[PATCH] boot: Warn users about fdt_high=~0 usage

[Marek Vasut]

On 11/17/25 3:32 PM, Tom Rini wrote:

Hello Tom,

>>> Maybe? I was thinking it should be around this in boot/image-fdt.c:
>>>           /* If fdt_high is set use it to select the relocation address */
>>>           fdt_high = env_get("fdt_high");
>>>           if (fdt_high) {
>>>                   ulong high_addr = hextoul(fdt_high, NULL);
>>>
>>>                   if (high_addr == ~0UL) {
>>>                           /* All ones means use fdt in place */
>>
>> This is exactly the code that this patch modifies . Where in the above do
>> you think this change should be ?
> 
> You're right and I did post the last line of context that's the first
> line of your patch. I really meant that we shouldn't even be checking if
> the current location is misaligned. We should tell people to stop
> disabling relocation.

I see, and in that case, I don't particularly care either way, whether 
we warn always, or only in case of actual problem.

>>> I think really. I am loath to break ABI like this but I'm not entirely
>>> sure we have a choice.
>>
>> We do, we simply warn users and remove the usage and fdt_high=~0 assignments
>> from the tree. The functionality itself does not have to be removed.
> 
> The problem is the functionality has always been for a hack workaround
> and bootm_low/bootm_size/etc were the right answer.
> 
> But maybe step one is just remove the in-tree usage and a big loud
> warning when it's set telling people to not do that.

This is fine by me.

>>> Maybe we detect disabled relocation and
>>> misaligned device tree and fall back to prompt? Or if it's too late,
>>> panic with an explanation? Or maybe we just move it 4 bytes higher. The
>>> device is in a going to fail state anyhow, so trying to recover it might
>>> be OK, and since the device tree needs to be modified by us it has to be
>>> in writable memory.
>> Keep in mind, on arm32 it may not necessarily fail to boot with 4-byte
>> aligned DT.
> 
> I'm pretty sure it is. It's not a problem about doing misaligned reads
> it's that the data structure and it's accessors require 8 byte
> alignment.

I think my knowledge of this corner case was from around this time, so 
old kernel versions might work, new ones will likely break. That is fine 
too.