[PATCH] riscv: mm: remove the BUG_ON check of mapping the last 4K bytes of memory

[Alexandre Ghiti]

Hi Palmer,

On 2/15/22 00:41, Palmer Dabbelt wrote:
> On Tue, 25 Jan 2022 08:10:41 PST (-0800), alex at ghiti.fr wrote:
>>
>> On 1/25/22 16:55, Jisheng Zhang wrote:
>>> remove the BUG_ON check of mapping the last 4K bytes of the addressable
>>> memory since "this is true for every kernel actually" as pointed out
>>> by Alexandre.
>>>
>>> Signed-off-by: Jisheng Zhang <jszhang at kernel.org>
>>> Reviewed-by: Alexandre Ghiti <alex at ghiti.fr>
>>> ---
>>>   arch/riscv/mm/init.c | 8 --------
>>>   1 file changed, 8 deletions(-)
>>>
>>> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
>>> index cf4d018b7d66..8347d0fda8cd 100644
>>> --- a/arch/riscv/mm/init.c
>>> +++ b/arch/riscv/mm/init.c
>>> @@ -811,14 +811,6 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
>>>       BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
>>>       BUG_ON((kernel_map.phys_addr % PMD_SIZE) != 0);
>>>
>>> -#ifdef CONFIG_64BIT
>>> -    /*
>>> -     * The last 4K bytes of the addressable memory can not be 
>>> mapped because
>>> -     * of IS_ERR_VALUE macro.
>>> -     */
>>> -    BUG_ON((kernel_map.virt_addr + kernel_map.size) > 
>>> ADDRESS_SPACE_END - SZ_4K);
>>> -#endif
>>
>>
>> This BUG_ON seems pretty legit to me: I re-read the exchanges we had,
>> and I see that I didn't notice that in your v2, you actually removed the
>> BUG_ON. So that's my bad, what I meant in the first place was that the
>> BUG_ON is true for 32-bit and 64-bit kernels actually.
>
> There's actually an ifndef 64BIT above that sort of handles this case 
> (though I didn't check to see if we're getting the limits correct, so 
> it may not work properly).  That's shrinking the memory, rather than 
> just firing a BUG, and it's not really any more code so we should go 
> that way for both.  I could see leaving a BUG in there, maybe just 
> explicitly using IS_ERR_VALUE as that's really what we're checking for 
> (though if that's not 4K a bunch of stuff will break, so maybe it just 
> doesn't matter).


I think you're talking about:

memory_limit = KERN_VIRT_SIZE - (IS_ENABLED(CONFIG_64BIT) ? SZ_4G : 0);

If yes, that's shrinking the available physical memory we can map in the 
linear mapping, whereas the BUG_ON here is triggered if the kernel 
mapping is very (very) large and overlaps the last *4K* of the address 
space, so I still think the BUG_ON is legit although very unlikely.

However, I'm wondering if the shrinking still holds with the recent move 
of KASAN, I may be back with a fix for that.

Alex


>
>> Sorry my RB was not right on this one :(
>>
>> Alex
>>
>>
>>> -
>>>       pt_ops_set_early();
>>>
>>>       /* Setup early PGD for fixmap */