On 23.04.2021 11:17, Lars-Peter Clausen wrote: > The device_config() and device_prep_slave_sg() APIs are meant for > memory-to-device or device-to-memory transfers. This means the device address > is fixed and usually maps to a FIFO in the device. > > What you are implementing is memory-to-memory, which means addresses are > incremented on both sides of transfer. And this works if you know what you are > doing, e.g. you know that you are using a specific dmaengine driver that > implements the dmaengine API in a way that does not comply with the > specification. But it breaks for generic dmaengine API clients that expect > compliant behavior. And it is the reason why we have an API in the first > place, to get consistent behavior. If you have to know about the driver > specific semantics you might as well just bypass the framework and use driver > specific functions. You're absolutely right about this. I think I might've been confused when picking the right DMA Engine API callback to write this new functionality into by the fact that, on our particular configuration, transfers were being done between main system memory and DDR memory on a PCI card. However, from the point of view of the semantics of the API, these are still memory-to-memory transfers, just like you said. > It seems to me what we are missing from the DMAengine API is the equivalent of > device_prep_dma_memcpy() that is able to take SG lists. There is already a > memset_sg, it should be possible to add something similar for memcpy. I wasn't aware of the existence of this callback. Grepping the sources reveals a single consumer at the moment, I guess this also might've led me to wrongly conclude device_prep_slave_sg was the right choice. Perhaps I can convert most of the code currently in xilinx_cdma_prep_slave_sg to fit the semantics of this DMAengine API function instead. Any thoughts on this? Cheers, Adrian
