[PATCH] habanalabs: halt the engines before hard-reset

[PATCH] habanalabs: halt the engines before hard-reset

[Oded Gabbay]

The driver must halt the engines before doing hard-reset, otherwise the
device can go into undefined state. There is a place where the driver
didn't do that and this patch fixes it.

Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
---
 drivers/misc/habanalabs/device.c    |  1 +
 drivers/misc/habanalabs/goya/goya.c | 42 +++++++++++++++++++++++++++++
 2 files changed, 43 insertions(+)

diff --git a/drivers/misc/habanalabs/device.c b/drivers/misc/habanalabs/device.c
index b155e9549076..166883b64725 100644
--- a/drivers/misc/habanalabs/device.c
+++ b/drivers/misc/habanalabs/device.c
@@ -1189,6 +1189,7 @@ int hl_device_init(struct hl_device *hdev, struct class *hclass)
 	if (hdev->asic_funcs->get_hw_state(hdev) == HL_DEVICE_HW_STATE_DIRTY) {
 		dev_info(hdev->dev,
 			"H/W state is dirty, must reset before initializing\n");
+		hdev->asic_funcs->halt_engines(hdev, true);
 		hdev->asic_funcs->hw_fini(hdev, true);
 	}
 
diff --git a/drivers/misc/habanalabs/goya/goya.c b/drivers/misc/habanalabs/goya/goya.c
index 72935851a7e8..5750294400f4 100644
--- a/drivers/misc/habanalabs/goya/goya.c
+++ b/drivers/misc/habanalabs/goya/goya.c
@@ -895,6 +895,11 @@ void goya_init_dma_qmans(struct hl_device *hdev)
  */
 static void goya_disable_external_queues(struct hl_device *hdev)
 {
+	struct goya_device *goya = hdev->asic_specific;
+
+	if (!(goya->hw_cap_initialized & HW_CAP_DMA))
+		return;
+
 	WREG32(mmDMA_QM_0_GLBL_CFG0, 0);
 	WREG32(mmDMA_QM_1_GLBL_CFG0, 0);
 	WREG32(mmDMA_QM_2_GLBL_CFG0, 0);
@@ -956,6 +961,11 @@ static int goya_stop_external_queues(struct hl_device *hdev)
 {
 	int rc, retval = 0;
 
+	struct goya_device *goya = hdev->asic_specific;
+
+	if (!(goya->hw_cap_initialized & HW_CAP_DMA))
+		return retval;
+
 	rc = goya_stop_queue(hdev,
 			mmDMA_QM_0_GLBL_CFG1,
 			mmDMA_QM_0_CP_STS,
@@ -1744,9 +1754,18 @@ void goya_init_tpc_qmans(struct hl_device *hdev)
  */
 static void goya_disable_internal_queues(struct hl_device *hdev)
 {
+	struct goya_device *goya = hdev->asic_specific;
+
+	if (!(goya->hw_cap_initialized & HW_CAP_MME))
+		goto disable_tpc;
+
 	WREG32(mmMME_QM_GLBL_CFG0, 0);
 	WREG32(mmMME_CMDQ_GLBL_CFG0, 0);
 
+disable_tpc:
+	if (!(goya->hw_cap_initialized & HW_CAP_TPC))
+		return;
+
 	WREG32(mmTPC0_QM_GLBL_CFG0, 0);
 	WREG32(mmTPC0_CMDQ_GLBL_CFG0, 0);
 
@@ -1782,8 +1801,12 @@ static void goya_disable_internal_queues(struct hl_device *hdev)
  */
 static int goya_stop_internal_queues(struct hl_device *hdev)
 {
+	struct goya_device *goya = hdev->asic_specific;
 	int rc, retval = 0;
 
+	if (!(goya->hw_cap_initialized & HW_CAP_MME))
+		goto stop_tpc;
+
 	/*
 	 * Each queue (QMAN) is a separate H/W logic. That means that each
 	 * QMAN can be stopped independently and failure to stop one does NOT
@@ -1810,6 +1833,10 @@ static int goya_stop_internal_queues(struct hl_device *hdev)
 		retval = -EIO;
 	}
 
+stop_tpc:
+	if (!(goya->hw_cap_initialized & HW_CAP_TPC))
+		return retval;
+
 	rc = goya_stop_queue(hdev,
 			mmTPC0_QM_GLBL_CFG1,
 			mmTPC0_QM_CP_STS,
@@ -1975,6 +2002,11 @@ static int goya_stop_internal_queues(struct hl_device *hdev)
 
 static void goya_dma_stall(struct hl_device *hdev)
 {
+	struct goya_device *goya = hdev->asic_specific;
+
+	if (!(goya->hw_cap_initialized & HW_CAP_DMA))
+		return;
+
 	WREG32(mmDMA_QM_0_GLBL_CFG1, 1 << DMA_QM_0_GLBL_CFG1_DMA_STOP_SHIFT);
 	WREG32(mmDMA_QM_1_GLBL_CFG1, 1 << DMA_QM_1_GLBL_CFG1_DMA_STOP_SHIFT);
 	WREG32(mmDMA_QM_2_GLBL_CFG1, 1 << DMA_QM_2_GLBL_CFG1_DMA_STOP_SHIFT);
@@ -1984,6 +2016,11 @@ static void goya_dma_stall(struct hl_device *hdev)
 
 static void goya_tpc_stall(struct hl_device *hdev)
 {
+	struct goya_device *goya = hdev->asic_specific;
+
+	if (!(goya->hw_cap_initialized & HW_CAP_TPC))
+		return;
+
 	WREG32(mmTPC0_CFG_TPC_STALL, 1 << TPC0_CFG_TPC_STALL_V_SHIFT);
 	WREG32(mmTPC1_CFG_TPC_STALL, 1 << TPC1_CFG_TPC_STALL_V_SHIFT);
 	WREG32(mmTPC2_CFG_TPC_STALL, 1 << TPC2_CFG_TPC_STALL_V_SHIFT);
@@ -1996,6 +2033,11 @@ static void goya_tpc_stall(struct hl_device *hdev)
 
 static void goya_mme_stall(struct hl_device *hdev)
 {
+	struct goya_device *goya = hdev->asic_specific;
+
+	if (!(goya->hw_cap_initialized & HW_CAP_MME))
+		return;
+
 	WREG32(mmMME_STALL, 0xFFFFFFFF);
 }
 
-- 
2.17.1

RE: [PATCH] habanalabs: halt the engines before hard-reset

[Tomer Tayar]

On Wed, Dec 25, 2019 at 15:19, Oded Gabbay <oded.gabbay@gmail.com> wrote:
> The driver must halt the engines before doing hard-reset, otherwise the
> device can go into undefined state. There is a place where the driver
> didn't do that and this patch fixes it.
> 
> Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>

Reviewed-by: Tomer Tayar <ttayar@habana.ai>