[v3] drm/i915/gt: Retry RING_HEAD reset until it sticks

Commit Message

Gote, Nitin R Oct. 14, 2024, 10:48 a.m. UTC

On Haswell, in particular, we see an issue where resets fails because
the engine resumes from an incorrect RING_HEAD. Since the RING_HEAD
doesn't point to the remaining requests to re-run, but may instead point
into the uninitialised portion of the ring, the GPU may be then fed
invalid instructions from a privileged context, oft pushing the GPU into
an unrecoverable hang.

If at first the write doesn't succeed, try, try again.

v2: Avoid unnecessary timeout macro (Andi)

v3: Correct comment format (Andi)

Link: https://gitlab.freedesktop.org/drm/intel/-/issues/5432
Testcase: igt/i915_selftest/hangcheck
Signed-off-by: Chris Wilson <[e-mail]>
Signed-off-by: Nitin Gote <[e-mail]>
---
 .../gpu/drm/i915/gt/intel_ring_submission.c   | 35 ++++++++++++++++---
 1 file changed, 31 insertions(+), 4 deletions(-)

Patch

diff --git a/drivers/gpu/drm/i915/gt/intel_ring_submission.c b/drivers/gpu/drm/i915/gt/intel_ring_submission.c
index 72277bc8322e..b5eb1b2915f3 100644
--- a/drivers/gpu/drm/i915/gt/intel_ring_submission.c
+++ b/drivers/gpu/drm/i915/gt/intel_ring_submission.c
@@ -192,6 +192,7 @@  static bool stop_ring(struct intel_engine_cs *engine)
 static int xcs_resume(struct intel_engine_cs *engine)
 {
 	struct intel_ring *ring = engine->legacy.ring;
+	ktime_t kt;
 
 	ENGINE_TRACE(engine, "ring:{HEAD:%04x, TAIL:%04x}\n",
 		     ring->head, ring->tail);
@@ -230,9 +231,31 @@  static int xcs_resume(struct intel_engine_cs *engine)
 	set_pp_dir(engine);
 
 	/* First wake the ring up to an empty/idle ring */
-	ENGINE_WRITE_FW(engine, RING_HEAD, ring->head);
+	if (IS_HASWELL(engine->i915)) {
+		/*
+		 * On Haswell, resets fails because engine resumes
+		 * from incorrect RING_HEAD and then GPU may be then
+		 * fed to invalid instrcutions, which may lead to unrecoverable hang.
+		 * So at first write doesn't succeed then try again.
+		 */
+		for ((kt) = ktime_get() + (2 * NSEC_PER_MSEC);
+				ktime_before(ktime_get(), (kt)); cpu_relax()) {
+			ENGINE_WRITE_FW(engine, RING_HEAD, ring->head);
+			if (ENGINE_READ_FW(engine, RING_HEAD) == ring->head)
+				break;
+		}
+	} else {
+		ENGINE_WRITE_FW(engine, RING_HEAD, ring->head);
+	}
+
 	ENGINE_WRITE_FW(engine, RING_TAIL, ring->head);
-	ENGINE_POSTING_READ(engine, RING_TAIL);
+	if (ENGINE_READ_FW(engine, RING_HEAD) != ENGINE_READ_FW(engine, RING_TAIL)) {
+		ENGINE_TRACE(engine, "failed to reset empty ring: [%x, %x]: %x\n",
+			     ENGINE_READ_FW(engine, RING_HEAD),
+			     ENGINE_READ_FW(engine, RING_TAIL),
+			     ring->head);
+		goto err;
+	}
 
 	ENGINE_WRITE_FW(engine, RING_CTL,
 			RING_CTL_SIZE(ring->size) | RING_VALID);
@@ -241,12 +264,16 @@  static int xcs_resume(struct intel_engine_cs *engine)
 	if (__intel_wait_for_register_fw(engine->uncore,
 					 RING_CTL(engine->mmio_base),
 					 RING_VALID, RING_VALID,
-					 5000, 0, NULL))
+					 5000, 0, NULL)) {
+		ENGINE_TRACE(engine, "failed to restart\n");
 		goto err;
+	}
 
-	if (GRAPHICS_VER(engine->i915) > 2)
+	if (GRAPHICS_VER(engine->i915) > 2) {
 		ENGINE_WRITE_FW(engine,
 				RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
+		ENGINE_POSTING_READ(engine, RING_MI_MODE);
+	}
 
 	/* Now awake, let it get started */
 	if (ring->tail != ring->head) {