On 23/10/2024 16:17, Rob Herring wrote:
> On Wed, Oct 23, 2024 at 9:43 AM Usama Arif <usamaarif642@xxxxxxxxx> wrote:
>>
>>
>>
>> On 23/10/2024 14:40, Rob Herring wrote:
>>> On Mon, Oct 7, 2024 at 10:30 AM Usama Arif <usamaarif642@xxxxxxxxx> wrote:
>>>>
>>>>
>>>>
>>>> On 07/10/2024 15:39, Rob Herring wrote:
>>>>> On Mon, Oct 7, 2024 at 9:06 AM Usama Arif <usamaarif642@xxxxxxxxx> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 04/10/2024 01:03, Rob Herring wrote:
>>>>>>> On Thu, Oct 03, 2024 at 12:38:40PM +0100, Usama Arif wrote:
>>>>>>>> __pa() is only intended to be used for linear map addresses and using
>>>>>>>> it for initial_boot_params which is in fixmap for arm64 will give an
>>>>>>>> incorrect value. Hence stash the physical address when it is known at
>>>>>>>> boot time and use it at kexec time instead of converting the virtual
>>>>>>>> address using __pa().
>>>>>>>>
>>>>>>>> Reported-by: Breno Leitao <leitao@xxxxxxxxxx>
>>>>>>>> Suggested-by: Mark Rutland <mark.rutland@xxxxxxx>
>>>>>>>> Signed-off-by: Usama Arif <usamaarif642@xxxxxxxxx>
>>>>>>>> Fixes: ac10be5cdbfa ("arm64: Use common of_kexec_alloc_and_setup_fdt()")
>>>>>>>> ---
>>>>>>>> arch/arm64/kernel/setup.c | 8 ++++++++
>>>>>>>> drivers/of/fdt.c | 6 ++++++
>>>>>>>> drivers/of/kexec.c | 8 ++++++--
>>>>>>>> include/linux/of_fdt.h | 2 ++
>>>>>>>> 4 files changed, 22 insertions(+), 2 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c
>>>>>>>> index b22d28ec8028..a4d96f5e2e05 100644
>>>>>>>> --- a/arch/arm64/kernel/setup.c
>>>>>>>> +++ b/arch/arm64/kernel/setup.c
>>>>>>>> @@ -194,6 +194,14 @@ static void __init setup_machine_fdt(phys_addr_t dt_phys)
>>>>>>>> /* Early fixups are done, map the FDT as read-only now */
>>>>>>>> fixmap_remap_fdt(dt_phys, &size, PAGE_KERNEL_RO);
>>>>>>>>
>>>>>>>> + /*
>>>>>>>> + * Save dt_phys address so that it can be used later for kexec. This
>>>>>>>> + * is done as __pa() is only intended to be used for linear map addresses
>>>>>>>> + * and using it for initial_boot_params which is in fixmap will give an
>>>>>>>> + * incorrect value.
>>>>>>>> + */
>>>>>>>> + set_initial_boot_params_pa(dt_phys);
>>>>>>>
>>>>>>> No new arch->dt functions please. If we need to save off the PA, then do
>>>>>>> that when we set initial_boot_params.
>>>>>>>
>>>>>>> Rob
>>>>>>
>>>>>>
>>>>>> initial_boot_params is set in early_init_dt_verify, called by early_init_dt_scan.
>>>>>> This is done in setup_machine_fdt in arm64 where the PA is available,
>>>>>> but in other functions in other architectures, where the PA is not available.
>>>>>
>>>>> Doesn't __pa() work for all the other architectures? That's what your
>>>>> patch indicates.
>>>>>
>>>>
>>>> Yes, __pa() works for all other architectures.
>>>>
>>>> But we would need to add initial_boot_params_pa of type phys_addr_t
>>>> as an argument for early_init_dt_scan, which is called by all other archs,
>>>> and we technically cant use 0 as an invalid value.
>>>>
>>>> We could convert initial_boot_params_pa to void *, and pass NULL for all
>>>> other archs. But again, I don't really think we should be changing the
>>>> early_init_dt_scan(dt_virt) call in all other archs to
>>>> early_init_dt_scan(dt_virt, NULL) just to save initial_boot_params_pa
>>>> in arm64?
>>>>
>>>>>> So it makes it quite messy to set it in the same place as initial_boot_params.
>>>>>> Its only needed for arm64 and making a change in all archs probably isnt a good idea?
>>>>>>
>>>>>> Any reason to not add a new function to make arch -> of/fdt call?
>>>>>
>>>>> Yes. It is the opposite direction I have reworked the interfaces to.
>>>>> We don't want each arch calling various early DT functions at random
>>>>> times and order. That's fragile when the DT functions make assumptions
>>>>> about when they are called or what's been initialized already.
>>>>>
>>>>> Another option is to make arm64 copy the DT as some arches do.
>>>>>
>>>>> Rob
>>>>
>>>> Ah maybe I didn't understand this properly, but isnt early_init_dt_scan an
>>>> arch -> of/fdt interfaces. set_initial_boot_params_pa is a similar interface
>>>> to early_init_dt_scan?
>>>
>>> Yes, and I don't want more APIs if they can be avoided. When is
>>> set_initial_boot_params_pa() supposed to be called? Is it before or
>>> after early_init_dt_scan()?
>>
>> Its only needed in arm64, and can be either before or after, as long as its
>> somewhere in setup_machine_fdt, where dt_phys is available.
>
> Maybe only arm64 today. What happens when riscv decides they too want
> to support the DT anywhere in memory including outside the linear
> address map and then they need the same thing.
>
>>> Can subsequent OF functions assume the PA
>>> is valid?
>>
>> After set_initial_boot_params_pa has been called, yes.
>
> How do I know it has been called? Do I have to go wade thru every arch
> to see? You could document the requirement to be immediately after
> early_init_dt_scan(), but then how do you enforce that? You can't
> unless you design the interface to just avoid the problem in the first
> place.
>
>>> If an arch doesn't call set_initial_boot_params_pa() is
>>> __pa() valid or did they just forget to call it?
>>
>> Only arm64 seems to do the fixmap as discussed in
>> https://lore.kernel.org/all/1ea5538f-7e96-4034-9af9-e2d5fd72e069@xxxxxxxxx/,
>> so __pa should work in others.
>>
>> Requiring the PA to
>>> be set at the same time as initial_boot_params avoids all those issues
>>> with any period of time having the PA incorrect.
>>>
>>
>> Are you recommending I send a patch which changes all archs to call
>> early_init_dt_scan(dt_virt, NULL)?
>> or maybe early_init_dt_scan(dt_virt, __pa(dt_virt))?
>> and arm to call early_init_dt_scan(dt_virt, dt_phys).
>
> I believe that's what I suggested already, so yes. Whether NULL or
> __pa(dt_virt))? __pa() would be better because then the arch has to
> think about whether that is right or not.
Sounds good! Will send a v2 with the change.
Thanks
>
>> Happy to do send a v2 with that if its the way forward, although I feel
>> set_initial_boot_params_pa() in just one arch might be better than
>> changing this for all archs.
>
> We don't work around kernel APIs if they don't meet changing needs. We
> change them.
>
> Rob
