Re: [PATCH 1/8] [I/OAT] DMA memcpy subsystem

["Andrew Grover" <andy.grover@gmail.com>]

On 3/16/06, Kumar Gala <galak@kernel.crashing.org> wrote:
> It would seem that when a client registers (or shortly there after
> when they call dma_async_client_chan_request()) they would expect to
> get the number of channels they need by some given time period.
>
> For example, lets say a client registers but no dma device exists.
> They will never get called to be aware of this condition.
>
> I would think most clients would either spin until they have all the
> channels they need or fall back to a non-async mechanism.

Clients *are* expected to fall back to non-async if they are not given
channels. The reason it was implemented with callbacks for
added/removed was that the client may be initializing before the
channels are enumerated. For example, the net subsystem will ask for
channels and not get them for a while, until the ioatdma PCI device is
found and its driver loads. In this scenario, we'd like the net
subsystem to be given these channels, instead of them going unused.

> Also, what do you think about adding an operation type (MEMCPY, XOR,
> CRYPTO_AES, etc).  We can than validate if the operation type
> expected is supported by the devices that exist.

No objections, but this speculative support doesn't need to be in our
initial patchset.

> Shouldn't we also have a dma_async_client_chan_free()?

Well we could just define it to be chan_request(0) but it doesn't seem
to be needed. Also, the allocation mechanism we have for channels is
different from alloc/free's semantics, so it may be best to not muddy
the water in this area.

Regards -- Andy