[PATCH v3] mtd: limit OTP NVMEM Cell parse to non Nand devices

[Rafał Miłecki]

On 27.03.2024 15:36, Christian Marangi wrote:
> On Wed, Mar 27, 2024 at 03:26:55PM +0100, Rafał Miłecki wrote:
>> On 2024-03-22 05:09, Christian Marangi wrote:
>>> MTD OTP logic is very fragile and can be problematic with some specific
>>> kind of devices.
>>>
>>> NVMEM across the years had various iteration on how Cells could be
>>> declared in DT and MTD OTP probably was left behind and
>>> add_legacy_fixed_of_cells was enabled without thinking of the
>>> consequences.
>>
>> Er... thank you?
>>
> 
> Probably made some bad assumption and sorry for it!

No problem :)


>>> That option enables NVMEM to scan the provided of_node and treat each
>>> child as a NVMEM Cell, this was to support legacy NVMEM implementation
>>> and don't cause regression.
>>>
>>> This is problematic if we have devices like Nand where the OTP is
>>> triggered by setting a special mode in the flash. In this context real
>>> partitions declared in the Nand node are registered as OTP Cells and
>>> this cause probe fail with -EINVAL error.
>>>
>>> This was never notice due to the fact that till now, no Nand supported
>>> the OTP feature. With commit e87161321a40 ("mtd: rawnand: macronix: OTP
>>> access for MX30LFxG18AC") this changed and coincidentally this Nand is
>>> used on an FritzBox 7530 supported on OpenWrt.
>>
>> So as you noticed this problem was *exposed* by adding OTP support for
>> Macronix NAND chips.
>>
>>
>>> Alternative and more robust way to declare OTP Cells are already
>>> prossible by using the fixed-layout node or by declaring a child node
>>> with the compatible set to "otp-user" or "otp-factory".
>>>
>>> To fix this and limit any regression with other MTD that makes use of
>>> declaring OTP as direct child of the dev node, disable
>>> add_legacy_fixed_of_cells if we detect the MTD type is Nand.
>>>
>>> With the following logic, the OTP NVMEM entry is correctly created with
>>> no Cells and the MTD Nand is correctly probed and partitions are
>>> correctly exposed.
>>>
>>> Fixes: 2cc3b37f5b6d ("nvmem: add explicit config option to read old
>>> syntax fixed OF cells")
>>
>> It's not that commit however that introduced the problem. Introducing
>> "add_legacy_fixed_of_cells" just added a clean way of enabling parsing
>> of old cells syntax. Even before my commit NVMEM subsystem was looking
>> for NVMEM cells in NAND devices.
>>
>> I booted kernel 6.6 which has commit e87161321a40 ("mtd: rawnand:
>> macronix: OTP > access for MX30LFxG18AC") but does NOT have commit
>> 2cc3b37f5b6d ("nvmem: add explicit config option to read old syntax
>> fixed OF cells").
>>
>> Look at this log from Broadcom Northstar (Linux 6.6):
>> [    0.410107] nand: device found, Manufacturer ID: 0xc2, Chip ID: 0xdc
>> [    0.416531] nand: Macronix MX30LF4G18AC
>> [    0.420409] nand: 512 MiB, SLC, erase size: 128 KiB, page size: 2048, OOB
>> size: 64
>> [    0.428022] iproc_nand 18028000.nand-controller: detected 512MiB total,
>> 128KiB blocks, 2KiB pages, 16B OOB, 8-bit, BCH-8
>> [    0.438991] Scanning device for bad blocks
>> [    0.873598] Bad eraseblock 738 at 0x000005c40000
>> [    1.030279] random: crng init done
>> [    1.854895] Bad eraseblock 2414 at 0x000012dc0000
>> [    2.657354] Bad eraseblock 3783 at 0x00001d8e0000
>> [    2.662967] Bad eraseblock 3785 at 0x00001d920000
>> [    2.848418] nvmem user-otp1: nvmem: invalid reg on
>> /nand-controller at 18028000/nand at 0
>> [    2.856126] iproc_nand 18028000.nand-controller: error -EINVAL: Failed to
>> register OTP NVMEM device
>>
>> So to summary it up:
>> 1. Problem exists since much earlier and wasn't introduced by 2cc3b37f5b6d
>> 2. Commit 2cc3b37f5b6d just gives you a clean way of solving this issue
>> 3. Problem was exposed by commit e87161321a40
>> 4. We miss fix for v6.6 which doesn't have 2cc3b37f5b6d (it hit v6.7)
>>
> 
> So the thing was broken all along? Maybe the regression was introduced
> when OF support for NVMEM cell was introduced? (and OF scan was enabled
> by default?)

I went back to the commit 4b361cfa8624 ("mtd: core: add OTP nvmem
provider support") from 2021 (it went into 5.14). It seems that even
back then nvmem_register() used to call nvmem_add_cells_from_of(). SO
maybe this problem is as old as that?


> Anyway Sorry for adding the wrong fixes, maybe Miquel can remote the
> commit from mtd/fixes and fix the problematic fixes tag?
> 
>>
>>> Cc: <stable at vger.kernel.org>
>>> Signed-off-by: Christian Marangi <ansuelsmth at gmail.com>
>