On Wed, Nov 18, 2015 at 5:45 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote: > On Wednesday 18 November 2015 17:29:19 Andy Shevchenko wrote: >> >> I understand most of the things here, what I don't is how a platform >> is supposed to work if you have the following: >> a) HW, that uses register space let's say higher than 32-bit; >> b) DMA engine, which should provide a DMA capability for above HW block; >> c) dma_addr_t which does not cover the HW register space. > > On this platform, the current code is obviously broken, because the pointer > is 32-bit wide and cannot reach the registers. I assume you agree on that > part. Yes. > With my patch, the 64-bit resource_size_t in dw_mci helps get the > correct FIFO address to this line: > > cfg.dst_addr = host->phy_regs + fifo_offset; > > There, it remains broken because of the dma_addr_t being too short, and > we also need Linus' patch from https://lkml.org/lkml/2013/4/26/120 in > addition to mine. Wow, never applied. > > >> For me it clearly looks like a platform (HW / SW) configuration issue. > > I think some people have argued in the past that we should always use > the same type for dma_addr_t, resource_size_t and phys_addr_t. That > would certainly fix the problem you describe as well. In practice, > everyone has that already, and my patch by itself fixes all the > cases where the FIFO is at a high address and dma_addr_t is already > 64-bit wide. Let me summarize. We have to have classification by address space 1) physical 2) virtual Therefore resource_addr_t must be equal to phys_addr_t since it may carry any possible physical address. dma_addr_t is a physical address wrt DMA mask. Correct? -- With Best Regards, Andy Shevchenko -- To unsubscribe from this list: send the line "unsubscribe linux-mmc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
