Re: [PATCH 11/13] dmaengine: kill struct dma_client and supporting infrastructure

["Dan Williams" <dan.j.williams@intel.com>]

On Fri, Dec 12, 2008 at 7:29 AM, Sosnowski, Maciej
<maciej.sosnowski@intel.com> wrote:
> Williams, Dan J wrote:
>> All users have been converted to either the general-purpose allocator,
>> dma_find_channel, or dma_request_channel.
>>
>> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
>> ---
> (...)
>>  /**
>> - * dma_chans_notify_available - broadcast available channels to the
>> clients
>> - */
>> -static void dma_clients_notify_available(void)
>> -{
>> -     struct dma_client *client;
>> -
>> -     mutex_lock(&dma_list_mutex);
>> -
>> -     list_for_each_entry(client, &dma_client_list, global_node)
>> -             dma_client_chan_alloc(client);
>> -
>> -     mutex_unlock(&dma_list_mutex);
>> -}
>
> I agree with Guennadi's concern about removing clients' notification
> of new devices available in the system.
> I understand that this design is based on polling instead,
> however polling is always less efficient approach.
> Do you think that restoring notifications in this redesigned dmaengine
> would be more painful than limiting clients to polling solution?
>

You are missing that net_dma has always "polled".  Consider the case
of how net_dma currently operates prior to ioatdma.ko being loaded.
It periodically calls get_softnet_dma() to see if a channel is
available, if it does not find one it falls back to a cpu copy.  All
that has changed is replacing this custom channel allocation routine
with the unified dma_find_channel() interface.  Channel notifications
are not required.  Either everything is built-in to guarantee that an
engine is available upon request[1], or the client is smart enough to
run without a channel for while like net_dma and raid offload.

Regards,
Dan

[1] http://marc.info/?l=linux-kernel&m=122835195303213&w=2