On Mon, May 18, 2020 at 02:01:50PM +0300, Andy Shevchenko wrote: > On Sat, May 16, 2020 at 05:33:53PM +0300, Serge Semin wrote: > > On Fri, May 15, 2020 at 05:01:29PM +0300, Andy Shevchenko wrote: > > > On Fri, May 15, 2020 at 01:47:48PM +0300, Serge Semin wrote: > > > > It isn't good to have numeric literals in the code especially if there > > > > are multiple of them and they are related. Moreover in current > > > > implementation the Tx DMA transfer activation level isn't optimal, > > > > since it's hardwired to be at 16-32 bytes level, while it's better > > > > to keep the SPI FIFO buffer as full as possible until all available > > > > data is submitted. So lets introduce the DMA burst level > > > > parametrization macros with optimal values - issue Rx transfer if at > > > > least 16 bytes are available in the buffer and execute Tx transaction > > > > if at least 16 bytes room is opened in SPI Tx FIFO. > > > > > > > - dw_writel(dws, DW_SPI_DMARDLR, 0xf); > > > > - dw_writel(dws, DW_SPI_DMATDLR, 0x10); > > > > + dw_writel(dws, DW_SPI_DMARDLR, RX_BURST_LEVEL - 1); > > > > + dw_writel(dws, DW_SPI_DMATDLR, dws->fifo_len - TX_BURST_LEVEL); > > > > > > ...and if FIFO length is less than TX_BURST_LEVEL? > > > > > > For the patch that introduces definitions, i.e. keeping the last line here as > > > > > > dw_writel(dws, DW_SPI_DMATDLR, TX_BURST_LEVEL); > > > > > > I'm good. You may put your tag in that case. For fifo_len case we need to > > > discuss in separate patch, perhaps. > > > > It's fixed in a consequent patch anyway. Though if v3 is required I'll remove > > this change from here. > > I consider that here you might have introduced a regression and actually doing > two things in one patch. Why not to split? Theoretically I could, but only for a hardware with FIFO smaller than 16 bytes. So did I, seeing this module has been dedicated for the Intel Medfield/Elkhart chips only? Anyway as I said this change is mostly redundant, since further in this patchset I'll replace the constants used here with burst length properly calculated based on the fifo-length and max-burst-length specific to the DMA. -Sergey > > -- > With Best Regards, > Andy Shevchenko > >
