[PATCH 0/4] spi-mem: Allow specifying the byte order in DTR mode

[Pratyush Yadav]

On 24/02/22 10:37AM, Michael Walle wrote:
> Am 2022-02-24 07:37, schrieb Tudor.Ambarus at microchip.com:
> > On 2/24/22 08:08, Tudor.Ambarus at microchip.com wrote:
> > > EXTERNAL EMAIL: Do not click links or open attachments unless you
> > > know the content is safe
> > > 
> > > On 2/23/22 20:38, Pratyush Yadav wrote:
> > > > EXTERNAL EMAIL: Do not click links or open attachments unless
> > > > you know the content is safe
> > > > 
> > > > Hi Tudor,
> > > > 
> > > > On 22/02/22 02:43PM, Tudor.Ambarus at microchip.com wrote:
> > > > > On 2/22/22 16:27, Michael Walle wrote:
> > > > > > EXTERNAL EMAIL: Do not click links or open attachments
> > > > > > unless you know the content is safe
> > > > > > 
> > > > > > Am 2022-02-22 15:23, schrieb Tudor.Ambarus at microchip.com:
> > > > > > > On 2/22/22 16:13, Michael Walle wrote:
> > > > > > > > EXTERNAL EMAIL: Do not click links or open
> > > > > > > > attachments unless you know
> > > > > > > > the content is safe
> > > > > > > > 
> > > > > > > > Am 2022-02-22 14:54, schrieb Tudor.Ambarus at microchip.com:
> > > > > > > > > On 2/21/22 09:44, Michael Walle wrote:
> > > > > > > > > > EXTERNAL EMAIL: Do not click links or
> > > > > > > > > > open attachments unless you
> > > > > > > > > > know
> > > > > > > > > > the content is safe
> > > > > > > > > > 
> > > > > > > > > > Am 2022-02-18 15:58, schrieb Tudor Ambarus:
> > > > > > > > > > > Fortunately there are controllers
> > > > > > > > > > > that can swap back the bytes at
> > > > > > > > > > > runtime, fixing the endiannesses.
> > > > > > > > > > > Provide
> > > > > > > > > > > a way for the upper layers to
> > > > > > > > > > > specify the byte order in DTR mode.
> > > > > > > > > > 
> > > > > > > > > > Are there any patches for the
> > > > > > > > > > atmel-quadspi yet? What happens if
> > > > > > > > > 
> > > > > > > > > not public, but will publish them these days.
> > > > > > > > > 
> > > > > > > > > > the controller doesn't support it? Will there be a software
> > > > > > > > > > fallback?
> > > > > > > > > 
> > > > > > > > > no need for a fallback, the controller can ignore
> > > > > > > > > op->data.dtr_bswap16
> > > > > > > > > if
> > > > > > > > > it can't swap bytes.
> > > > > > > > 
> > > > > > > > I don't understand. If the controller doesn't
> > > > > > > > swap the 16bit values,
> > > > > > > > you will read the wrong content, no?
> > > > > > > > 
> > > > > > > 
> > > > > > > In linux no, because macronix swaps bytes on a 2
> > > > > > > byte boundary both on
> > > > > > > reads and on page program. The problem is when you
> > > > > > > mix 8D-8D-8D mode
> > > > > > > and
> > > > > > > 1-1-1 mode along the boot stages. Let's assume you write all boot
> > > > > > > binaries
> > > > > > > in 1-1-1 mode. When reaching u-boot if you enable
> > > > > > > 8D-8D-8D mode, when
> > > > > > > u-boot
> > > > > > > will try to get the kernel it will fail, as the
> > > > > > > flash swaps the bytes
> > > > > > > compared
> > > > > > > to what was written with 1-1-1 mode. You write D0 D1
> > > > > > > D2 D3 in 1-1-1
> > > > > > > mode and
> > > > > > > when reaching u-boot you will read D1 D0 D3 D2 and
> > > > > > > it will mess the
> > > > > > > kernel image.
> > > > > > 
> > > > > > But you have to consider also 3rd parties, like an
> > > > > > external programmer
> > > > > > or
> > > > > 
> > > > > Why? If you use the same mode when reading and writing,
> > > > > everything is fine.
> > > > > I'm not sure what's your suggestion here.
> > > > 
> > > > So our stance here is that we don't care about external programs?>
> > > > If that is the case then why bother with all this anyway? Since
> > > > the swap
> > > > happens at both page program and read, what you write is what
> > > > you read
> > > > back. Who cares the order stored in the actual flash memory as
> > > > long as
> > > > the data read is correct?
> > > > 
> > > > If we do care about external programs, then what would happen if the
> > > > external program writes data in 8D-8D-8D mode _without_ swapping the
> > > > bytes? This would also cause data corruption. You can't control what
> > > > they mode they use, and you can't detect it later either.
> > > > 
> > > > I think there is no winning here. You just have to say that external
> > > > programs should write in 8D-8D-8D mode or it won't boot.
> 
> IMHO it should just work that you can use 1S-1S-1S mode and 8D-8D-8D on the
> same flash. After all, that is Tudor's use case. The ROM access the flash
> in single bit mode and linux in 8D-8D-8D mode. Maybe u-boot will use quad

But you don't know that ROM will always access the flash in single bit 
mode. For example, ROM on some TI SoC can read SFDP and use 8D-8D-8D 
mode for reading images from flash. If you want to flash data from 
Linux, and it byte swaps, ROM won't be able to read the images properly.

This can only work when everything that reads/writes in 8D mode does 
byte swapping. Otherwise it will lead to a mess where data is read 
correctly by some software but not by some other software. I don't know 
how practical it is to make this assumption.

> mode in between. All of these accesses should return the same flash
> content.
> 
> > > How about swapping the bytes just at user request? Maybe with a
> > > Kconfig
> > > option.
> > 
> > Michael has suggested on #irc to always swap the bytes: if the SPI
> > controller
> > can't do it, to do it in software at SPI NOR level. I don't know what to
> > say
> > about this, because JEDEC216 just informs the reader I guess:
> > "Byte order of 16-bit words is swapped when read in 8D-8D-8D mode
> > compared to
> > 1-1-1 mode.", this doesn't look like a hard request. The downside to
> > doing
> > the swapping in software is performance penalty which will make macronix
> > users have second thoughts. I don't have a strong opinion, but I lean
> > towards
> > doing the swap just at user request, regardless if I do it via the SPI
> > controller
> > or in software.
> 
> Just having and opt-in will be a mess in the future with flashes containing
> byte swapped content and we can't even fix it and we will have to live with
> that forever. IMHO right now is the best time to circumvent that scenario.
> I don't have anything against make it user configurable, but it should be
> an opt-out.
> 
> I haven't looked at any controllers who can do 8D-8D-8D accesses, maybe most
> of them can do the swapping on their own? So if you don't want to support a

I checked the datasheet of the Cadence Quadspi (spi-cadence-quadspi.c) 
controller. I don't see any such option.

> software fallback, then we should just say this mode isn't supported if
> the controller can't do the byte swapping and we fall back to a slower mode.

>From all I understand of this, it looks to me that this can't really be 
solved completely. If you want to allow compatibility with 1S-1S-1S mode 
then you lose compatibility with 8D-8D-8D software that doesn't do this 
swap. So the question really is which one we consider "more important". 
In my eyes the choice is arbitrary.

But I am not convinced that adding a Kconfig option is the right thing 
to do. I think that would cause too much confusion. It is entirely 
possible that your data gets corrupted going from one kernel version to 
another depending on how it was compiled. Us SPI NOR developers know 
this tiny detail but other people won't, and it would be hard to explain 
this to them.

-- 
Regards,
Pratyush Yadav
Texas Instruments Inc.