[igt-dev] [PATCH i-g-t v2 0/2] Add pinned to perf_pmu and remove store dword

[igt-dev] [PATCH i-g-t v2 0/2] Add pinned to perf_pmu and remove store dword

[Zbigniew Kempczyński]

In short:

1. Add pinned flag for softpin 
2. Add additional store dword for looping condition + assertions
   for BDW/BSW.

Zbigniew Kempczyński (2):
  tests/i915/perf_pmu: Add softpin flag for no-reloc path
  tests/i915/perf_pmu: Add second store dword for looping

 tests/i915/perf_pmu.c | 32 ++++++++++++++++++++++++--------
 1 file changed, 24 insertions(+), 8 deletions(-)

-- 
2.34.1

[igt-dev] [PATCH i-g-t v2 1/2] tests/i915/perf_pmu: Add softpin flag for no-reloc path

[Zbigniew Kempczyński]

As Chris noticed semaphore object doesn't have EXEC_OBJECT_PINNED flag.
It is likely not a problem as we provide offsets from reloc allocator
thus there's no risk of overlapping offsets. But for purity ensure
we passed pinned flag when we're in softpin mode.

Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
---
 tests/i915/perf_pmu.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/tests/i915/perf_pmu.c b/tests/i915/perf_pmu.c
index 1e385bebc7..39e9fc5fef 100644
--- a/tests/i915/perf_pmu.c
+++ b/tests/i915/perf_pmu.c
@@ -861,6 +861,7 @@ __sema_busy(int gem_fd, uint64_t ahnd, int pmu, const intel_ctx_t *ctx,
 		.handle = create_sema(gem_fd, ahnd, reloc, &obj.offset),
 		.relocation_count = !ahnd ? 2 : 0,
 		.relocs_ptr = to_user_pointer(reloc),
+		.flags = !ahnd ? 0 : EXEC_OBJECT_PINNED,
 	};
 	struct drm_i915_gem_execbuffer2 eb = {
 		.batch_start_offset = 64,
-- 
2.34.1

[igt-dev] [PATCH i-g-t v2 2/2] tests/i915/perf_pmu: Add second store dword for looping

[Zbigniew Kempczyński]

On Broadwell/Braswell we observe semaphore wait isn't starting sometimes.
It's hard to say what's the main reason of this behavior - it happens
rare (each 2-3 months). To find out is this delayed write lets add
additional store dword for loop condition. Add also some assertions
to catch unexpected sw/hw behavior.

Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
---
 tests/i915/perf_pmu.c | 31 +++++++++++++++++++++++--------
 1 file changed, 23 insertions(+), 8 deletions(-)

diff --git a/tests/i915/perf_pmu.c b/tests/i915/perf_pmu.c
index 39e9fc5fef..fa6dddaf40 100644
--- a/tests/i915/perf_pmu.c
+++ b/tests/i915/perf_pmu.c
@@ -812,6 +812,12 @@ create_sema(int gem_fd, uint64_t ahnd,
 		0,
 		1,
 
+		/* Separate write for loop */
+		MI_STORE_DWORD_IMM,
+		0,
+		0,
+		2,
+
 		/* Wait until the semaphore value is set to 0 [by caller] */
 		MI_SEMAPHORE_WAIT | MI_SEMAPHORE_POLL | MI_SEMAPHORE_SAD_NEQ_SDD,
 		1,
@@ -827,17 +833,22 @@ create_sema(int gem_fd, uint64_t ahnd,
 	handle = gem_create(gem_fd, 4096);
 	*poffset = get_offset(ahnd, handle, 4096, 0);
 
-	memset(reloc, 0, 2 * sizeof(*reloc));
+	memset(reloc, 0, 3 * sizeof(*reloc));
 	reloc[0].target_handle = handle;
 	reloc[0].offset = 64 + 1 * sizeof(uint32_t);
 	reloc[1].target_handle = handle;
-	reloc[1].offset = 64 + 6 * sizeof(uint32_t);
+	reloc[1].offset = 64 + 5 * sizeof(uint32_t);
+	reloc[1].delta = 4;
+	reloc[2].target_handle = handle;
+	reloc[2].offset = 64 + 10 * sizeof(uint32_t);
 
 	if (ahnd) {
 		cs[1] = *poffset;
 		cs[2] = *poffset >> 32;
-		cs[6] = *poffset;
-		cs[7] = *poffset >> 32;
+		cs[5] = *poffset + 4;
+		cs[6] = (*poffset + 4) >> 32;
+		cs[10] = *poffset;
+		cs[11] = *poffset >> 32;
 	}
 
 	gem_write(gem_fd, handle, 64, cs, sizeof(cs));
@@ -856,10 +867,10 @@ __sema_busy(int gem_fd, uint64_t ahnd, int pmu, const intel_ctx_t *ctx,
 	};
 	uint64_t total, sema, busy;
 	uint64_t start[2], val[2];
-	struct drm_i915_gem_relocation_entry reloc[2];
+	struct drm_i915_gem_relocation_entry reloc[3];
 	struct drm_i915_gem_exec_object2 obj = {
 		.handle = create_sema(gem_fd, ahnd, reloc, &obj.offset),
-		.relocation_count = !ahnd ? 2 : 0,
+		.relocation_count = !ahnd ? 3 : 0,
 		.relocs_ptr = to_user_pointer(reloc),
 		.flags = !ahnd ? 0 : EXEC_OBJECT_PINNED,
 	};
@@ -878,13 +889,17 @@ __sema_busy(int gem_fd, uint64_t ahnd, int pmu, const intel_ctx_t *ctx,
 
 	gem_quiescent_gpu(gem_fd);
 
-	map = gem_mmap__device_coherent(gem_fd, obj.handle, 0, 4096, PROT_WRITE);
+	map = gem_mmap__device_coherent(gem_fd, obj.handle, 0, 4096, PROT_READ | PROT_WRITE);
 	gem_execbuf(gem_fd, &eb);
+	igt_assert(gem_bo_busy(gem_fd, obj.handle));
 	spin = igt_spin_new(gem_fd, .ahnd = ahnd, .ctx = ctx, .engine = e->flags);
 
 	/* Wait until the batch is executed and the semaphore is busy-waiting */
-	while (!READ_ONCE(*map) && gem_bo_busy(gem_fd, obj.handle))
+	while (READ_ONCE(map[1]) != 2 && gem_bo_busy(gem_fd, obj.handle))
 		;
+	if (!gem_bo_busy(gem_fd, obj.handle))
+		igt_warn("Bo != busy, map[0] = %u, map[1] = %u\n", map[0], map[1]);
+	igt_assert(map[0] == 1);
 	igt_assert(gem_bo_busy(gem_fd, obj.handle));
 	gem_close(gem_fd, obj.handle);
 
-- 
2.34.1