[PATCH v3 15/19] lib/amdgpu: use a memory fence to serialize write

[Sunil Khatri]

Use a memory fence to serialize all the writes in the
queue before updating wptr and ringing the doorbell

This ensures memory is written in same sequence as
intended to.

Signed-off-by: Sunil Khatri <sunil.khatri at amd.com>
---
 lib/amdgpu/amd_user_queue.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/lib/amdgpu/amd_user_queue.c b/lib/amdgpu/amd_user_queue.c
index d1763f5d6..0235b07ac 100644
--- a/lib/amdgpu/amd_user_queue.c
+++ b/lib/amdgpu/amd_user_queue.c
@@ -148,8 +148,17 @@ void amdgpu_user_queue_submit(amdgpu_device_handle device, struct amdgpu_ring_co
 	/* empty dword is needed for fence signal pm4 */
 	amdgpu_pkt_add_dw(0);
 
+#if DETECT_CC_GCC && (DETECT_ARCH_X86 || DETECT_ARCH_X86_64)
+	asm volatile ("mfence" : : : "memory");
+#endif
+
+	/* Below call update the wptr address so will wait till all writes are completed */
 	amdgpu_pkt_end();
 
+#if DETECT_CC_GCC && (DETECT_ARCH_X86 || DETECT_ARCH_X86_64)
+	asm volatile ("mfence" : : : "memory");
+#endif
+
 	/* Update the door bell */
 	ring_context->doorbell_cpu[DOORBELL_INDEX] = *ring_context->wptr_cpu;
 
-- 
2.43.0