[PATCH 2/3] drm/radeon: fix VM flush on SI (v3)

[Alex Deucher]

We need to wait for the GPUVM flush to complete.  There
was some confusion as to how this mechanism was supposed
to work.  The operation is not atomic.  For GPU initiated
invalidations you need to read back a VM register to
introduce enough latency for the update to complete.

v2: drop gart changes
v3: just read back rather than polling

Reviewed-by: Christian König <christian.koenig at amd.com>
Signed-off-by: Alex Deucher <alexander.deucher at amd.com>
Cc: stable at vger.kernel.org
---
 drivers/gpu/drm/radeon/si.c     | 10 ++++++++++
 drivers/gpu/drm/radeon/si_dma.c |  8 ++++++++
 drivers/gpu/drm/radeon/sid.h    | 18 ++++++++++++++++++
 3 files changed, 36 insertions(+)

diff --git a/drivers/gpu/drm/radeon/si.c b/drivers/gpu/drm/radeon/si.c
index 60df444..5d89b87 100644
--- a/drivers/gpu/drm/radeon/si.c
+++ b/drivers/gpu/drm/radeon/si.c
@@ -5057,6 +5057,16 @@ void si_vm_flush(struct radeon_device *rdev, struct radeon_ring *ring,
 	radeon_ring_write(ring, 0);
 	radeon_ring_write(ring, 1 << vm_id);
 
+	/* wait for the invalidate to complete */
+	radeon_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
+	radeon_ring_write(ring, (WAIT_REG_MEM_FUNCTION(0) |  /* always */
+				 WAIT_REG_MEM_ENGINE(0))); /* me */
+	radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
+	radeon_ring_write(ring, 0);
+	radeon_ring_write(ring, 0); /* ref */
+	radeon_ring_write(ring, 0); /* mask */
+	radeon_ring_write(ring, 0x20); /* poll interval */
+
 	/* sync PFP to ME, otherwise we might get invalid PFP reads */
 	radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
 	radeon_ring_write(ring, 0x0);
diff --git a/drivers/gpu/drm/radeon/si_dma.c b/drivers/gpu/drm/radeon/si_dma.c
index f5cc777..aa7b872 100644
--- a/drivers/gpu/drm/radeon/si_dma.c
+++ b/drivers/gpu/drm/radeon/si_dma.c
@@ -206,6 +206,14 @@ void si_dma_vm_flush(struct radeon_device *rdev, struct radeon_ring *ring,
 	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
 	radeon_ring_write(ring, (0xf << 16) | (VM_INVALIDATE_REQUEST >> 2));
 	radeon_ring_write(ring, 1 << vm_id);
+
+	/* wait for invalidate to complete */
+	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_POLL_REG_MEM, 0, 0, 0, 0));
+	radeon_ring_write(ring, VM_INVALIDATE_REQUEST);
+	radeon_ring_write(ring, 0xff << 16); /* retry */
+	radeon_ring_write(ring, 1 << vm_id); /* mask */
+	radeon_ring_write(ring, 0); /* value */
+	radeon_ring_write(ring, (0 << 28) | 0x20); /* func(always) | poll interval */
 }
 
 /**
diff --git a/drivers/gpu/drm/radeon/sid.h b/drivers/gpu/drm/radeon/sid.h
index 4069be89..8499924 100644
--- a/drivers/gpu/drm/radeon/sid.h
+++ b/drivers/gpu/drm/radeon/sid.h
@@ -1632,6 +1632,23 @@
 #define	PACKET3_MPEG_INDEX				0x3A
 #define	PACKET3_COPY_DW					0x3B
 #define	PACKET3_WAIT_REG_MEM				0x3C
+#define		WAIT_REG_MEM_FUNCTION(x)                ((x) << 0)
+                /* 0 - always
+		 * 1 - <
+		 * 2 - <=
+		 * 3 - ==
+		 * 4 - !=
+		 * 5 - >=
+		 * 6 - >
+		 */
+#define		WAIT_REG_MEM_MEM_SPACE(x)               ((x) << 4)
+                /* 0 - reg
+		 * 1 - mem
+		 */
+#define		WAIT_REG_MEM_ENGINE(x)                  ((x) << 8)
+                /* 0 - me
+		 * 1 - pfp
+		 */
 #define	PACKET3_MEM_WRITE				0x3D
 #define	PACKET3_COPY_DATA				0x40
 #define	PACKET3_CP_DMA					0x41
@@ -1835,6 +1852,7 @@
 #define	DMA_PACKET_TRAP					  0x7
 #define	DMA_PACKET_SRBM_WRITE				  0x9
 #define	DMA_PACKET_CONSTANT_FILL			  0xd
+#define	DMA_PACKET_POLL_REG_MEM				  0xe
 #define	DMA_PACKET_NOP					  0xf
 
 #define VCE_STATUS					0x20004
-- 
1.8.3.1