[dm-devel] [PATCH 2/7] mm: introduce kvmalloc and kvmalloc_node

[Mike Snitzer]

On Tue, Jul 14 2015 at  5:13pm -0400,
David Rientjes <rientjes at google.com> wrote:

> On Thu, 9 Jul 2015, Mikulas Patocka wrote:
> 
> > > > Index: linux-4.2-rc1/mm/util.c
> > > > ===================================================================
> > > > --- linux-4.2-rc1.orig/mm/util.c	2015-07-07 15:58:11.000000000 +0200
> > > > +++ linux-4.2-rc1/mm/util.c	2015-07-08 19:22:26.000000000 +0200
> > > > @@ -316,6 +316,61 @@ unsigned long vm_mmap(struct file *file,
> > > >  }
> > > >  EXPORT_SYMBOL(vm_mmap);
> > > >  
> > > > +void *kvmalloc_node(size_t size, gfp_t gfp, int node)
> > > > +{
> > > > +	void *p;
> > > > +	unsigned uninitialized_var(noio_flag);
> > > > +
> > > > +	/* vmalloc doesn't support no-wait allocations */
> > > > +	WARN_ON_ONCE(!(gfp & __GFP_WAIT));
> > > > +
> > > > +	if (likely(size <= KMALLOC_MAX_SIZE)) {
> > > > +		/*
> > > > +		 * Use __GFP_NORETRY so that we don't loop waiting for the
> > > > +		 *	allocation - we don't have to loop here, if the memory
> > > > +		 *	is too fragmented, we fallback to vmalloc.
> > > 
> > > I'm not sure about this decision.  The direct reclaim retry code is the
> > > normal default behaviour and becomes more important with larger allocation
> > > attempts.  So why turn it off, and make it more likely that we return
> > > vmalloc memory?
> > 
> > It can avoid triggering the OOM killer in case of fragmented memory.
> > 
> > This is general question - if the code can handle allocation failure 
> > gracefully, what gfp flags should it use? Maybe add some flag 
> > __GFP_MAYFAIL instead of __GFP_NORETRY that changes the behavior in 
> > desired way?
> > 
> 
> There's a misunderstanding in regards to the comment: __GFP_NORETRY 
> doesn't turn direct reclaim or compaction off, it is still attempted and 
> with the same priority as any other allocation.  This only stops the page 
> allocator from calling the oom killer, which will free memory or panic the 
> system, and looping when memory is available.
> 
> In regards to the proposal in general, I think it's unnecessary because we 
> are still left behind with other users who open code their call to 
> vmalloc.  I was interested in commit 058504edd026 ("fs/seq_file: fallback 
> to vmalloc allocation") since it solved an issue with high memory 
> fragmentation.  Note how it falls back to vmalloc(): _without_ this 
> __GFP_NORETRY.  That's because we only want to fallback when high-order 
> allocations fail and the page allocator doesn't implicitly loop due to the 
> order.  ext4_kvmalloc(), ext4_kzmalloc() does the same.
> 
> The differences in implementations between those that do kmalloc() and 
> fallback to vmalloc() are different enough that I don't think we need this 
> addition.

Wouldn't mm benefit from acknowledging the pattern people are
open-coding and switching existing code over to official methods for
accomplishing the same?

It is always easier to shoehorn utility functions locally within a
subsystem (be it ext4, dm, etc) but once enough do something in a
similar but different way it really should get elevated.