Skip to contentSkip to conversation

Scheduler tests

PATCH New Derek Morton 1 participant 6 messages Mar 30, 2016 Library code

lore
Share & tools
lore
#0 lore raw

v4 cover Scheduler tests

This patch set adds scheduler tests.
Patch 1 adds library code used by the tests. There are other tests under development which are planned to reuse some of these libraries. Patch 2 adds some basic tests, read dependency tests and write dependency tests. Patch 3 Is the patch previously submitted by John Harrison to update gem_ctx_param_basic with ioctls to set context priorities. It is included as part of this patch set as Patch 6 & 7 are dependant on it.
Patch 4 adds subtests to check sheduler behaviour for batch buffers submitted at differing priorities.
Patch 5 adds subtests to check priority starving behaviour.

v2: Updates for comments from and Daniel Vetter
Added tests requested by Joonas Lahtinen

v3: Removed the patch to update gem_has_ring() and updated the test to
use BSD if there is 1 ring and BSD1 / BSD2 if there are two.

v4: Removed patch to make gem_has_ring() non static as another patch has already done so.
Addressed more comments from Daniele Ceraolo Spurio

Derek Morton (4):
lib/intel_batchbuffer: Add functions to be used in the scheduler test tests/gem_scheduler: Add gem_scheduler test tests/gem_scheduler: Add subtests to test batch priority behaviour
gem_scheduler: Added subtests to test priority starving

John Harrison (1):
igt/gem_ctx_param_basic: Updated to support scheduler priority
interface

 lib/intel_batchbuffer.c     | 372 +++++++++++++++++++++++++-
 lib/intel_batchbuffer.h     |  14 +
 lib/ioctl_wrappers.h        |   1 +
 tests/Makefile.sources      |   1 +
 tests/gem_ctx_param_basic.c |  34 ++-
 tests/gem_scheduler.c       | 622 ++++++++++++++++++++++++++++++++++++++++++++
 6 files changed, 1038 insertions(+), 6 deletions(-)
 create mode 100644 tests/gem_scheduler.c
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

PATCH v4 1/5 lib/intel_batchbuffer: Add functions to be used in the scheduler test

Adds functions to create a number of different batch buffers to perform several functions including:
Batch buffer which will run for a long duration to provide a delay on a specified ring.
Function to calibrate the delay batch buffer to run for a specified period of time.
Function to create a batch buffer which writes timestamps to a buffer object. Function to compare timestamps allowing for wrapping of the values.

v2: Moved code to intel_batchbuffer (Daniel Vetter)
Addressed review comments from Daniele Ceraolo Spurio

v4: Addressed more comments from Daniele Ceraolo Spurio

---
 lib/intel_batchbuffer.c | 372 +++++++++++++++++++++++++++++++++++++++++++++++-
 lib/intel_batchbuffer.h |  14 ++
 2 files changed, 381 insertions(+), 5 deletions(-)
lib/intel_batchbuffer.c +367 -5
index 692521f..aa4eeaf 100644
@@ -1,8 +1,8 @@
 /**************************************************************************
- * 
+ *
  * Copyright 2006 Tungsten Graphics, Inc., Cedar Park, Texas.
  * All Rights Reserved.
- * 
+ *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the
  * "Software"), to deal in the Software without restriction, including
@@ -10,11 +10,11 @@
  * distribute, sub license, and/or sell copies of the Software, and to
  * permit persons to whom the Software is furnished to do so, subject to
  * the following conditions:
- * 
+ *
  * The above copyright notice and this permission notice (including the
  * next paragraph) shall be included in all copies or substantial portions
  * of the Software.
- * 
+ *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
@@ -22,7 +22,7 @@
  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
- * 
+ *
  **************************************************************************/
 
 #include <inttypes.h>
@@ -30,6 +30,8 @@
 #include <stdio.h>
 #include <string.h>
 #include <assert.h>
+#include <stdint.h>
+#include <time.h>
 
 #include "drm.h"
 #include "drmtest.h"
@@ -42,6 +44,7 @@
 #include "ioctl_wrappers.h"
 #include "media_spin.h"
 #include "gpgpu_fill.h"
+#include "igt_gt.h"
 
 #include <i915_drm.h>
 
@@ -817,3 +820,362 @@ igt_media_spinfunc_t igt_get_media_spinfunc(int devid)
 
 	return spin;
 }
+
+#define SEC_TO_NSEC (1000 * 1000 * 1000)
+#define DWORDS_TO_BYTES(x) ((x)*4)
+
+#define MI_STORE_REGISTER_MEM(LENGTH)   ((0x024 << 23) | ((LENGTH - 2) & 0xff))
+#define MI_MATH(NrInst)                 ((0x01A << 23) | ((NrInst - 1) & 0x3f))
+#define MI_CONDITIONAL_BATCH_BUFFER_END ((0x036 << 23) | (1 << 21) | 2)
+#define MI_COPY_MEM_MEM                 ((0x02E << 23) | (3))
+
+#define ALU_LOAD(TO, FROM)  ((0x080 << 20) | ((TO) << 10) | (FROM))
+#define ALU_SUB             ( 0x101 << 20)
+#define ALU_STORE(TO, FROM) ((0x180 << 20) | ((TO) << 10) | (FROM))
+
+#define TIMESTAMP_offset      (0x358) /* Elapsed time from system start */
+#define ALU_GPU_R0_LSB_offset (0x600)
+#define ALU_GPU_R0_MSB_offset (0x604)
+#define ALU_GPU_R1_LSB_offset (0x608)
+#define ALU_GPU_R1_MSB_offset (0x60C)
+#define ALU_GPU_R2_LSB_offset (0x610)
+#define ALU_GPU_R2_MSB_offset (0x614)
+
+#define ALU_R0_ENCODING   (0x00)
+#define ALU_R1_ENCODING   (0x01)
+#define ALU_SRCA_ENCODING (0x20)
+#define ALU_SRCB_ENCODING (0x21)
+#define ALU_ACCU_ENCODING (0x31)
+
+static int bb_address_size_dw(int gen)
+{
+	if (gen >= 8)
+		return 2;
+	else
+		return 1;
+}
+
+static uint32_t get_mmio_base(int ringid)
+{
+	switch (ringid) {
+	case I915_EXEC_RENDER:
+		return 0x02000;
+	case I915_EXEC_BSD:
+	case I915_EXEC_BSD | 1<<13: /* BSD1 */
+		return 0x12000;
+	case I915_EXEC_BSD | 2<<13: /* BSD2 */
+		return 0x1c000;
+	case I915_EXEC_BLT:
+		return 0x22000;
+	case I915_EXEC_VEBOX:
+		return 0x1A000;
+	default:
+		igt_assert_f(0, "Invalid ringid %d passed to get_mmio_base()\n", ringid);
+	}
+}
+
+/**
+ * igt_batch_used
+ * @batch batchbuffer to get offset from
+ *
+ * This returns the number of bytes of the batchbuffer that have been used.
+ * e.g. The offset into the batchbuffer that the next OUT_BATCH would write to.
+ *
+ * Returns:
+ * The number of bytes of the batchbuffer that have been used.
+ */
+uint32_t igt_batch_used(struct intel_batchbuffer *batch)
+{
+	return batch->ptr - batch->buffer;
+}
+
+/**
+ * igt_create_delay_bb:
+ * @batch: Batch buffer to write to
+ * @ringid: Execution flag of the ring. e.g. I915_EXEC_RENDER
+ * @loops: Number of times to loop
+ * @dest: Buffer to use for saving the current loop count and timestamp.
+ *
+ * This creates a batch buffer which will iterate a loop a specified number
+ * of times. Intended for creating batch buffers which take an arbitarily
+ * long time to execute. This can be useful to keep a ring busy while
+ * constructing a test scenario.
+ *
+ * The dest buffer will have a number of Dwords written by the batch buffer
+ * when it runs. They are:
+ * DW0 & DW1 - These are loaded with the value of 'loops' and are decremented
+ *             as the batch buffer executes. They will be 0 after the batch
+ *             buffer completes if it finished succesfully.
+ * DW2 Timestamp - An indication of when the batch buffer ran allowing a
+ *                 comparison between batch buffers to show execution order.
+ *                 May wrap so igt_compare_timestamps() should be used to
+ *                 compare timestamps.
+ *                 The timestamp will wrap every few minutes.
+ *
+ */
+void igt_create_delay_bb(struct intel_batchbuffer *batch, int ringid,
+                         uint32_t loops, drm_intel_bo *dest)
+{
+	int addr_size_dw;
+	uint32_t mmio_base, jump_offset;
+
+	/* CMD parser blocks reading TIMESTAMP register on gen 7.5 */
+	igt_require(batch->gen >= 8);
+
+	addr_size_dw = bb_address_size_dw(batch->gen);
+	mmio_base = get_mmio_base(ringid);
+	igt_assert(batch);
+	BEGIN_BATCH(32, 5);
+
+	/* store current timestamp in DW2 */
+	OUT_BATCH(MI_STORE_REGISTER_MEM(addr_size_dw + 2));
+	OUT_BATCH(mmio_base + TIMESTAMP_offset);
+	OUT_RELOC(dest, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, DWORDS_TO_BYTES(2));
+
+	/* Load R0 with loops */
+	OUT_BATCH(MI_LOAD_REGISTER_IMM);
+	OUT_BATCH(mmio_base + ALU_GPU_R0_LSB_offset);
+	OUT_BATCH(loops);
+	OUT_BATCH(MI_LOAD_REGISTER_IMM);
+	OUT_BATCH(mmio_base + ALU_GPU_R0_MSB_offset);
+	OUT_BATCH(0x00000000);
+	/* Load R1 with 1 */
+	OUT_BATCH(MI_LOAD_REGISTER_IMM);
+	OUT_BATCH(mmio_base + ALU_GPU_R1_LSB_offset);
+	OUT_BATCH(0x00000001);
+	OUT_BATCH(MI_LOAD_REGISTER_IMM);
+	OUT_BATCH(mmio_base + ALU_GPU_R1_MSB_offset);
+	OUT_BATCH(0x00000000);
+	/* Save loop jump offset */
+	jump_offset = igt_batch_used(batch);
+	/* Copy R0 / R1 into SRCA / SRCB, Perform R0 - R1, Store result in R0 */
+	/* e.g. R0 -= 1 */
+	OUT_BATCH(MI_MATH(4));
+	OUT_BATCH(ALU_LOAD(ALU_SRCA_ENCODING, ALU_R0_ENCODING));
+	OUT_BATCH(ALU_LOAD(ALU_SRCB_ENCODING, ALU_R1_ENCODING));
+	OUT_BATCH(ALU_SUB);
+	OUT_BATCH(ALU_STORE(ALU_R0_ENCODING, ALU_ACCU_ENCODING));
+	/* Copy R0 to dest
+	 * On Gen8 MI_CONDITIONAL_BATCH_BUFFER_END BSD ring Compare address
+	 * points to 2 Dwords, a mask (DW0) and data (DW1) which are ANDed
+	 * together.
+	 * On Gen9+, and the other rings on Gen8 Compare address points to
+	 * just Data (DW0). For simplicity always copy R0 LSB to DW0 and DW1.
+	 */
+	OUT_BATCH(MI_STORE_REGISTER_MEM(addr_size_dw + 2));
+	OUT_BATCH(mmio_base + ALU_GPU_R0_LSB_offset);
+	OUT_RELOC(dest, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, 0);
+	OUT_BATCH(MI_STORE_REGISTER_MEM(addr_size_dw + 2));
+	OUT_BATCH(mmio_base + ALU_GPU_R0_LSB_offset);
+	OUT_RELOC(dest, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, DWORDS_TO_BYTES(1));
+	/* Repeat until R0 == 0 */
+	OUT_BATCH(MI_CONDITIONAL_BATCH_BUFFER_END);
+	OUT_BATCH(0x00000000);
+	OUT_RELOC(dest, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, 0);
+	OUT_BATCH(MI_BATCH_BUFFER_START | (addr_size_dw - 1));
+	OUT_RELOC(batch->bo, I915_GEM_DOMAIN_INSTRUCTION, 0, jump_offset);
+
+	/* Should never get here, but end if it happens */
+	OUT_BATCH(MI_BATCH_BUFFER_END);
+	ADVANCE_BATCH();
+}
+
+/* Emit instructions to store timestamp of specified ring at offset (in bytes) into bo */
+static void emit_store_timestamp_instruction(struct intel_batchbuffer *batch,
+                                             drm_intel_bo *bo,
+                                             int ringid, int offset)
+{
+	BEGIN_BATCH(3, 1);
+	OUT_BATCH(MI_STORE_REGISTER_MEM(bb_address_size_dw(batch->gen) + 2));
+	OUT_BATCH(get_mmio_base(ringid) + TIMESTAMP_offset);
+	OUT_RELOC(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, offset);
+
+	ADVANCE_BATCH();
+}
+
+/**
+ * igt_create_timestamp_bb:
+ * @batch: Batch buffer to write to
+ * @ringid: Execution flag of the ring. e.g. I915_EXEC_RENDER
+ * @dest: Buffer to use for saving the timestamps.
+ * @load: Buffer to access. Set NULL if not required.
+ * @write: If true and load is not NULL, will also write a timestamp to load
+ * buffer. If false and load is not NULL, will read from load buffer into dest.
+ * Intended for dependency checking.
+ *
+ * This creates a batch buffer which writes timestamps into a buffer object.
+ * If 'load' is non null, data is either written to 'load' or copied from 'load'
+ * depending on whether 'write' is set.
+ *
+ * The dest buffer will have a number of Dwords written by the batch buffer
+ * when it runs. They are:
+ * DW0 Reported timestamp - An indication of when the batch buffer ran allowing a
+ *                          comparison between batch buffers to show execution order.
+ *                          May wrap so igt_compare_timestamps() should be used to
+ *                          compare timestamps.
+ *                          The timestamp will wrap every few minutes.
+ * DW2 Value copied from DW0 of load if write == false
+ *
+ */
+void igt_create_timestamp_bb(struct intel_batchbuffer *batch, int ringid,
+                             drm_intel_bo *dest, drm_intel_bo *load, bool write)
+{
+	/* CMD parser blocks reading TIMESTAMP register on gen 7.5 */
+	igt_require(batch->gen >= 8);
+
+	igt_assert(batch);
+
+	/* store current reported timestamp in DW0 */
+	emit_store_timestamp_instruction(batch, dest, ringid, 0);
+
+	if (load != NULL) {
+		if (write) {
+			emit_store_timestamp_instruction(batch, load, ringid, 0);
+		} else {
+			BEGIN_BATCH(3, 2);
+			OUT_BATCH(MI_COPY_MEM_MEM);
+			OUT_RELOC(dest, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, DWORDS_TO_BYTES(2));
+			OUT_RELOC(load, I915_GEM_DOMAIN_INSTRUCTION, 0, DWORDS_TO_BYTES(0));
+			ADVANCE_BATCH();
+		}
+	}
+
+	BEGIN_BATCH(1, 0);
+	OUT_BATCH(MI_BATCH_BUFFER_END);
+	ADVANCE_BATCH();
+}
+
+/**
+ * igt_create_noop_bb:
+ * @batch: Batch buffer to write to
+ * @noops: Number of MI_NOOP instructions to add to the batch buffer.
+ *
+ * This creates a batch buffer with a specified number of MI_NOOP instructions.
+ */
+void igt_create_noop_bb(struct intel_batchbuffer *batch, int noops)
+{
+	int loop;
+
+	igt_assert(batch);
+
+	BEGIN_BATCH(noops + 1, 0);
+	for (loop = 0; loop < noops; loop++)
+		OUT_BATCH(MI_NOOP);
+	OUT_BATCH(MI_BATCH_BUFFER_END);
+	ADVANCE_BATCH();
+}
+
+/* Store calibrated values so they only need calculating once.
+ * Use intel_execution_engines array as list of supported rings
+ */
+static uint32_t *calibrated_ring_value = NULL;
+
+/**
+ * igt_calibrate_delay_bb:
+ * @fd: file descriptor for i915 driver instance
+ * @bufmgr: Buffer manager to be used for creation of batch buffers
+ * @ringid: Ring to calibrate. e.g. I915_EXEC_RENDER
+ *
+ * This calculates the value of loops that would need to be passed to
+ * igt_create_delay_bb() to create a delay of about 1 second on the specified
+ * ring.
+ *
+ * Returns:
+ * uint32_t to be passed to igt_create_delay_bb().
+ */
+/* 0x100000 will run for about 0.6 - 0.8 seconds (dependant on ring) on BXT HW */
+#define CAL_SEED (0x100000)
+uint32_t igt_calibrate_delay_bb(int fd, drm_intel_bufmgr *bufmgr, int ringid)
+{
+	uint32_t buf[2];
+	struct intel_batchbuffer *bb;
+	struct timespec start, end;
+	uint64_t duration;
+	uint64_t calibrated;
+	drm_intel_bo *target_bo;
+	int ring_index;
+	uint32_t data = 0xffffffff;
+
+	if (calibrated_ring_value == NULL) {
+		int count;
+		/* intel_execution_engines is external, so need to determine
+		 * the number of elements at runtime to create the same number
+		 * of elements in calibrated_ring_value
+		 */
+		for (count = 0; intel_execution_engines[count].name != NULL; count++) {}
+		calibrated_ring_value = calloc(count, sizeof(uint32_t));
+		igt_assert(calibrated_ring_value);
+	}
+
+	/* Check if there is already a calibration value for this ring */
+	for (ring_index = 0; intel_execution_engines[ring_index].name != NULL; ring_index++) {
+		if ((intel_execution_engines[ring_index].exec_id |
+		    intel_execution_engines[ring_index].flags) == ringid) {
+			if (calibrated_ring_value[ring_index] != 0) {
+				return calibrated_ring_value[ring_index];
+			}
+			break;
+		}
+	}
+
+	target_bo = drm_intel_bo_alloc(bufmgr, "target bo", BATCH_SZ, BATCH_SZ);
+	igt_assert(target_bo);
+
+	/* Put some non zero values in the target bo */
+	do_or_die(drm_intel_bo_subdata(target_bo, 0, 4, &data));
+
+
+	bb = intel_batchbuffer_alloc(bufmgr, intel_get_drm_devid(fd));
+	igt_assert(bb);
+	igt_create_delay_bb(bb, ringid, CAL_SEED, target_bo);
+
+	gem_quiescent_gpu(fd);
+	clock_gettime(CLOCK_MONOTONIC, &start);
+	intel_batchbuffer_flush_on_ring(bb, ringid);
+	/* This will not return until the bo has finished executing */
+	drm_intel_bo_wait_rendering(target_bo);
+	clock_gettime(CLOCK_MONOTONIC, &end);
+
+	do_or_die(drm_intel_bo_get_subdata(target_bo, 0, 4, (void*)buf));
+
+	/* buf[0] in the target buffer should be 0 if the batch buffer completed */
+	igt_assert_f(buf[0] == 0, "buf[0] expected 0x0, got 0x%x\n", buf[0]);
+
+	duration = ((((uint64_t)end.tv_sec - (uint64_t)start.tv_sec) * SEC_TO_NSEC)
+	           + (uint64_t)end.tv_nsec) - (uint64_t)start.tv_nsec;
+
+	calibrated = (((uint64_t)(CAL_SEED) * SEC_TO_NSEC) / duration);
+	igt_debug("Uncalibrated run took %" PRIu64 ".%04" PRIu64 "s\n",
+	          duration / SEC_TO_NSEC,
+	          (duration % SEC_TO_NSEC) / 100000);
+	drm_intel_bo_unreference(target_bo);
+	intel_batchbuffer_free(bb);
+
+	/* Sanity check. If duration < 100ms, something has clearly gone wrong */
+	igt_assert(duration > (SEC_TO_NSEC  / 10));
+
+	igt_assert_f(calibrated <= UINT32_MAX, "Calibrated value > UINT32_MAX\n");
+
+	if (intel_execution_engines[ring_index].name != NULL)
+		calibrated_ring_value[ring_index] = (uint32_t)calibrated;
+	return (uint32_t)calibrated;
+}
+
+/**
+ * igt_compare_timestamps:
+ * @ts1: timestamp 1
+ * @ts2: timestamp 2
+ *
+ * This compares two uint32_t timestamps. To handle wrapping it assumes the
+ * difference between the two timestamps is less than 1/2 the max elapsed time
+ * represented by the counters.
+ * It also assumes the timestamps are samples from the same counter.
+ *
+ * Returns:
+ * True if ts2 > ts1, allowing for counter wrapping, false otherwise.
+ */
+
+bool igt_compare_timestamps(uint32_t ts1, uint32_t ts2)
+{
+	return (int32_t)(ts2 - ts1) > 0;
+}
lib/intel_batchbuffer.h +14 -0
index 869747d..839892e 100644
@@ -323,4 +323,18 @@ typedef void (*igt_media_spinfunc_t)(struct intel_batchbuffer *batch,
 
 igt_media_spinfunc_t igt_get_media_spinfunc(int devid);
 
+uint32_t igt_batch_used(struct intel_batchbuffer *batch);
+
+void igt_create_delay_bb(struct intel_batchbuffer *batch, int ringid,
+                         uint32_t loops, drm_intel_bo *dest);
+
+void igt_create_timestamp_bb(struct intel_batchbuffer *batch, int ringid,
+			     drm_intel_bo *dest, drm_intel_bo *load, bool write);
+
+void igt_create_noop_bb(struct intel_batchbuffer *batch, int noops);
+
+uint32_t igt_calibrate_delay_bb(int fd, drm_intel_bufmgr *bufmgr, int ringid);
+
+bool igt_compare_timestamps(uint32_t ts1, uint32_t ts2);
+
 #endif
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw

PATCH v4 2/5 tests/gem_scheduler: Add gem_scheduler test

In reply to Derek Morton

This is intended to test the scheduler behaviour is correct.
The subtests are
<ring>-basic
Tests that batch buffers of the same priority submitted to a ring execute in the order they are submitted.
<ring>-read
Submits a batch buffer with a read dependency to a buffer object to a ring which is held in the scheduler queue by a long running batch buffer. Submit batch buffers to other rings that have a read dependency to the same buffer object. Ensure they execute before the batch buffer being held up behind the long running batch buffer.
<ring>-write
Submits a batch buffer with a write dependency to a buffer object to a ring which is held in the scheduler queue by a long running batch buffer. Submit batch buffers to other rings that have a write dependency to the same buffer object. Submit batch buffers with no interdependencies to all rings. Ensure the batch buffers that have write dependencies are executed in submission order but the batch buffers without interdependencies do not get held up.

v2: Addressed review comments from Daniele Ceraolo Spurio

v3: Added logic to use generic BSD ring if there is 1 and BSD1 / BSD2 if there

    are 2.

v4: Addressed more review comments from Daniele Ceraolo Spurio

---
 tests/Makefile.sources |   1 +
 tests/gem_scheduler.c  | 451 +++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 452 insertions(+)
 create mode 100644 tests/gem_scheduler.c
tests/Makefile.sources +1 -0
index 26e487a..cdbb37c 100644
@@ -68,6 +68,7 @@ TESTS_progs_M = \
 	gem_request_retire \
 	gem_reset_stats \
 	gem_ringfill \
+	gem_scheduler \
 	gem_set_tiling_vs_blt \
 	gem_shrink \
 	gem_softpin \
tests/gem_scheduler.c new file +451 -0

451 lines not on this page · raw message

Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#3 lore raw

PATCH v4 3/5 igt/gem_ctx_param_basic: Updated to support scheduler priority interface

In reply to Derek Morton

From: John Harrison <[email protected]>

The GPU scheduler has added an execution priority level to the context object. There is an IOCTL interface to allow user apps/libraries to set this priority. This patch updates the context paramter IOCTL test to include the new interface.

For: VIZ-1587

---
 lib/ioctl_wrappers.h        |  1 +
 tests/gem_ctx_param_basic.c | 34 +++++++++++++++++++++++++++++++++-
 2 files changed, 34 insertions(+), 1 deletion(-)
lib/ioctl_wrappers.h +1 -0
index d986f61..3b8a015 100644
@@ -109,6 +109,7 @@ struct local_i915_gem_context_param {
 #define LOCAL_CONTEXT_PARAM_BAN_PERIOD	0x1
 #define LOCAL_CONTEXT_PARAM_NO_ZEROMAP	0x2
 #define LOCAL_CONTEXT_PARAM_GTT_SIZE	0x3
+#define LOCAL_CONTEXT_PARAM_PRIORITY	0x4
 	uint64_t value;
 };
 void gem_context_require_ban_period(int fd);
tests/gem_ctx_param_basic.c +33 -1
index b75800c..585a1a8 100644
@@ -147,10 +147,42 @@ igt_main
 		TEST_SUCCESS(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM);
 	}
 
+	ctx_param.param = LOCAL_CONTEXT_PARAM_PRIORITY;
+
+	igt_subtest("priority-root-set") {
+		ctx_param.context = ctx;
+		ctx_param.value = 2048;
+		TEST_FAIL(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM, EINVAL);
+		ctx_param.value = -2048;
+		TEST_FAIL(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM, EINVAL);
+		ctx_param.value = 512;
+		TEST_SUCCESS(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM);
+		ctx_param.value = -512;
+		TEST_SUCCESS(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM);
+		ctx_param.value = 0;
+		TEST_SUCCESS(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM);
+	}
+
+	igt_subtest("priority-non-root-set") {
+		igt_fork(child, 1) {
+			igt_drop_root();
+
+			ctx_param.context = ctx;
+			ctx_param.value = 512;
+			TEST_FAIL(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM, EPERM);
+			ctx_param.value = -512;
+			TEST_SUCCESS(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM);
+			ctx_param.value = 0;
+			TEST_SUCCESS(LOCAL_IOCTL_I915_GEM_CONTEXT_SETPARAM);
+		}
+
+		igt_waitchildren();
+	}
+
 	/* NOTE: This testcase intentionally tests for the next free parameter
 	 * to catch ABI extensions. Don't "fix" this testcase without adding all
 	 * the tests for the new param first. */
-	ctx_param.param = LOCAL_CONTEXT_PARAM_GTT_SIZE + 1;
+	ctx_param.param = LOCAL_CONTEXT_PARAM_PRIORITY + 1;
 
 	igt_subtest("invalid-param-get") {
 		ctx_param.context = ctx;
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw

PATCH v4 4/5 tests/gem_scheduler: Add subtests to test batch priority behaviour

In reply to Derek Morton

Add subtests to test each ring to check batch buffers of a higher priority will be executed before batch buffers of a lower priority.

v2: Addressed review comments from Daniele Ceraolo Spurio

v4: Changed priorities to +/-200 - Daniele Ceraolo Spurio

---
 tests/gem_scheduler.c | 57 ++++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 47 insertions(+), 10 deletions(-)
tests/gem_scheduler.c +47 -10
index 704dbb1..504607f 100644
@@ -39,7 +39,8 @@
 
 IGT_TEST_DESCRIPTION("Check scheduler behaviour. Basic tests ensure independant "
                      "batch buffers of the same priority are executed in "
-                     "submission order. Read-read tests ensure "
+                     "submission order. Priority tests ensure higher priority "
+                     "batch buffers are executed first. Read-read tests ensure "
                      "batch buffers with a read dependency to the same buffer "
                      "object do not block each other. Write-write dependency "
                      "tests ensure batch buffers with a write dependency to a "
@@ -61,7 +62,7 @@ static struct ring {
 } rings[] = {
 	{ "render", I915_EXEC_RENDER, false },
 	{ "bsd",    I915_EXEC_BSD , false },
-	{ "bsd2",    I915_EXEC_BSD | 2<<13, false },
+	{ "bsd2",   I915_EXEC_BSD | 2<<13, false },
 	{ "blt",    I915_EXEC_BLT, false },
 	{ "vebox",  I915_EXEC_VEBOX, false },
 };
@@ -70,7 +71,7 @@ static struct ring {
 
 static void check_rings(int fd) {
 	int loop;
-	for (loop=0; loop < NBR_RINGS; loop++) {
+	for (loop = 0; loop < NBR_RINGS; loop++) {
 		if (gem_has_ring(fd, rings[loop].id)) {
 			if (rings[loop].id == (I915_EXEC_BSD | 2<<13)) {
 				rings[loop].exists = gem_has_bsd2(fd);
@@ -145,11 +146,23 @@ static void init_context(int *fd, drm_intel_bufmgr **bufmgr, int ringid)
 	intel_batchbuffer_free(noop_bb);
 }
 
-/* Basic test. Check batch buffers of the same priority and with no dependencies
- * are executed in the order they are submitted.
+static void set_priority(int fd, int value)
+{
+	struct local_i915_gem_context_param param;
+	param.context = 0; /* Default context */
+	param.size = 0;
+	param.param = LOCAL_CONTEXT_PARAM_PRIORITY;
+	param.value = (uint64_t)value;
+	gem_context_set_param(fd, &param);
+}
+
+/* If 'priority' is 0, check batch buffers of the same priority and with
+ * no dependencies are executed in the order they are submitted.
+ * If 'priority' is set !0, check batch buffers of higher priority are
+ * executed before batch buffers of lower priority.
  */
 #define NBR_BASIC_FDs (3)
-static void run_test_basic(int in_flight, int ringid)
+static void run_test_basic(int in_flight, int ringid, int priority)
 {
 	int fd[NBR_BASIC_FDs];
 	int loop;
@@ -169,6 +182,13 @@ static void run_test_basic(int in_flight, int ringid)
 	for (loop = 0; loop < NBR_BASIC_FDs; loop++)
 		init_context(&(fd[loop]), &(bufmgr[loop]), ringid);
 
+	/* For high priority set priority of second context to overtake first
+	 * For low priority set priority of first context to be overtaxen by second
+	 */
+	if(priority > 0)
+		set_priority(fd[2], priority);
+	else if(priority < 0)
+		set_priority(fd[1], priority);
 
 	/* Create buffer objects */
 	delay_bo = create_and_check_bo(bufmgr[0], "delay bo");
@@ -212,9 +232,14 @@ static void run_test_basic(int in_flight, int ringid)
 	igt_assert_f(igt_compare_timestamps(delay_buf[2], ts1_buf[0]),
 	             "Delay ts (0x%08" PRIx32 ") > TS1 ts (0x%08" PRIx32 ")\n",
 	             delay_buf[2], ts1_buf[0]);
-	igt_assert_f(igt_compare_timestamps(ts1_buf[0], ts2_buf[0]),
-	             "TS1 ts (0x%08" PRIx32 ") > TS2 ts (0x%08" PRIx32 ")\n",
-	             ts1_buf[0], ts2_buf[0]);
+	if(priority)
+		igt_assert_f(igt_compare_timestamps(ts2_buf[0], ts1_buf[0]),
+		             "TS2 ts (0x%08" PRIx32 ") > TS1 ts (0x%08" PRIx32 ")\n",
+		             ts2_buf[0], ts1_buf[0]);
+	else
+		igt_assert_f(igt_compare_timestamps(ts1_buf[0], ts2_buf[0]),
+		             "TS1 ts (0x%08" PRIx32 ") > TS2 ts (0x%08" PRIx32 ")\n",
+		             ts1_buf[0], ts2_buf[0]);
 
 	/* Cleanup */
 	for (loop = 0; loop < in_flight; loop++)
@@ -430,7 +455,19 @@ igt_main
 	for (loop = 0; loop < NBR_RINGS; loop++)
 		igt_subtest_f("%s-basic", rings[loop].name) {
 			gem_require_ring(fd, rings[loop].id);
-			run_test_basic(in_flight, rings[loop].id);
+			run_test_basic(in_flight, rings[loop].id, false);
+		}
+
+	for (loop = 0; loop < NBR_RINGS; loop++)
+		igt_subtest_f("%s-priority-high", rings[loop].name) {
+			gem_require_ring(fd, rings[loop].id);
+			run_test_basic(in_flight, rings[loop].id, 200);
+		}
+
+	for (loop = 0; loop < NBR_RINGS; loop++)
+		igt_subtest_f("%s-priority-low", rings[loop].name) {
+			gem_require_ring(fd, rings[loop].id);
+			run_test_basic(in_flight, rings[loop].id, -200);
 		}
 
 	for (loop = 0; loop < NBR_RINGS; loop++)
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#5 lore raw

PATCH v4 5/5 gem_scheduler: Added subtests to test priority starving

In reply to Derek Morton

When a batch is selected for execution all other batch buffers in the scheduler queue get a small priority increase to prevent starving due to continuous submissions by a high priority user.
Added a subtest to check this behaviour.

Requested by Joonas Lahtinen during scheduler code review

v4: Addressed review comments from Daniele Ceraolo Spurio

---
 tests/gem_scheduler.c | 134 ++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 134 insertions(+)
tests/gem_scheduler.c +134 -0
index 504607f..5d8fecf 100644
@@ -257,6 +257,134 @@ static void run_test_basic(int in_flight, int ringid, int priority)
 	free(in_flight_bbs);
 }
 
+
+/* Test priority starving behaviour.
+   When a batch is selected for execution all other batch buffers in the
+   scheduler queue get a small priority increase to prevent starving due
+   to continuous submissions by a high priority user.
+ * Submit two delay batch buffers then fill the in flight queue. Submit a
+ * slightly reduced priority batch buffer and a normal priority batch buffer.
+ * When the first delay batch buffer completes and the next is put in flight all
+ * other batch buffers in the scheduler queue should get a small priority
+ * increase.
+ * Submit another normal priority batch buffer. It should not overtake the
+ * reduced priority batch buffer whose priority has been increased.
+ */
+#define NBR_PRIO_STARVE_FDs (4)
+static void run_test_priority_starve(int in_flight, int ringid)
+{
+	int fd[NBR_PRIO_STARVE_FDs];
+	int loop;
+	drm_intel_bufmgr *bufmgr[NBR_PRIO_STARVE_FDs];
+	uint32_t *delay_buf, *delay2_buf, *ts1_buf, *ts2_buf, *ts3_buf;
+	struct intel_batchbuffer *ts1_bb, *ts2_bb, *ts3_bb;
+	struct intel_batchbuffer **in_flight_bbs;
+	uint32_t calibrated_1s;
+	drm_intel_bo *delay_bo, *delay2_bo ,*ts1_bo, *ts2_bo, *ts3_bo;
+
+	in_flight_bbs = malloc(in_flight * sizeof(struct intel_batchbuffer *));
+	igt_assert(in_flight_bbs);
+
+	/* Need multiple i915 fd's. Scheduler will not change execution order of
+	 * batch buffers from the same context.
+	 */
+	for(loop = 0; loop < NBR_PRIO_STARVE_FDs; loop++)
+		init_context(&(fd[loop]), &(bufmgr[loop]), ringid);
+
+	/* Lower priority of first timestamp batch buffer
+	 * Second timestamp batch buffer should overtake the first
+	 * Overtaken batch should then get a priority increase so the third does
+	 * not overtake it.
+	 */
+	set_priority(fd[1], -1);
+
+	/* Create buffer objects */
+	delay_bo = create_and_check_bo(bufmgr[0], "delay bo");
+	delay2_bo = create_and_check_bo(bufmgr[0], "delay bo2");
+	ts1_bo = create_and_check_bo(bufmgr[1], "ts1 bo");
+	ts2_bo = create_and_check_bo(bufmgr[2], "ts2 bo");
+	ts3_bo = create_and_check_bo(bufmgr[3], "ts3 bo");
+
+	calibrated_1s = igt_calibrate_delay_bb(fd[0], bufmgr[0], ringid);
+
+	/* Batch buffers to fill the in flight queue */
+	in_flight_bbs[0] = create_delay_bb(fd[0], bufmgr[0], ringid, calibrated_1s, delay_bo);
+	in_flight_bbs[1] = create_delay_bb(fd[0], bufmgr[0], ringid, calibrated_1s, delay2_bo);
+	for(loop = 2; loop < in_flight; loop++)
+		in_flight_bbs[loop] = create_noop_bb(fd[0], bufmgr[0], 5);
+
+	/* Extra batch buffers in the scheduler queue */
+	ts1_bb = create_timestamp_bb(fd[1], bufmgr[1], ringid, ts1_bo, NULL, false);
+	ts2_bb = create_timestamp_bb(fd[2], bufmgr[2], ringid, ts2_bo, NULL, false);
+	ts3_bb = create_timestamp_bb(fd[3], bufmgr[3], ringid, ts3_bo, NULL, false);
+
+	/* Flush batchbuffers */
+	for(loop = 0; loop < in_flight; loop++)
+		intel_batchbuffer_flush_on_ring(in_flight_bbs[loop], ringid);
+	intel_batchbuffer_flush_on_ring(ts1_bb, ringid);
+	intel_batchbuffer_flush_on_ring(ts2_bb, ringid);
+
+	/* This will not return until the bo has finished executing */
+	drm_intel_bo_map(delay_bo, 0);
+	/* Once the first delay is complete and any bumping has occured, submit
+	 * the final batch buffer
+	 */
+	intel_batchbuffer_flush_on_ring(ts3_bb, ringid);
+
+	drm_intel_bo_map(delay2_bo, 0);
+	drm_intel_bo_map(ts1_bo, 0);
+	drm_intel_bo_map(ts2_bo, 0);
+	drm_intel_bo_map(ts3_bo, 0);
+
+	delay_buf = delay_bo->virtual;
+	delay2_buf = delay2_bo->virtual;
+	ts1_buf = ts1_bo->virtual;
+	ts2_buf = ts2_bo->virtual;
+	ts3_buf = ts3_bo->virtual;
+
+	igt_debug("Delay Timestamp = 0x%08" PRIx32 "\n", delay_buf[2]);
+	igt_debug("Delay Timestamp = 0x%08" PRIx32 "\n", delay2_buf[2]);
+	igt_debug("TS1 Timestamp = 0x%08" PRIx32 "\n", ts1_buf[0]);
+	igt_debug("TS2 Timestamp = 0x%08" PRIx32 "\n", ts2_buf[0]);
+	igt_debug("TS3 Timestamp = 0x%08" PRIx32 "\n", ts3_buf[0]);
+
+	/* buf[0] in the target buffer should be 0 if the batch buffer completed */
+	igt_assert_f(delay_buf[0] == 0,
+	             "delay_buf[0] expected 0x0, got 0x%" PRIx32 "\n", delay_buf[0]);
+
+	/* Delay BB should be first */
+	igt_assert_f(igt_compare_timestamps(delay_buf[2], ts1_buf[0]),
+	             "Delay ts (0x%08" PRIx32 ") > TS1 ts (0x%08" PRIx32 ")\n",
+	             delay_buf[2], ts1_buf[0]);
+
+	/* TS2 BB should overtake TS1 BB */
+	igt_assert_f(igt_compare_timestamps(ts2_buf[0], ts1_buf[0]),
+	             "TS2 ts (0x%08" PRIx32 ") > TS1 ts (0x%08" PRIx32 ")\n",
+	             ts2_buf[0], ts1_buf[0]);
+
+	/* TS3 BB should not overtake TS1 BB */
+	igt_assert_f(igt_compare_timestamps(ts1_buf[0], ts3_buf[0]),
+	             "TS1 ts (0x%08" PRIx32 ") > TS3 ts (0x%08" PRIx32 ")\n",
+	             ts1_buf[0], ts3_buf[0]);
+
+	/* Cleanup */
+	for(loop = 0; loop < in_flight; loop++)
+		intel_batchbuffer_free(in_flight_bbs[loop]);
+	intel_batchbuffer_free(ts1_bb);
+	intel_batchbuffer_free(ts2_bb);
+
+	drm_intel_bo_unreference(delay_bo);
+	drm_intel_bo_unreference(delay2_bo);
+	drm_intel_bo_unreference(ts1_bo);
+	drm_intel_bo_unreference(ts2_bo);
+	drm_intel_bo_unreference(ts3_bo);
+	for(loop = 0; loop < NBR_PRIO_STARVE_FDs; loop++) {
+		drm_intel_bufmgr_destroy(bufmgr[loop]);
+		close(fd[loop]);
+	}
+	free(in_flight_bbs);
+}
+
 /* Dependency test.
  * write=0, Submit batch buffers with read dependencies to all rings. Delay one
  * with a long executing batch buffer. Check the others are not held up.
@@ -482,6 +610,12 @@ igt_main
 			run_test_dependency(in_flight, loop, true);
 		}
 
+	for (loop = 0; loop < NBR_RINGS; loop++)
+		igt_subtest_f("%s-priority-starve", rings[loop].name) {
+			gem_require_ring(fd, rings[loop].id);
+			run_test_priority_starve(in_flight, rings[loop].id);
+		}
+
 	igt_fixture {
 		close(fd);
 	}
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx