On Monday, August 22, 2016 1:56:46 PM CEST Martin Blumenstingl wrote:

On Mon, Aug 22, 2016 at 1:47 PM, Arnd Bergmann [off-list ref] wrote:

quoted

On Sunday, August 21, 2016 4:49:03 PM CEST Martin Blumenstingl wrote:

quoted

We will default to the system's native endianness for the eepmisc value.
This may be overwritten by the actual calibration data. If it is not
overwritten we interpret the template data in it's native endianness,
meaning that no swapping is required.

I'm still skeptical about this one. What is the significance of "native
endianess" here? You are keying the endianess of the eeprom tables off the
way the CPU operates, but for a PCI device there is no correlation between
those two.

(the ar9003 eeprom format and handling is different compared to 9287,
def and 4k)
ar9003_eeprom.c contains EEPROM templates -> these are compiled into
the ath9k kernel module. Values from these templates can be
overwritten by the EEPROM found on the actual hardware.
This change tries to handle the case where the values in the hardware
EEPROM do not override any of the template values (means final EEPROM
data = template data). In this case the we can simply rely on the
endianness which was used to compile ath9k.ko.

Ok, I see what you mean now. However, looking at the source now, I
also see

#define LE16(x) cpu_to_le16(x)
#define LE32(x) cpu_to_le32(x)
        .baseEepHeader = {
                .regDmn = { LE16(0), LE16(0x1f) },

suggesting that the fields are meant to be little-endian in object
code, and your patch does not change that. In fact, Felix's
ffdc4cbe5b17 ("ath9k_hw: clean up EEPROM endian handling on AR9003")
seems to have corrected this already.

	Arnd