Re: [PATCH v4] tee: Use iov_iter to better support shared buffer registration

[Jens Axboe <axboe@kernel.dk>]

On 12/5/23 9:55 AM, Arnaud POULIQUEN wrote:
> hi Jens Axboe, Al Viro,
> 
> On 12/4/23 18:13, Jens Axboe wrote:
>> On 12/4/23 10:02 AM, Arnaud POULIQUEN wrote:
>>> Hi,
>>>
>>> On 12/4/23 17:40, Jens Axboe wrote:
>>>> On 12/4/23 9:36 AM, Jens Axboe wrote:
>>>>> On 12/4/23 5:42 AM, Sumit Garg wrote:
>>>>>> IMO, access_ok() should be the first thing that import_ubuf() or
>>>>>> import_single_range() should do, something as follows:
>>>>>>
>>>>>> diff --git a/lib/iov_iter.c b/lib/iov_iter.c
>>>>>> index 8ff6824a1005..4aee0371824c 100644
>>>>>> --- a/lib/iov_iter.c
>>>>>> +++ b/lib/iov_iter.c
>>>>>> @@ -1384,10 +1384,10 @@ EXPORT_SYMBOL(import_single_range);
>>>>>>
>>>>>>  int import_ubuf(int rw, void __user *buf, size_t len, struct iov_iter *i)
>>>>>>  {
>>>>>> -       if (len > MAX_RW_COUNT)
>>>>>> -               len = MAX_RW_COUNT;
>>>>>>         if (unlikely(!access_ok(buf, len)))
>>>>>>                 return -EFAULT;
>>>>>> +       if (len > MAX_RW_COUNT)
>>>>>> +               len = MAX_RW_COUNT;
>>>>>>
>>>>>>         iov_iter_ubuf(i, rw, buf, len);
>>>>>>         return 0;
>>>>>>
>>>>>> Jens A., Al Viro,
>>>>>>
>>>>>> Was there any particular reason which I am unaware of to perform
>>>>>> access_ok() check on modified input length?
>>>>>
>>>>> This change makes sense to me, and seems consistent with what is done
>>>>> elsewhere too.
>>>>
>>>> For some reason I missed import_single_range(), which does it the same
>>>> way as import_ubuf() currently does - cap the range before the
>>>> access_ok() check. The vec variants sum as they go, but access_ok()
>>>> before the range.
>>>>
>>>> I think part of the issue here is that the single range imports return 0
>>>> for success and -ERROR otherwise. This means that the caller does not
>>>> know if the full range was imported or not. OTOH, we always cap any data
>>>> transfer at MAX_RW_COUNT, so may make more sense to fix up the caller
>>>> here.
>>>>
>>>
>>> Should we limit to MAX_RW_COUNT or return an error? Seems to me that
>>> limiting could generate side effect later that could be not simple to
>>> debug.
>>
>> We've traditionally just truncated the length, so principle of least
>> surprise says we should continue doing that.
>>
> 
> 
> As Jens Wiklander has proposed using iov_iter_ubuf() instead of
> import_ubuf(), should I propose a patch updating import_ubuf() and
> import_single_range()? Or would you prefer that we keep the functions
> unchanged for the time being?

Arguably it should be consistent with iovec imports, which return the
length (or error). But it might be safer to just check access_ok()
first and then truncate at least, vs what is there now.

Note that for 6.8 import_single_range() is gone as it was really just
doing the same thing that import_ubuf() is. Any further changes in this
area should CC Christian Brauner as well, as he has that staged in his
tree.

-- 
Jens Axboe