From mboxrd@z Thu Jan 1 00:00:00 1970 From: Arnd Bergmann Date: Thu, 14 Jul 2016 14:33:11 +0200 Subject: [ath9k-devel] [PATCH v4 1/3] Documentation: dt: net: add ath9k wireless device binding In-Reply-To: References: <20160624123430.4097-1-martin.blumenstingl@googlemail.com> <1877260.9mAQ7eY9me@wuerfel> Message-ID: <4660464.54jApCBV7y@wuerfel> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: ath9k-devel@lists.ath9k.org On Monday, July 11, 2016 11:21:26 PM CEST Martin Blumenstingl wrote: > On Mon, Jul 11, 2016 at 12:01 AM, Arnd Bergmann wrote: > >> ath9k reads the data from the EEPROM into memory. With that property > >> disabled ath9k simply assumes that the endianness of the values in the > >> EEPROM are having the correct endianness for the host system (in other > >> words: no swap is being applied). > >> I am not sure I understand you correctly, but isn't what you are > >> explaining an issue in the ath9k code, rather than in this > >> documentation? > > > > I looked at the code more to find that out now, but I'm more confused > > now, as the eeprom seems to be read as a byte stream, and the endianess > > conversion that the driver performs is not on the data values in it, > > but seems to instead swap the bytes in each 16-bit word, regardless > > of the contents (the values inside of the byte stream are always > > interpreted as big-endian). Is that a correct observation? > that seems to be the case for the ar9003 eeprom. Other implementations > are doing it different, look at ath9k_hw_ar9287_check_eeprom for > example: first ath9k_hw_nvram_swap_data checks the two magic bytes at > the beginning of the data and swaps the bytes in each 16-bit word if > the magic bytes don't match the magic bytes for the "native system > endianness" (see AR5416_EEPROM_MAGIC). Then more swapping is applied. > I asked for more details about the EEPROM format (specifically the > endianness part) here [0] as I don't have access to the datasheets > (all I have is the ath9k code) Ok. > > What I see in ath_pci_eeprom_read() is that the 16-bit words are taken > > from the lower 16 bit of the little-endian AR_EEPROM_STATUS_DATA > > register. As this is coming from a PCI register, it must have a device > > specific endianess that is identical on all CPUs, so in the description > > above, mentioning CPU endianess is a bit confusing. I could not find > > the code that does the conditional byteswap, instead this function > > > > static bool ar9300_eeprom_read_byte(struct ath_hw *ah, int address, > > u8 *buffer) > > { > > u16 val; > > > > if (unlikely(!ath9k_hw_nvram_read(ah, address / 2, &val))) > > return false; > > > > *buffer = (val >> (8 * (address % 2))) & 0xff; > > return true; > > } > > > > evidently assumes that the lower 8 bit of the 16-bit data from PCI > > come first, i.e. it byteswaps on big-endian CPUs to get the bytestream > > back into the order in which it is stored in the EEPROM. > Please have a look at the ath9k_hw_nvram_swap_data function and > eeprom_ops.check_eeprom (for example ath9k_hw_ar9287_check_eeprom). > These are performing the conditional swapping (in addition to whatever > previous swapping there was). > The basic code works like this: read the full EEPROM data into memory > (either from PCI bus, ath9k_platform_data or request_firmware), then > eeprom_ops.check_eeprom will call ath9k_hw_nvram_swap_data for 16-bit > word swapping and afterwards the check_eeprom implementation will doe > further swapping. > Apart from that your findings seem correct (at least this is identical > to how I would interpret the code). Ok, so my interpretation of what this is done for is that the swap in ath9k_hw_nvram_swap_data() is done to compensate for the data that is read byte-reversed from the PCI bus and it is does not swap when the data is read from a file. The result is a structure with big-endian 16-bit and 32-bit members but all fields in the right place. The swapping in ath9k_hw_ar9287_check_eeprom() then turns the big-endian fields into little-endian fields so it can be used on little-endian CPUs without going through le16_to_cpu(). However, the whole thing still looks fragile to me as it doesn't seem to handle the case where we want to swap the values but not the bus. My guess is that we still want to fix the driver to handle this more consistently in order to decide whether a DT property is needed or not. > > Interestingly, this also seems to happen for ath_ahb_eeprom_read() > > even though on that one the value does not get swapped by the bus > > accessor, so presumably big-endian machines with a ahb based ath9k > > store their eeprom byte-reversed? > on AHB the eeprom data has to be provided via ath9k_platform_data / > request_firmware mechanism. Thus there is no bus specific swapping, > only the ath9k_hw_nvram_swap_data / eeprom_ops.check_eeprom swapping > is applied in this case. I guess the header then indicates that none of the swapping is performed. Arnd