[PATCH 1/2] habanalabs: block scalar load_and_exe on external queue

[PATCH 1/2] habanalabs: block scalar load_and_exe on external queue

[Oded Gabbay]

In Gaudi, the user can't execute scalar load_and_exe on external queue
because it can be a security hole. The driver doesn't parse the commands
being loaded and it can be msg_prot, which the user isn't allowed to use.

Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
---
 drivers/misc/habanalabs/gaudi/gaudi.c         | 25 ++++++++++++++++++-
 .../habanalabs/include/gaudi/gaudi_packets.h  |  3 +++
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/drivers/misc/habanalabs/gaudi/gaudi.c b/drivers/misc/habanalabs/gaudi/gaudi.c
index 61f88e9884ce..f34ac8d35a14 100644
--- a/drivers/misc/habanalabs/gaudi/gaudi.c
+++ b/drivers/misc/habanalabs/gaudi/gaudi.c
@@ -3790,6 +3790,25 @@ static int gaudi_validate_dma_pkt_no_mmu(struct hl_device *hdev,
 						src_in_host);
 }
 
+static int gaudi_validate_load_and_exe_pkt(struct hl_device *hdev,
+					struct hl_cs_parser *parser,
+					struct packet_load_and_exe *user_pkt)
+{
+	u32 cfg;
+
+	cfg = le32_to_cpu(user_pkt->cfg);
+
+	if (cfg & GAUDI_PKT_LOAD_AND_EXE_CFG_DST_MASK) {
+		dev_err(hdev->dev,
+			"User not allowed to use Load and Execute\n");
+		return -EPERM;
+	}
+
+	parser->patched_cb_size += sizeof(struct packet_load_and_exe);
+
+	return 0;
+}
+
 static int gaudi_validate_cb(struct hl_device *hdev,
 			struct hl_cs_parser *parser, bool is_mmu)
 {
@@ -3838,6 +3857,11 @@ static int gaudi_validate_cb(struct hl_device *hdev,
 			rc = -EPERM;
 			break;
 
+		case PACKET_LOAD_AND_EXE:
+			rc = gaudi_validate_load_and_exe_pkt(hdev, parser,
+				(struct packet_load_and_exe *) user_pkt);
+			break;
+
 		case PACKET_LIN_DMA:
 			parser->contains_dma_pkt = true;
 			if (is_mmu)
@@ -3855,7 +3879,6 @@ static int gaudi_validate_cb(struct hl_device *hdev,
 		case PACKET_FENCE:
 		case PACKET_NOP:
 		case PACKET_ARB_POINT:
-		case PACKET_LOAD_AND_EXE:
 			parser->patched_cb_size += pkt_size;
 			break;
 
diff --git a/drivers/misc/habanalabs/include/gaudi/gaudi_packets.h b/drivers/misc/habanalabs/include/gaudi/gaudi_packets.h
index 9a5800b0086b..0f0cd067bb43 100644
--- a/drivers/misc/habanalabs/include/gaudi/gaudi_packets.h
+++ b/drivers/misc/habanalabs/include/gaudi/gaudi_packets.h
@@ -197,6 +197,9 @@ struct packet_wait {
 	__le32 ctl;
 };
 
+#define GAUDI_PKT_LOAD_AND_EXE_CFG_DST_SHIFT	0
+#define GAUDI_PKT_LOAD_AND_EXE_CFG_DST_MASK	0x00000001
+
 struct packet_load_and_exe {
 	__le32 cfg;
 	__le32 ctl;
-- 
2.17.1

[PATCH 2/2] habanalabs: use PI in MMU cache invalidation

[Oded Gabbay]

From: Omer Shpigelman <oshpigelman@habana.ai>

The PS flow for MMU cache invalidation caused timeouts in stress tests.
Use PS + PI flow so no timeouts should happen whatsoever.

Signed-off-by: Omer Shpigelman <oshpigelman@habana.ai>
Reviewed-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
---
 drivers/misc/habanalabs/gaudi/gaudi.c  | 8 ++++++++
 drivers/misc/habanalabs/gaudi/gaudiP.h | 3 +++
 2 files changed, 11 insertions(+)

diff --git a/drivers/misc/habanalabs/gaudi/gaudi.c b/drivers/misc/habanalabs/gaudi/gaudi.c
index f34ac8d35a14..211547d4f8a7 100644
--- a/drivers/misc/habanalabs/gaudi/gaudi.c
+++ b/drivers/misc/habanalabs/gaudi/gaudi.c
@@ -2725,6 +2725,12 @@ static int gaudi_mmu_init(struct hl_device *hdev)
 	WREG32(mmSTLB_HOP_CONFIGURATION,
 			hdev->mmu_huge_page_opt ? 0x30440 : 0x40440);
 
+	/*
+	 * The H/W expects the first PI after init to be 1. After wraparound
+	 * we'll write 0.
+	 */
+	gaudi->mmu_cache_inv_pi = 1;
+
 	gaudi->hw_cap_initialized |= HW_CAP_MMU;
 
 	return 0;
@@ -6017,6 +6023,8 @@ static int gaudi_mmu_invalidate_cache(struct hl_device *hdev, bool is_hard,
 	mutex_lock(&hdev->mmu_cache_lock);
 
 	/* L0 & L1 invalidation */
+	WREG32(mmSTLB_INV_PS, 3);
+	WREG32(mmSTLB_CACHE_INV, gaudi->mmu_cache_inv_pi++);
 	WREG32(mmSTLB_INV_PS, 2);
 
 	rc = hl_poll_timeout(
diff --git a/drivers/misc/habanalabs/gaudi/gaudiP.h b/drivers/misc/habanalabs/gaudi/gaudiP.h
index a46530d375fa..41a8d9bff6bf 100644
--- a/drivers/misc/habanalabs/gaudi/gaudiP.h
+++ b/drivers/misc/habanalabs/gaudi/gaudiP.h
@@ -229,6 +229,8 @@ struct gaudi_internal_qman_info {
  * @multi_msi_mode: whether we are working in multi MSI single MSI mode.
  *                  Multi MSI is possible only with IOMMU enabled.
  * @ext_queue_idx: helper index for external queues initialization.
+ * @mmu_cache_inv_pi: PI for MMU cache invalidation flow. The H/W expects an
+ *                    8-bit value so use u8.
  */
 struct gaudi_device {
 	int (*armcp_info_get)(struct hl_device *hdev);
@@ -248,6 +250,7 @@ struct gaudi_device {
 	u32				hw_cap_initialized;
 	u8				multi_msi_mode;
 	u8				ext_queue_idx;
+	u8				mmu_cache_inv_pi;
 };
 
 void gaudi_init_security(struct hl_device *hdev);
-- 
2.17.1

RE: [PATCH 1/2] habanalabs: block scalar load_and_exe on external queue

[Tomer Tayar]

On Sat, Jun 6, 2020 at 0:29, Oded Gabbay <oded.gabbay@gmail.com> wrote:
> In Gaudi, the user can't execute scalar load_and_exe on external queue
> because it can be a security hole. The driver doesn't parse the commands
> being loaded and it can be msg_prot, which the user isn't allowed to use.
> 
> Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>

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