Re: usb/gadget: warning in ep_write_iter/__alloc_pages_nodemask

[Alan Stern]

On Wed, 14 Dec 2016, Michal Hocko wrote:

> On Tue 13-12-16 08:33:34, Alan Stern wrote:
> > On Tue, 13 Dec 2016, Michal Hocko wrote:
> > 
> > > > > That being said, what ep_write_iter does sounds quite stupit. It just
> > > > > allocates a large continuous buffer which seems to be under user
> > > > > control...  Aka no good! It should do that per pages or something like
> > > > > that. Something worth fixing
> > > > 
> > > > It's not important enough to make the driver do all this work.  If
> > > > users want to send large amounts of data, they can send it a page at a
> > > > time (or something like that).
> > > 
> > > Is it really necessary to allocate the full iov_iter_count? Why cannot
> > > we process the from buffer one page at a time?
> > 
> > We could (although one page is really too small -- USB 3.1 can transfer
> > 800 KB per ms so we ought to handle at least 128 KB at a time).
> 
> Is there any problem to submit larger transfers without having the
> buffer physically contiguous?

Async I/O would be rather awkward; it would have to use a work queue
routine.  But it could be done.

And we would still end up allocating the same total space (more
actually, because we would need to store the scatter-gather table too).  
It just wouldn't be contiguous.

> > But
> > turn the argument around: If the user wants to transfer that much data,
> > why can't he _submit_ it one page at a time?
> 
> Not sure I understand.
>  
> > > > If you really want to prevent the driver from attempting to allocate a
> > > > large buffer, all that's needed is an upper limit on the total size.  
> > > > For example, 64 KB.
> > > 
> > > Well, my point was that it is not really hard to imagine to deplete
> > > larger contiguous memory blocks (say PAGE_ALLOC_COSTLY_ORDER). Those are
> > > still causing the OOM killer and chances are that a controlled flood of
> > > these requests could completely DoS the system.
> > 
> > Putting a limit on the total size of a single transfer would prevent 
> > this.
> 
> Dunno, putting a limit to the user visible interface sounds wrong to me.

In practice, I think the data transfer sizes tend to be not very large.  
But I could be wrong about that.

Alan Stern