@@ -108,6 +108,9 @@ sysfs_heartbeat_interval_SOURCES = i915/sysfs_heartbeat_interval
TESTS_progs += sysfs_preempt_timeout
sysfs_preempt_timeout_SOURCES = i915/sysfs_preempt_timeout
+TESTS_progs += sysfs_timeslice_duration
+sysfs_timeslice_duration_SOURCES = i915/sysfs_timeslice_duration
+
TESTS_progs += gem_bad_reloc
gem_bad_reloc_SOURCES = i915/gem_bad_reloc.c
new file mode 100644
@@ -0,0 +1,503 @@
+/*
+ * Copyright © 2019 Intel Corporation
+ *
+ * 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 without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * 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 NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR 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 <dirent.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <inttypes.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include "drmtest.h" /* gem_quiescent_gpu()! */
+#include "i915/gem_engine_topology.h"
+#include "i915/gem_mman.h"
+#include "igt_dummyload.h"
+#include "igt_sysfs.h"
+#include "ioctl_wrappers.h" /* igt_require_gem()! */
+#include "intel_chipset.h"
+#include "intel_reg.h"
+#include "sw_sync.h"
+
+#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 bool __enable_hangcheck(int dir, bool state)
+{
+ return igt_sysfs_set(dir, "enable_hangcheck", state ? "1" : "0");
+}
+
+static bool enable_hangcheck(int i915, bool state)
+{
+ bool success;
+ int dir;
+
+ dir = igt_sysfs_open_parameters(i915);
+ if (dir < 0) /* no parameters, must be default! */
+ return false;
+
+ success = __enable_hangcheck(dir, state);
+ close(dir);
+
+ return success;
+}
+
+static void set_timeslice(int engine, unsigned int value)
+{
+ unsigned int delay;
+
+ igt_sysfs_printf(engine, "timeslice_duration_ms", "%u", value);
+ igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &delay);
+ igt_assert_eq(delay, value);
+}
+
+static void test_idempotent(int i915, int engine)
+{
+ const unsigned int delays[] = { 0, 1, 1234, 654321 };
+ unsigned int saved;
+
+ /* Quick test to verify the kernel reports the same values as we write */
+
+ igt_assert(igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &saved) == 1);
+ igt_debug("Initial timeslice_duration_ms:%u\n", saved);
+
+ for (int i = 0; i < ARRAY_SIZE(delays); i++)
+ set_timeslice(engine, delays[i]);
+
+ set_timeslice(engine, saved);
+}
+
+static void test_invalid(int i915, int engine)
+{
+ unsigned int saved, delay;
+
+ /* Quick test that non-representable delays are rejected */
+
+ igt_assert(igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &saved) == 1);
+ igt_debug("Initial timeslice_duration_ms:%u\n", saved);
+
+ igt_sysfs_printf(engine, "timeslice_duration_ms", PRIu64, -1);
+ igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &delay);
+ igt_assert_eq(delay, saved);
+
+ igt_sysfs_printf(engine, "timeslice_duration_ms", "%d", -1);
+ igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &delay);
+ igt_assert_eq(delay, saved);
+
+ igt_sysfs_printf(engine, "timeslice_duration_ms", PRIu64, 123ull << 32);
+ igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &delay);
+ igt_assert_eq(delay, saved);
+}
+
+static void set_unbannable(int i915, uint32_t ctx)
+{
+ struct drm_i915_gem_context_param p = {
+ .ctx_id = ctx,
+ .param = I915_CONTEXT_PARAM_BANNABLE,
+ };
+
+ igt_assert_eq(__gem_context_set_param(i915, &p), 0);
+}
+
+static uint32_t create_context(int i915, unsigned int class, unsigned int inst, int prio)
+{
+ uint32_t ctx;
+
+ ctx = gem_context_create_for_engine(i915, class, inst);
+ set_unbannable(i915, ctx);
+ gem_context_set_priority(i915, ctx, prio);
+
+ return ctx;
+}
+
+static int cmp_u32(const void *_a, const void *_b)
+{
+ const uint32_t *a = _a, *b = _b;
+
+ return *a - *b;
+}
+
+static double clockrate(int i915)
+{
+ int freq;
+ drm_i915_getparam_t gp = {
+ .value = &freq,
+ .param = I915_PARAM_CS_TIMESTAMP_FREQUENCY,
+ };
+
+ igt_require(igt_ioctl(i915, DRM_IOCTL_I915_GETPARAM, &gp) == 0);
+ return 1e9 / freq;
+}
+
+static uint64_t __test_duration(int i915, int engine, unsigned int timeout)
+{
+ struct drm_i915_gem_exec_object2 obj[3] = {
+ {
+ .handle = gem_create(i915, 4096),
+ .offset = 0,
+ .flags = EXEC_OBJECT_PINNED,
+ },
+ {
+ .handle = gem_create(i915, 4096),
+ .offset = 4096,
+ .flags = EXEC_OBJECT_PINNED,
+ },
+ { gem_create(i915, 4096) }
+ };
+ struct drm_i915_gem_execbuffer2 eb = {
+ .buffer_count = ARRAY_SIZE(obj),
+ .buffers_ptr = to_user_pointer(obj),
+ };
+ const int gen = intel_gen(intel_get_drm_devid(i915));
+ double duration = clockrate(i915);
+ unsigned int class, inst, mmio;
+ uint32_t *cs, *map;
+ uint32_t ctx[2];
+ int start;
+ int i;
+
+ igt_require(gem_scheduler_has_preemption(i915));
+ igt_require(gen >= 8); /* MI_SEMAPHORE_WAIT */
+
+ igt_assert(igt_sysfs_scanf(engine, "class", "%u", &class) == 1);
+ igt_assert(igt_sysfs_scanf(engine, "instance", "%u", &inst) == 1);
+ igt_require(igt_sysfs_scanf(engine, "mmio_base", "%x", &mmio) == 1);
+
+ set_timeslice(engine, timeout);
+
+ ctx[0] = create_context(i915, class, inst, 0);
+ ctx[1] = create_context(i915, class, inst, 0);
+
+ map = gem_mmap__cpu(i915, obj[2].handle, 0, 4096, PROT_WRITE);
+
+ cs = map;
+ for (i = 0; i < 10; i++) {
+ *cs++ = MI_SEMAPHORE_WAIT |
+ MI_SEMAPHORE_POLL |
+ MI_SEMAPHORE_SAD_NEQ_SDD |
+ (4 - 2 + (gen >= 12));
+ *cs++ = 0;
+ *cs++ = obj[0].offset + sizeof(uint32_t) * i;
+ *cs++ = 0;
+ if (gen >= 12)
+ *cs++ = 0;
+
+ *cs++ = 0x24 << 23 | 2; /* SRM */
+ *cs++ = mmio + 0x358;
+ *cs++ = obj[1].offset + sizeof(uint32_t) * i;
+ *cs++ = 0;
+
+ *cs++ = MI_STORE_DWORD_IMM;
+ *cs++ = obj[0].offset +
+ 4096 - sizeof(uint32_t) * i - sizeof(uint32_t);
+ *cs++ = 0;
+ *cs++ = 1;
+ }
+ *cs++ = MI_BATCH_BUFFER_END;
+
+ cs += 16 - ((cs - map) & 15);
+ start = (cs - map) * sizeof(*cs);
+ for (i = 0; i < 10; i++) {
+ *cs++ = MI_STORE_DWORD_IMM;
+ *cs++ = obj[0].offset + sizeof(uint32_t) * i;
+ *cs++ = 0;
+ *cs++ = 1;
+
+ *cs++ = MI_SEMAPHORE_WAIT |
+ MI_SEMAPHORE_POLL |
+ MI_SEMAPHORE_SAD_NEQ_SDD |
+ (4 - 2 + (gen >= 12));
+ *cs++ = 0;
+ *cs++ = obj[0].offset +
+ 4096 - sizeof(uint32_t) * i - sizeof(uint32_t);
+ *cs++ = 0;
+ if (gen >= 12)
+ *cs++ = 0;
+ }
+ *cs++ = MI_BATCH_BUFFER_END;
+ igt_assert(cs - map < 4096 / sizeof(*cs));
+ munmap(map, 4096);
+
+ eb.rsvd1 = ctx[0];
+ gem_execbuf(i915, &eb);
+
+ eb.rsvd1 = ctx[1];
+ eb.batch_start_offset = start;
+ gem_execbuf(i915, &eb);
+
+ gem_sync(i915, obj[2].handle);
+
+ gem_set_domain(i915, obj[1].handle,
+ I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
+ map = gem_mmap__cpu(i915, obj[1].handle, 0, 4096, PROT_WRITE);
+ for (i = 0; i < 9; i++)
+ map[i] = map[i + 1] - map[i];
+ qsort(map, 9, sizeof(*map), cmp_u32);
+ duration *= map[4] / 2; /* 2 sema-waits between timestamp updates */
+ munmap(map, 4096);
+
+ for (i = 0; i < ARRAY_SIZE(ctx); i++)
+ gem_context_destroy(i915, ctx[i]);
+
+ for (i = 0; i < ARRAY_SIZE(obj); i++)
+ gem_close(i915, obj[i].handle);
+
+ return duration;
+}
+
+static void test_duration(int i915, int engine)
+{
+ int delays[] = { 1, 50, 100, 500 };
+ unsigned int saved;
+
+ /*
+ * Timeslicing at its very basic level is sharing the GPU by
+ * running one context for interval before running another. After
+ * each interval the running context is swapped for another runnable
+ * context.
+ *
+ * We can measure this directly by watching the xCS_TIMESTAMP and
+ * recording its value every time we switch into the context, using
+ * a couple of semaphores to busyspin for the timeslice.
+ */
+
+ igt_assert(igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &saved) == 1);
+ igt_debug("Initial timeslice_duration_ms:%u\n", saved);
+
+ gem_quiescent_gpu(i915);
+
+ for (int i = 0; i < ARRAY_SIZE(delays); i++) {
+ uint64_t elapsed;
+
+ elapsed = __test_duration(i915, engine, delays[i]);
+ igt_info("timeslice_duration_ms:%d, elapsed=%.3fms\n",
+ delays[i], elapsed * 1e-6);
+
+ /*
+ * We need to give a couple of jiffies slack for the scheduler timeouts
+ * and then a little more slack fr the overhead in submitting and
+ * measuring. 50ms should cover all of our sins and be useful
+ * tolerance.
+ */
+ igt_assert_f(elapsed / 1000 / 1000 < delays[i] + 50,
+ "Timeslice exceeded request!\n");
+ }
+
+ gem_quiescent_gpu(i915);
+ set_timeslice(engine, saved);
+}
+
+static uint64_t __test_timeout(int i915, int engine, unsigned int timeout)
+{
+ unsigned int class, inst;
+ struct timespec ts = {};
+ igt_spin_t *spin[2];
+ uint64_t elapsed;
+ uint32_t ctx[2];
+
+ igt_assert(igt_sysfs_scanf(engine, "class", "%u", &class) == 1);
+ igt_assert(igt_sysfs_scanf(engine, "instance", "%u", &inst) == 1);
+
+ set_timeslice(engine, timeout);
+
+ ctx[0] = create_context(i915, class, inst, 0);
+ spin[0] = igt_spin_new(i915, ctx[0],
+ .flags = (IGT_SPIN_NO_PREEMPTION |
+ IGT_SPIN_POLL_RUN |
+ IGT_SPIN_FENCE_OUT));
+ igt_spin_busywait_until_started(spin[0]);
+
+ ctx[1] = create_context(i915, class, inst, 0);
+ igt_nsec_elapsed(&ts);
+ spin[1] = igt_spin_new(i915, ctx[1], .flags = IGT_SPIN_POLL_RUN);
+ igt_spin_busywait_until_started(spin[1]);
+ elapsed = igt_nsec_elapsed(&ts);
+
+ igt_spin_free(i915, spin[1]);
+
+ igt_assert_eq(sync_fence_wait(spin[0]->out_fence, 1), 0);
+ igt_assert_eq(sync_fence_status(spin[0]->out_fence), -EIO);
+
+ igt_spin_free(i915, spin[0]);
+
+ gem_context_destroy(i915, ctx[1]);
+ gem_context_destroy(i915, ctx[0]);
+ gem_quiescent_gpu(i915);
+
+ return elapsed;
+}
+
+static void test_timeout(int i915, int engine)
+{
+ int delays[] = { 1, 50, 100, 500 };
+ unsigned int saved;
+
+ /*
+ * Timeslicing requires us to preempt the running context in order to
+ * switch into its contemporary. If we couple a unpreemptable hog
+ * with a fast forced reset, we can measure the timeslice by how long
+ * it takes for the hog to be reset and the high priority context
+ * to complete.
+ */
+
+ igt_require(igt_sysfs_printf(engine, "preempt_timeout_ms", "%u", 1) == 1);
+ igt_assert(igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &saved) == 1);
+ igt_debug("Initial timeslice_duration_ms:%u\n", saved);
+
+ gem_quiescent_gpu(i915);
+ igt_require(enable_hangcheck(i915, false));
+
+ for (int i = 0; i < ARRAY_SIZE(delays); i++) {
+ uint64_t elapsed;
+
+ elapsed = __test_timeout(i915, engine, delays[i]);
+ igt_info("timeslice_duration_ms:%d, elapsed=%.3fms\n",
+ delays[i], elapsed * 1e-6);
+
+ /*
+ * We need to give a couple of jiffies slack for the scheduler timeouts
+ * and then a little more slack fr the overhead in submitting and
+ * measuring. 50ms should cover all of our sins and be useful
+ * tolerance.
+ */
+ igt_assert_f(elapsed / 1000 / 1000 < delays[i] + 50,
+ "Timeslice exceeded request!\n");
+ }
+
+ igt_assert(enable_hangcheck(i915, true));
+ gem_quiescent_gpu(i915);
+ set_timeslice(engine, saved);
+}
+
+static void test_off(int i915, int engine)
+{
+ unsigned int class, inst;
+ unsigned int saved;
+ igt_spin_t *spin[2];
+ uint32_t ctx[2];
+
+ /*
+ * As always, there are some who must run uninterrupted and simply do
+ * not want to share the GPU even for a microsecond. Those greedy
+ * clients can disable timeslicing entirely, and so set the timeslice
+ * to 0. We test that a hog is not preempted within the 150s of
+ * our boredom threshold.
+ */
+
+ igt_require(igt_sysfs_printf(engine, "preempt_timeout_ms", "%u", 1) == 1);
+ igt_assert(igt_sysfs_scanf(engine, "timeslice_duration_ms", "%u", &saved) == 1);
+ igt_debug("Initial timeslice_duration_ms:%u\n", saved);
+
+ gem_quiescent_gpu(i915);
+ igt_require(enable_hangcheck(i915, false));
+
+ igt_assert(igt_sysfs_scanf(engine, "class", "%u", &class) == 1);
+ igt_assert(igt_sysfs_scanf(engine, "instance", "%u", &inst) == 1);
+
+ set_timeslice(engine, 0);
+
+ ctx[0] = create_context(i915, class, inst, 0);
+ spin[0] = igt_spin_new(i915, ctx[0],
+ .flags = (IGT_SPIN_NO_PREEMPTION |
+ IGT_SPIN_POLL_RUN |
+ IGT_SPIN_FENCE_OUT));
+ igt_spin_busywait_until_started(spin[0]);
+
+ ctx[1] = create_context(i915, class, inst, 0);
+ spin[1] = igt_spin_new(i915, ctx[1], .flags = IGT_SPIN_POLL_RUN);
+
+ for (int i = 0; i < 150; i++) {
+ igt_assert_eq(sync_fence_status(spin[0]->out_fence), 0);
+ sleep(1);
+ }
+
+ set_timeslice(engine, 1);
+
+ igt_spin_busywait_until_started(spin[1]);
+ igt_spin_free(i915, spin[1]);
+
+ igt_assert_eq(sync_fence_wait(spin[0]->out_fence, 1), 0);
+ igt_assert_eq(sync_fence_status(spin[0]->out_fence), -EIO);
+
+ igt_spin_free(i915, spin[0]);
+
+ gem_context_destroy(i915, ctx[1]);
+ gem_context_destroy(i915, ctx[0]);
+
+ igt_assert(enable_hangcheck(i915, true));
+ gem_quiescent_gpu(i915);
+
+ set_timeslice(engine, saved);
+}
+
+igt_main
+{
+ int i915 = -1, engines = -1;
+
+ igt_fixture {
+ int sys;
+
+ i915 = drm_open_driver(DRIVER_INTEL);
+ igt_require_gem(i915);
+ igt_allow_hang(i915, 0, 0);
+
+ sys = igt_sysfs_open(i915);
+ igt_require(sys != -1);
+
+ engines = openat(sys, "engine", O_RDONLY);
+ igt_require(engines != -1);
+
+ close(sys);
+ }
+
+ igt_subtest_with_dynamic("idempotent")
+ dyn_sysfs_engines(i915, engines, "timeslice_duration_ms",
+ test_idempotent);
+
+ igt_subtest_with_dynamic("invalid")
+ dyn_sysfs_engines(i915, engines, "timeslice_duration_ms",
+ test_invalid);
+
+ igt_subtest_with_dynamic("duration")
+ dyn_sysfs_engines(i915, engines, "timeslice_duration_ms",
+ test_duration);
+
+ igt_subtest_with_dynamic("timeout")
+ dyn_sysfs_engines(i915, engines, "timeslice_duration_ms",
+ test_timeout);
+
+ igt_subtest_with_dynamic("off")
+ dyn_sysfs_engines(i915, engines, "timeslice_duration_ms",
+ test_off);
+
+ igt_fixture {
+ close(engines);
+ close(i915);
+ }
+}
@@ -240,6 +240,7 @@ i915_progs = [
'i915_suspend',
'sysfs_heartbeat_interval',
'sysfs_preempt_timeout',
+ 'sysfs_timeslice_duration',
]
test_deps = [ igt_deps ]
We [will] expose various per-engine scheduling controls. One of which, 'timeslice_duration_ms', defines the scheduling quantum. If a context exhausts its timeslice, it will be preempted in favour of running one of its compatriots. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> --- tests/Makefile.sources | 3 + tests/i915/sysfs_timeslice_duration.c | 503 ++++++++++++++++++++++++++ tests/meson.build | 1 + 3 files changed, 507 insertions(+) create mode 100644 tests/i915/sysfs_timeslice_duration.c