accel/habanalabs: add critical-event bit in notifier

Enhance the existing user notifications by adding a HW and FW critical
event bits to be used when a HW or FW event occur that requires
both SW abort and hard-resetting the chip.

Signed-off-by: Moti Haimovski <mhaimovski@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Stanislaw Gruszka <stanislaw.gruszka@linux.intel.com>

6 files changed, 170 insertions, 6 deletions

diff --git a/drivers/accel/habanalabs/common/device.c b/drivers/accel/habanalabs/common/device.c
index b8c74185eabd..f91f3509336f 100644
--- a/drivers/accel/habanalabs/common/device.c
+++ b/drivers/accel/habanalabs/common/device.c
@@ -998,6 +998,8 @@ static void hl_device_heartbeat(struct work_struct *work)
 {
 	struct hl_device *hdev = container_of(work, struct hl_device,
 						work_heartbeat.work);
+	struct hl_info_fw_err_info info = {0};
+	u64 event_mask = HL_NOTIFIER_EVENT_DEVICE_RESET | HL_NOTIFIER_EVENT_DEVICE_UNAVAILABLE;
 
 	if (!hl_device_operational(hdev, NULL))
 		goto reschedule;
@@ -1008,7 +1010,10 @@ static void hl_device_heartbeat(struct work_struct *work)
 	if (hl_device_operational(hdev, NULL))
 		dev_err(hdev->dev, "Device heartbeat failed!\n");
 
-	hl_device_reset(hdev, HL_DRV_RESET_HARD | HL_DRV_RESET_HEARTBEAT);
+	info.err_type = HL_INFO_FW_HEARTBEAT_ERR;
+	info.event_mask = &event_mask;
+	hl_handle_fw_err(hdev, &info);
+	hl_device_cond_reset(hdev, HL_DRV_RESET_HARD | HL_DRV_RESET_HEARTBEAT, event_mask);
 
 	return;
 
@@ -2626,3 +2631,49 @@ void hl_handle_page_fault(struct hl_device *hdev, u64 addr, u16 eng_id, bool is_
 	if (event_mask)
 		*event_mask |=  HL_NOTIFIER_EVENT_PAGE_FAULT;
 }
+
+void hl_capture_hw_err(struct hl_device *hdev, u16 event_id)
+{
+	struct hw_err_info *info = &hdev->captured_err_info.hw_err;
+
+	/* Capture only the first HW err */
+	if (atomic_cmpxchg(&info->event_detected, 0, 1))
+		return;
+
+	info->event.timestamp = ktime_to_ns(ktime_get());
+	info->event.event_id = event_id;
+
+	info->event_info_available = true;
+}
+
+void hl_handle_critical_hw_err(struct hl_device *hdev, u16 event_id, u64 *event_mask)
+{
+	hl_capture_hw_err(hdev, event_id);
+
+	if (event_mask)
+		*event_mask |= HL_NOTIFIER_EVENT_CRITICL_HW_ERR;
+}
+
+void hl_capture_fw_err(struct hl_device *hdev, struct hl_info_fw_err_info *fw_info)
+{
+	struct fw_err_info *info = &hdev->captured_err_info.fw_err;
+
+	/* Capture only the first FW error */
+	if (atomic_cmpxchg(&info->event_detected, 0, 1))
+		return;
+
+	info->event.timestamp = ktime_to_ns(ktime_get());
+	info->event.err_type = fw_info->err_type;
+	if (fw_info->err_type == HL_INFO_FW_REPORTED_ERR)
+		info->event.event_id = fw_info->event_id;
+
+	info->event_info_available = true;
+}
+
+void hl_handle_fw_err(struct hl_device *hdev, struct hl_info_fw_err_info *info)
+{
+	hl_capture_fw_err(hdev, info);
+
+	if (info->event_mask)
+		*info->event_mask |= HL_NOTIFIER_EVENT_CRITICL_FW_ERR;
+}
diff --git a/drivers/accel/habanalabs/common/habanalabs.h b/drivers/accel/habanalabs/common/habanalabs.h
index afdae5775eaa..176a2e2c050d 100644
--- a/drivers/accel/habanalabs/common/habanalabs.h
+++ b/drivers/accel/habanalabs/common/habanalabs.h
@@ -3032,17 +3032,55 @@ struct razwi_info {
 };
 
 /**
+ * struct hw_err_info - HW error information.
+ * @event: holds information on the event.
+ * @event_detected: if set as 1, then a HW event was discovered for the
+ *                  first time after the driver has finished booting-up.
+ *                  currently we assume that only fatal events (that require hard-reset) are
+ *                  reported so we don't care of the others that might follow it.
+ *                  so once changed to 1, it will remain that way.
+ *                  TODO: support multiple events.
+ * @event_info_available: indicates that a HW event info is now available.
+ */
+struct hw_err_info {
+	struct hl_info_hw_err_event	event;
+	atomic_t			event_detected;
+	bool				event_info_available;
+};
+
+/**
+ * struct fw_err_info - FW error information.
+ * @event: holds information on the event.
+ * @event_detected: if set as 1, then a FW event was discovered for the
+ *                  first time after the driver has finished booting-up.
+ *                  currently we assume that only fatal events (that require hard-reset) are
+ *                  reported so we don't care of the others that might follow it.
+ *                  so once changed to 1, it will remain that way.
+ *                  TODO: support multiple events.
+ * @event_info_available: indicates that a HW event info is now available.
+ */
+struct fw_err_info {
+	struct hl_info_fw_err_event	event;
+	atomic_t			event_detected;
+	bool				event_info_available;
+};
+
+/**
  * struct hl_error_info - holds information collected during an error.
  * @cs_timeout: CS timeout error information.
  * @razwi_info: RAZWI information.
  * @undef_opcode: undefined opcode information.
  * @page_fault_info: page fault information.
+ * @hw_err: (fatal) hardware error information.
+ * @fw_err: firmware error information.
  */
 struct hl_error_info {
 	struct cs_timeout_info		cs_timeout;
 	struct razwi_info		razwi_info;
 	struct undefined_opcode_info	undef_opcode;
 	struct page_fault_info		page_fault_info;
+	struct hw_err_info		hw_err;
+	struct fw_err_info		fw_err;
 };
 
 /**
@@ -3453,6 +3491,20 @@ struct hl_cs_encaps_sig_handle {
 	u32  count;
 };
 
+/**
+ * struct hl_info_fw_err_info - firmware error information structure
+ * @err_type: The type of error detected (or reported).
+ * @event_mask: Pointer to the event mask to be modified with the detected error flag
+ *              (can be NULL)
+ * @event_id: The id of the event that reported the error
+ *            (applicable when err_type is HL_INFO_FW_REPORTED_ERR).
+ */
+struct hl_info_fw_err_info {
+	enum hl_info_fw_err_type err_type;
+	u64 *event_mask;
+	u16 event_id;
+};
+
 /*
  * IOCTLs
  */
@@ -3883,6 +3935,8 @@ void hl_handle_razwi(struct hl_device *hdev, u64 addr, u16 *engine_id, u16 num_o
 void hl_capture_page_fault(struct hl_device *hdev, u64 addr, u16 eng_id, bool is_pmmu);
 void hl_handle_page_fault(struct hl_device *hdev, u64 addr, u16 eng_id, bool is_pmmu,
 				u64 *event_mask);
+void hl_handle_critical_hw_err(struct hl_device *hdev, u16 event_id, u64 *event_mask);
+void hl_handle_fw_err(struct hl_device *hdev, struct hl_info_fw_err_info *info);
 
 #ifdef CONFIG_DEBUG_FS
 
diff --git a/drivers/accel/habanalabs/common/habanalabs_drv.c b/drivers/accel/habanalabs/common/habanalabs_drv.c
index 8ccc3d6519b5..0cb6e52a1192 100644
--- a/drivers/accel/habanalabs/common/habanalabs_drv.c
+++ b/drivers/accel/habanalabs/common/habanalabs_drv.c
@@ -221,12 +221,9 @@ int hl_device_open(struct inode *inode, struct file *filp)
 
 	hl_debugfs_add_file(hpriv);
 
+	memset(&hdev->captured_err_info, 0, sizeof(hdev->captured_err_info));
 	atomic_set(&hdev->captured_err_info.cs_timeout.write_enable, 1);
-	atomic_set(&hdev->captured_err_info.razwi_info.razwi_detected, 0);
-	atomic_set(&hdev->captured_err_info.page_fault_info.page_fault_detected, 0);
 	hdev->captured_err_info.undef_opcode.write_enable = true;
-	hdev->captured_err_info.razwi_info.razwi_info_available = false;
-	hdev->captured_err_info.page_fault_info.page_fault_info_available = false;
 
 	hdev->open_counter++;
 	hdev->last_successful_open_jif = jiffies;
diff --git a/drivers/accel/habanalabs/common/habanalabs_ioctl.c b/drivers/accel/habanalabs/common/habanalabs_ioctl.c
index 5005e6fca691..13cd5013c72a 100644
--- a/drivers/accel/habanalabs/common/habanalabs_ioctl.c
+++ b/drivers/accel/habanalabs/common/habanalabs_ioctl.c
@@ -830,6 +830,50 @@ static int user_mappings_info(struct hl_fpriv *hpriv, struct hl_info_args *args)
 	return copy_to_user(out, pgf_info->user_mappings, actual_size) ? -EFAULT : 0;
 }
 
+static int hw_err_info(struct hl_fpriv *hpriv, struct hl_info_args *args)
+{
+	void __user *user_buf = (void __user *) (uintptr_t) args->return_pointer;
+	struct hl_device *hdev = hpriv->hdev;
+	u32 user_buf_size = args->return_size;
+	struct hw_err_info *info;
+	int rc;
+
+	if ((!user_buf_size) || (!user_buf))
+		return -EINVAL;
+
+	if (user_buf_size < sizeof(struct hl_info_hw_err_event))
+		return -ENOMEM;
+
+	info = &hdev->captured_err_info.hw_err;
+	if (!info->event_info_available)
+		return -ENOENT;
+
+	rc = copy_to_user(user_buf, &info->event, sizeof(struct hl_info_hw_err_event));
+	return rc ? -EFAULT : 0;
+}
+
+static int fw_err_info(struct hl_fpriv *hpriv, struct hl_info_args *args)
+{
+	void __user *user_buf = (void __user *) (uintptr_t) args->return_pointer;
+	struct hl_device *hdev = hpriv->hdev;
+	u32 user_buf_size = args->return_size;
+	struct fw_err_info *info;
+	int rc;
+
+	if ((!user_buf_size) || (!user_buf))
+		return -EINVAL;
+
+	if (user_buf_size < sizeof(struct hl_info_fw_err_event))
+		return -ENOMEM;
+
+	info = &hdev->captured_err_info.fw_err;
+	if (!info->event_info_available)
+		return -ENOENT;
+
+	rc = copy_to_user(user_buf, &info->event, sizeof(struct hl_info_fw_err_event));
+	return rc ? -EFAULT : 0;
+}
+
 static int send_fw_generic_request(struct hl_device *hdev, struct hl_info_args *info_args)
 {
 	void __user *buff = (void __user *) (uintptr_t) info_args->return_pointer;
@@ -950,6 +994,12 @@ static int _hl_info_ioctl(struct hl_fpriv *hpriv, void *data,
 	case HL_INFO_UNREGISTER_EVENTFD:
 		return eventfd_unregister(hpriv, args);
 
+	case HL_INFO_HW_ERR_EVENT:
+		return hw_err_info(hpriv, args);
+
+	case HL_INFO_FW_ERR_EVENT:
+		return fw_err_info(hpriv, args);
+
 	default:
 		break;
 	}
diff --git a/drivers/accel/habanalabs/gaudi/gaudi.c b/drivers/accel/habanalabs/gaudi/gaudi.c
index 4ba5352cb7cb..0e02aebb6ea6 100644
--- a/drivers/accel/habanalabs/gaudi/gaudi.c
+++ b/drivers/accel/habanalabs/gaudi/gaudi.c
@@ -7634,6 +7634,7 @@ static void gaudi_print_clk_change_info(struct hl_device *hdev, u16 event_type,
 static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entry)
 {
 	struct gaudi_device *gaudi = hdev->asic_specific;
+	struct hl_info_fw_err_info fw_err_info;
 	u64 data = le64_to_cpu(eq_entry->data[0]), event_mask = 0;
 	u32 ctl = le32_to_cpu(eq_entry->hdr.ctl);
 	u32 fw_fatal_err_flag = 0, flags = 0;
@@ -7912,7 +7913,10 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entr
 	case GAUDI_EVENT_FW_ALIVE_S:
 		gaudi_print_irq_info(hdev, event_type, false, &event_mask);
 		gaudi_print_fw_alive_info(hdev, &eq_entry->fw_alive);
-		event_mask |= HL_NOTIFIER_EVENT_GENERAL_HW_ERR;
+		fw_err_info.err_type = HL_INFO_FW_REPORTED_ERR;
+		fw_err_info.event_id = event_type;
+		fw_err_info.event_mask = &event_mask;
+		hl_handle_fw_err(hdev, &fw_err_info);
 		goto reset_device;
 
 	default:
@@ -7943,6 +7947,10 @@ reset_device:
 	}
 
 	if (reset_required) {
+		/* escalate general hw errors to critical/fatal error */
+		if (event_mask & HL_NOTIFIER_EVENT_GENERAL_HW_ERR)
+			hl_handle_critical_hw_err(hdev, event_type, &event_mask);
+
 		hl_device_cond_reset(hdev, flags, event_mask);
 	} else {
 		hl_fw_unmask_irq(hdev, event_type);
diff --git a/drivers/accel/habanalabs/gaudi2/gaudi2.c b/drivers/accel/habanalabs/gaudi2/gaudi2.c
index d250c6f01acb..6926af5b5ed1 100644
--- a/drivers/accel/habanalabs/gaudi2/gaudi2.c
+++ b/drivers/accel/habanalabs/gaudi2/gaudi2.c
@@ -9444,6 +9444,10 @@ reset_device:
 	} else {
 		reset_flags |= HL_DRV_RESET_DELAY;
 	}
+	/* escalate general hw errors to critical/fatal error */
+	if (event_mask & HL_NOTIFIER_EVENT_GENERAL_HW_ERR)
+		hl_handle_critical_hw_err(hdev, event_type, &event_mask);
+
 	event_mask |= HL_NOTIFIER_EVENT_DEVICE_RESET;
 	hl_device_cond_reset(hdev, reset_flags, event_mask);
 }