Re: [PATCH] mm: Warn about costly page allocation

[Mel Gorman <mgorman@suse.de>]

On Mon, Jul 09, 2012 at 05:46:57PM +0900, Minchan Kim wrote:
> > > <SNIP>
> > > +#if defined(CONFIG_DEBUG_VM) && !defined(CONFIG_COMPACTION)
> > > +static inline void check_page_alloc_costly_order(unsigned int order)
> > > +{
> > > +	if (unlikely(order > PAGE_ALLOC_COSTLY_ORDER)) {
> > > +		printk_once("WARNING: You are tring to allocate %d-order page."
> > > +		" You might need to turn on CONFIG_COMPACTION\n", order);
> > > +	}
> > 
> > WARN_ON_ONCE would tell you what is trying to satisfy the allocation.
> 
> Do you mean that it would be better to use WARN_ON_ONCE rather than raw printk?

Yes.

> If so, I would like to insist raw printk because WARN_ON_ONCE could be disabled
> by !CONFIG_BUG.
> If I miss something, could you elaborate it more?
> 

Ok, but all this will tell you is that *something* tried a high-order
allocation. It will not tell you who and because it's a printk_once, it
will also not tell you how often it's happening. You could add a
dump_stack to capture that information.

> > 
> > It should further check if this is a GFP_MOVABLE allocation or not and if
> > not, then it should either be documented that compaction may only delay
> > allocation failures and that they may need to consider reserving the memory
> > in advance or doing something like forcing MIGRATE_RESERVE to only be used
> > for high-order allocations.
> 
> Okay. but I got confused you want to add above description in code directly
> like below or write it down in comment of check_page_alloc_costly_order?
> 

You're aiming this at embedded QA people according to your changelog so
do whatever you think is going to be the most effective. It's already
"known" that high-order kernel allocations are meant to be unreliable and
apparently this is being ignored. The in-code warning could look
something like

if (unlikely(order > PAGE_ALLOC_COSTLY_ORDER)) {
	printk_once("%s: page allocation high-order stupidity: order:%d, mode:0x%x\n",
                   current->comm, order, gfp_mask);
	if (gfp_flags & __GFP_MOVABLE) {
		printk_once("Enable compaction or whatever\n");
		dump_stack();
	} else {
		printk_once("Regular high-order kernel allocations like this will eventually start failing.");
		dump_stack();
	}
}

There should be a comment above it giving more information if you think
the embedded people will actually read it. Of course, if this warning
triggers during driver initialisation then it might be a completely useless.
You could rate limit the warning (printk_ratelimit()) instead to be more
effective. As I don't know what sort of device drivers you are seeing this
problem with I can't judge what the best style of warning would be.

-- 
Mel Gorman
SUSE Labs

--
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>