2020-06-05 20:24:55

by Oded Gabbay

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

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 <[email protected]>
---
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


2020-06-05 20:25:11

by Oded Gabbay

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

From: Omer Shpigelman <[email protected]>

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 <[email protected]>
Reviewed-by: Oded Gabbay <[email protected]>
Signed-off-by: Oded Gabbay <[email protected]>
---
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

2020-06-07 05:42:48

by Tomer Tayar

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

On Sat, Jun 6, 2020 at 0:29, Oded Gabbay <[email protected]> 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 <[email protected]>

Reviewed-by: Tomer Tayar <[email protected]>