Re: [RFC] bio: fix page leak bio_add_hw_page failure

[Pavel Begunkov <asml.silence@gmail.com>]

On 7/20/21 3:11 AM, Ming Lei wrote:
> On Mon, Jul 19, 2021 at 06:06:49PM +0100, Pavel Begunkov wrote:
>> On 7/19/21 4:34 PM, Ming Lei wrote:
>>> On Mon, Jul 19, 2021 at 11:53:00AM +0100, Pavel Begunkov wrote:
>>>> __bio_iov_append_get_pages() doesn't put not appended pages on
>>>> bio_add_hw_page() failure, so potentially leaking them, fix it. Also, do
>>>> the same for __bio_iov_iter_get_pages(), even though it looks like it
>>>> can't be triggered by userspace in this case.
>>>>
>>>> Fixes: 0512a75b98f8 ("block: Introduce REQ_OP_ZONE_APPEND")
>>>> Cc: stable@vger.kernel.org # 5.8+
>>>> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
>>>> ---
>>>>
>>>> I haven't tested the fail path, thus RFC. Would be great if someone can
>>>> do it or take over the fix.
>>>>
>>>>  block/bio.c | 15 +++++++++++++--
>>>>  1 file changed, 13 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/block/bio.c b/block/bio.c
>>>> index 1fab762e079b..d95e3456ba0c 100644
>>>> --- a/block/bio.c
>>>> +++ b/block/bio.c
>>>> @@ -979,6 +979,14 @@ static int bio_iov_bvec_set_append(struct bio *bio, struct iov_iter *iter)
>>>>  	return 0;
>>>>  }
>>>>  
>>>> +static void bio_put_pages(struct page **pages, size_t size, size_t off)
>>>> +{
>>>> +	size_t i, nr = DIV_ROUND_UP(size + (off & ~PAGE_MASK), PAGE_SIZE);
>>>> +
>>>> +	for (i = 0; i < nr; i++)
>>>> +		put_page(pages[i]);
>>>> +}
>>>> +
>>>>  #define PAGE_PTRS_PER_BVEC     (sizeof(struct bio_vec) / sizeof(struct page *))
>>>>  
>>>>  /**
>>>> @@ -1023,8 +1031,10 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
>>>>  			if (same_page)
>>>>  				put_page(page);
>>>>  		} else {
>>>> -			if (WARN_ON_ONCE(bio_full(bio, len)))
>>>> -                                return -EINVAL;
>>>> +			if (WARN_ON_ONCE(bio_full(bio, len))) {
>>>> +				bio_put_pages(pages + i, left, offset);
>>>> +				return -EINVAL;
>>>> +			}
>>>
>>> It is unlikely to happen:
>>>
>>>         unsigned short nr_pages = bio->bi_max_vecs - bio->bi_vcnt;
>>>         struct bio_vec *bv = bio->bi_io_vec + bio->bi_vcnt;
>>>         struct page **pages = (struct page **)bv;
>>>
>>> 		pages += entries_left * (PAGE_PTRS_PER_BVEC - 1);
>>> 		size = iov_iter_get_pages(iter, pages, LONG_MAX, nr_pages, &offset);
>>
>> Agree, mentioned in the commit, however ...
>>
>>>>  			__bio_add_page(bio, page, len, offset);
>>>>  		}
>>>>  		offset = 0;
>>>> @@ -1069,6 +1079,7 @@ static int __bio_iov_append_get_pages(struct bio *bio, struct iov_iter *iter)
>>>>  		len = min_t(size_t, PAGE_SIZE - offset, left);
>>>>  		if (bio_add_hw_page(q, bio, page, len, offset,
>>>>  				max_append_sectors, &same_page) != len) {
>>>> +			bio_put_pages(pages + i, left, offset);
>>>
>>> Same with above.
>>
>> ... bio_add_hw_page() is more complex and additionally does checks
>> against queue_max_zone_append_sectors(), queue_max_segments(), and
>> queue_virt_boundary() in of bvec_gap_to_prev().
>>
>> It may be unlikely, but are you sure that those are just safety
>> checks? It's not so obvious to me, so would be great if you could
>> point out the other place where the verification is done.
> 
> OK, bio_add_hw_page() is special, and it needs the handling, but 
> __bio_iov_iter_get_pages() needn't that since it is so obvious.

Right. I don't mind to drop the first chunk, but it doesn't hurt, and
I'd guess the bug came from copy-pasting and editing
__bio_iov_iter_get_pages(). That's the reason I added it in the first
place.

-- 
Pavel Begunkov