Re: [PATCH v2 1/3] kasan: Avoid sleepable page allocation from atomic context

[Andrey Ryabinin <ryabinin.a.a@xxxxxxxxx>]

On 4/9/25 4:25 PM, Alexander Gordeev wrote:
> On Wed, Apr 09, 2025 at 04:10:58PM +0200, Andrey Ryabinin wrote:
> 
> Hi Andrey,
> 
>>> @@ -301,7 +301,7 @@ static int kasan_populate_vmalloc_pte(pte_t *ptep, 
>>> unsigned long addr,
>>>     if (likely(!pte_none(ptep_get(ptep))))
>>>             return 0;
>>>  
>>> -   page = __get_free_page(GFP_KERNEL);
>>> +   page = __get_free_page(GFP_ATOMIC);
>>>     if (!page)
>>>             return -ENOMEM;
>>>  
>>
>> I think a better way to fix this would be moving out allocation from atomic 
>> context. Allocate page prior
>> to apply_to_page_range() call and pass it down to 
>> kasan_populate_vmalloc_pte().
> 
> I think the page address could be passed as the parameter to 
> kasan_populate_vmalloc_pte().

We'll need to pass it as 'struct page **page' or maybe as pointer to some 
struct, e.g.:
struct page_data {
 struct page *page;
};


So, the kasan_populate_vmalloc_pte() would do something like this:

kasan_populate_vmalloc_pte() {
        if (!pte_none)
                return 0;
        if (!page_data->page)
                return -EAGAIN;

        //use page to set pte

        //NULLify pointer so that next kasan_populate_vmalloc_pte() will bail
        // out to allocate new page
        page_data->page = NULL; 
}

And it might be good idea to add 'last_addr' to page_data, so that we know 
where we stopped
so that the next apply_to_page_range() call could continue, instead of starting 
from the beginning. 


> 
>> Whenever kasan_populate_vmalloc_pte() will require additional page we could 
>> bail out with -EAGAIN,
>> and allocate another one.
> 
> When would it be needed? kasan_populate_vmalloc_pte() handles just one page.
> 

apply_to_page_range() goes over range of addresses and calls 
kasan_populate_vmalloc_pte()
multiple times (each time with different 'addr' but the same '*unused' arg). 
Things will go wrong
if you'll use same page multiple times for different addresses.


> Thanks!