Re: [PATCH v2 3/3] riscv: Fix crash when flushing executable ioremap regions

[Jan Kiszka <jan.kiszka@web.de>]

On 20.02.20 06:49, Alex Ghiti wrote:
> Hi Jan,
>
> On 2/16/20 2:56 PM, Alex Ghiti wrote:
>> On 2/16/20 11:05 AM, Jan Kiszka wrote:
>>> On 16.02.20 15:41, Alex Ghiti wrote:
>>>> Hi Jan,
>>>>
>>>> On 2/15/20 6:49 AM, Jan Kiszka wrote:
>>>>> From: Jan Kiszka <jan.kiszka@siemens.com>
>>>>>
>>>>> Those are not backed by page structs, and pte_page is returning an
>>>>> invalid pointer.
>>>>>
>>>>> Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
>>>>> =2D--
>>>>>   arch/riscv/mm/cacheflush.c | 3 ++-
>>>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/arch/riscv/mm/cacheflush.c b/arch/riscv/mm/cacheflush.c
>>>>> index 8930ab7278e6..9ee2c1a387cc 100644
>>>>> =2D-- a/arch/riscv/mm/cacheflush.c
>>>>> +++ b/arch/riscv/mm/cacheflush.c
>>>>> @@ -84,7 +84,8 @@ void flush_icache_pte(pte_t pte)
>>>>>   {
>>>>>       struct page *page =3D pte_page(pte);
>>>>>
>>>>> -    if (!test_and_set_bit(PG_dcache_clean, &page->flags))
>>>>> +    if (!pfn_valid(pte_pfn(pte)) ||
>>>>> +        !test_and_set_bit(PG_dcache_clean, &page->flags))
>>>>>           flush_icache_all();
>>>>>   }
>>>>>   #endif /* CONFIG_MMU */
>>>>> =2D-
>>>>> 2.16.4
>>>>>
>>>>>
>>>>
>>>> When did you encounter such a situation ? i.e. executable code that is
>>>> not backed by struct page ?
>>>>
>>>> Riscv uses the generic implementation of ioremap and the way
>>>> _PAGE_IOREMAP is defined does not allow to map executable memory region
>>>> using ioremap, so I'm interested to understand how we end up in
>>>> flush_icache_pte for an executable region not backed by any struct
>>>> page.
>>>
>>> You can create executable mappings of memory that Linux does not
>>> initially consider as RAM via ioremap_prot or ioremap_page_range. We are
>>> using that in Jailhouse to load the hypervisor code into reserved memory
>>> that is ioremapped for the purpose. Works fine on x86, arm and arm64.
>>>
>>> Jan
>>
>> Ok thanks, I had missed this API.
>>
>> Regarding your patch, I find it weird to do anything if the pfn is
>> invalid, we could have garbage in pte pointing to an invalid region
>> for example (I admit that the effect of flushing the icache would not
>> be catastrophic in that situation).
>>
>> I'm not saying I will come with a better solution but I'll take a
>> deeper look tomorrow.
>>
>> Alex
>>
>
> I took a look at the Jailhouse driver. After loading the hypervisor into
> the ioremapped region, it explicitly ensures icache/dcache consistency
> by calling flush_icache_range here:
>
> https://github.com/siemens/jailhouse/blob/master/driver/main.c#L505
>

Yeah, the arm64 port needed this.

> There seems to be an implicit (?) rule that states that in-kernel code
> modification must handle icache/dcache consistency:
>
> In arm64 set_pte_at definition, they do not sync icache/dcache when the
> pte is kernel:
>
> https://elixir.bootlin.com/linux/latest/source/arch/arm64/include/asm/pgtable.h#L271
>
>
> In mips, they do the same:
>
> https://elixir.bootlin.com/linux/latest/source/arch/mips/mm/cache.c#L137
>
> So funnily, I'd do the contrary of what you have done, the mips way:
>
> diff --git a/arch/riscv/mm/cacheflush.c b/arch/riscv/mm/cacheflush.c
> index 8930ab7278e6..c90c8bb49109 100644
> --- a/arch/riscv/mm/cacheflush.c
> +++ b/arch/riscv/mm/cacheflush.c
> @@ -84,6 +84,9 @@ void flush_icache_pte(pte_t pte)
>   {
>          struct page *page = pte_page(pte);
>
> +       if (unlikely(!pfn_valid(pte_pfn(pte))))
> +               return;
> +
>          if (!test_and_set_bit(PG_dcache_clean, &page->flags))
>                  flush_icache_all();
>   }
>
> What do you think ?
>

I wouldn't mind doing it like above. I suspect that became the common
simple pattern because no one expected a use case like with Jailhouse.
But I'm by far not an expert in mm topics in the kernel.

Jan