Re: [PATCH v2 -next] riscv: mm: remove redundant trampoline PGD for 64bit

[Nanyong Sun <sunnanyong@huawei.com>]

On 2021/9/8 16:56, Alex Ghiti wrote:
> Hi Nanyong,
>
> Le 8/09/2021 à 08:42, Nanyong Sun a écrit :
>>
>> On 2021/8/14 6:08, Palmer Dabbelt wrote:
>>> On Mon, 02 Aug 2021 05:43:02 PDT (-0700), alex@ghiti.fr wrote:
>>>> Hi Nanyong,
>>>>
>>>> Le 28/07/2021 à 13:55, Alex Ghiti a écrit :
>>>>>
>>>>>
>>>>> Le 28/07/2021 à 04:49, Nanyong Sun a écrit :
>>>>>> Remove redundant trampoline PGD for 64bit and add more comment
>>>>>> for why 32bit systems need trampoline PGD.
>>>>>>
>>>>>>
>>>>>> +load_kernel_pgd:
>>>>>> +        /*
>>>>>> +         * Switch to kernel page tables.  A full fence is necessary
>>>>>> in order to
>>>>>> +         * avoid using the trampoline translations, which are only
>>>>>> correct for
>>>>>> +         * the first superpage.  Fetching the fence is guarnteed 
>>>>>> to work
>>>>>> +         * because that first superpage is translated the same way.
>>>>>> +         */
>>>>>> +        csrw CSR_SATP, a2
>>>>>> +        sfence.vma
>>>>>> +
>>>>>> +load_done:
>>>>>>       /* Set trap vector to spin forever to help debug */
>>>>>>       la a0, .Lsecondary_park
>>>>>>       csrw CSR_TVEC, a0
>>>>
>>>>
>>>> I suppose stvec was set this way to catch any problem with 
>>>> early_pg_dir,
>>>> you moved that and then this defeats this original purpose.
>>>
>> Hi Alex,
>>
>>      I don't think so, before set early_pg_dir to satp, it's the 
>> physical address world, we must set stvec as
>>
>> the first place in virtual address world we want jump to. And I don't 
>> think ".Lsecondary_park " can catch
>>
>> problem of bad early_pg_dir, if the basic page table is wrong, CPU 
>> also can not go to the virtual address stored in stvec correctly.
>
> But I think then that it loops forever at the stvec address which 
> allows to know where the boot failed.

If satp had a problem, then cpu can not fetch instruction where stvec 
pointing to, as what palmer said: if you end up in a position where the 
processer is unable to commit an instruction you also

lose the ability to do anything meaningful with the debugger, thus 
essentially locking up the system.

>
>>
>> More, in the original code, before set trampoline_pg_dir, what if the 
>> trampoline_pg_dir had a problem?
>
> You're right but this debug 'feature' was not installed, I guess 
> somebody had a hard time at some point with the early page table and 
> not the trampoline :)
>
> Anyway, I was just pointing that you 'broke' the current way things 
> work and unless this is for an explicit good reason, that should not 
> happen.
>
The design logic is: at the first time cpu convert to virtual address 
world from physical world, actually stvec is a real "trampoline",  it 
can not be set

as a pointer to spin trap, it should be set to the first place in 
virtual world where we wanna go. After that, then we set stvec as a spin 
trap to catch

any problem in later running.

So, for 64bit system, if we want to delete  trampoline_pg_dir, the 
design principle is not broken here. For 32bit system, I really need 
change back.

>>
>>> Essentially.
>>>
>>> The specific issue is that the JTAG debug spec is defined (or at 
>>> least was when I was using it, it's been years since I've needed to 
>>> do that) in terms of committed instructions.  Thus if you end up in 
>>> a position where the processer is unable to commit an instruction 
>>> you also lose the ability to do anything meaningful with the 
>>> debugger, thus essentially locking up the system.
>>>
>>> The most common way to end up in a situation where the processor is 
>>> unable to commit an instruction is to have a fault with an invalid 
>>> trap vector: maybe dangling from M-mode, the last boot, reset, 
>>> whatever.  Then as soon as you take a trap the system locks up.  Any 
>>> trap before we have a working trap handler is a bug, but it's way 
>>> harder to debug things when the debugger doesn't function.
>>>
>>> There is of course no way to fundamentally prevent these sort of 
>>> no-commitable-instruction situations, but I got into the habbit of 
>>> just setting up a trivial trap entry point ASAP -- it probably took 
>>> a dozen rounds of trying to debug the debugger only to realize it 
>>> was per spec to hang, but that idiom eventually crept into pretty 
>>> much everything.
>>>
>>> Not sure if the debug spec is still written this way (or if 
>>> debuggers respect it), as I haven't had to use one in a while.
>>>
>>>>
>>>>
>>>>>>
>>>>>> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
>>>>>> index ac48742fa6fc..306fcb2334fa 100644
>>>>>> --- a/arch/riscv/mm/init.c
>>>>>> +++ b/arch/riscv/mm/init.c
>>>>>> @@ -219,13 +219,17 @@ unsigned long pfn_base __ro_after_init;
>>>>>>   EXPORT_SYMBOL(pfn_base);
>>>>>>   pgd_t swapper_pg_dir[PTRS_PER_PGD] __page_aligned_bss;
>>>>>> +#ifndef CONFIG_64BIT
>>>>>>   pgd_t trampoline_pg_dir[PTRS_PER_PGD] __page_aligned_bss;
>>>>>> +#endif /* CONFIG_64BIT */
>>>>
>>>>
>>>> As stated in Documentation/process/coding-style.rst, it is better 
>>>> to use
>>>> __maybe_unused rather than #ifdefs.
>>>>
>>>>
>> I'm afraid that __maybe_unused can not save one page memory here.
>
> What do you mean?
>
I mean trampoline_pg_dir cost 4096 bytes here and __maybe_unused only 
tell compiler don't raise a warning, but it still cost memory.