sonic-orchagent与syncd之间的请求与应答

23次阅读

共计 14354 个字符,预计需要花费 36 分钟才能阅读完成。

​ syncd 进程是介于 orchagent 与 driver 之间的进程。syncd 从 asic-db 中读取的数据经转换后调用驱动提供的 sai 接口进行下硬件,同时需要将驱动的应答进行一定的处理,还需要处理驱动的事件通知 (比如端口 up/down,mac 老化等信息)。处理的消息如下图所示:

orchagent 与 syncd 之间的操作

orchagent 与 syncd 之间会进行如下几种操作:

  • create:创建一个对象
  • remove: 删除一个对象
  • set: 设置对象属性
  • get: 获取对象属性
  • notify:driver 事件通知

对于 create,remove,set 请求,orchagent 会在 sairedis 层构建一个虚拟的 sai 层:sairedis。orchagent 执行 sai 接口只是对 asic-db 进行操作,生成或者删除虚拟对象 (vid)。默认所有操作都是成功的,直接返回,不等待 syncd 的应答。执行上图的 1 和 6。syncd 从 asic-db 中读出请求执行上图的 2,3,4。如果 4 步骤返回成功,则整个请求运行结束,否则 syncd 将会发送 shutdown 通知给 orchagent。orchagent 会退出,如上图的 5,6.

对于 get 操作,orchagent 执行 1 后会使用 select 阻塞等待 syncd 的应答,如果 syncd 在 60 分钟内没有应答,那么 orchagent 会产生 segment 退出。get 操作执行顺序为 1 ->2->3->4->5->6。

对于 driver 的 notify,orchagent 会在主进程的 select 中监听 asic-db。驱动检测到硬件事件后,调用 syncd 注册的回调函数通知 syncd。syncd 中有一个专门处理 driver-notify 的线程 ntf-thread。ntf-thread 解析 driver 的 notify,然后通过 asic-db 通知 orchagent。执行顺序 7 ->8->9。

注:orchagent 与 syncd 关于 sai 这一层非常相似。它们会调用大量的同名函数。这些函数只是名字相同,orchagent 调用的是 sai-redis 库中的函数,而 syncd 调用的是 driver 提供的 sai 库

get 操作阻塞等待

orchagent 执行 sai 的 get 操作时会调用到 redis_generic_get 函数。

std::shared_ptr<swss::ConsumerTable> g_redisGetConsumer;
sai_status_t redis_generic_get(
_In_ sai_object_type_t object_type,
_In_ sai_object_id_t object_id,
_In_ uint32_t attr_count,
_Out_ sai_attribute_t *attr_list)
{SWSS_LOG_ENTER();
std::string str_object_id = sai_serialize_object_id(object_id);
return internal_redis_generic_get(
object_type,
str_object_id,
attr_count,
attr_list);
}
sai_status_t internal_redis_generic_get(
_In_ sai_object_type_t object_type,
_In_ const std::string &serialized_object_id,
_In_ uint32_t attr_count,
_Out_ sai_attribute_t *attr_list)
{SWSS_LOG_ENTER();
/*
* Since user may reuse buffers, then oid list buffers maybe not cleared
* and contain som garbage, let's clean them so we send all oids as null to
* syncd.
*/
clear_oid_values(object_type, attr_count, attr_list);
std::vector<swss::FieldValueTuple> entry = SaiAttributeList::serialize_attr_list(
object_type,
attr_count,
attr_list,
false);
std::string str_object_type = sai_serialize_object_type(object_type);
std::string key = str_object_type + ":" + serialized_object_id;
SWSS_LOG_DEBUG("generic get key: %s, fields: %lu", key.c_str(), entry.size());
if (g_record)
{recordLine("g|" + key + "|" + joinFieldValues(entry));
}
// get is special, it will not put data
// into asic view, only to message queue
// 写入本次 get 事件
g_asicState->set(key, entry, "get");
// wait for response
// 创建临时 select
swss::Select s;
// 添加事件
s.addSelectable(g_redisGetConsumer.get());
// 循环等待 syncd 的应答
while (true)
{SWSS_LOG_DEBUG("wait for response");
swss::Selectable *sel;
// 阻塞等待,时间为 GET_RESPONSE_TIMEOUT
int result = s.select(&sel, GET_RESPONSE_TIMEOUT);
// 只处理应答情况 OBJECT
if (result == swss::Select::OBJECT)
{
swss::KeyOpFieldsValuesTuple kco;
g_redisGetConsumer->pop(kco);
const std::string &op = kfvOp(kco);
const std::string &opkey = kfvKey(kco);
SWSS_LOG_DEBUG("response: op = %s, key = %s", opkey.c_str(), op.c_str());
if (op != "getresponse") // ignore non response messages
{continue;}
sai_status_t status = internal_redis_get_process(
object_type,
attr_count,
attr_list,
kco);
if (g_record)
{const std::string &str_status = kfvKey(kco);
const std::vector<swss::FieldValueTuple> &values = kfvFieldsValues(kco);
// first serialized is status
recordLine("G|" + str_status + "|" + joinFieldValues(values));
}
SWSS_LOG_DEBUG("generic get status: %d", status);
return status;
}
SWSS_LOG_ERROR("generic get failed due to SELECT operation result: %s", getSelectResultAsString(result).c_str());
break;
}
// 超时和异常都返回 SAI_STATUS_FAILURE
if (g_record)
{recordLine("G|SAI_STATUS_FAILURE");
}
SWSS_LOG_ERROR("generic get failed to get response");
return SAI_STATUS_FAILURE;
}

对于 get 操作,当 syncd 比较忙的时候,极端情况下会导致 orchagent 异常退出。

notify 处理

syncd 向驱动注册回调函数

syncd 定义了几个 notify 全局函数指针:

sai_switch_state_change_notification_fn on_switch_state_change_ntf = on_switch_state_change;
sai_switch_shutdown_request_notification_fn on_switch_shutdown_request_ntf = on_switch_shutdown_request;
sai_fdb_event_notification_fn on_fdb_event_ntf = on_fdb_event;
sai_port_state_change_notification_fn on_port_state_change_ntf = on_port_state_change;
sai_packet_event_notification_fn on_packet_event_ntf = on_packet_event;
sai_queue_pfc_deadlock_notification_fn on_queue_deadlock_ntf = on_queue_deadlock;

syncd 和 sai 共享命名空间,所以驱动直接使用这些函数指针即可调用对应的函数,在初始化的时候将这些全局函数指针通过 sai_set_switch_attribute 函数设置到 sai 层。

syncd 设置 sai 层

void check_notifications_pointers(
_In_ uint32_t attr_count,
_In_ sai_attribute_t *attr_list)
{SWSS_LOG_ENTER();
/*
* This function should only be called on CREATE/SET api when object is
* SWITCH.
*
* Notifications pointers needs to be corrected since those we receive from
* sairedis are in sairedis memory space and here we are using those ones
* we declared in syncd memory space.
*
* Also notice that we are using the same pointers for ALL switches.
*/
for (uint32_t index = 0; index < attr_count; ++index)
{sai_attribute_t &attr = attr_list[index];
auto meta = sai_metadata_get_attr_metadata(SAI_OBJECT_TYPE_SWITCH, attr.id);
if (meta->attrvaluetype != SAI_ATTR_VALUE_TYPE_POINTER)
{continue;}
/*
* Does not matter if pointer is valid or not, we just want the
* previous value.
*/
sai_pointer_t prev = attr.value.ptr;
if (prev == NULL)
{
/*
* If pointer is NULL, then fine, let it be.
*/
continue;
}
switch (attr.id)
{
case SAI_SWITCH_ATTR_SWITCH_STATE_CHANGE_NOTIFY:
attr.value.ptr = (void*)on_switch_state_change_ntf;
break;
case SAI_SWITCH_ATTR_SHUTDOWN_REQUEST_NOTIFY:
attr.value.ptr = (void*)on_switch_shutdown_request_ntf;
break;
case SAI_SWITCH_ATTR_FDB_EVENT_NOTIFY:
attr.value.ptr = (void*)on_fdb_event_ntf;
break;
case SAI_SWITCH_ATTR_PORT_STATE_CHANGE_NOTIFY:
attr.value.ptr = (void*)on_port_state_change_ntf;
break;
case SAI_SWITCH_ATTR_PACKET_EVENT_NOTIFY:
attr.value.ptr = (void*)on_packet_event_ntf;
break;
case SAI_SWITCH_ATTR_QUEUE_PFC_DEADLOCK_NOTIFY:
attr.value.ptr = (void*)on_queue_deadlock_ntf;
break;
default:
SWSS_LOG_ERROR("pointer for %s is not handled, FIXME!", meta->attridname);
continue;
}
/*
* Here we translated pointer, just log it.
*/
SWSS_LOG_INFO("%s: %lp (orch) => %lp (syncd)", meta->attridname, prev, attr.value.ptr);
}
}
/*
* Routine Description:
* Set switch attribute value
*
* Arguments:
* [in] switch_id Switch id
* [in] attr - switch attribute
*
* Return Values:
* SAI_STATUS_SUCCESS on success
* Failure status code on error
*/
sai_status_t sai_set_switch_attribute(_In_ sai_object_id_t switch_id,
_In_ const sai_attribute_t *attr) {SAI_LOG_ENTER();
sai_status_t status = SAI_STATUS_SUCCESS;
switch_status_t switch_status = SWITCH_STATUS_SUCCESS;
switch_uint64_t flags = 0;
switch_api_device_info_t api_device_info;
sai_packet_action_t sai_packet_action;
switch_acl_action_t switch_packet_action;
switch_packet_type_t switch_packet_type = SWITCH_PACKET_TYPE_UNICAST;
bool cut_through = false;
if (!attr) {
status = SAI_STATUS_INVALID_PARAMETER;
SAI_LOG_ERROR("null attribute: %s", sai_status_to_string(status));
return status;
}
memset(&api_device_info, 0x0, sizeof(api_device_info));
if (status != SAI_STATUS_SUCCESS) {return status;}
if (attr->id <= SAI_SWITCH_ATTR_ACL_STAGE_EGRESS) { // Unsupported
SAI_LOG_DEBUG("Switch attribute set: %s", switch_attr_name[attr->id]);
}
switch (attr->id) {
......
case SAI_SWITCH_ATTR_FDB_EVENT_NOTIFY:
sai_switch_notifications.on_fdb_event = attr->value.ptr;
break;
case SAI_SWITCH_ATTR_PORT_STATE_CHANGE_NOTIFY:
sai_switch_notifications.on_port_state_change = attr->value.ptr;
break;
case SAI_SWITCH_ATTR_PACKET_EVENT_NOTIFY:
sai_switch_notifications.on_packet_event = attr->value.ptr;
break;
case SAI_SWITCH_ATTR_SWITCH_STATE_CHANGE_NOTIFY:
sai_switch_notifications.on_switch_state_change = attr->value.ptr;
break;
case SAI_SWITCH_ATTR_SHUTDOWN_REQUEST_NOTIFY:
sai_switch_notifications.on_switch_shutdown_request = attr->value.ptr;
break;
......
default:
SAI_LOG_ERROR("Unsupported Switch attribute: %d", attr->id);
// Unsupported: Temporary hack till all attrs are supported
switch_status = SWITCH_STATUS_SUCCESS;
}
......
}

sai 接口初始化的时候会向驱动注册回调函数,回调函数中会调用我们注册的全局函数指针,我们以 fdb 为例进行说明:

sai_status_t sai_fdb_initialize(sai_api_service_t *sai_api_service) {SAI_LOG_DEBUG("initializing fdb");
sai_api_service->fdb_api = fdb_api;
switch_uint16_t mac_event_flags = 0;
mac_event_flags |= SWITCH_MAC_EVENT_LEARN | SWITCH_MAC_EVENT_AGE |
SWITCH_MAC_EVENT_MOVE | SWITCH_MAC_EVENT_DELETE;
switch_api_mac_notification_register(device, SWITCH_SAI_APP_ID, mac_event_flags, &sai_mac_notify_cb);
switch_api_mac_table_set_learning_timeout(device, SAI_L2_LEARN_TIMEOUT);
return SAI_STATUS_SUCCESS;
}
// 初始化 fdb 的 sai 接口的时候,向驱动注册了 sai_mac_notify_cb 回调函数。static void sai_mac_notify_cb(const switch_device_t device,
const uint16_t num_entries,
const switch_api_mac_entry_t *mac_entry,
const switch_mac_event_t mac_event,
void *app_data) {SAI_LOG_ENTER();
sai_fdb_event_notification_data_t fdb_event[num_entries];
sai_attribute_t attr_lists[num_entries][2];
uint16_t entry = 0;
// 判断回调函数是否为空
if (!sai_switch_notifications.on_fdb_event) {return;}
if (!mac_entry) {SAI_LOG_ERROR("invalid argument");
return;
}
if (!num_entries) {SAI_LOG_DEBUG("sai mac notify callback with null entries");
return;
}
for (entry = 0; entry < num_entries; entry++) {memset(&fdb_event[entry], 0, sizeof(fdb_event[entry]));
fdb_event[entry].event_type = switch_mac_event_to_sai_fdb_event(mac_event);
memcpy(fdb_event[entry].fdb_entry.mac_address,
mac_entry[entry].mac.mac_addr,
ETH_ALEN);
fdb_event[entry].fdb_entry.switch_id =
(((unsigned long)SWITCH_HANDLE_TYPE_DEVICE)
<< SWITCH_HANDLE_TYPE_SHIFT) |
0x1;
fdb_event[entry].fdb_entry.bv_id = mac_entry[entry].network_handle;
memset(attr_lists[entry], 0, sizeof(attr_lists[entry]));
attr_lists[entry][0].id = SAI_FDB_ENTRY_ATTR_TYPE;
attr_lists[entry][0].value.s32 = SAI_FDB_ENTRY_TYPE_DYNAMIC;
attr_lists[entry][1].id = SAI_FDB_ENTRY_ATTR_BRIDGE_PORT_ID;
attr_lists[entry][1].value.oid = mac_entry->handle;
fdb_event[entry].attr_count = 2;
if (fdb_event[entry].event_type == SAI_FDB_EVENT_FLUSHED) {
// Overwriting now for SONiC to be able to process it correctly
fdb_event[entry].event_type = SAI_FDB_EVENT_AGED;
}
fdb_event[entry].attr = attr_lists[entry];
}
// 调用 syncd 的回调函数
sai_switch_notifications.on_fdb_event(num_entries, fdb_event);
SAI_LOG_EXIT();
return;
}

syncd 启动 notify 线程

std::shared_ptr<std::thread> ntf_process_thread;
void startNotificationsProcessingThread()
{SWSS_LOG_ENTER();
runThread = true;
ntf_process_thread = std::make_shared<std::thread>(ntf_process_function);
}
void ntf_process_function()
{SWSS_LOG_ENTER();
while (runThread)
{cv.wait(ulock);
// this is notifications processing thread context, which is different
// from SAI notifications context, we can safe use g_mutex here,
// processing each notification is under same mutex as processing main
// events, counters and reinit
swss::KeyOpFieldsValuesTuple item;
while (tryDequeue(item))// 从队列中取出 notify
{processNotification(item);// 处理 notify
}
}
}
bool tryDequeue(_Out_ swss::KeyOpFieldsValuesTuple &item)
{std::lock_guard<std::mutex> lock(queue_mutex);
SWSS_LOG_ENTER();
if (ntf_queue.empty())
{return false;}
item = ntf_queue.front();
ntf_queue.pop();
return true;
}
void processNotification(_In_ const swss::KeyOpFieldsValuesTuple &item)
{std::lock_guard<std::mutex> lock(g_mutex);
SWSS_LOG_ENTER();
std::string notification = kfvKey(item);
std::string data = kfvOp(item);
if (notification == "switch_state_change")
{handle_switch_state_change(data);
}
else if (notification == "fdb_event")
{handle_fdb_event(data);
}
else if (notification == "port_state_change")
{handle_port_state_change(data);
}
else if (notification == "switch_shutdown_request")
{handle_switch_shutdown_request(data);
}
else if (notification == "queue_deadlock")
{handle_queue_deadlock(data);
}
else
{SWSS_LOG_ERROR("unknow notification: %s", notification.c_str());
}
}
void handle_fdb_event(_In_ const std::string &data)
{SWSS_LOG_ENTER();
uint32_t count;
sai_fdb_event_notification_data_t *fdbevent = NULL;
sai_deserialize_fdb_event_ntf(data, count, &fdbevent);
process_on_fdb_event(count, fdbevent);
sai_deserialize_free_fdb_event_ntf(count, fdbevent);
}
void process_on_fdb_event(
_In_ uint32_t count,
_In_ sai_fdb_event_notification_data_t *data)
{SWSS_LOG_ENTER();
SWSS_LOG_DEBUG("fdb event count: %d", count);
for (uint32_t i = 0; i < count; i++)
{sai_fdb_event_notification_data_t *fdb = &data[i];
SWSS_LOG_DEBUG("fdb %u: type: %d", i, fdb->event_type);
fdb->fdb_entry.switch_id = translate_rid_to_vid(fdb->fdb_entry.switch_id, SAI_NULL_OBJECT_ID);
fdb->fdb_entry.bv_id = translate_rid_to_vid(fdb->fdb_entry.bv_id, fdb->fdb_entry.switch_id);
translate_rid_to_vid_list(SAI_OBJECT_TYPE_FDB_ENTRY, fdb->fdb_entry.switch_id, fdb->attr_count, fdb->attr);
/*
* Currently because of bcrm bug, we need to install fdb entries in
* asic view and currently this event don't have fdb type which is
* required on creation.
*/
redisPutFdbEntryToAsicView(fdb);
}
std::string s = sai_serialize_fdb_event_ntf(count, data);
send_notification("fdb_event", s);
}
void send_notification(
_In_ std::string op,
_In_ std::string data,
_In_ std::vector<swss::FieldValueTuple> &entry)
{SWSS_LOG_ENTER();
SWSS_LOG_INFO("%s %s", op.c_str(), data.c_str());
// 写入数据库
notifications->send(op, data, entry);
SWSS_LOG_DEBUG("notification send successfull");
}
void send_notification(
_In_ std::string op,
_In_ std::string data)
{SWSS_LOG_ENTER();
std::vector<swss::FieldValueTuple> entry;
send_notification(op, data, entry);
}

orchagent 启动 notify 线程

// 启动线程
sai_status_t sai_api_initialize(
_In_ uint64_t flags,
_In_ const sai_service_method_table_t* services)
{
......
notification_thread = std::make_shared<std::thread>(std::thread(ntf_thread));
......
}
// 线程主函数
void ntf_thread()
{SWSS_LOG_ENTER();
swss::Select s;
s.addSelectable(g_redisNotifications.get());
s.addSelectable(&g_redisNotificationTrheadEvent);
while (g_run)
{
swss::Selectable *sel;
int result = s.select(&sel);
if (sel == &g_redisNotificationTrheadEvent)
{
// user requested shutdown_switch
break;
}
if (result == swss::Select::OBJECT)
{
swss::KeyOpFieldsValuesTuple kco;
std::string op;
std::string data;
std::vector<swss::FieldValueTuple> values;
g_redisNotifications->pop(op, data, values);
SWSS_LOG_DEBUG("notification: op = %s, data = %s", op.c_str(), data.c_str());
handle_notification(op, data, values);
}
}
}
void handle_fdb_event(_In_ const std::string &data)
{SWSS_LOG_ENTER();
SWSS_LOG_DEBUG("data: %s", data.c_str());
uint32_t count;
sai_fdb_event_notification_data_t *fdbevent = NULL;
sai_deserialize_fdb_event_ntf(data, count, &fdbevent);
{std::lock_guard<std::mutex> lock(g_apimutex);
// NOTE: this meta api must be under mutex since
// it will access meta DB and notification comes
// from different thread
meta_sai_on_fdb_event(count, fdbevent);
}
if (on_fdb_event != NULL)
{on_fdb_event(count, fdbevent);
}
sai_deserialize_free_fdb_event_ntf(count, fdbevent);
}

syncd 的回调函数

std::mutex queue_mutex;
std::queue<swss::KeyOpFieldsValuesTuple> ntf_queue;
void on_fdb_event(
_In_ uint32_t count,
_In_ const sai_fdb_event_notification_data_t *data)
{SWSS_LOG_ENTER();
std::string s = sai_serialize_fdb_event_ntf(count, data);
enqueue_notification("fdb_event", s);
}
void enqueue_notification(
_In_ std::string op,
_In_ std::string data,
_In_ std::vector<swss::FieldValueTuple> &entry)
{SWSS_LOG_ENTER();
SWSS_LOG_INFO("%s %s", op.c_str(), data.c_str());
swss::KeyOpFieldsValuesTuple item(op, data, entry);
// this is notification context, so we need to protect queue
std::lock_guard<std::mutex> lock(queue_mutex);
// 压入队列
ntf_queue.push(item);
cv.notify_all();}
void enqueue_notification(
_In_ std::string op,
_In_ std::string data)
{SWSS_LOG_ENTER();
std::vector<swss::FieldValueTuple> entry;
enqueue_notification(op, data, entry);
}

上面三部分就是硬件触发的异步事件从硬件层同步到 syncd 层,再到 orchagent 层的处理过程。涉及一个回调函数,两个 notify 处理线程。

正文完
 0