Re: USBFS Memory allocation Bug

[Mel Gorman <mel@csn.ul.ie>]

On Wed, Mar 10, 2010 at 10:49:11AM -0500, Alan Stern wrote:
> On Wed, 10 Mar 2010, Mel Gorman wrote:
> 
> > On Wed, Mar 10, 2010 at 10:16:09AM -0500, Alan Stern wrote:
> > > On Wed, 10 Mar 2010, Mel Gorman wrote:
> > > 
> > > > On Tue, Mar 09, 2010 at 10:09:57PM +0100, Markus Rechberger wrote:
> > > > > Hi,
> > > > > 
> > > > > I just stumbled around following issue:
> > > > > 
> > > > > <SNIP to have one full message>
> > > > >
> > > > > [275835.491094] mediasrv: page allocation failure. order:6, mode:0x40d0
> > > > 
> > > > This is massive. Why is usbdev_ioctl requiring an order-6 allocation?
> > > > Does it have the option to fall back to vmalloc() for example?
> > > 
> > > We needed to implement an upper limit on the buffer size, and the value
> > > chosen was sufficient for 8 ms of data.  There is no fallback to
> > > vmalloc().
> > > 
> > > Of course, the mediasrv program could always make multiple smaller
> > > data submissions instead of a single big one.
> > > 
> > 
> > Is there any means for the driver to take the large request, break it up
> > into multiple smaller requests and submit them one at a time?
> 
> In theory almost anything is possible.  But it would be a big effort
> and not consistent with the way the rest of the driver works.
> 

Then about the only other suggestion would be a mempool containing a small
number of largest-possible buffers that is enabled if there is no swap
available.

> > > > So, it's a high-order allocation but no swap is configured. As the
> > > > system has a big mix of both anonymous and file memory, it is likely it
> > > > was unable to find a large enough contiguous range of file-backed memory
> > > > it could discard. There is pretty much nothing the memory manager could
> > > > do here. 
> > > > 
> > > > FWIW, if swap was configured I'd think there was more the memory manager
> > > > could have done.
> > > > 
> > > > Is usb falling back to vmalloc() or order-0 pages possible?
> > > 
> > > No.  It's not possible since the buffer has to be contiguous for DMA
> > > purposes (this is a requirement of the driver interface).
> > > 
> > 
> > Regrettably, there is not much the VM can do in this situation. Without
> > swap, it cannot magic up contiguous memory. If memory compaction gets
> > merged then it could do more but the risk of failure is still non-zero.
> 
> Agreed.  The best way to fix the allocation failure is to provide swap 
> space.
> 

Or a mempool so at least it can always make some forward progress.
Whether it is worth it or not depends on what happens the application
when this allocation fails.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab