On Wed, Jul 4, 2018 at 8:16 PM, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote: > On Wed, Jul 4, 2018 at 8:13 PM, Logan Gunthorpe <logang@xxxxxxxxxxxx> wrote: >> On 7/4/2018 11:10 AM, Andy Shevchenko wrote: >>> We have an iDMA 32-bit hardware (see drivers/dma/dw/) which has an >>> extension 64-bit registers where only one of them has a specific bit >>> to "commit" the changes written to all of them. And by some very >>> unknown reason that bit is in lo part which automatically means we >>> must to write it last. >> >> And it supports both BE and LE? And in both cases it's the lo part? > > It's only LE for now. > > P.S. If you more interested in code in kernel look for idma32_fifo_partition() > (While the bit is set in each of 32-bit part, it's actually present in > only one place) I think it doesn't contradict with what you are saying rather supports it. I would expect to have lo-hi and hi-lo semantics done according to the data flow, not to the address. -- With Best Regards, Andy Shevchenko
