[xcvrd] Add bitmap support for SFP error event (#184)

Support SFP error bitmap.

Currently, SONiC use a single value to represent SFP error, however, multiple SFP errors could exist at the same time. This PR is aimed to support it   

Signed-off-by: Stephen Sun <stephens@nvidia.com>

sonic-xcvrd/tests/test_xcvrd.py

@@ -1,17 +1,15 @@
 import os
 import sys
-import subprocess
 
-import pytest
 import unittest
-from imp import load_source
 if sys.version_info >= (3, 3):
     from unittest.mock import MagicMock, patch
 else:
     from mock import MagicMock, patch
 
 from sonic_py_common import daemon_base
 from swsscommon import swsscommon
+from sonic_platform_base.sfp_base import SfpBase
 from .mock_swsscommon import Table
 
 
@@ -24,13 +22,12 @@
 test_path = os.path.dirname(os.path.abspath(__file__))
 modules_path = os.path.dirname(test_path)
 scripts_path = os.path.join(modules_path, "xcvrd")
-helper_file_path = os.path.join(scripts_path, "xcvrd_utilities"+"/y_cable_helper.py")
 sys.path.insert(0, modules_path)
 
 os.environ["XCVRD_UNIT_TESTING"] = "1"
-load_source('y_cable_helper', scripts_path + '/xcvrd_utilities/y_cable_helper.py')
-from y_cable_helper import *
 from xcvrd.xcvrd import *
+from xcvrd.xcvrd_utilities.y_cable_helper import *
+from xcvrd.xcvrd_utilities.sfp_status_helper import *
 
 
 class TestXcvrdScript(object):
@@ -219,9 +216,9 @@ def test_init_port_sfp_status_tbl(self):
         init_port_sfp_status_tbl(stop_event)
 
     @patch('xcvrd.xcvrd_utilities.y_cable_helper.y_cable_platform_sfputil', MagicMock(return_value=[0]))
-    @patch('y_cable_helper.logical_port_name_to_physical_port_list', MagicMock(return_value=[0]))
-    @patch('y_cable_helper._wrapper_get_presence', MagicMock(return_value=True))
-    @patch('y_cable_helper.get_muxcable_info', MagicMock(return_value={'tor_active': 'self',
+    @patch('xcvrd.xcvrd_utilities.y_cable_helper.logical_port_name_to_physical_port_list', MagicMock(return_value=[0]))
+    @patch('xcvrd.xcvrd_utilities.y_cable_helper._wrapper_get_presence', MagicMock(return_value=True))
+    @patch('xcvrd.xcvrd_utilities.y_cable_helper.get_muxcable_info', MagicMock(return_value={'tor_active': 'self',
                                                                        'mux_direction': 'self',
                                                                        'manual_switch_count': '7',
                                                                        'auto_switch_count': '71',
@@ -258,9 +255,9 @@ def test_post_port_mux_info_to_db(self):
         assert(rc != -1)
 
     @patch('xcvrd.xcvrd_utilities.y_cable_helper.y_cable_platform_sfputil', MagicMock(return_value=[0]))
-    @patch('y_cable_helper.logical_port_name_to_physical_port_list', MagicMock(return_value=[0]))
-    @patch('y_cable_helper._wrapper_get_presence', MagicMock(return_value=True))
-    @patch('y_cable_helper.get_muxcable_static_info', MagicMock(return_value={'read_side': 'self',
+    @patch('xcvrd.xcvrd_utilities.y_cable_helper.logical_port_name_to_physical_port_list', MagicMock(return_value=[0]))
+    @patch('xcvrd.xcvrd_utilities.y_cable_helper._wrapper_get_presence', MagicMock(return_value=True))
+    @patch('xcvrd.xcvrd_utilities.y_cable_helper.get_muxcable_static_info', MagicMock(return_value={'read_side': 'self',
                                                                               'nic_lane1_precursor1': '1',
                                                                               'nic_lane1_precursor2': '-7',
                                                                               'nic_lane1_maincursor': '-1',
@@ -318,3 +315,23 @@ def test_get_media_settings_key(self):
         result = get_media_settings_key(0, xcvr_info_dict)
         assert result == ['MOLEX-1064141421', 'QSFP+-*']
         # TODO: Ensure that error message was logged
+
+    def test_detect_port_in_error_status(self):
+        class MockTable:
+            def get(self, key):
+                pass
+
+        status_tbl = MockTable()
+        status_tbl.get = MagicMock(return_value=(True, {'error': 'N/A'}))
+        assert not detect_port_in_error_status(None, status_tbl)
+
+        status_tbl.get = MagicMock(return_value=(True, {'error': SfpBase.SFP_ERROR_DESCRIPTION_BLOCKING}))
+        assert detect_port_in_error_status(None, status_tbl)
+
+    def test_is_error_sfp_status(self):
+        error_values = [7, 11, 19, 35]
+        for error_value in error_values:
+            assert is_error_block_eeprom_reading(error_value)
+
+        assert not is_error_block_eeprom_reading(int(SFP_STATUS_INSERTED))
+        assert not is_error_block_eeprom_reading(int(SFP_STATUS_REMOVED))

sonic-xcvrd/xcvrd/xcvrd.py

@@ -15,11 +15,11 @@
     import threading
     import time
 
-    from enum import Enum
     from sonic_py_common import daemon_base, device_info, logger
     from sonic_py_common import multi_asic
     from swsscommon import swsscommon
 
+    from .xcvrd_utilities import sfp_status_helper
     from .xcvrd_utilities import y_cable_helper
 except ImportError as e:
     raise ImportError(str(e) + " - required module not found")
@@ -43,18 +43,6 @@
 TIME_FOR_SFP_READY_SECS = 1
 XCVRD_MAIN_THREAD_SLEEP_SECS = 60
 
-# SFP status definition, shall be aligned with the definition in get_change_event() of ChassisBase
-SFP_STATUS_REMOVED = '0'
-SFP_STATUS_INSERTED = '1'
-
-# SFP error code enum, new elements can be added to the enum if new errors need to be supported.
-SFP_STATUS_ERR_ENUM = Enum('SFP_STATUS_ERR_ENUM', ['SFP_STATUS_ERR_I2C_STUCK', 'SFP_STATUS_ERR_BAD_EEPROM',
-                                                   'SFP_STATUS_ERR_UNSUPPORTED_CABLE', 'SFP_STATUS_ERR_HIGH_TEMP',
-                                                   'SFP_STATUS_ERR_BAD_CABLE'], start=2)
-
-# Convert the error code to string and store them in a set for convenience
-errors_block_eeprom_reading = set(str(error_code.value) for error_code in SFP_STATUS_ERR_ENUM)
-
 EVENT_ON_ALL_SFP = '-1'
 # events definition
 SYSTEM_NOT_READY = 'system_not_ready'
@@ -188,11 +176,13 @@ def _wrapper_get_transceiver_change_event(timeout):
     if platform_chassis is not None:
         try:
             status, events = platform_chassis.get_change_event(timeout)
-            sfp_events = events['sfp']
-            return status, sfp_events
+            sfp_events = events.get('sfp')
+            sfp_errors = events.get('sfp_error')
+            return status, sfp_events, sfp_errors
         except NotImplementedError:
             pass
-    return platform_sfputil.get_transceiver_change_event(timeout)
+    status, events = platform_sfputil.get_transceiver_change_event(timeout)
+    return status, events, None
 
 
 def _wrapper_get_sfp_type(physical_port):
@@ -203,6 +193,14 @@ def _wrapper_get_sfp_type(physical_port):
             pass
     return None
 
+
+def _wrapper_get_sfp_error_description(physical_port):
+    if platform_chassis:
+        try:
+            return platform_chassis.get_sfp(physical_port).get_error_description()
+        except NotImplementedError:
+            pass
+    return None
 # Remove unnecessary unit from the raw data
 
 
@@ -553,7 +551,7 @@ def recover_missing_sfp_table_entries(sfp_util, int_tbl, status_tbl, stop_event)
             continue
 
         keys = int_tbl[asic_index].getKeys()
-        if logical_port_name not in keys and not detect_port_in_error_status(logical_port_name, status_tbl[asic_index]):
+        if logical_port_name not in keys and not sfp_status_helper.detect_port_in_error_status(logical_port_name, status_tbl[asic_index]):
             post_port_sfp_info_to_db(logical_port_name, int_tbl[asic_index], transceiver_dict, stop_event)
 
 
@@ -791,30 +789,17 @@ def waiting_time_compensation_with_sleep(time_start, time_to_wait):
 # Update port SFP status table on receiving SFP change event
 
 
-def update_port_transceiver_status_table(logical_port_name, status_tbl, status):
-    fvs = swsscommon.FieldValuePairs([('status', status)])
+def update_port_transceiver_status_table(logical_port_name, status_tbl, status, error_descriptions='N/A'):
+    fvs = swsscommon.FieldValuePairs([('status', status), ('error', error_descriptions)])
     status_tbl.set(logical_port_name, fvs)
 
+
 # Delete port from SFP status table
 
 
 def delete_port_from_status_table(logical_port_name, status_tbl):
     status_tbl._del(logical_port_name)
 
-# Check whether port in error status
-
-
-def detect_port_in_error_status(logical_port_name, status_tbl):
-    rec, fvp = status_tbl.get(logical_port_name)
-    if rec:
-        status_dict = dict(fvp)
-        if status_dict['status'] in errors_block_eeprom_reading:
-            return True
-        else:
-            return False
-    else:
-        return False
-
 # Init TRANSCEIVER_STATUS table
 
 
@@ -844,16 +829,16 @@ def init_port_sfp_status_tbl(stop_event=threading.Event()):
         physical_port_list = logical_port_name_to_physical_port_list(logical_port_name)
         if physical_port_list is None:
             helper_logger.log_error("No physical ports found for logical port '{}'".format(logical_port_name))
-            update_port_transceiver_status_table(logical_port_name, status_tbl[asic_index], SFP_STATUS_REMOVED)
+            update_port_transceiver_status_table(logical_port_name, status_tbl[asic_index], sfp_status_helper.SFP_STATUS_REMOVED)
 
         for physical_port in physical_port_list:
             if stop_event.is_set():
                 break
 
             if not _wrapper_get_presence(physical_port):
-                update_port_transceiver_status_table(logical_port_name, status_tbl[asic_index], SFP_STATUS_REMOVED)
+                update_port_transceiver_status_table(logical_port_name, status_tbl[asic_index], sfp_status_helper.SFP_STATUS_REMOVED)
             else:
-                update_port_transceiver_status_table(logical_port_name, status_tbl[asic_index], SFP_STATUS_INSERTED)
+                update_port_transceiver_status_table(logical_port_name, status_tbl[asic_index], sfp_status_helper.SFP_STATUS_INSERTED)
 
 #
 # Helper classes ===============================================================
@@ -892,7 +877,7 @@ def task_worker(self, y_cable_presence):
                     logger.log_warning("Got invalid asic index for {}, ignored".format(logical_port_name))
                     continue
 
-                if not detect_port_in_error_status(logical_port_name, status_tbl[asic_index]):
+                if not sfp_status_helper.detect_port_in_error_status(logical_port_name, status_tbl[asic_index]):
                     post_port_dom_info_to_db(logical_port_name, dom_tbl[asic_index], self.task_stopping_event)
                     post_port_dom_threshold_info_to_db(logical_port_name, dom_tbl[asic_index], self.task_stopping_event)
                     if y_cable_presence[0] is True:
@@ -1035,7 +1020,7 @@ def task_worker(self, stopping_event, sfp_error_event, y_cable_presence):
         while not stopping_event.is_set():
             next_state = state
             time_start = time.time()
-            status, port_dict = _wrapper_get_transceiver_change_event(timeout)
+            status, port_dict, error_dict = _wrapper_get_transceiver_change_event(timeout)
             if not port_dict:
                 continue
             helper_logger.log_debug("Got event {} {} in state {}".format(status, port_dict, state))
@@ -1095,11 +1080,11 @@ def task_worker(self, stopping_event, sfp_error_event, y_cable_presence):
                                 logger.log_warning("Got invalid asic index for {}, ignored".format(logical_port))
                                 continue
 
-                            if value == SFP_STATUS_INSERTED:
+                            if value == sfp_status_helper.SFP_STATUS_INSERTED:
                                 helper_logger.log_info("Got SFP inserted event")
                                 # A plugin event will clear the error state.
                                 update_port_transceiver_status_table(
-                                    logical_port, status_tbl[asic_index], SFP_STATUS_INSERTED)
+                                    logical_port, status_tbl[asic_index], sfp_status_helper.SFP_STATUS_INSERTED)
                                 helper_logger.log_info("receive plug in and update port sfp status table.")
                                 rc = post_port_sfp_info_to_db(logical_port, int_tbl[asic_index], transceiver_dict)
                                 # If we didn't get the sfp info, assuming the eeprom is not ready, give a try again.
@@ -1111,28 +1096,36 @@ def task_worker(self, stopping_event, sfp_error_event, y_cable_presence):
                                 post_port_dom_threshold_info_to_db(logical_port, dom_tbl[asic_index])
                                 notify_media_setting(logical_port, transceiver_dict, app_port_tbl[asic_index])
                                 transceiver_dict.clear()
-                            elif value == SFP_STATUS_REMOVED:
+                            elif value == sfp_status_helper.SFP_STATUS_REMOVED:
                                 helper_logger.log_info("Got SFP removed event")
                                 update_port_transceiver_status_table(
-                                    logical_port, status_tbl[asic_index], SFP_STATUS_REMOVED)
-                                helper_logger.log_info("receive plug out and pdate port sfp status table.")
+                                    logical_port, status_tbl[asic_index], sfp_status_helper.SFP_STATUS_REMOVED)
+                                helper_logger.log_info("receive plug out and update port sfp status table.")
                                 del_port_sfp_dom_info_from_db(logical_port, int_tbl[asic_index], dom_tbl[asic_index])
-                            elif value in errors_block_eeprom_reading:
-                                helper_logger.log_info("Got SFP Error event")
-                                # Add port to error table to stop accessing eeprom of it
-                                # If the port already in the error table, the stored error code will
-                                # be updated to the new one.
-                                update_port_transceiver_status_table(logical_port, status_tbl[asic_index], value)
-                                helper_logger.log_info("receive error update port sfp status table.")
-                                # In this case EEPROM is not accessible, so remove the DOM info
-                                # since it will be outdated if long time no update.
-                                # but will keep the interface info in the DB since it static.
-                                del_port_sfp_dom_info_from_db(logical_port, None, dom_tbl[asic_index])
-
                             else:
-                                # SFP return unkown event, just ignore for now.
-                                helper_logger.log_warning("Got unknown event {}, ignored".format(value))
-                                continue
+                                try:
+                                    error_bits = int(value)
+                                    helper_logger.log_info("Got SFP error event {}".format(value))
+
+                                    error_descriptions = sfp_status_helper.fetch_generic_error_description(error_bits)
+
+                                    if sfp_status_helper.has_vendor_specific_error(error_bits):
+                                        if error_dict:
+                                            vendor_specific_error_description = error_dict.get(key)
+                                        else:
+                                            vendor_specific_error_description = _wrapper_get_sfp_error_description(key)
+                                        error_descriptions.append(vendor_specific_error_description)
+
+                                    # Add error info to database
+                                    # Any existing error will be replaced by the new one.
+                                    update_port_transceiver_status_table(logical_port, status_tbl[asic_index], value, '|'.join(error_descriptions))
+                                    helper_logger.log_info("Receive error update port sfp status table.")
+                                    # In this case EEPROM is not accessible. The DOM info will be removed since it can be out-of-date.
+                                    # The interface info remains in the DB since it is static.
+                                    if sfp_status_helper.is_error_block_eeprom_reading(error_bits):
+                                        del_port_sfp_dom_info_from_db(logical_port, None, dom_tbl[asic_index])
+                                except (TypeError, ValueError) as e:
+                                    logger.log_error("Got unrecognized event {}, ignored".format(value))
 
                     # Since ports could be connected to a mux cable, if there is a change event process the change for being on a Y cable Port
                     y_cable_helper.change_ports_status_for_y_cable_change_event(

sonic-xcvrd/xcvrd/xcvrd_utilities/sfp_status_helper.py

@@ -0,0 +1,37 @@
+from sonic_platform_base.sfp_base import SfpBase
+
+# SFP status definition, shall be aligned with the definition in get_change_event() of ChassisBase
+SFP_STATUS_REMOVED = '0'
+SFP_STATUS_INSERTED = '1'
+
+# SFP error code dictinary, new elements can be added if new errors need to be supported.
+SFP_ERRORS_BLOCKING_MASK = 0x02
+SFP_ERRORS_GENERIC_MASK = 0x0000FFFE
+SFP_ERRORS_VENDOR_SPECIFIC_MASK = 0xFFFF0000
+
+def is_error_block_eeprom_reading(error_bits):
+    return 0 != (error_bits & SFP_ERRORS_BLOCKING_MASK)
+
+
+def has_vendor_specific_error(error_bits):
+    return 0 != (error_bits & SFP_ERRORS_VENDOR_SPECIFIC_MASK)
+
+
+def fetch_generic_error_description(error_bits):
+    generic_error_bits = (error_bits & SFP_ERRORS_GENERIC_MASK)
+    error_descriptions = []
+    if generic_error_bits:
+        for error_bit, error_description in SfpBase.SFP_ERROR_BIT_TO_DESCRIPTION_DICT.items():
+            if error_bit & generic_error_bits:
+                error_descriptions.append(error_description)
+    return error_descriptions
+
+
+def detect_port_in_error_status(logical_port_name, status_tbl):
+    rec, fvp = status_tbl.get(logical_port_name)
+    if rec:
+        status_dict = dict(fvp)
+        error = status_dict.get('error')
+        return SfpBase.SFP_ERROR_DESCRIPTION_BLOCKING in error
+    return False
+