On Mon, Nov 09, 2020 at 10:19:09PM +0300, Serge Semin wrote: > On Mon, Nov 09, 2020 at 02:14:22PM +0000, Mark Brown wrote: > > I'd expect it'd perform better, especially on systems that are > > apparently struggling for CPU bandwidth like yours seems to. > CPU-wise. RO-mode won't help in that case. Moreover it will be even > more errors-prone for the systems with small CPU bandwidth. As I said Right, these are two separate issues - one is that the client device has requirements on the transmit data at times when the driver isn't defining what should be transmitted, the other is that the controller driver is using full duplex mode even for single duplex data. I just happened to notice the second issue while reviewing the change - there shouldn't be any code for setting the dummy transmit pattern in the driver in the first place. > 2) Rx-only with atomic CPU utilization. In order to make sure that the > CPU keeps up with fetching the data from the Rx FIFO, we have to > disable the local CPU IRQs while performing the Rx-only transfers, so > to prevent the Rx FIFO overflow while the CPU is doing something else. ... > So in all other cases for normal CPU-based SPI-transfers when > GPIO-based chip-select is available the safest solution would be to > use a normal Push-Pull mode. In this case we have no risk in getting > the Rx FIFO overflow unless there is a bug in the code, which is > fixable anyway. I'm not clear why we would have issues with the FIFO overflowing in PIO mode, especially with a GPIO chip select - even if we're forced to tell the controller how big the transfer is if we're using a GPIO chip select we could just tell it we're doing a series of FIFO sized transfers? > Another possible solution for the problem would be to fix the SPI core > so aside with tx_buf being set to the NULL-pointer, a client driver > would provide a default level or some specific value being put to the > SPI bus on Rx-only transfers. If an SPI-controller is capable of > satisfying the request, then it will accept the transfer. If it's not, > then the SPI core may try to convert the Rx-only transfer into the > Full-duplex transfer with the Tx-buffer being initialized with the > requested level. We do have support in the core for creating dummy data buffers for controllers that can't do half duplex - that's the SPI_MUST_TX and matching SPI_MUST_RX that I mentioned in my initial reply. Currently we always zero fill transmit buffers, the expected semantic is that if the client driver isn't supplying data that means the device doesn't care what gets sent and it's not clear to me that it isn't sensible to just keep that like I said earlier, I don't know how common it's going to be to need this since most half duplex transfers generally are half duplex. The whole point with the SPI_MUST_ flags was to remove the need for controller drivers to open code handling this, it was adding complication and supporting configuration of the dummy data feels like it's adding room for things to go wrong.
Attachment:
signature.asc
Description: PGP signature
