[PATCH 1/4] accel/habanalabs: ignore false positive razwi

[PATCH 1/4] accel/habanalabs: ignore false positive razwi

[Oded Gabbay]

From: Tal Cohen <talcohen@habana.ai>

In Gaudi2 asic, PSOC RAZWI may cause in HBW or LBW. The address that
caused the error is read from HW register and printed by the Driver.
There are cases where the Driver receives an indication on PSOC
RAZWI error but the address value is zero. In that case, the indication
is a false positive.
The Driver should not "count" a PSOC RAZWI event error when the
caused the address is zeroed.

Signed-off-by: Tal Cohen <talcohen@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
---
 drivers/accel/habanalabs/gaudi2/gaudi2.c | 43 +++++++++++++++---------
 1 file changed, 27 insertions(+), 16 deletions(-)

diff --git a/drivers/accel/habanalabs/gaudi2/gaudi2.c b/drivers/accel/habanalabs/gaudi2/gaudi2.c
index b778cf764a68..6f05aa230376 100644
--- a/drivers/accel/habanalabs/gaudi2/gaudi2.c
+++ b/drivers/accel/habanalabs/gaudi2/gaudi2.c
@@ -8251,6 +8251,7 @@ static bool gaudi2_handle_psoc_razwi_happened(struct hl_device *hdev, u32 razwi_
 	u16 num_of_eng, eng_id[PSOC_RAZWI_MAX_ENG_PER_RTR];
 	char eng_name_str[PSOC_RAZWI_ENG_STR_SIZE];
 	bool razwi_happened = false;
+	u64 addr;
 	int i;
 
 	num_of_eng = gaudi2_psoc_razwi_get_engines(common_razwi_info, ARRAY_SIZE(common_razwi_info),
@@ -8269,43 +8270,53 @@ static bool gaudi2_handle_psoc_razwi_happened(struct hl_device *hdev, u32 razwi_
 		if (RREG32(base[i] + DEC_RAZWI_HBW_AW_SET)) {
 			addr_hi = RREG32(base[i] + DEC_RAZWI_HBW_AW_ADDR_HI);
 			addr_lo = RREG32(base[i] + DEC_RAZWI_HBW_AW_ADDR_LO);
-			dev_err(hdev->dev,
+			addr = ((u64)addr_hi << 32) + addr_lo;
+			if (addr) {
+				dev_err(hdev->dev,
 					"PSOC HBW AW RAZWI: %s, address (aligned to 128 byte): 0x%llX\n",
-					eng_name_str, ((u64)addr_hi << 32) + addr_lo);
-			hl_handle_razwi(hdev, ((u64)addr_hi << 32) + addr_lo, &eng_id[0],
+					eng_name_str, addr);
+				hl_handle_razwi(hdev, addr, &eng_id[0],
 					num_of_eng, HL_RAZWI_HBW | HL_RAZWI_WRITE, event_mask);
-			razwi_happened = true;
+				razwi_happened = true;
+			}
 		}
 
 		if (RREG32(base[i] + DEC_RAZWI_HBW_AR_SET)) {
 			addr_hi = RREG32(base[i] + DEC_RAZWI_HBW_AR_ADDR_HI);
 			addr_lo = RREG32(base[i] + DEC_RAZWI_HBW_AR_ADDR_LO);
-			dev_err(hdev->dev,
+			addr = ((u64)addr_hi << 32) + addr_lo;
+			if (addr) {
+				dev_err(hdev->dev,
 					"PSOC HBW AR RAZWI: %s, address (aligned to 128 byte): 0x%llX\n",
-					eng_name_str, ((u64)addr_hi << 32) + addr_lo);
-			hl_handle_razwi(hdev, ((u64)addr_hi << 32) + addr_lo, &eng_id[0],
+					eng_name_str, addr);
+				hl_handle_razwi(hdev, addr, &eng_id[0],
 					num_of_eng, HL_RAZWI_HBW | HL_RAZWI_READ, event_mask);
-			razwi_happened = true;
+				razwi_happened = true;
+			}
 		}
 
 		if (RREG32(base[i] + DEC_RAZWI_LBW_AW_SET)) {
 			addr_lo = RREG32(base[i] + DEC_RAZWI_LBW_AW_ADDR);
-			dev_err(hdev->dev,
+			if (addr_lo) {
+				dev_err(hdev->dev,
 					"PSOC LBW AW RAZWI: %s, address (aligned to 128 byte): 0x%X\n",
 					eng_name_str, addr_lo);
-			hl_handle_razwi(hdev, addr_lo, &eng_id[0],
+				hl_handle_razwi(hdev, addr_lo, &eng_id[0],
 					num_of_eng, HL_RAZWI_LBW | HL_RAZWI_WRITE, event_mask);
-			razwi_happened = true;
+				razwi_happened = true;
+			}
 		}
 
 		if (RREG32(base[i] + DEC_RAZWI_LBW_AR_SET)) {
 			addr_lo = RREG32(base[i] + DEC_RAZWI_LBW_AR_ADDR);
-			dev_err(hdev->dev,
-					"PSOC LBW AR RAZWI: %s, address (aligned to 128 byte): 0x%X\n",
-					eng_name_str, addr_lo);
-			hl_handle_razwi(hdev, addr_lo, &eng_id[0],
+			if (addr_lo) {
+				dev_err(hdev->dev,
+						"PSOC LBW AR RAZWI: %s, address (aligned to 128 byte): 0x%X\n",
+						eng_name_str, addr_lo);
+				hl_handle_razwi(hdev, addr_lo, &eng_id[0],
 					num_of_eng, HL_RAZWI_LBW | HL_RAZWI_READ, event_mask);
-			razwi_happened = true;
+				razwi_happened = true;
+			}
 		}
 		/* In common case the loop will break, when there is only one engine id, or
 		 * several engines with the same router. The exceptional case is with psoc razwi
-- 
2.40.0

[PATCH 2/4] accel/habanalabs: allow user to modify EDMA RL register

[Oded Gabbay]

From: Rakesh Ughreja <rughreja@habana.ai>

EDMA transpose workload requires to signal for every activation.
User FW sends all the dummy signals to RD_LBW_RATE_LIM_CFG, to save
lbw bandwidth. We need the user to be able to access that register to
configure it.

Signed-off-by: Rakesh Ughreja <rughreja@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
---
 drivers/accel/habanalabs/gaudi2/gaudi2_security.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/accel/habanalabs/gaudi2/gaudi2_security.c b/drivers/accel/habanalabs/gaudi2/gaudi2_security.c
index 694735f9e6e6..acd33130e7f9 100644
--- a/drivers/accel/habanalabs/gaudi2/gaudi2_security.c
+++ b/drivers/accel/habanalabs/gaudi2/gaudi2_security.c
@@ -479,6 +479,7 @@ static const u32 gaudi2_pb_dcr0_edma0_unsecured_regs[] = {
 	mmDCORE0_EDMA0_CORE_CTX_TE_NUMROWS,
 	mmDCORE0_EDMA0_CORE_CTX_IDX,
 	mmDCORE0_EDMA0_CORE_CTX_IDX_INC,
+	mmDCORE0_EDMA0_CORE_RD_LBW_RATE_LIM_CFG,
 	mmDCORE0_EDMA0_QM_CQ_CFG0_0,
 	mmDCORE0_EDMA0_QM_CQ_CFG0_1,
 	mmDCORE0_EDMA0_QM_CQ_CFG0_2,
-- 
2.40.0

[PATCH 3/4] accel/habanalabs: remove commented code that won't be used

[Oded Gabbay]

From: Koby Elbaz <kelbaz@habana.ai>

Once it was decided that these security settings are to be done by FW
rather than by the driver, there's no reason to keep them in the code.

Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
---
 drivers/accel/habanalabs/gaudi2/gaudi2_security.c | 9 ---------
 1 file changed, 9 deletions(-)

diff --git a/drivers/accel/habanalabs/gaudi2/gaudi2_security.c b/drivers/accel/habanalabs/gaudi2/gaudi2_security.c
index acd33130e7f9..52a12c2bb58e 100644
--- a/drivers/accel/habanalabs/gaudi2/gaudi2_security.c
+++ b/drivers/accel/habanalabs/gaudi2/gaudi2_security.c
@@ -3443,15 +3443,6 @@ static int gaudi2_init_protection_bits(struct hl_device *hdev)
 				ARRAY_SIZE(gaudi2_pb_thermal_sensor0), NULL, HL_PB_NA);
 	}
 
-	/* HBM */
-	/* Temporarily skip until SW-63348 is solved
-	 * instance_offset = mmHBM1_MC0_BASE - mmHBM0_MC0_BASE;
-	 * rc |= hl_init_pb_with_mask(hdev, HL_PB_SHARED, HL_PB_NA, GAUDI2_HBM_NUM,
-	 *		instance_offset, gaudi2_pb_hbm,
-	 *		ARRAY_SIZE(gaudi2_pb_hbm), NULL, HL_PB_NA,
-	 *		prop->dram_enabled_mask);
-	 */
-
 	/* Scheduler ARCs */
 	instance_offset = mmARC_FARM_ARC1_AUX_BASE - mmARC_FARM_ARC0_AUX_BASE;
 	rc |= hl_init_pb_ranges(hdev, HL_PB_SHARED, HL_PB_NA,
-- 
2.40.0

[PATCH 4/4] accel/habanalabs: fix bug in free scratchpad memory

[Oded Gabbay]

From: Moti Haimovski <mhaimovski@habana.ai>

This commit fixes a bug in Gaudi2 when freeing the scratchpad memory
in case software init fails.

Signed-off-by: Moti Haimovski <mhaimovski@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
---
 drivers/accel/habanalabs/gaudi2/gaudi2.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/accel/habanalabs/gaudi2/gaudi2.c b/drivers/accel/habanalabs/gaudi2/gaudi2.c
index 6f05aa230376..acc304ebca82 100644
--- a/drivers/accel/habanalabs/gaudi2/gaudi2.c
+++ b/drivers/accel/habanalabs/gaudi2/gaudi2.c
@@ -3630,8 +3630,8 @@ static int gaudi2_sw_init(struct hl_device *hdev)
 special_blocks_free:
 	gaudi2_special_blocks_iterator_free(hdev);
 free_scratchpad_mem:
-	hl_asic_dma_pool_free(hdev, gaudi2->scratchpad_kernel_address,
-				gaudi2->scratchpad_bus_address);
+	hl_asic_dma_free_coherent(hdev, PAGE_SIZE, gaudi2->scratchpad_kernel_address,
+				  gaudi2->scratchpad_bus_address);
 free_virt_msix_db_mem:
 	hl_cpu_accessible_dma_pool_free(hdev, prop->pmmu.page_size, gaudi2->virt_msix_db_cpu_addr);
 free_cpu_accessible_dma_pool:
-- 
2.40.0