i915/gem_exec_schedule: Try to spot unfairness

An important property for multi-client systems is that each client gets
a 'fair' allotment of system time. (Where fairness is at the whim of the
context properties, such as priorities.) This test forks N independent
clients (albeit they happen to share a single vm), and does an equal
amount of work in client and asserts that they take an equal amount of
time.

Though we have never claimed to have a completely fair scheduler, that
is what is expected.

v2: igt_assert_f and more commentary; exclude vip from client stats,
include range of frame intervals from each individual client
v3: Write down what the test actually does!

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Ramalingam C <ramalingam.c@intel.com>

1 files changed, 960 insertions, 0 deletions

diff --git a/tests/i915/gem_exec_schedule.c b/tests/i915/gem_exec_schedule.c
index f23d63ac..3c950b06 100644
--- a/tests/i915/gem_exec_schedule.c
+++ b/tests/i915/gem_exec_schedule.c
@@ -29,6 +29,7 @@
 #include <sys/poll.h>
 #include <sys/ioctl.h>
 #include <sys/mman.h>
+#include <sys/resource.h>
 #include <sys/syscall.h>
 #include <sched.h>
 #include <signal.h>
@@ -2516,6 +2517,932 @@ static void measure_semaphore_power(int i915)
 	rapl_close(&pkg);
 }
 
+static int read_timestamp_frequency(int i915)
+{
+	int value = 0;
+	drm_i915_getparam_t gp = {
+		.value = &value,
+		.param = I915_PARAM_CS_TIMESTAMP_FREQUENCY,
+	};
+	ioctl(i915, DRM_IOCTL_I915_GETPARAM, &gp);
+	return value;
+}
+
+static uint64_t div64_u64_round_up(uint64_t x, uint64_t y)
+{
+	return (x + y - 1) / y;
+}
+
+static uint64_t ns_to_ctx_ticks(int i915, uint64_t ns)
+{
+	int f = read_timestamp_frequency(i915);
+	if (intel_gen(intel_get_drm_devid(i915)) == 11)
+		f = 12500000; /* icl!!! are you feeling alright? CTX vs CS */
+	return div64_u64_round_up(ns * f, NSEC_PER_SEC);
+}
+
+static uint64_t ticks_to_ns(int i915, uint64_t ticks)
+{
+	return div64_u64_round_up(ticks * NSEC_PER_SEC,
+				  read_timestamp_frequency(i915));
+}
+
+#define MI_INSTR(opcode, flags) (((opcode) << 23) | (flags))
+
+#define MI_MATH(x)                      MI_INSTR(0x1a, (x) - 1)
+#define MI_MATH_INSTR(opcode, op1, op2) ((opcode) << 20 | (op1) << 10 | (op2))
+/* Opcodes for MI_MATH_INSTR */
+#define   MI_MATH_NOOP                  MI_MATH_INSTR(0x000, 0x0, 0x0)
+#define   MI_MATH_LOAD(op1, op2)        MI_MATH_INSTR(0x080, op1, op2)
+#define   MI_MATH_LOADINV(op1, op2)     MI_MATH_INSTR(0x480, op1, op2)
+#define   MI_MATH_LOAD0(op1)            MI_MATH_INSTR(0x081, op1)
+#define   MI_MATH_LOAD1(op1)            MI_MATH_INSTR(0x481, op1)
+#define   MI_MATH_ADD                   MI_MATH_INSTR(0x100, 0x0, 0x0)
+#define   MI_MATH_SUB                   MI_MATH_INSTR(0x101, 0x0, 0x0)
+#define   MI_MATH_AND                   MI_MATH_INSTR(0x102, 0x0, 0x0)
+#define   MI_MATH_OR                    MI_MATH_INSTR(0x103, 0x0, 0x0)
+#define   MI_MATH_XOR                   MI_MATH_INSTR(0x104, 0x0, 0x0)
+#define   MI_MATH_STORE(op1, op2)       MI_MATH_INSTR(0x180, op1, op2)
+#define   MI_MATH_STOREINV(op1, op2)    MI_MATH_INSTR(0x580, op1, op2)
+/* Registers used as operands in MI_MATH_INSTR */
+#define   MI_MATH_REG(x)                (x)
+#define   MI_MATH_REG_SRCA              0x20
+#define   MI_MATH_REG_SRCB              0x21
+#define   MI_MATH_REG_ACCU              0x31
+#define   MI_MATH_REG_ZF                0x32
+#define   MI_MATH_REG_CF                0x33
+
+#define MI_LOAD_REGISTER_REG    MI_INSTR(0x2A, 1)
+
+static void delay(int i915,
+		  const struct intel_execution_engine2 *e,
+		  uint32_t handle,
+		  uint64_t addr,
+		  uint64_t ns)
+{
+	const int use_64b = intel_gen(intel_get_drm_devid(i915)) >= 8;
+	const uint32_t base = gem_engine_mmio_base(i915, e->name);
+#define CS_GPR(x) (base + 0x600 + 8 * (x))
+#define RUNTIME (base + 0x3a8)
+	enum { START_TS, NOW_TS };
+	uint32_t *map, *cs, *jmp;
+
+	igt_require(base);
+
+	/* Loop until CTX_TIMESTAMP - initial > @ns */
+
+	cs = map = gem_mmap__device_coherent(i915, handle, 0, 4096, PROT_WRITE);
+
+	*cs++ = MI_LOAD_REGISTER_IMM;
+	*cs++ = CS_GPR(START_TS) + 4;
+	*cs++ = 0;
+	*cs++ = MI_LOAD_REGISTER_REG;
+	*cs++ = RUNTIME;
+	*cs++ = CS_GPR(START_TS);
+
+	while (offset_in_page(cs) & 63)
+		*cs++ = 0;
+	jmp = cs;
+
+	*cs++ = 0x5 << 23; /* MI_ARB_CHECK */
+
+	*cs++ = MI_LOAD_REGISTER_IMM;
+	*cs++ = CS_GPR(NOW_TS) + 4;
+	*cs++ = 0;
+	*cs++ = MI_LOAD_REGISTER_REG;
+	*cs++ = RUNTIME;
+	*cs++ = CS_GPR(NOW_TS);
+
+	/* delta = now - start; inverted to match COND_BBE */
+	*cs++ = MI_MATH(4);
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(NOW_TS));
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(START_TS));
+	*cs++ = MI_MATH_SUB;
+	*cs++ = MI_MATH_STOREINV(MI_MATH_REG(NOW_TS), MI_MATH_REG_ACCU);
+
+	/* Save delta for reading by COND_BBE */
+	*cs++ = 0x24 << 23 | (1 + use_64b); /* SRM */
+	*cs++ = CS_GPR(NOW_TS);
+	*cs++ = addr + 4000;
+	*cs++ = addr >> 32;
+
+	/* Delay between SRM and COND_BBE to post the writes */
+	for (int n = 0; n < 8; n++) {
+		*cs++ = MI_STORE_DWORD_IMM;
+		if (use_64b) {
+			*cs++ = addr + 4064;
+			*cs++ = addr >> 32;
+		} else {
+			*cs++ = 0;
+			*cs++ = addr + 4064;
+		}
+		*cs++ = 0;
+	}
+
+	/* Break if delta [time elapsed] > ns */
+	*cs++ = MI_COND_BATCH_BUFFER_END | MI_DO_COMPARE | (1 + use_64b);
+	*cs++ = ~ns_to_ctx_ticks(i915, ns);
+	*cs++ = addr + 4000;
+	*cs++ = addr >> 32;
+
+	/* Otherwise back to recalculating delta */
+	*cs++ = MI_BATCH_BUFFER_START | 1 << 8 | use_64b;
+	*cs++ = addr + offset_in_page(jmp);
+	*cs++ = addr >> 32;
+
+	munmap(map, 4096);
+}
+
+static struct drm_i915_gem_exec_object2
+delay_create(int i915, uint32_t ctx,
+	     const struct intel_execution_engine2 *e,
+	     uint64_t target_ns)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = batch_create(i915),
+		.flags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS,
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.rsvd1 = ctx,
+		.flags = e->flags,
+	};
+
+	obj.offset = obj.handle << 12;
+	gem_execbuf(i915, &execbuf);
+	gem_sync(i915, obj.handle);
+
+	delay(i915, e, obj.handle, obj.offset, target_ns);
+
+	obj.flags |= EXEC_OBJECT_PINNED;
+	return obj;
+}
+
+static void tslog(int i915,
+		  const struct intel_execution_engine2 *e,
+		  uint32_t handle,
+		  uint64_t addr)
+{
+	const int use_64b = intel_gen(intel_get_drm_devid(i915)) >= 8;
+	const uint32_t base = gem_engine_mmio_base(i915, e->name);
+#define CS_GPR(x) (base + 0x600 + 8 * (x))
+#define CS_TIMESTAMP (base + 0x358)
+	enum { INC, MASK, ADDR };
+	uint32_t *timestamp_lo, *addr_lo;
+	uint32_t *map, *cs;
+
+	igt_require(base);
+
+	map = gem_mmap__device_coherent(i915, handle, 0, 4096, PROT_WRITE);
+	cs = map + 512;
+
+	/* Record the current CS_TIMESTAMP into a journal [a 512 slot ring]. */
+	*cs++ = 0x24 << 23 | (1 + use_64b); /* SRM */
+	*cs++ = CS_TIMESTAMP;
+	timestamp_lo = cs;
+	*cs++ = addr;
+	*cs++ = addr >> 32;
+
+	/* Load the address + inc & mask variables */
+	*cs++ = MI_LOAD_REGISTER_IMM;
+	*cs++ = CS_GPR(ADDR);
+	addr_lo = cs;
+	*cs++ = addr;
+	*cs++ = MI_LOAD_REGISTER_IMM;
+	*cs++ = CS_GPR(ADDR) + 4;
+	*cs++ = addr >> 32;
+
+	*cs++ = MI_LOAD_REGISTER_IMM;
+	*cs++ = CS_GPR(INC);
+	*cs++ = 4;
+	*cs++ = MI_LOAD_REGISTER_IMM;
+	*cs++ = CS_GPR(INC) + 4;
+	*cs++ = 0;
+
+	*cs++ = MI_LOAD_REGISTER_IMM;
+	*cs++ = CS_GPR(MASK);
+	*cs++ = 0xfffff7ff;
+	*cs++ = MI_LOAD_REGISTER_IMM;
+	*cs++ = CS_GPR(MASK) + 4;
+	*cs++ = 0xffffffff;
+
+	/* Increment the [ring] address for saving CS_TIMESTAMP */
+	*cs++ = MI_MATH(8);
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(INC));
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(ADDR));
+	*cs++ = MI_MATH_ADD;
+	*cs++ = MI_MATH_STORE(MI_MATH_REG(ADDR), MI_MATH_REG_ACCU);
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(ADDR));
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(MASK));
+	*cs++ = MI_MATH_AND;
+	*cs++ = MI_MATH_STORE(MI_MATH_REG(ADDR), MI_MATH_REG_ACCU);
+
+	/* Rewrite the batch buffer for the next execution */
+	*cs++ = 0x24 << 23 | (1 + use_64b); /* SRM */
+	*cs++ = CS_GPR(ADDR);
+	*cs++ = addr + offset_in_page(timestamp_lo);
+	*cs++ = addr >> 32;
+	*cs++ = 0x24 << 23 | (1 + use_64b); /* SRM */
+	*cs++ = CS_GPR(ADDR);
+	*cs++ = addr + offset_in_page(addr_lo);
+	*cs++ = addr >> 32;
+
+	*cs++ = MI_BATCH_BUFFER_END;
+
+	munmap(map, 4096);
+}
+
+static struct drm_i915_gem_exec_object2
+tslog_create(int i915, uint32_t ctx, const struct intel_execution_engine2 *e)
+{
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = batch_create(i915),
+		.flags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS,
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.rsvd1 = ctx,
+		.flags = e->flags,
+	};
+
+	obj.offset = obj.handle << 12;
+	gem_execbuf(i915, &execbuf);
+	gem_sync(i915, obj.handle);
+
+	tslog(i915, e, obj.handle, obj.offset);
+
+	obj.flags |= EXEC_OBJECT_PINNED;
+	return obj;
+}
+
+static int cmp_u32(const void *A, const void *B)
+{
+	const uint32_t *a = A, *b = B;
+
+	if (*a < *b)
+		return -1;
+	else if (*a > *b)
+		return 1;
+	else
+		return 0;
+}
+
+static bool has_ctx_timestamp(int i915, const struct intel_execution_engine2 *e)
+{
+	const int gen = intel_gen(intel_get_drm_devid(i915));
+
+	if (gen == 8 && e->class == I915_ENGINE_CLASS_VIDEO)
+		return false; /* looks fubar */
+
+	return true;
+}
+
+static struct intel_execution_engine2
+pick_random_engine(int i915, const struct intel_execution_engine2 *not)
+{
+	const struct intel_execution_engine2 *e;
+	unsigned int count = 0;
+
+	__for_each_physical_engine(i915, e) {
+		if (e->flags == not->flags)
+			continue;
+		if (!gem_class_has_mutable_submission(i915, e->class))
+			continue;
+		count++;
+	}
+	if (!count)
+		return *not;
+
+	count = rand() % count;
+	__for_each_physical_engine(i915, e) {
+		if (e->flags == not->flags)
+			continue;
+		if (!gem_class_has_mutable_submission(i915, e->class))
+			continue;
+		if (!count--)
+			break;
+	}
+
+	return *e;
+}
+
+static void fair_child(int i915, uint32_t ctx,
+		       const struct intel_execution_engine2 *e,
+		       uint64_t frame_ns,
+		       int timeline,
+		       uint32_t common,
+		       unsigned int flags,
+		       unsigned long *ctl,
+		       unsigned long *median,
+		       unsigned long *iqr)
+#define F_SYNC		(1 << 0)
+#define F_PACE		(1 << 1)
+#define F_FLOW		(1 << 2)
+#define F_HALF		(1 << 3)
+#define F_SOLO		(1 << 4)
+#define F_SPARE		(1 << 5)
+#define F_NEXT		(1 << 6)
+#define F_VIP		(1 << 7)
+#define F_RRUL		(1 << 8)
+#define F_SHARE		(1 << 9)
+#define F_PING		(1 << 10)
+#define F_THROTTLE	(1 << 11)
+#define F_ISOLATE	(1 << 12)
+{
+	const int batches_per_frame = flags & F_SOLO ? 1 : 3;
+	struct drm_i915_gem_exec_object2 obj[4] = {
+		{},
+		{
+			.handle = common ?: gem_create(i915, 4096),
+		},
+		delay_create(i915, ctx, e, frame_ns / batches_per_frame),
+		delay_create(i915, ctx, e, frame_ns / batches_per_frame),
+	};
+	struct intel_execution_engine2 ping = *e;
+	int p_fence = -1, n_fence = -1;
+	unsigned long count = 0;
+	int n;
+
+	srandom(getpid());
+	if (flags & F_PING)
+		ping = pick_random_engine(i915, e);
+	obj[0] = tslog_create(i915, ctx, &ping);
+
+	while (!READ_ONCE(*ctl)) {
+		struct drm_i915_gem_execbuffer2 execbuf = {
+			.buffers_ptr = to_user_pointer(obj),
+			.buffer_count = 3,
+			.rsvd1 = ctx,
+			.rsvd2 = -1,
+			.flags = e->flags,
+		};
+
+		if (flags & F_FLOW) {
+			unsigned int seq;
+
+			seq = count;
+			if (flags & F_NEXT)
+				seq++;
+
+			execbuf.rsvd2 =
+				sw_sync_timeline_create_fence(timeline, seq);
+			execbuf.flags |= I915_EXEC_FENCE_IN;
+		}
+
+		execbuf.flags |= I915_EXEC_FENCE_OUT;
+		gem_execbuf_wr(i915, &execbuf);
+		n_fence = execbuf.rsvd2 >> 32;
+		execbuf.flags &= ~(I915_EXEC_FENCE_OUT | I915_EXEC_FENCE_IN);
+		for (n = 1; n < batches_per_frame; n++)
+			gem_execbuf(i915, &execbuf);
+		close(execbuf.rsvd2);
+
+		execbuf.buffer_count = 1;
+		execbuf.batch_start_offset = 2048;
+		execbuf.flags = ping.flags | I915_EXEC_FENCE_IN;
+		execbuf.rsvd2 = n_fence;
+		gem_execbuf(i915, &execbuf);
+
+		if (flags & F_PACE && p_fence != -1) {
+			struct pollfd pfd = {
+				.fd = p_fence,
+				.events = POLLIN,
+			};
+			poll(&pfd, 1, -1);
+		}
+		close(p_fence);
+
+		if (flags & F_SYNC) {
+			struct pollfd pfd = {
+				.fd = n_fence,
+				.events = POLLIN,
+			};
+			poll(&pfd, 1, -1);
+		}
+
+		if (flags & F_THROTTLE)
+			igt_ioctl(i915, DRM_IOCTL_I915_GEM_THROTTLE, 0);
+
+		igt_swap(obj[2], obj[3]);
+		igt_swap(p_fence, n_fence);
+		count++;
+	}
+	close(p_fence);
+
+	gem_close(i915, obj[3].handle);
+	gem_close(i915, obj[2].handle);
+	if (obj[1].handle != common)
+		gem_close(i915, obj[1].handle);
+
+	gem_sync(i915, obj[0].handle);
+	if (median) {
+		uint32_t *map;
+
+		/*
+		 * We recorded the CS_TIMESTAMP of each frame, and if
+		 * the GPU is being shared completely fairly, we expect
+		 * each frame to be at the same interval from the last.
+		 *
+		 * Compute the interval between frames and report back
+		 * both the median interval and the range for this client.
+		 */
+
+		map = gem_mmap__device_coherent(i915, obj[0].handle,
+						0, 4096, PROT_WRITE);
+		for (n = 1; n < min(count, 512); n++) {
+			igt_assert(map[n]);
+			map[n - 1] = map[n] - map[n - 1];
+		}
+		qsort(map, --n, sizeof(*map), cmp_u32);
+		*iqr = ticks_to_ns(i915, map[(3 * n + 3) / 4] - map[n / 4]);
+		*median = ticks_to_ns(i915, map[n / 2]);
+		munmap(map, 4096);
+	}
+	gem_close(i915, obj[0].handle);
+}
+
+static int cmp_ul(const void *A, const void *B)
+{
+	const unsigned long *a = A, *b = B;
+
+	if (*a < *b)
+		return -1;
+	else if (*a > *b)
+		return 1;
+	else
+		return 0;
+}
+
+static uint64_t d_cpu_time(const struct rusage *a, const struct rusage *b)
+{
+	uint64_t cpu_time = 0;
+
+	cpu_time += (a->ru_utime.tv_sec - b->ru_utime.tv_sec) * NSEC_PER_SEC;
+	cpu_time += (a->ru_utime.tv_usec - b->ru_utime.tv_usec) * 1000;
+
+	cpu_time += (a->ru_stime.tv_sec - b->ru_stime.tv_sec) * NSEC_PER_SEC;
+	cpu_time += (a->ru_stime.tv_usec - b->ru_stime.tv_usec) * 1000;
+
+	return cpu_time;
+}
+
+static void timeline_advance(int timeline, int delay_ns)
+{
+	struct timespec tv = { .tv_nsec = delay_ns };
+	nanosleep(&tv, NULL);
+	sw_sync_timeline_inc(timeline, 1);
+}
+
+static void fairness(int i915,
+		     const struct intel_execution_engine2 *e,
+		     int timeout, unsigned int flags)
+{
+	const int frame_ns = 16666 * 1000;
+	const int fence_ns = flags & F_HALF ? 2 * frame_ns : frame_ns;
+	unsigned long *result, *iqr;
+	uint32_t common = 0;
+
+	igt_require(has_ctx_timestamp(i915, e));
+	igt_require(gem_class_has_mutable_submission(i915, e->class));
+
+	if (flags & F_SHARE)
+		common = gem_create(i915, 4095);
+
+	result = mmap(NULL, 4096, PROT_WRITE, MAP_SHARED | MAP_ANON, -1, 0);
+	igt_assert(result != MAP_FAILED);
+	iqr = mmap(NULL, 4096, PROT_WRITE, MAP_SHARED | MAP_ANON, -1, 0);
+	igt_assert(iqr != MAP_FAILED);
+
+	/*
+	 * The combined workload always runs at a 60fps target (unless F_HALF!).
+	 * This gives a frame of interval of 16ms that is evenly split across
+	 * all the clients, so simulating a system with a bunch of clients that
+	 * are perfectly balanced and can sustain 60fps. Our job is to ensure
+	 * that each client does run at a smooth 60fps.
+	 *
+	 * Each client runs a fixed length delay loop (as a single request,
+	 * or split into 3) and then records the CS_TIMESTAMP after completing
+	 * its delay. Given a fair allotment of GPU time to each client,
+	 * that timestamp will [ideally] be at a precise 16ms intervals.
+	 * In practice, time is wasted on context switches, so as the number
+	 * of clients increases, the proprotion of time spent on context
+	 * switches grows. As we get to 64 render clients, we will be spending
+	 * as much time in context switches as executing the client workloads.
+	 *
+	 * Each client frame may be paced by some throttling technique found
+	 * in the wild. i.e. each client may wait until a simulated vblank
+	 * to indicate the start of a new frame, or it may wait until the
+	 * completion of a previous frame. This causes submission from each
+	 * client and across the system to be chunky and uneven.
+	 *
+	 * We look at the variation of frame intervals within each client, and
+	 * the variation of the medians across the clients to see if the
+	 * distribution (budget) of GPU time was fair enough.
+	 *
+	 * Alternative (and important) metrics will be more latency centric;
+	 * looking at how well we can sustain meeting deadline given competition
+	 * by clients for the GPU.
+	 */
+
+	for (int n = 2; n <= 256; n <<= 1) { /* 32 == 500us per client */
+		int timeline = sw_sync_timeline_create();
+		int nfences = timeout * NSEC_PER_SEC / fence_ns + 1;
+		int nchild = n - 1; /* odd for easy medians */
+		const int child_ns = frame_ns / (nchild + !!(flags & F_SPARE));
+		const int lo = nchild / 4;
+		const int hi = (3 * nchild + 3) / 4 - 1;
+		struct rusage old_usage, usage;
+		uint64_t cpu_time, d_time;
+		struct timespec tv;
+		struct igt_mean m;
+
+		memset(result, 0, (nchild + 1) * sizeof(result[0]));
+
+		if (flags & F_PING) { /* fill the others with light bg load */
+			struct intel_execution_engine2 *ping;
+
+			__for_each_physical_engine(i915, ping) {
+				if (ping->flags == e->flags)
+					continue;
+
+				igt_fork(child, 1) {
+					uint32_t ctx = gem_context_clone_with_engines(i915, 0);
+
+					fair_child(i915, ctx, ping,
+						   child_ns / 8,
+						   -1, common,
+						   F_SOLO | F_PACE | F_SHARE,
+						   &result[nchild],
+						   NULL, NULL);
+
+					gem_context_destroy(i915, ctx);
+				}
+			}
+		}
+
+		getrusage(RUSAGE_CHILDREN, &old_usage);
+		igt_nsec_elapsed(memset(&tv, 0, sizeof(tv)));
+		igt_fork(child, nchild) {
+			uint32_t ctx;
+
+			if (flags & F_ISOLATE) {
+				int clone, dmabuf = -1;
+
+				if (common)
+					dmabuf = prime_handle_to_fd(i915, common);
+
+				clone = gem_reopen_driver(i915);
+				gem_context_copy_engines(i915, 0, clone, 0);
+				i915 = clone;
+
+				if (dmabuf != -1)
+					common = prime_fd_to_handle(i915, dmabuf);
+			}
+
+			ctx = gem_context_clone_with_engines(i915, 0);
+
+			if (flags & F_VIP && child == 0) {
+				gem_context_set_priority(i915, ctx, MAX_PRIO);
+				flags |= F_FLOW;
+			}
+			if (flags & F_RRUL && child == 0)
+				flags |= F_SOLO | F_FLOW | F_SYNC;
+
+			fair_child(i915, ctx, e, child_ns,
+				   timeline, common, flags,
+				   &result[nchild],
+				   &result[child], &iqr[child]);
+
+			gem_context_destroy(i915, ctx);
+		}
+
+		while (nfences--)
+			timeline_advance(timeline, fence_ns);
+
+		result[nchild] = 1;
+		for (int child = 0; child < nchild; child++) {
+			while (!READ_ONCE(result[child]))
+				timeline_advance(timeline, fence_ns);
+		}
+
+		igt_waitchildren();
+		close(timeline);
+
+		/*
+		 * Are we running out of CPU time, and fail to submit frames?
+		 *
+		 * We try to rule out any undue impact on the GPU scheduling
+		 * from the CPU scheduler by looking for core saturation. If
+		 * we may be in a situation where the clients + kernel are
+		 * taking a whole core (think lockdep), then it is increasingly
+		 * likely that our measurements include delays from the CPU
+		 * scheduler. Err on the side of caution.
+		 */
+		d_time = igt_nsec_elapsed(&tv);
+		getrusage(RUSAGE_CHILDREN, &usage);
+		cpu_time = d_cpu_time(&usage, &old_usage);
+		igt_debug("CPU usage: %.0f%%\n", 100. * cpu_time / d_time);
+		if (4 * cpu_time > 3 * d_time) {
+			if (nchild > 7) /* good enough to judge pass/fail */
+				break;
+
+			igt_skip_on_f(4 * cpu_time > 3 * d_time,
+				      "%.0f%% CPU usage, presuming capacity exceeded\n",
+				      100. * cpu_time / d_time);
+		}
+
+		/* With no contention, we should match our target frametime */
+		if (nchild == 1) {
+			igt_assert(4 * result[0] > 3 * fence_ns &&
+				   3 * result[0] < 4 * fence_ns);
+			continue;
+		}
+
+		/*
+		 * The VIP should always be able to hit the target frame rate;
+		 * regardless of budget contention from lessor clients.
+		 */
+		if (flags & (F_VIP | F_RRUL)) {
+			igt_info("VIP interval %.2fms, range %.2fms\n",
+				 1e-6 * result[0], 1e-6 * iqr[0]);
+			igt_assert_f(4 * result[0] > 3 * fence_ns &&
+				     3 * result[0] < 4 * fence_ns,
+				     "VIP expects to run exactly when it wants, expects an interval of %.2fms, was %.2fms\n",
+				     1e-6 * fence_ns, 1e-6 * result[0]);
+			igt_assert_f(2 * iqr[0] < result[0],
+				     "VIP frame IQR %.2fms exceeded median threshold %.2fms\n",
+				     1e-6 * iqr[0],
+				     1e-6 * result[0] / 2);
+			if (!--nchild)
+				continue;
+
+			/* Exclude the VIP result from the plebian statistics */
+			memmove(result, result + 1, nchild * sizeof(*result));
+			memmove(iqr, iqr + 1, nchild * sizeof(*iqr));
+		}
+
+		igt_mean_init(&m);
+		for (int child = 0; child < nchild; child++)
+			igt_mean_add(&m, result[child]);
+
+		qsort(result, nchild, sizeof(*result), cmp_ul);
+		qsort(iqr, nchild, sizeof(*iqr), cmp_ul);
+
+		/*
+		 * The target interval for median/mean is 16ms (fence_ns).
+		 * However, this work is evenly split across the clients so
+		 * the range (and median) of client medians may be much less
+		 * than 16ms [16/3N]. We present median of medians to try
+		 * and avoid any instability while running in CI; at the cost
+		 * of insensitivity!
+		 */
+		igt_info("%3d clients, range: [%.1f, %.1f], iqr: [%.1f, %.1f], median: %.1f [%.1f, %.1f], mean: %.1f ± %.2f ms, cpu: %.0f%%\n",
+			 nchild,
+			 1e-6 * result[0],  1e-6 * result[nchild - 1],
+			 1e-6 * result[lo], 1e-6 * result[hi],
+			 1e-6 * result[nchild / 2],
+			 1e-6 * iqr[lo], 1e-6 * iqr[hi],
+			 1e-6 * igt_mean_get(&m),
+			 1e-6 * sqrt(igt_mean_get_variance(&m)),
+			 100. * cpu_time / d_time);
+
+		igt_assert_f(iqr[nchild / 2] < 2 * result[nchild / 2],
+			     "Child frame IQR %.2fms exceeded median threshold %.2fms\n",
+			     1e-6 * iqr[nchild / 2],
+			     1e-6 * result[nchild / 2] * 2);
+
+		igt_assert_f(4 * igt_mean_get(&m) > 3 * result[nchild / 2] &&
+			     3 * igt_mean_get(&m) < 4 * result[nchild / 2],
+			     "Mean of client interval %.2fms differs from median %.2fms, distribution is skewed\n",
+
+			     1e-6 * igt_mean_get(&m), 1e-6 * result[nchild / 2]);
+
+		igt_assert_f(result[nchild / 2] > frame_ns / 2,
+			     "Median client interval %.2fms did not match target interval %.2fms\n",
+			     1e-6 * result[nchild / 2], 1e-6 * frame_ns);
+
+
+		igt_assert_f(result[hi] - result[lo] < result[nchild / 2],
+			     "Interquartile range of client intervals %.2fms is as large as the median threshold %.2fms, clients are not evenly distributed!\n",
+			     1e-6 * (result[hi] - result[lo]),
+			     1e-6 * result[nchild / 2]);
+
+		/* May be slowed due to sheer volume of context switches */
+		if (result[0] > 2 * fence_ns)
+			break;
+	}
+
+	munmap(iqr, 4096);
+	munmap(result, 4096);
+	if (common)
+		gem_close(i915, common);
+}
+
+static void test_fairness(int i915, int timeout)
+{
+	static const struct {
+		const char *name;
+		unsigned int flags;
+	} fair[] = {
+		/*
+		 * none - maximal greed in each client
+		 *
+		 * Push as many frames from each client as fast as possible
+		 */
+		{ "none",       0 },
+		{ "none-vip",   F_VIP }, /* one vip client must meet deadlines */
+		{ "none-solo",  F_SOLO }, /* 1 batch per frame per client */
+		{ "none-share", F_SHARE }, /* read from a common buffer */
+		{ "none-rrul",  F_RRUL }, /* "realtime-response under load" */
+		{ "none-ping",  F_PING }, /* measure inter-engine fairness */
+
+		/*
+		 * throttle - original per client throttling
+		 *
+		 * Used for front buffering rendering where there is no
+		 * extenal frame marker. Each client tries to only keep
+		 * 20ms of work submitted, though that measurement is
+		 * flawed...
+		 *
+		 * This is used by Xorg to try and maintain some resembalance
+		 * of input/output consistency when being feed a continuous
+		 * stream of X11 draw requests straight into scanout, where
+		 * the clients may submit the work faster than can be drawn.
+		 *
+		 * Throttling tracks requests per-file (and assumes that
+		 * all requests are in submission order across the whole file),
+		 * so we split each child to its own fd.
+		 */
+		{ "throttle",       F_THROTTLE | F_ISOLATE },
+		{ "throttle-vip",   F_THROTTLE | F_ISOLATE | F_VIP },
+		{ "throttle-solo",  F_THROTTLE | F_ISOLATE | F_SOLO },
+		{ "throttle-share", F_THROTTLE | F_ISOLATE | F_SHARE },
+		{ "throttle-rrul",  F_THROTTLE | F_ISOLATE | F_RRUL },
+
+		/*
+		 * pace - mesa "submit double buffering"
+		 *
+		 * Submit a frame, wait for previous frame to start. This
+		 * prevents each client from getting too far ahead of its
+		 * rendering, maintaining a consistent input/output latency.
+		 */
+		{ "pace",       F_PACE },
+		{ "pace-solo",  F_PACE | F_SOLO },
+		{ "pace-share", F_PACE | F_SOLO | F_SHARE },
+		{ "pace-ping",  F_PACE | F_SOLO | F_SHARE | F_PING},
+
+		/* sync - only submit a frame at a time */
+		{ "sync",      F_SYNC },
+		{ "sync-vip",  F_SYNC | F_VIP },
+		{ "sync-solo", F_SYNC | F_SOLO },
+
+		/* flow - synchronise execution against the clock (vblank) */
+		{ "flow",       F_PACE | F_FLOW },
+		{ "flow-solo",  F_PACE | F_FLOW | F_SOLO },
+		{ "flow-share", F_PACE | F_FLOW | F_SHARE },
+		{ "flow-ping",  F_PACE | F_FLOW | F_SHARE | F_PING },
+
+		/* next - submit ahead of the clock (vblank double buffering) */
+		{ "next",       F_PACE | F_FLOW | F_NEXT },
+		{ "next-solo",  F_PACE | F_FLOW | F_NEXT | F_SOLO },
+		{ "next-share", F_PACE | F_FLOW | F_NEXT | F_SHARE },
+		{ "next-ping",  F_PACE | F_FLOW | F_NEXT | F_SHARE | F_PING },
+
+		/* spare - underutilise by a single client timeslice */
+		{ "spare",      F_PACE | F_FLOW | F_SPARE },
+		{ "spare-solo", F_PACE | F_FLOW | F_SPARE | F_SOLO },
+
+		/* half - run at half pace (submit 16ms of work every 32ms) */
+		{ "half",       F_PACE | F_FLOW | F_HALF },
+		{ "half-solo",  F_PACE | F_FLOW | F_HALF | F_SOLO },
+
+		{}
+	};
+
+	igt_fixture {
+		igt_info("CS timestamp frequency: %d\n",
+			 read_timestamp_frequency(i915));
+
+		igt_require(intel_gen(intel_get_drm_devid(i915)) >= 8);
+	}
+
+	for (typeof(*fair) *f = fair; f->name; f++) {
+		igt_subtest_with_dynamic_f("fair-%s", f->name)  {
+			const struct intel_execution_engine2 *e;
+
+			__for_each_physical_engine(i915, e) {
+				if (!gem_class_can_store_dword(i915, e->class))
+					continue;
+
+				igt_dynamic_f("%s", e->name)
+					fairness(i915, e, timeout, f->flags);
+			}
+		}
+	}
+}
+
+static uint32_t read_ctx_timestamp(int i915,
+				   uint32_t ctx,
+				   const struct intel_execution_engine2 *e)
+{
+	const int use_64b = intel_gen(intel_get_drm_devid(i915)) >= 8;
+	const uint32_t base = gem_engine_mmio_base(i915, e->name);
+	struct drm_i915_gem_relocation_entry reloc;
+	struct drm_i915_gem_exec_object2 obj = {
+		.handle = gem_create(i915, 4096),
+		.offset = 32 << 20,
+		.relocs_ptr = to_user_pointer(&reloc),
+		.relocation_count = 1,
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&obj),
+		.buffer_count = 1,
+		.flags = e->flags,
+		.rsvd1 = ctx,
+	};
+#define RUNTIME (base + 0x3a8)
+	uint32_t *map, *cs;
+	uint32_t ts;
+
+	igt_require(base);
+
+	cs = map = gem_mmap__device_coherent(i915, obj.handle,
+					     0, 4096, PROT_WRITE);
+
+	*cs++ = 0x24 << 23 | (1 + use_64b); /* SRM */
+	*cs++ = RUNTIME;
+	memset(&reloc, 0, sizeof(reloc));
+	reloc.target_handle = obj.handle;
+	reloc.presumed_offset = obj.offset;
+	reloc.offset = offset_in_page(cs);
+	reloc.delta = 4000;
+	*cs++ = obj.offset + 4000;
+	*cs++ = obj.offset >> 32;
+
+	*cs++ = MI_BATCH_BUFFER_END;
+
+	gem_execbuf(i915, &execbuf);
+	gem_sync(i915, obj.handle);
+	gem_close(i915, obj.handle);
+
+	ts = map[1000];
+	munmap(map, 4096);
+
+	return ts;
+}
+
+static void fairslice(int i915,
+		      const struct intel_execution_engine2 *e,
+		      unsigned long flags)
+{
+	igt_spin_t *spin = NULL;
+	uint32_t ctx[3];
+	uint32_t ts[3];
+
+	for (int i = 0; i < ARRAY_SIZE(ctx); i++) {
+		ctx[i] = gem_context_clone_with_engines(i915, 0);
+		if (spin == NULL) {
+			spin = __igt_spin_new(i915,
+					      .ctx = ctx[i],
+					      .engine = e->flags,
+					      .flags = flags);
+		} else {
+			struct drm_i915_gem_execbuffer2 eb = {
+				.buffer_count = 1,
+				.buffers_ptr = to_user_pointer(&spin->obj[IGT_SPIN_BATCH]),
+				.flags = e->flags,
+				.rsvd1 = ctx[i],
+			};
+			gem_execbuf(i915, &eb);
+		}
+	}
+
+	sleep(2); /* over the course of many timeslices */
+
+	igt_assert(gem_bo_busy(i915, spin->handle));
+	igt_spin_end(spin);
+	for (int i = 0; i < ARRAY_SIZE(ctx); i++)
+		ts[i] = read_ctx_timestamp(i915, ctx[i], e);
+
+	for (int i = 0; i < ARRAY_SIZE(ctx); i++)
+		gem_context_destroy(i915, ctx[i]);
+	igt_spin_free(i915, spin);
+
+	qsort(ts, 3, sizeof(*ts), cmp_u32);
+	igt_info("%s: [%.1f, %.1f, %.1f] ms\n", e->name,
+		 1e-6 * ticks_to_ns(i915, ts[0]),
+		 1e-6 * ticks_to_ns(i915, ts[1]),
+		 1e-6 * ticks_to_ns(i915, ts[2]));
+
+	igt_assert_f(ts[2], "CTX_TIMESTAMP not reported!\n");
+	igt_assert_f((ts[2] - ts[0]) * 6 < ts[1],
+		     "Range of timeslices greater than tolerable: %.2fms > %.2fms; unfair!\n",
+		     1e-6 * ticks_to_ns(i915, ts[2] - ts[0]),
+		     1e-6 * ticks_to_ns(i915, ts[1]) / 6);
+}
+
 #define test_each_engine(T, i915, e) \
 	igt_subtest_with_dynamic(T) __for_each_physical_engine(i915, e) \
 		igt_dynamic_f("%s", e->name)
@@ -2582,6 +3509,35 @@ igt_main
 		test_each_engine("lateslice", fd, e)
 			lateslice(fd, e->flags);
 
+		igt_subtest_group {
+			igt_fixture {
+				igt_require(gem_scheduler_has_semaphores(fd));
+				igt_require(gem_scheduler_has_preemption(fd));
+				igt_require(intel_gen(intel_get_drm_devid(fd)) >= 8);
+			}
+
+			test_each_engine("fairslice", fd, e)
+				fairslice(fd, e, 0);
+
+			test_each_engine("u-fairslice", fd, e)
+				fairslice(fd, e, IGT_SPIN_USERPTR);
+
+			igt_subtest("fairslice-all")  {
+				__for_each_physical_engine(fd, e) {
+					igt_fork(child, 1)
+						fairslice(fd, e, 0);
+				}
+				igt_waitchildren();
+			}
+			igt_subtest("u-fairslice-all")  {
+				__for_each_physical_engine(fd, e) {
+					igt_fork(child, 1)
+						fairslice(fd, e, IGT_SPIN_USERPTR);
+				}
+				igt_waitchildren();
+			}
+		}
+
 		test_each_engine("submit-early-slice", fd, e)
 			submit_slice(fd, e, EARLY_SUBMIT);
 		test_each_engine("submit-golden-slice", fd, e)
@@ -2611,6 +3567,10 @@ igt_main
 			promotion(fd, e->flags);
 
 		igt_subtest_group {
+			test_fairness(fd, 2);
+		}
+
+		igt_subtest_group {
 			igt_fixture {
 				igt_require(gem_scheduler_has_preemption(fd));
 			}