Re: [PATCH 4/8] snd: make snd_map_bufs() deal with ITER_UBUF

[Jens Axboe <axboe@kernel.dk>]

On 3/28/23 12:52 PM, Al Viro wrote:
> On Tue, Mar 28, 2023 at 11:52:10AM -0600, Jens Axboe wrote:
>> On 3/28/23 11:50 AM, Linus Torvalds wrote:
>>> On Tue, Mar 28, 2023 at 10:36 AM Jens Axboe <axboe@kernel.dk> wrote:
>>>>
>>>> @@ -3516,23 +3516,28 @@ static void __user **snd_map_bufs(struct snd_pcm_runtime *runtime,
>>>>                                   struct iov_iter *iter,
>>>>                                   snd_pcm_uframes_t *frames, int max_segs)
>>>>  {
>>>> +       int nr_segs = iovec_nr_user_vecs(iter);
>>>
>>> This has a WARN_ON_ONCE() for !user_backed, but then..
>>>
>>>>         void __user **bufs;
>>>> +       struct iovec iov;
>>>>         unsigned long i;
>>>>
>>>>         if (!iter->user_backed)
>>>>                 return ERR_PTR(-EFAULT);
>>>
>>> here the code tries to deal with it.
>>>
>>> So I think the two should probably be switched around.
>>
>> True, it was actually like that before I refactored it to include
>> that common helper. I'll swap them around, thanks.
> 
> Umm...  That looks really weird - if nothing else, it seems that this
> thing quietly ignores the ->iov_len on all but the first iovec.

I agree, but this is how it currently works...

> Might make sense to ask ALSA folks what the hell is going on there;
> it's readv()/writev() on pcm device, and it looks like userland ABI
> is really perverted here... ;-/

I have sent them email separately to confirm that the only cases
that makes sense here is nr_segs >= 2. But the ABI is what it is,
however horrible it may be :/

-- 
Jens Axboe