Re: [Qemu-devel] [PATCH v4 2/3] The support for block queue

[Zhi Yong Wu]

On Tue, Aug 2, 2011 at 4:21 AM, Ryan Harper <address@hidden> wrote:
> * Zhi Yong Wu <address@hidden> [2011-08-01 01:30]:
>
>> +static AIOPool block_queue_pool = {
>> +    .aiocb_size         = sizeof(struct BlockDriverAIOCB),
>> +    .cancel             = qemu_block_queue_cancel,
>> +};
>> +
>> +static void qemu_block_queue_callback(void *opaque, int ret)
>> +{
>> +    BlockDriverAIOCB *acb = opaque;
>> +
>> +    qemu_aio_release(acb);
>> +}
>> +
>
> So, here we really want to invoke the original commands callback, and
> then free the request here (via qemu_aio_release()).  see below.
>
>
>> +BlockDriverAIOCB *qemu_block_queue_enqueue(BlockQueue *queue,
>> +                        BlockDriverState *bs,
>> +                        BlockRequestHandler *handler,
>> +                        int64_t sector_num,
>> +                        QEMUIOVector *qiov,
>> +                        int nb_sectors,
>> +                        BlockDriverCompletionFunc *cb,
>> +                        void *opaque)
>> +{
>> +    BlockIORequest *request;
>> +    BlockDriverAIOCB *acb;
>> +
>> +    request = qemu_malloc(sizeof(BlockIORequest));
>> +    request->bs = bs;
>> +    request->handler = handler;
>> +    request->sector_num = sector_num;
>> +    request->qiov = qiov;
>> +    request->nb_sectors = nb_sectors;
>> +    request->cb = cb;
>> +    request->opaque = opaque;
>> +
>> +    QTAILQ_INSERT_TAIL(&queue->requests, request, entry);
>> +
>> +    acb = qemu_aio_get(&block_queue_pool, bs,
>> +                       qemu_block_queue_callback, opaque);
>> +    request->acb = acb;
>> +
>> +    return acb;
>> +}
>> +
>> +int qemu_block_queue_handler(BlockIORequest *request)
>> +{
>> +    int ret;
>> +    BlockDriverAIOCB *res;
>> +
>> +    /* indicate this req is from block queue */
>> +    request->bs->req_from_queue = true;
>> +
>> +    res = request->handler(request->bs, request->sector_num,
>> +                           request->qiov, request->nb_sectors,
>> +                           request->cb, request->opaque);
>> +
>> +    if (request->acb) {
>> +        qemu_block_queue_callback(request->acb, 0);
>> +    }
>> +
>> +    ret = (res == NULL) ? 0 : 1;
>> +
>> +    return ret;
>> +}
>
>
> You don't want to malloc the BlockIORequest directly in _enqueue(), rather
> you want to allocate that from your AIOPool via qemu_aio_get().  As it
> is now, we're allocating two BlockIORequests (malloc and then a
> aio_get()).  You'll need a BlockDriverAIOCB in the BlockIORequest
> structure.  Then in your queue_handler, instead of passing the original
> read or write callback (request->cb), you want to hook the block_queue 
> callback
> (qemu_block_queue_callback()), in that callback you can then invoke the
> request callback and then release the request.
Right
>
> The request should be, only one malloc (via the pool which will re-use
> the memory instead of incuring a malloc on every request), and then you
> release the memory back to the pool once your request is complete, which
> you'll know after wiring up the block_queue callback to the completion
> of the request's handler.  And then since we don't double allocate, you
> won't need to do the qemu_free(request) in block.c in block_timer...
No, please see the comments in another patch. We need to return a
valid acb to upper layer in order to tell it the request has been
succeeded in starting. But actually the request is enqueued, so normal
acb can not be returned immediately because this request is not
handled by the driver; so we need to malloc an immediate acb struct,
and return it to upper layer.
>
>
> --
> Ryan Harper
> Software Engineer; Linux Technology Center
> IBM Corp., Austin, Tx
> address@hidden
>



-- 
Regards,

Zhi Yong Wu