Re: [PATCH v2 2/2] dmaengine: apple-sio: Add Apple SIO driver

[Vinod Koul]

On 04-10-23, 15:52, Martin Povišer wrote:
> 
> > On 4. 10. 2023, at 15:46, Vinod Koul <vkoul@xxxxxxxxxx> wrote:
> > 
> > On 04-10-23, 15:32, Martin Povišer wrote:
> > 
> >>>> + * There are two kinds of 'transaction descriptors' in play here.
> >>>> + *
> >>>> + * There's the struct sio_tx, and the struct dma_async_tx_descriptor embedded
> >>>> + * inside, which jointly represent a transaction to the dmaengine subsystem.
> >>>> + * At this time we only support those transactions to be cyclic.
> >>>> + *
> >>>> + * Then there are the coprocessor descriptors, which is what the coprocessor
> >>>> + * knows and understands. These don't seem to have a cyclic regime, so we can't
> >>>> + * map the dmaengine transaction on an exact coprocessor counterpart. Instead
> >>>> + * we continually queue up many coprocessor descriptors to implement a cyclic
> >>>> + * transaction.
> >>>> + *
> >>>> + * The number below is the maximum of how far ahead (how many) coprocessor
> >>>> + * descriptors we should be queuing up, per channel, for a cyclic transaction.
> >>>> + * Basically it's a made-up number.
> >>>> + */
> >>>> +#define SIO_MAX_NINFLIGHT 4
> >>> 
> >>> you meant SIO_MAX_INFLIGHT if not what is NINFLIGHT?
> >> 
> >> I mean the number is arbitrary, it doesn’t reflect any coprocessor limit since
> >> I haven’t run the tests to figure one out. It's supposed to be a small reasonable
> >> number.
> > 
> > Sorry that was not my question. Should this macro be SIO_MAX_NINFLIGHT
> > or SIO_MAX_INFLIGHT..?
> 
> Yeah, I realized after I sent the reply, sorry. I don’t know what you would
> interpret to be the difference between NINFLIGHT and INFLIGHT, in my book
> both would be the "number of inflight” in the context here.

No worries, I wante to check that was on purpose, it is fine


> >>>> +static int sio_device_config(struct dma_chan *chan,
> >>>> +      struct dma_slave_config *config)
> >>>> +{
> >>>> + struct sio_chan *siochan = to_sio_chan(chan);
> >>>> + struct sio_data *sio = siochan->host;
> >>>> + bool is_tx = sio_chan_direction(siochan->no) == DMA_MEM_TO_DEV;
> >>>> + struct sio_shmem_chan_config *cfg = sio->shmem;
> >>>> + int ret;
> >>>> +
> >>>> + switch (is_tx ? config->dst_addr_width : config->src_addr_width) {
> >>>> + case DMA_SLAVE_BUSWIDTH_1_BYTE:
> >>>> + cfg->datashape = 0;
> >>>> + break;
> >>>> + case DMA_SLAVE_BUSWIDTH_2_BYTES:
> >>>> + cfg->datashape = 1;
> >>>> + break;
> >>>> + case DMA_SLAVE_BUSWIDTH_4_BYTES:
> >>>> + cfg->datashape = 2;
> >>>> + break;
> >>>> + default:
> >>>> + return -EINVAL;
> >>>> + }
> >>>> +
> >>>> + cfg->fifo = 0x800;
> >>>> + cfg->limit = 0x800;
> >>>> + cfg->threshold = 0x800;
> >>>> + dma_wmb();
> >>> 
> >>> ??
> >> 
> >> Again, shared memory
> >> 
> >>>> +
> >>>> + ret = sio_call(sio, FIELD_PREP(SIOMSG_TYPE, MSG_CONFIGURE) |
> >>>> +     FIELD_PREP(SIOMSG_EP, siochan->no));
> >>> 
> >>> this does not sound okay, can you explain why this call is here
> >> 
> >> We are sending the configuration to the coprocessor, it will NACK
> >> it if invalid, seems very fitting here.
> > 
> > I dont this so, purpose of the device_config() is to send peripheral
> > config to driver for use on the next descriptor which is submitted. So
> > sending to co-processor now (when we might even have a txn going on)
> > does not seem right
> > 
> > What would be the behaviour if already a txn is progressing on the
> > co-processor
> 
> I have no idea.
> 
> OK, though is that necessarily part of the dmaengine interface? I ask

Yes, the configuration is applied on next descriptor that is prepared

See: Documentation/driver-api/dmaengine/provider.rst
"  - This command should NOT perform synchronously, or on any
    currently queued transfers, but only on subsequent ones"

> because the other driver I have written (apple-admac.c) does basically
> the same, only it applies the new configuration in MMIO registers rather
> than sending it to a coprocessor, but the end result is the same:
> the configuration gets checked for validity, and applied right away.

That should be fixed too :-)

-- 
~Vinod