Re: [PATCH v2 0/3] userfaultfd: convert userfaultfd functions to use folios

["zhangpeng (AS)" <zhangpeng362@huawei.com>]

On 2023/3/14 21:23, Matthew Wilcox wrote:

> On Tue, Mar 14, 2023 at 01:13:47PM +0000, Peng Zhang wrote:
>> From: ZhangPeng<zhangpeng362@huawei.com>
>>
>> This patch series converts several userfaultfd functions to use folios.
>> And this series pass the userfaultfd selftests and the LTP userfaultfd
>> test cases.
> That's what you said about the earlier patchset too.  Assuming you
> ran the tests, they need to be improved to fid the bug that was in the
> earlier version of the patches.

I did run the tests both times before sending the patches. However, the
bug in the earlier version patches[1] is a hard corner case[2] to
trigger. To trigger it, we need to call copy_large_folio_from_user()
with allow_pagefault == true, which requires hugetlb_mcopy_atomic_pte()
to return -ENOENT. This means that calling copy_large_folio_from_user()
with allow_pagefault == false failed, i.e. copy_from_user() failed.
Building a self-test that copy_from_user() fails could be difficult.

__mcopy_atomic()
	__mcopy_atomic_hugetlb()
		hugetlb_mcopy_atomic_pte()
			copy_large_folio_from_user(..., ..., false);
			// if ret_val > 0, return -ENOENT
				copy_from_user()
				// copy_from_user() needs to fail
		if (err == -ENOENT) copy_large_folio_from_user(..., ..., true);


[1] https://lore.kernel.org/all/20230314033734.481904-3-zhangpeng362@huawei.com/

> -long copy_huge_page_from_user(struct page *dst_page,
> +long copy_large_folio_from_user(struct folio *dst_folio,
>   				const void __user *usr_src,
> -				unsigned int pages_per_huge_page,
>   				bool allow_pagefault)
>   {
>   	void *page_kaddr;
>   	unsigned long i, rc = 0;
> -	unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
> +	unsigned int nr_pages = folio_nr_pages(dst_folio);
> +	unsigned long ret_val = nr_pages * PAGE_SIZE;
>   	struct page *subpage;
> +	struct folio *inner_folio;
>   
> -	for (i = 0; i < pages_per_huge_page; i++) {
> -		subpage = nth_page(dst_page, i);
> +	for (i = 0; i < nr_pages; i++) {
> +		subpage = folio_page(dst_folio, i);
> +		inner_folio = page_folio(subpage);
>   		if (allow_pagefault)
> -			page_kaddr = kmap(subpage);
> +			page_kaddr = kmap_local_folio(inner_folio, 0);
>   		else
>   			page_kaddr = kmap_atomic(subpage);
>   		rc = copy_from_user(page_kaddr,
>   				usr_src + i * PAGE_SIZE, PAGE_SIZE);
>   		if (allow_pagefault)
> -			kunmap(subpage);
> +			kunmap_local(page_kaddr);
>   		else
>   			kunmap_atomic(page_kaddr);


Thanks,
Peng.