Re: patch to give kmalloc a chance (RFC)

[Andrew Morton <akpm@xxxxxxxxx>]

Steve Lord wrote:
> 
> I am picturing Rik with shivers going down his spine right now ;-)
> 

Take a look at __jbd_kmalloc() ;)

For a NOFS allocation, simply retrying is a crude-but-effective
fix.  We have kicked kswapd and kswapd _does_ have __GFP_FS.

So things work out, as long as the filesystem doesn't hold locks which
would prevent kswapd from being able to reclaim some memory.

It's not just a lock ordering thing either - the GFP_NOFS allocator
is not able to use the traditional IO-based throttling.  It's
probably just that there's plenty of writeback in flight but the
allocator needs to be more patient.