7 @@ static igt_spin_t * __spin_poll(int fd, uint32_t ctx, unsigned long flags)
 	struct igt_spin_factory opts = {
 		.ctx = ctx,
 		.engine = flags,
+		.flags = IGT_SPIN_FAST,
 	};
 
 	if (gem_can_store_dword(fd, opts.engine))
diff --git a/tests/gem_exec_latency.c b/tests/gem_exec_latency.c
index 75811f325..de16322a6 100644
--- a/tests/gem_exec_latency.c
+++ b/tests/gem_exec_latency.c
@@ -65,7 +65,7 @@ poll_ring(int fd, unsigned ring, const char *name)
 {
 	const struct igt_spin_factory opts = {
 		.engine = ring,
-		.flags = IGT_SPIN_POLL_RUN,
+		.flags = IGT_SPIN_POLL_RUN | IGT_SPIN_FAST,
 	};
 	struct timespec tv = {};
 	unsigned long cycles;
@@ -464,7 +464,7 @@ rthog_latency_on_ring(int fd, unsigned int engine, const char *name, unsigned in
 #define MMAP_SZ (64 << 10)
 	const struct igt_spin_factory opts = {
 		.engine = engine,
-		.flags = IGT_SPIN_POLL_RUN,
+		.flags = IGT_SPIN_POLL_RUN | IGT_SPIN_FAST,
 	};
 	struct rt_pkt *results;
 	unsigned int engines[16];
-- 
2.18.0