Re: allocation failure in b43/ssb

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 7 Nov 2008 14:54:03 -0500 "John W. Linville" <linville@tuxdriver.com> wrote:

> On Fri, Nov 07, 2008 at 02:09:44PM -0500, Dave Jones wrote:
> > We just had a report from a user who spotted a bunch of messages
> > of the form below in his logs..
> 
> <snip>
> 
> > The kernel is a little old (2.6.26).
> > 
> > Looking at it though, I'm puzzled..
> > 
> > It seems we failed to allocate an order-1 allocation, even though
> > we had memory available in the DMA and normal zones.
> > 
> > What am I missing?
> > 
> > Full logs at https://bugzilla.redhat.com/attachment.cgi?id=322882
> 
> Most of those traces look a little scrambled, but as best as I can
> tell that call to setup_rx_descbuffer comes from dma_rx.  That means
> that the call to __dev_alloc_skb is done with GFP_ATOMIC.  I don't
> know if that factors into the answer to what is happening or not.
> 

It does.  If there were not any two physically-contiguous free pages in
the page allocator reserves, __alloc_pages() will not be able to
satisfy that order-1 allocation request.

You can have plenty of free memory, but if it's all fragmented,
non-order-0 atomic allocations can still fail.

GFP_KERNEL allocations can run page reclaim to _make_ the allocation
request succeed in this situation.