Re: [RFC PATCH V5] mm readahead: Fix readahead fail for no local memory and limit readahead pages

[David Rientjes <rientjes@google.com>]

On Thu, 6 Feb 2014, Andrew Morton wrote:

> On Thu, 6 Feb 2014 14:58:21 -0800 (PST) David Rientjes <rientjes@google.com> wrote:
> 
> > > > +#define MAX_REMOTE_READAHEAD   4096UL
> > > >  /*
> > > >   * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
> > > >   * sensible upper limit.
> > > >   */
> > > >  unsigned long max_sane_readahead(unsigned long nr)
> > > >  {
> > > > -	return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
> > > > -		+ node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
> > > > +	unsigned long local_free_page;
> > > > +	int nid;
> > > > +
> > > > +	nid = numa_node_id();
> > 
> > If you're intending this to be cached for your calls into 
> > node_page_state() you need nid = ACCESS_ONCE(numa_node_id()).
> 
> ugh.  That's too subtle and we didn't even document it.
> 
> We could put the ACCESS_ONCE inside numa_node_id() I assume but we
> still have the same problem as smp_processor_id(): the numa_node_id()
> return value is wrong as soon as you obtain it if running preemptibly. 
> 
> We could plaster Big Fat Warnings all over the place or we could treat
> numa_node_id() and derivatives in the same way as smp_processor_id()
> (which is a huge pain).  Or something else, but we've left a big hand
> grenade here and Raghavendra won't be the last one to pull the pin?
> 

Normally it wouldn't matter because there's no significant downside to it 
racing, things like mempolicies which use numa_node_id() extensively would 
result in, oops, a page allocation on the wrong node.

This stands out to me, though, because you're expecting the calculation to 
be correct for a specific node.

The patch is still wrong, though, it should just do

	int node = ACCESS_ONCE(numa_mem_id());
	return min(nr, (node_page_state(node, NR_INACTIVE_FILE) +
		        node_page_state(node, NR_FREE_PAGES)) / 2);

since we want to readahead based on the cpu's local node, the comment 
saying we're reading ahead onto "remote memory" is wrong since a 
memoryless node has local affinity to numa_mem_id().

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>