Fixed a bug where the storage service might crash (#583) #595
Conversation
summary: When the storage service deamon is started, the nebula::meta::MetaClient will use the MetaServerBasedPartManager object(MetaClient::registerListener), but the MetaServerBasedPartManager object will be released before nebula::meta::MetaClient(through KVStore object). Added the MetaClient::unRegisterListener method to stop the background thread and reset the listener_ object, And the unRegisterListener is called when the MetaServerBasedPartManager object was released.
|
Unit testing passed. |
|
I am a little confused. Since PartManager |
That's because two threads will be started in metaclient, and MetaServerBasedPartManager will be used in these two threads. |
src/meta/client/MetaClient.h
Outdated
| listener_ = listener; | ||
| } | ||
|
|
||
| void unRegisterListener() { | ||
| deInit(); |
There was a problem hiding this comment.
I understood now. Maybe you should not call deInit here? Because other's may still use the bgThread in MetaClient, e.g. SchemaManager.
There was a problem hiding this comment.
Here we need to stop the background thread first, otherwise it will have race conditions in listener_.
There was a problem hiding this comment.
We'd better protect listener_ in MetaClient instead of stop bgThread directly. Because we don't want meta client is controlled by kvstore.
There was a problem hiding this comment.
According to my understanding.
1 metaclient itself is dependent on the listener_ object.
2 If listener_ is released, both two threads of metaclient will not work, then there is no need for metaclient at this time.
There was a problem hiding this comment.
Not exactly. Listener_ is not necessary for meta client.
Use a read-write lock in metaclient to protect the listener_.
|
Unit testing passed. |
src/meta/client/MetaClient.h
Outdated
| @@ -69,9 +69,16 @@ class MetaClient { | |||
| void init(); | |||
|
|
|||
| void registerListener(MetaChangedListener* listener) { | |||
| CHECK(listener_ == nullptr); | |||
There was a problem hiding this comment.
Move the check under the lock
There was a problem hiding this comment.
Totally, the pr lgtm. Does the crash problem fixed after unRegisterListener?
Yes, this PR will fix the crash.
Fixed possible thread competition for checking listener_ .
|
Unit testing passed. |
|
Unit testing passed. |
…esoft-inc#595) * Fixed a bug where the storage service might crash (vesoft-inc#583) summary: When the storage service deamon is started, the nebula::meta::MetaClient will use the MetaServerBasedPartManager object(MetaClient::registerListener), but the MetaServerBasedPartManager object will be released before nebula::meta::MetaClient(through KVStore object). Added the MetaClient::unRegisterListener method to stop the background thread and reset the listener_ object, And the unRegisterListener is called when the MetaServerBasedPartManager object was released. * Fixed a bug where the storage service might crash (vesoft-inc#583) Use a read-write lock in metaclient to protect the listener_. * Fixed a bug where the storage service might crash (vesoft-inc#583) Fixed possible thread competition for checking listener_ .
* uuid * uuid Co-authored-by: jakevin <30525741+jackwener@users.noreply.github.com>
summary:
When the storage service deamon is started, the nebula::meta::MetaClient
will use the MetaServerBasedPartManager object(MetaClient::registerListener),
but the MetaServerBasedPartManager object will be released before
nebula::meta::MetaClient(through KVStore object).
Added the MetaClient::unRegisterListener method to stop the background thread
and reset the listener_ object, And the unRegisterListener is called
when the MetaServerBasedPartManager object was released.