[GSoC2012]the problem of one-rbtree approach in POSIX Key project

[Ashi]

2012/6/15 Gedare Bloom <gedare at rtems.org>

> On Fri, Jun 15, 2012 at 6:02 AM, Ashi <ashi08104 at gmail.com> wrote:
> > Hi, all.
> > Last two weeks, I'm working on the one-rbtree approach in the POSIX Key
> > project. There are several reasons that I start to code without deciding
> > which approach is best(there are several approaches have been discussed
> > before, here[0] is the summary I have done before): first, I want to make
> > myself more clear about the problem lies behind the POSIX Key by coding
> > first. Second, we may also need to implement several different approaches
> > and evaluate them finally. This post is a partly summary of one-rbtree
> > approach and also with many problems in it. By the way, after try to
> > implement the one-rbtree approach, I find the approach manage POSIX Key
> data
> > only by Object manager(also described in [0]) is really immature. I've
> tried
> > to implemented that before starting the one-rbtree approach, however, I'm
> > stuck and turn to one-rbtree approach.
> >
> Can you expand on why the object manager approach does not work?

Yeah, I should explain it earlier:
When I came to the Key getspecific stage of this approach, I found using
object manager to manage all key's value didn't  simplify the key
getspecific at all. The key getspecific interface is:
void *pthread_getspecific( pthread_key_t  key )
what we want to do is return the key's data by using the input key and the
implicit imput the executing thread id. Now if we used object manager to
manage all key's value, that is allocating one object such as:
typedef struct {
Object object;
void *value;
}
then the work becomes finding a kind of mapping between input parameter(the
key and executing thread id) and object.id. That can be done like the
one-rbtree approah or something other. After get the object.id, we can get
the value pointer by the object manager, work is done. This kind design
didn't simplify anything compared with one-rbtree approach, so I give up
this design. I have thought about other reason like the limitation of
object number, because the number of key's value can be very large, but not
clear whether that's a limitation.

> >
> > There are 6 scenarios which we need deal with in POSIX key:
> > 0. Key manage initialization
> > 1. Key create
> > 2. Key setspecific
> > 3. Key getspecific
> > 4. Key delete
> > 5. Thread delete
> >
> > In the one-rbtree approach, I've done the 0~4(I've committed the code to
> > one_rbtree branch in github repo[1]) , and the 5th thread delete hasn't
> been
> > done, and needs more discussion(I'll list the problem below):
> > 0. Key manage initialization
> > I added two data structures in key.h, the POSIX_Keys_Rbtree_node is the
> node
> > of the rbtree which holds all keys' values, and use Key and Thread_id
> member
> > as rbtree key. POSIX_Keys_List_node is the node of a list which is used
> to
> > save all nodes in one key or all nodes in one thread.
> >
> > typedef struct {
> >   /** This field is the rbtree node structure. */
> >   RBTree_Node Node;
> Structure field names are usually lower-case, so "node".
>
OK.

>
> >   /** This field is the POSIX key used as an rbtree key */
> >   pthread_key_t Key;
> >   /** This field is the Thread id also used as an rbtree key */
> >   Objects_Id Thread_id;
> >   /** This field points to the POSIX key value of specific thread */
> >   void *Value;
> >  }  POSIX_Keys_Rbtree_node;
> >
> > typedef struct POSIX_Keys_List_node_ {
> >   /** This field is the pointer which points to the next node in the
> list */
> >   struct POSIX_Keys_List_node_ *Next;
> >   /** This field is the key of list node */
> >   POSIX_Keys_Rbtree_node *Rbnode;
> >  }  POSIX_Keys_List_node;
> >
> Instead of writing your own linked list you should use RTEMS Chains
> (if you need a linked list).
>
I'll check that.

>
> > and revises the POSIX_Keys_Control structure, it has a member Head, which
> > points to the list of all node in specific key. The reason I attach a
> list
> > of nodes in one key is that when key deletion happens, I can delete all
> the
> > key nodes from the rbtree in that key by traverse the list. I wonder is
> > there any better way to get all nodes belong one key from the rbtree?
> >
> Does key delete remove all the keys having the same pthread_key_t, or
> only those keys belonging to the executing thread? If only the one key
> needs to be removed then I don't see the need for this list at all,
> you would just "extract" the one key.
>
I think key delete will remove all the key's value having the same
pthread_key_t, so we need track one key's all value. Here is the POSIX Key
delete standard(
http://pubs.opengroup.org/onlinepubs/009604499/functions/pthread_key_delete.html
).
I wonder if we can search the rbtree(here every node has two keys:
pthread_key_t key and Object_Id thread_id) only specify one key of two? If
so, we needn't the list to keep track.

>
> > typedef struct {
> >    /** This field is the Object control structure. */
> >    Objects_Control     Object;
> >    /** This field is the data destructor. */
> >    void (*destructor) (void *);
> >    /** This field is the head of key's node list */
> >    POSIX_Keys_List_node *Head;
> >  }  POSIX_Keys_Control;
> >
> > The work during the POSIX Key manager initialization is:
> > - besides the _Objects_Initialize_information, the global rbtree
> > _POSIX_Keys_Rbtree is initialized to an empty rbtree.
> >
> > 1. Key create
> > work in key create:
> > - allocate the POSIX_Keys_Control object and initialize it.
> >
> > 2. Key setspecific
> > work in key setspecific:
> > - allocate a POSIX_Keys_Rbtree_node object and a POSIX_Keys_List_node
> object
> > from RTEMS workspace
> > - add the rbtree node to the rbtree
> > - add the list node to the key's list
> > problem:
> > I'm not really clear about the workspace and heap memory, I just refer
> some
> > code in other POSIX manager, all of them used _Workspace_Allocate and no
> > malloc there. Could anyone explain the difference between them? Or is
> there
> > any doc about that?
> >
> Workspace is for allocation of kernel objects. Not sure if there is an
> explanation somewhere, although probably there ought to be...
> Workspace is basically like heap (malloc), except that the workspace
> is constrained in size based on application configuration. These
> objects are rightly allocated from the workspace.
>


> > 3. Key getspecific
> > work in key getspecific:
> > - get the node in the rbtree
> > problem:
> > - I find a _Thread_Enable_dispatch() function in the current
> keygetspecfic.c
> > file, and don't know what is it used for.
> >
> _POSIX_Keys_Get implicitly obtains a dispatch disable lock (via
> Objects_Get). This is releasing the lock. If your code does not use
> Objects_Get then probably it does not need to enable dispatching...
> although you are likely having to protect the global rbtree somehow
> with a critical section?
>
I haven't consided any about the critical section before, I think key
create and key setspecific which write something need protection, right? Is
dispatch disable lock enable to do the protection?

> 4. Key delete
> work in key delete:
> - deallocate all nodes of specific key by iterate the key node list, and
> then delete it from rbtree, also delete it from the list
> - deallocate the POSIX_Key_Control object
>

> Again unclear to me if the key delete should apply to all keys with
> the same pthread_key_t or only the one for a particular thread. We
> need to get an answer to this question before any decision can be made
> about the best path forward.
>
> > 5. Thread delete
> > I find when a posix thread deleted, a _POSIX_Keys_Run_destructors()
> function
> > runs, then all key's data in the deleted thread is destructed. I have
> some
> > idea about it, but haven't implement it. I'm thinking add the necessary
> > information about key to the thread's API_Extensions area. Actually, I
> add a
> > list head pointer to the POSIX_API_Control structure, the pointer points
> to
> > a list of nodes which contains all the key's value pointer in one thread.
> > When thread exits, we can deallocate all the key data by traverse this
> list
> > and then delete the whole list. And we need add node to this list when
> key
> > setspecific. When Key deletes nothing should be done on this list. I
> think
> > we needn't delete key nodes of deleted thread from rbtree when thread
> > deleted, and the only work should be done when thread deleted is
> > deallocating the key's data in that thread, right?
> >
> I think a chain pointing to all of a thread's keys could work fine...
> although this thread-deletion requirement makes me think that every
> thread could just have an rbtree of its own keys. The overhead of an
> rbtree versus a chain is not a lot, and this approach would save
> having a separate list for each thread's keys. You would just use a
> thread's "local" keys rbtree. I'm not sure you would even need locking
> in this case... interesting.
>
Yeah, I'm also thinking about one rbtree per thread design. I'll dig more.

>
> >
> > thanks for any reply on this!
> >
> > links:
> > [0]
> https://docs.google.com/document/d/1orDfUJdnxbxizClFNcOVbh0t3T0vtmt5FYJETSNb6KE/edit
> > [1]https://github.com/ashi08104/rtems/tree/one_rbtree#
> > --
> > Best wishes!
> > Zhongwei Yao
> >
> >
> > _______________________________________________
> > rtems-devel mailing list
> > rtems-devel at rtems.org
> > http://www.rtems.org/mailman/listinfo/rtems-devel
> >
>



-- 
Best wishes!
Zhongwei Yao
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.rtems.org/pipermail/devel/attachments/20120617/73614e8b/attachment-0001.html>