Re: Oops in 3.7-rc8 isolate_free_pages_block()

[Mel Gorman <mgorman@suse.de>]

On Thu, Dec 06, 2012 at 11:38:47AM -0800, Linus Torvalds wrote:
> Ok, I've applied the patch.
> 

Thanks.

> Mel, some grepping shows that there is an old line that does
> 
>     end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
> 
> which looks bogus.

It's bogus. The impact is that multiple compaction attempts may be needed
to clear a particular block for allocation. THP allocation success rate
under stress will be lower and the latency before a range of pages is
collapsed by khugepaged to a huge page will be higher. The impact of this
is less and it should not result in a bug like Henrik's

An attentive reviewer is going to exclaim that GFP_ATOMIC allocations for
jumbo frames is impacted by this but it isn't. Even with this bogus walk,
compaction will be clearing SWAP_CLUSTER_MAX contiguous chunks which is
enough for jumbo frames.

> That should probably also use "+ 1" instead. But
> I'll consider that an independent issue, so I applied the one patch
> regardless.
> 
> There is also a
> 
>     low_pfn += pageblock_nr_pages;
>     low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
> 
> that looks suspicious for similar reasons. Maybe
> 
>     low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
> 

This one is working by co-incidence because the low_pfn will be aligned
in most cases. If it was outright broken then CMA would never work either.

> instead? Although that *can* result in the same low_pfn in the end, so
> maybe that one was correct after all? I just did some grepping, no
> actual semantic analysis...
> 

They need fixing but the impact is much less severe and does not justify
delaying 3.8 over unlike the other last-minute fixes. My performance
writing patches during talks was less than stellar yesterday so I'll avoid
a repeat performance and follow up with Andrew early next week with a cc
to -stable. It'll also give me a chance to run the patches through the
highalloc stress tests and confirm that allocation success rate is higher
and latency lower as would be expected by such a fix.

Thanks.

-- 
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>