On 01/10/2019 10:51, Chris Wilson wrote:
If execlists's lite-restore is based on the common GEM context tag
rather than the per-intel_context LRCA, then a context switch between
two intel_contexts on the same engine derived from the same GEM context
will perform a lite-restore instead of a full context switch. We can
exploit this by poisoning the ringbuffer of the first context and trying
to trick a simple RING_TAIL update (i.e. lite-restore)
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
---
Then switch back to ce[0] for fun.
---
drivers/gpu/drm/i915/gt/selftest_lrc.c | 132 +++++++++++++++++++++++++
1 file changed, 132 insertions(+)
diff --git a/drivers/gpu/drm/i915/gt/selftest_lrc.c b/drivers/gpu/drm/i915/gt/selftest_lrc.c
index 93f2fcdc49bf..b90b970a44b9 100644
--- a/drivers/gpu/drm/i915/gt/selftest_lrc.c
+++ b/drivers/gpu/drm/i915/gt/selftest_lrc.c
@@ -79,6 +79,137 @@ static int live_sanitycheck(void *arg)
return err;
}
+static int live_unlite_restore(void *arg)
+{
+ struct drm_i915_private *i915 = arg;
+ struct intel_engine_cs *engine;
+ struct i915_gem_context *ctx;
+ enum intel_engine_id id;
+ intel_wakeref_t wakeref;
+ struct igt_spinner spin;
+ int err = -ENOMEM;
+
+ /*
+ * Check that we can correctly context switch between 2 instances
+ * on the same engine from the same parent context.
+ */
+
+ mutex_lock(&i915->drm.struct_mutex);
+ wakeref = intel_runtime_pm_get(&i915->runtime_pm);
+
+ if (igt_spinner_init(&spin, &i915->gt))
+ goto err_unlock;
+
+ ctx = kernel_context(i915);
+ if (!ctx)
+ goto err_spin;
+
+ for_each_engine(engine, i915, id) {
+ struct intel_context *ce[2] = {};
+ struct i915_request *rq[3];
+ struct igt_live_test t;
+ int n;
+
+ if (!intel_engine_can_store_dword(engine))
+ continue;
+
+ if (igt_live_test_begin(&t, i915, __func__, engine->name)) {
+ err = -EIO;
+ break;
+ }
+
+ for (n = 0; n < ARRAY_SIZE(ce); n++) {
+ ce[n] = intel_context_create(ctx, engine);
+ if (IS_ERR(ce[n])) {
+ err = PTR_ERR(ce[n]);
+ goto err_ce;
+ }
+
+ err = intel_context_pin(ce[n]);
+ if (err)
+ goto err_ce;
If pinning fails err_ce path will underflow the unpin. Perhaps you need
to only store in ce[] when both steps have passed and keep it in a local
until then.
+
+ /*
+ * Setup the pair of contexts such that if we
+ * lite-restore using the RING_TAIL from ce[1] it
+ * will execute garbage from ce[0]->ring.
+ */
+ memset(ce[n]->ring->vaddr,
+ POISON_INUSE,
+ ce[n]->ring->vma->size);
+ }
+ intel_ring_reset(ce[1]->ring, ce[1]->ring->vma->size / 2);
+ __execlists_update_reg_state(ce[1], engine);
+
+ rq[0] = igt_spinner_create_request(&spin, ce[0], MI_ARB_CHECK);
+ if (IS_ERR(rq[0])) {
+ err = PTR_ERR(rq[0]);
+ goto err_ce;
+ }
+
+ GEM_BUG_ON(rq[0]->tail > ce[1]->ring->emit);
+ i915_request_get(rq[0]);
+ i915_request_add(rq[0]);
+
+ if (!igt_wait_for_spinner(&spin, rq[0])) {
+ i915_request_put(rq[0]);
+ goto err_ce;
+ }
+
+ rq[1] = i915_request_create(ce[1]);
+ if (IS_ERR(rq[1])) {
+ err = PTR_ERR(rq[1]);
+ i915_request_put(rq[0]);
+ goto err_ce;
+ }
+ GEM_BUG_ON(rq[1]->tail <= rq[0]->tail);
+
+ /* Ensure we do a completion switch from ce[0] to ce[1] */
+ i915_request_await_dma_fence(rq[1], &rq[0]->fence);
What do you mean by completion switch? You are setting up a dependency
so rq[1] (and rq[2]) won't be put into the elsp until spinner is ended
so it may not even be a context switch. Wouldn't you actually need the
opposite? I was expecting you would let the spinner run, make sure rq[1]
is in elsp and then count on time slicing to trigger a context switch.
Regards,
Tvrtko
+ i915_request_put(rq[0]);
+
+ i915_request_get(rq[1]);
+ i915_request_add(rq[1]);
+
+ /* And switch back to ce[0] for good measure */
+ rq[2] = i915_request_create(ce[0]);
+ if (IS_ERR(rq[2])) {
+ err = PTR_ERR(rq[2]);
+ i915_request_put(rq[1]);
+ goto err_ce;
+ }
+ GEM_BUG_ON(rq[2]->tail > rq[1]->tail);
+
+ i915_request_await_dma_fence(rq[2], &rq[1]->fence);
+ i915_request_put(rq[1]);
+
+ i915_request_add(rq[2]);
+
+err_ce:
+ igt_spinner_end(&spin);
+ for (n = 0; n < ARRAY_SIZE(ce); n++) {
+ if (IS_ERR_OR_NULL(ce[n]))
+ break;
+
+ intel_context_unpin(ce[n]);
+ intel_context_put(ce[n]);
+ }
+
+ if (igt_live_test_end(&t))
+ err = -EIO;
+ if (err)
+ break;
+ }
+
+ kernel_context_close(ctx);
+err_spin:
+ igt_spinner_fini(&spin);
+err_unlock:
+ intel_runtime_pm_put(&i915->runtime_pm, wakeref);
+ mutex_unlock(&i915->drm.struct_mutex);
+ return err;
+}
+
static int
emit_semaphore_chain(struct i915_request *rq, struct i915_vma *vma, int idx)
{
@@ -2178,6 +2309,7 @@ int intel_execlists_live_selftests(struct drm_i915_private *i915)
{
static const struct i915_subtest tests[] = {
SUBTEST(live_sanitycheck),
+ SUBTEST(live_unlite_restore),
SUBTEST(live_timeslice_preempt),
SUBTEST(live_busywait_preempt),
SUBTEST(live_preempt),
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
