[389-devel] Review of plugin code

Fri Aug 7 23:33:19 UTC 2015

On 08/07/2015 05:18 PM, William Brown wrote:
> On Thu, 2015-08-06 at 14:25 -0700, Noriko Hosoi wrote:
>> Hi William,
>>
>> Very interesting plug-in!
> Thanks. As a plugin, it's value is quite useless due to the nsDS5ReplicaType
> flags. But it's a nice simple exercise to get ones head around how the plugin
> architecture works from scratch. It's one thing to patch a plugin, compared to
> writing one from nothing.
>
>> Regarding betxn plug-in, it is for putting the entire operation -- the
>> primary update + associated updates by the enabled plug-ins -- in one
>> transaction.  By doing so, the entire updates are committed to the DB if
>> and only if all of the updates are successful. Otherwise, all of them
>> are rolled back.  That guarantees there will be no consistency among
>> entries.
> Okay, so if I can be a pain, how to betxn handle reads? Do reads come from
> within the transaction?

Yes.

> Or is there a way to read from the database outside the
> transaction.
>
> Say for example:
>
> begin
> add some object Y
> read Y
> commit
>
> Does read Y see the object within the transaction?

Yes.

> Is there a way to make the
> search happen so that it occurs outside the transaction, IE it doesn't see Y?

Not a nested search operation.  A nested search operation will always 
use the parent/context transaction.

>
>> In that sense, your read-only plug-in is not a good example for betxn
>> since it does not do any updates. :)  Considering the purpose of the
>> "read-only" plug-in, invoking it at the pre-op timing (before the
>> transaction) would be the best.
> Very true! I kind of knew what betxn did, but I wanted to confirm more
> completely in my mind. So I think what my read-only plugin does at the moment
> works quite nicely then outside of betxn.
>
> Is there a piece of documentation (perhaps the plugin guide) that lists the
> order in which these operations are called?

Not sure, but in general it is:

incoming operation from client
front end processing
preoperation
call backend
bepreoperation
start transaction
betxnpreoperation
do operation in the database
betxnpostoperation
end transaction
bepostoperation
return from backend
send result to client
postoperation

>
>> Since MEP requires the updates on the DB, it's supposed to be called in
>> betxn.  That way, what was done in the MEP plug-in is committed or
>> rolled back together with the primary updates.
> Makes sense.
>
>> The toughest part is the deadlock prevention.  At the start transaction,
>> it holds a DB lock.  And most plug-ins maintain its own mutex to protect
>> its resource.  It'd easily cause deadlock situation especially when
>> multiple plug-ins are enabled (which is common :). So, please be careful
>> not to acquire/free locks in the wrong order...
> Of course. This is always an issue in multi-threaded code and anything with
> locking. Stress tests are probably good to find these deadlocks, no?

Yes.  There is some code in dblayer.c that will stress the transaction 
code by locking/unlocking many db pages concurrently with external 
operations.
https://git.fedorahosted.org/cgit/389/ds.git/tree/ldap/servers/slapd/back-ldbm/dblayer.c#n210
https://git.fedorahosted.org/cgit/389/ds.git/tree/ldap/servers/slapd/back-ldbm/dblayer.c#n4131

>
>> About your commented out code in read_only.c, I guess you copied the
>> part from mep.c and are wondering what it is for?
>> There are various type of plug-ins.
>>
>>      $ egrep nsslapd-pluginType dse.ldif | sort | uniq
>>      nsslapd-pluginType: accesscontrol
>>      nsslapd-pluginType: bepreoperation
>>      nsslapd-pluginType: betxnpostoperation
>>      nsslapd-pluginType: betxnpreoperation
>>      nsslapd-pluginType: database
>>      nsslapd-pluginType: extendedop
>>      nsslapd-pluginType: internalpreoperation
>>      nsslapd-pluginType: matchingRule
>>      nsslapd-pluginType: object
>>      nsslapd-pluginType: preoperation
>>      nsslapd-pluginType: pwdstoragescheme
>>      nsslapd-pluginType: reverpwdstoragescheme
>>      nsslapd-pluginType: syntax
>>
>> The reason why slapi_register_plugin and slapi_register_plugin_ext were
>> implemented was:
>>
>>      /*
>>        * Allows a plugin to register a plugin.
>>        * This was added so that 'object' plugins could register all
>>        * the plugin interfaces that it supports.
>>        */
>>
>> On the other hand, MEP has this type.
>>
>>      nsslapd-pluginType: betxnpreoperation
>>
>> The type is not "object", but the MEP plug-in is implemented as having
>> the type.  Originally, it might have been "object"...  Then, we
>> introduced the support for "betxn".  To make the transition to "betxn"
>> smoothly, we put the code to check "betxn" is in the type. If there is
>> "betxn" as in "betxnpreoperation", call the plug-in in betxn, otherwise
>> call them outside of the transaction.  Having the switch in the
>> configuration, we could go back to the original position without
>> rebuilding the plug-in.
>>
>> Since we do not go back to pre-betxn era, the switch may not be too
>> important.  But keeping it would be a good idea for the consistency with
>> the other plug-ins.
>>
>> Does this answer you question?  Please feel free to let us know if it
>> does not.
> That answers some of my question. I guess the larger part of the question is how
> the plugin subsystem treats each pluginType differently and the value of having
> a plugin register to more than one pluginType. Are there some documents you can
> point me to about this?

I'm not sure if there are docs which answer your specific question. But 
a plugin may want to perform tasks at different points in the 
operation.  For example, DNA may want to generate a unique value in the 
pretxn phase, then roll back that value in the posttxn phase if the 
operation failed.  Replication wants to many tasks at many different 
points in the operation.

>
> Additionally, with betxn, this seems quite black-or-white. It's either on a ds
> instance that has betxn support, so every update will be betxn capable, or it's
> not on such a system so you fall back to other methods. Is this correct? With
> new plugins is it even worth writing them without betxn support?

Correct.  It is not even worth writing a new plugin without betxn 
support, if it does any update to the database that depends on other 
updates to the database triggered by the incoming operation from the client.

>
>
>> I'm sure our team is interested in your idea and work, so let me share
>> your test plug-in with them.
> Sure. It's not really that useful, like I said, nsDS5ReplicaType already does
> this job. It was just an exercise to get my head more into the framework before
> I work on http://directory.fedoraproject.org/docs/389ds/design/mep-rework.html
>
> Thanks for the review, and for answering my questions. Your advice has helped a
> lot!
>
> Sincerely,
>
> --
> 389-devel mailing list
> 389-devel at lists.fedoraproject.org
> https://admin.fedoraproject.org/mailman/listinfo/389-devel