On Wed, 19 Apr 2006 11:28:49 +0200 David Jander <david.jander at protonic.nl> wrote:
> > Hi Kenneth, > > On Tuesday 18 April 2006 21:19, Kenneth Poole wrote: > > Regarding DMA allocation for CPM uarts, we had a similar issue with our > > MPC857T and MPC885 boards. I think the real problem is that > > bus_to_virt() and virt_to_bus() don't work on 8xx platforms for > > addresses allocated using dma_alloc_coherent(). We had a previous > > discussion Pantelis Antoniou and he agreed that the use of bus_to_virt() > > and virt_to_bus() should be deprecated. This is also recommended in the > > whitepaper series that discussed porting device drivers from 2.4 to 2.6. > > You are right! I hadn't noticed the (implicit) use of virt_to_bus() in > cpm_uart_core.c. In the case that "((unsigned long)addr >= CPM_ADDR)", the > address isn't translated, that's why the uart keeps sending lots of 0xff's > instead of real data: There is empty flash at that adress (after 0xff100000)! > This also explains why the system hangs when I set CONFIG_CONSISTENT_START to > 0xfd100000, because then the CPM will access unadressed space. > > Vitaly: Is it true that your ADS boards both have > IMMAP_ADDR < CONFIG_CONSISTENT_START ? > That would explain why those do work. > On the PQ2's dma_alloc_coherent(), behaves differently, so they might also > work. > Right. > In fact, bus_to_virt(), which is also used, does nothing more than adding > KERNELBASE to the address, so a!=bus_to_virt(virt_to_bus(a)) if 'a' is > obtained by anything other than kmalloc(). So if 'a' is obtained from > dma_alloc_coherent(), will bus_to_virt(virt_to_bus(a)) at least yield a > correctly mapped virtual address? > I am asking, because line 263 of cpm_uart_core.c does exactly this, and it > smells ;-) > Well I was aware about no-good there, but since it works fine with my HW, and no other complaints, there were no reason to overhaul. > > What we did was use dma_alloc_coherent() in cpm_uart_cpm1.c, saving > > dma_addr in the pinfo structure. The in cpm_uart_core.c, we use dma_addr > > as the base address for the cbd_bufaddr values in each of the > > descriptors. The software, when accessing the DMA buffers, uses mem_addr > > as the base, applying the same offset computed previously for the dma > > addresses. This technique is used in other drivers all over 2.6. > > Sounds like the correct thing to do, but do you (by any chance) have a patch > for this change, to share? > If there is any code to reference, I'd appreciate it (and merge to the driver). -- Sincerely, Vitaly
