On Fri, Dec 15, 2017 at 9:05 AM, Mario Six mario.six@gdsys.cc wrote:

Hi Simon,

On Thu, Dec 14, 2017 at 9:36 PM, Simon Glass sjg@chromium.org wrote:

...

+Stephen, Tom

Hi Mario,

I've had to drop this since it breaks tegra. Stephen feels that this is likely a bug in the patch rather than anything wrong with Tegra. Do you have any thoughts? I can potentially try things out on the Tegra boards I have.

Regards, Simon

I think I've found the problem: It's an endianness issue. The of_translate_address function seems to return a raw big-endian address, while the of_read_number function translates it to CPU endianness before returning it. On the MPC8308 I tested on this made no difference, but little-endian systems will have problems. When I put a bogus dev_read_addr in the sandbox serial driver, I see the effect as well.

Here's a hopefully fixed version (just the diff; I'll post a proper patch when your tests succeeded); it corrects the issue on sandbox, so I hope it will work on tegra as well:

----- >8 -----

diff --git a/drivers/core/ofnode.c b/drivers/core/ofnode.c index 0030ab962e..ca22d09ff2 100644 --- a/drivers/core/ofnode.c +++ b/drivers/core/ofnode.c @@ -200,13 +200,22 @@ fdt_addr_t ofnode_get_addr_index(ofnode node, int index) uint flags; u64 size; int na;

•           __be32 addr;
  
            prop_val = of_get_address(ofnode_to_np(node), index, &size,
                                      &flags);
            if (!prop_val)
                    return FDT_ADDR_T_NONE;
            na = of_n_addr_cells(ofnode_to_np(node));
  

•           return of_read_number(prop_val, na);
  

•           addr = of_read_number(prop_val, na);
  

•           if (IS_ENABLED(CONFIG_OF_TRANSLATE)) {
  

•                   u64 paddr =
  

of_translate_address(ofnode_to_np(node), &addr);

•                   return (fdt_addr_t)(be64_to_cpu(paddr) >> 32);
  

•           } else {
  

•                   return addr;
  

•           }
    } else {
            return fdt_get_base_address(gd->fdt_blob,
                                        ofnode_to_offset(node));
  

----- >8 -----

Thanks for the help, and best regards,

Mario

Small correction:

----- >8 -----

diff --git a/drivers/core/ofnode.c b/drivers/core/ofnode.c index 0030ab962e..09f0aeab4f 100644 --- a/drivers/core/ofnode.c +++ b/drivers/core/ofnode.c @@ -200,13 +200,22 @@ fdt_addr_t ofnode_get_addr_index(ofnode node, int index) uint flags; u64 size; int na; + __be32 addr;

prop_val = of_get_address(ofnode_to_np(node), index, &size, &flags); if (!prop_val) return FDT_ADDR_T_NONE; na = of_n_addr_cells(ofnode_to_np(node)); - return of_read_number(prop_val, na); + addr = of_read_number(prop_val, na); + + if (IS_ENABLED(CONFIG_OF_TRANSLATE)) { + u64 paddr = of_translate_address(ofnode_to_np(node), &addr); + + return be32_to_cpu((fdt_addr_t)paddr); + } else { + return addr; + } } else { return fdt_get_base_address(gd->fdt_blob, ofnode_to_offset(node));

----- >8 -----

Regards,

Mario

...

On 8 December 2017 at 10:11, sjg@google.com wrote:

...

On 23 November 2017 at 23:51, Mario Six mario.six@gdsys.cc wrote:

...

The function dev_read_addr calls ofnode_get_addr_index in the live tree case, which does not apply bus translations to the address read from the device tree. This results in illegal addresses on boards that rely on bus translations being applied.

Fix this situation by applying bus translations in the live tree case as well.

Signed-off-by: Mario Six mario.six@gdsys.cc

Changes v1 -> v2:

• Added IS_ENABLED(CONFIG_OF_TRANSLATE) case distinction

---

drivers/core/ofnode.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)

Reviewed-by: Simon Glass sjg@chromium.org

Applied to u-boot-dm thanks!