Re: [PATCH v3 6/6] mtd: rawnand: meson: rename node for chip select

[Arseniy Krasnov]

On 11.05.2023 15:11, Miquel Raynal wrote:
> Hi Arseniy,
> 
> avkrasnov@xxxxxxxxxxxxxx wrote on Thu, 11 May 2023 13:16:59 +0300:
> 
>> On 11.05.2023 12:17, Arseniy Krasnov wrote:
>>>
>>>
>>> On 11.05.2023 12:12, Miquel Raynal wrote:  
>>>> Hi Arseniy,
>>>>
>>>> avkrasnov@xxxxxxxxxxxxxx wrote on Thu, 11 May 2023 11:59:07 +0300:
>>>>  
>>>>> On 10.05.2023 23:53, Miquel Raynal wrote:
>>>>>
>>>>> Hello Martin, Miquel
>>>>>  
>>>>>> Hi Martin & Arseniy,
>>>>>>
>>>>>> martin.blumenstingl@xxxxxxxxxxxxxx wrote on Wed, 10 May 2023 22:40:37
>>>>>> +0200:
>>>>>>     
>>>>>>> Hello Arseniy,
>>>>>>>
>>>>>>> On Wed, May 10, 2023 at 1:13 PM Arseniy Krasnov
>>>>>>> <AVKrasnov@xxxxxxxxxxxxxx> wrote:    
>>>>>>>>
>>>>>>>> This renames node with values for chip select from "reg" to "cs". It is
>>>>>>>> needed because when OTP access is enabled on the attached storage, MTD
>>>>>>>> subsystem registers this storage in the NVMEM subsystem. NVMEM in turn
>>>>>>>> tries to use "reg" node in its own manner, supposes that it has another
>>>>>>>> layout. All of this leads to device initialization failure.      
>>>>>>> In general: if we change the device-tree interface (in this case:
>>>>>>> replacing a "reg" with a "cs" property) the dt-bindings have to be
>>>>>>> updated as well.    
>>>>>>
>>>>>> True, and I would add, bindings should not be broken.    
>>>>>
>>>>> I see, that's true. That is bad way to change bindings.
>>>>>  
>>>>>>     
>>>>>>> Documentation/devicetree/bindings/mtd/nand-controller.yaml and
>>>>>>> Documentation/devicetree/bindings/mtd/amlogic,meson-nand.yaml show
>>>>>>> that the chip select of a NAND chip is specified with a "reg"
>>>>>>> property.    
>>>>>>
>>>>>> All NAND controller binding expect the chip-select to be in the
>>>>>> 'reg' property, very much like a spi device would use reg to store the
>>>>>> cs as well: the reg property tells you how you address the device.
>>>>>>
>>>>>> I also fully agree with Martin's comments below. Changing reg is likely
>>>>>> a wrong approach :)
>>>>>>     
>>>>>>> Also the code has to be backwards compatible with old .dtbs.
>>>>>>>    
>>>>>>>> Example:
>>>>>>>>
>>>>>>>> [...] nvmem mtd0-user-otp: nvmem: invalid reg on /soc/bus@ffe00000/...
>>>>>>>> [...] mtd mtd0: Failed to register OTP NVMEM device
>>>>>>>> [...] meson-nand ffe07800.nfc: failed to register MTD device: -22
>>>>>>>> [...] meson-nand ffe07800.nfc: failed to init NAND chips
>>>>>>>> [...] meson-nand: probe of ffe07800.nfc failed with error -22      
>>>>>>> This is odd - can you please share your definition of the &nfc node?    
>>>>>
>>>>> Sure, here it is:
>>>>>
>>>>> mtd_nand: nfc@7800 {                            
>>>>> 	compatible = "amlogic,meson-axg-nfc";
>>>>> 	...
>>>>> 	nand@0 {                                
>>>>>         	reg = <0>;
>>>>>         };
>>>>> }
>>>>>
>>>>> I checked, that 'nand_set_flash_node()' is called with 'nand@0' and i suppose
>>>>> that it is correct (as You mentioned below). But, 'nvmem_add_cells_from_of()' is called
>>>>> with parent: 'nfc@7800', then it iterates over its childs, e.g. 'nand@0' and thus i get such
>>>>> situation. I guess, that 'nvmem_add_cells_from_of()' must be called with 'nand@0' ?  
>>>>
>>>> We recently had issues with nvmem parsing, but I believe a mainline
>>>> kernel should now be perfectly working on this regard. What version of
>>>> the Linux kernel are you using?  
>>>
>>> My current version is:
>>>
>>> VERSION = 6                                                             
>>> PATCHLEVEL = 2                                                          
>>> SUBLEVEL = 0                                                            
>>> EXTRAVERSION = -rc8 
>>>
>>> Fix was in drivers/nvmem/* ?
>>>
>>> Thanks, Arseniy  
>>
>> Upd: I resolved problem in the following way:
>>
>> nand@0 {                                
>> 	reg = <0>;//chip select
>>
> 	partitions {
> 		compatible = ...
> 
>> 	otp@0 {                         
>> 		#address-cells = <2>;   
>> 		#size-cells = <0>;      
> 
> #address/size-cells is not needed here
> 
>> 		compatible = "user-otp";
>> 		reg = <A B>;            
>> 	};                              
>> 	otp@1 {                         
>> 		#address-cells = <2>;   
>> 		#size-cells = <0>;      
> 
> Ditto
> 
>> 		compatible = "factory-otp";
>> 		reg = <C D>;            
>> 	};                              
> 
> 	};
> 
>> };
>>
>> Now nvmem subsystem parses 'otp@0' and 'otp@1' and error is gone. 'compatible' values are
>> the same as in drivers/mtd/mtdcore.c:mtd_otp_nvmem_add(). 'reg' in 'nand@0' is used as
>> chip select as supposed.
> 
> I don't fully get it. The parsing on the nvmem side should not fail if
> there is no subpartition/otp-region defined. Can you confirm an empty
> NAND device node works? Because your last e-mail suggested the opposite.

Ok, so i'll describe what happens in my case. Let's NAND node be like this (IIUC this is
considered as empty NAND device):

mtd_nand: nfc@7800 {                            
	compatible = "amlogic,meson-axg-nfc";
	...
	nand@0 {                                
	       	reg = <0>;
	};
}

I see, that

1) 'mtd_otp_nvmem_add()' calls 'mtd_otp_nvmem_register()' twice for two types of
   OTP memory "user-otp" and "factory-otp". Let's take a look only on "user-otp".
2) 'mtd_otp_nvmem_register()' tries to lookup for node in 'nand@0' which is compatible with
   "user-otp" and then passes found (or not found, e.g. NULL) node to  'nvmem_register()'.
3) 'nvmem_register()' uses this node iterating over its childs and searching value "reg" in
   each child. If "user-otp" node is not found in 2), 'nvmem_register()' uses node 'nfc@7800'
   also looking for "reg" value in each of its child. In this case it found "reg" in 'nand@0'
   and fails.

Now, if i add "compatible = "user-otp";" to 'nand@0', in step 2) search will be successful,
and "reg" value will be used from this new node (or we remove "reg" from it - nothing happens
as You wrote). So, problem is that nvmem tries to parse node with invalid "reg" value.

Also I see, that 'nvmem_register()' is called earlier in 'mtd_nvmem_add()', but with no effect.
I think, that it is not related with enabled OTP feature.

Thanks, Arseniy

> 
>>
>> I think, this patch should be abandoned in the next version.
>>
>> Thanks, Arseniy
>>
>>>   
>>>>
>>>> Thanks,
>>>> Miquèl  
> 
> 
> Thanks,
> Miquèl