On Fri, 28 Feb 2020 15:06:58 +0530
Pratyush Yadav [off-list ref] wrote:

Hi Boris,

On 27/02/20 05:58PM, Boris Brezillon wrote:

quoted

On Wed, 26 Feb 2020 15:06:58 +0530
Pratyush Yadav [off-list ref] wrote:
  

quoted

Double Transfer Rate (DTR) is SPI protocol in which data is transferred
on each clock edge as opposed to on each clock cycle. Make
framework-level changes to allow supporting flashes in DTR mode.

Right now, mixed DTR modes are not supported. So, for example a mode
like 4S-4D-4D will not work. All phases need to be either DTR or STR.  

Didn't go deep into the patch but at first glance you don't seem to
extend the framework to support stateful modes as I tried to do here
[1]. That's really something we should address before considering
supporting xD-xD-xD modes, unless the SPI-NOR only supports one
stateful mode. If we don't do that first, we might face all sort of
unpleasant issues:

* kexec not working correctly because the previous kernel left the NOR
  in an unknown state
* suspend/resume not working properly
* linux not booting properly because the bootloader left the device in
  its non-default mode
* ...  

Correct. I am working on a follow-up series that takes care of these 
problems. The series will allow spi-nor to detect what mode the flash is 
in and then run the SFPD procedure in that mode (or maybe switch to 
single SPI mode and then go about its business as usual? I haven't 
figured out all the details yet).

So for the context of this series, assume we are handed the flash in 
single SPI mode.
 

quoted

[1]https://patchwork.kernel.org/cover/10638055/  

BTW, I took a quick look at this series but I don't see any code that 
tries to detect which mode the flash is in (which is the troublesome 
part [0]). So, for example, if the bootloader leaves the flash in 
8D-8D-8D mode, how would your series handle that situation?

Oh, it's definitely not taking care of that, it was just paving the
road for spi-nor state tracking. You'd need to extend it to support
8D-8D-8D to 1-1-1 transitions at boot time (if that's even possible).

[0] There are multiple problems to take care of when trying to detect 
    which mode a flash is in. We can try reading SFDP in each mode and 
    whichever mode gives us the correct "SFDP" signature is the mode the 
    flash is in. But the problem is that even in xSPI standard Read SFDP 
    command is optional in 8D-8D-8D mode, let alone non-xSPI flashes.
    Another problem is that the address bytes and dummy cycles for Read 
    SFDP are not the same for every flash. The xSPI standard says 
    address bytes can be 3/4 and dummy cycles can be 8/20. So, for 
    example, Cypress s28hs/s28ht family and Micron Xccela (mt35x) family 
    use 4 address bytes, but the Adesto ATXP032/ATXP032R flashes use 3 
    address bytes.

I'd rather go with something simpler and more widely supported than SFDP
reads. Don't we have a simple command that's supported by all flashes
and returns well known data. Isn't there an EXIT sequence that allows
NORs to return to a single SPI state?

    Say that a flash supports Read SFDP in 8D-8D-8D mode and we try all 
    the combinations to find out which mode the flash is in, we now have 
    the problem of actually identifying the flash. Unfortunately, the 
    Read ID command is not uniform across flash vendors. The Micron 
    Xccela flashes use 8 dummy cycles and no address bytes for Read ID. 
    The Cypress s28hs/t family uses configurable dummy cycles 
    (defaulting to 3) and needs 4 dummy address bytes all of which are 
    0.

Yep, that's what I complained about when I tried to support the
Macronix flash. They didn't plan for a reliable RETURN-TO-SINGLE-SPI
sequence which would not conflict with any other existing SPI commands,
and that's a real problem.

    If we can't find out which flash it is, we can't run its fixup 
    hooks, and might end up running it with incorrect settings. And all 
    this is assuming a flash even has SFDP and has it available in all 
    modes.

Absolutely.

    So, the only solution I can now think of is having the flash name in 
    its compatible string in the device tree. This way we can skip all 
    the Read ID ugliness and can have flash-specific hooks to make it 
    easier to detect the mode it is in (though I wonder if it is even 
    possible to detect the mode in a flash that doesn't have SFDP in 
    8D-8D-8D).

Hm, I'd really like to avoid that if possible.

    Thoughts? Is there a better way to solve this problem that I didn't 
    think of?

Nope, except maybe mandate that the bootloader always put the NOR in
single SPI mode before booting Linux (and Linux should do the same,
which is what my series was trying to address IIRC).