On 04/15/2020 12:35 PM, Mike Christie wrote:
> On 04/14/2020 09:30 PM, Bart Van Assche wrote:
>> On 2020-04-13 22:15, Mike Christie wrote:
>>> @@ -537,8 +538,15 @@ void transport_deregister_session_configfs(struct se_session *se_sess)
>>> }
>>> EXPORT_SYMBOL(transport_deregister_session_configfs);
>>>
>>> +
>>
>> A single blank line is probably sufficient here?
>>
>
> Yes. That was a cut and paste mistake when I was separating the code
> into patches. Will fix.
>
>
>>> void transport_free_session(struct se_session *se_sess)
>>> {
>>> + kobject_put(&se_sess->kobj);
>>> +}
>>> +EXPORT_SYMBOL(transport_free_session);
>>> +
>>> +void __target_free_session(struct se_session *se_sess)
>>> +{
>>> struct se_node_acl *se_nacl = se_sess->se_node_acl;
>>>
>>> /*
>>> @@ -582,7 +590,6 @@ void transport_free_session(struct se_session *se_sess)
>>> percpu_ref_exit(&se_sess->cmd_count);
>>> kmem_cache_free(se_sess_cache, se_sess);
>>> }
>>> -EXPORT_SYMBOL(transport_free_session);
>>
>> Does this patch defer execution of the code inside
>> transport_free_session() from when transport_free_session() is called to
>> when the last reference to a session is dropped? Can that have
>
> Yes.
>
>> unintended side effects? How about keeping most of the code that occurs
>
> Yes. For example, we drop the refcount on the ACL in
> __target_free_session so that is now not done until the last session
> rerfcount is done. I did this because we reference the acl in a sysfs file.
>
>
>> in transport_free_session() in that function and only freeing the memory
>> associated with the session if the last reference is dropped?
>>
>
> I tried to minimize it already.
>
> That is why I have the new session->fabric_free_cb in the next patch.
> That way we do not need refcounts on structs like the tpg and can detach
> that like normal in
> transport_deregister_session/transport_deregister_session_configfs.
>
Oh yeah, James and Bart, while investigating Bart's comment I noticed
there is a bug where in the session->fabric_free_cb the fabric module
needs the fabric_sess_ptr but that will already have been cleared in
transport_deregister_session.
So James, you will hit a bug in there if you try to adapt elx to these
patches right now.
I will resend with that fixed and Bart's comments handled.
