On Wed, Jul 16, 2014 at 03:48:22PM +0200, Vlastimil Babka wrote:
> Compaction uses watermark checking to determine if it succeeded in creating
> a high-order free page. My testing has shown that this is quite racy and it
> can happen that watermark checking in compaction succeeds, and moments later
> the watermark checking in page allocation fails, even though the number of
> free pages has increased meanwhile.
>
> It should be more reliable if direct compaction captured the high-order free
> page as soon as it detects it, and pass it back to allocation. This would
> also reduce the window for somebody else to allocate the free page.
>
> Capture has been implemented before by 1fb3f8ca0e92 ("mm: compaction: capture
> a suitable high-order page immediately when it is made available"), but later
> reverted by 8fb74b9f ("mm: compaction: partially revert capture of suitable
> high-order page") due to a bug.
>
> This patch differs from the previous attempt in two aspects:
>
> 1) The previous patch scanned free lists to capture the page. In this patch,
> only the cc->order aligned block that the migration scanner just finished
> is considered, but only if pages were actually isolated for migration in
> that block. Tracking cc->order aligned blocks also has benefits for the
> following patch that skips blocks where non-migratable pages were found.
>
> 2) The operations done in buffered_rmqueue() and get_page_from_freelist() are
> closely followed so that page capture mimics normal page allocation as much
> as possible. This includes operations such as prep_new_page() and
> page->pfmemalloc setting (that was missing in the previous attempt), zone
> statistics are updated etc. Due to subtleties with IRQ disabling and
> enabling this cannot be simply factored out from the normal allocation
> functions without affecting the fastpath.
>
> This patch has tripled compaction success rates (as recorded in vmstat) in
> stress-highalloc mmtests benchmark, although allocation success rates increased
> only by a few percent. Closer inspection shows that due to the racy watermark
> checking and lack of lru_add_drain(), the allocations that resulted in direct
> compactions were often failing, but later allocations succeeeded in the fast
> path. So the benefit of the patch to allocation success rates may be limited,
> but it improves the fairness in the sense that whoever spent the time
> compacting has a higher change of benefitting from it, and also can stop
> compacting sooner, as page availability is detected immediately. With better
> success detection, the contribution of compaction to high-order allocation
> success success rates is also no longer understated by the vmstats.
>
> Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
> Cc: Minchan Kim <minchan@xxxxxxxxxx>
> Cc: Mel Gorman <mgorman@xxxxxxx>
> Cc: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
> Cc: Michal Nazarewicz <mina86@xxxxxxxxxx>
> Cc: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx>
> Cc: Christoph Lameter <cl@xxxxxxxxx>
> Cc: Rik van Riel <riel@xxxxxxxxxx>
> Cc: David Rientjes <rientjes@xxxxxxxxxx>
> <SNIP>
> @@ -2279,14 +2307,43 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
> */
> count_vm_event(COMPACTSTALL);
>
> - /* Page migration frees to the PCP lists but we want merging */
> - drain_pages(get_cpu());
> - put_cpu();
> + /* Did we capture a page? */
> + if (page) {
> + struct zone *zone;
> + unsigned long flags;
> + /*
> + * Mimic what buffered_rmqueue() does and capture_new_page()
> + * has not yet done.
> + */
> + zone = page_zone(page);
> +
> + local_irq_save(flags);
> + zone_statistics(preferred_zone, zone, gfp_mask);
> + local_irq_restore(flags);
>
> - page = get_page_from_freelist(gfp_mask, nodemask,
> - order, zonelist, high_zoneidx,
> - alloc_flags & ~ALLOC_NO_WATERMARKS,
> - preferred_zone, classzone_idx, migratetype);
> + VM_BUG_ON_PAGE(bad_range(zone, page), page);
> + if (!prep_new_page(page, order, gfp_mask))
> + /* This is normally done in get_page_from_freelist() */
> + page->pfmemalloc = !!(alloc_flags &
> + ALLOC_NO_WATERMARKS);
> + else
> + page = NULL;
> + }
> +
> + /* No capture but let's try allocating anyway */
> + if (!page) {
> + /*
> + * Page migration frees to the PCP lists but we want
> + * merging
> + */
> + drain_pages(get_cpu());
> + put_cpu();
> +
Would the attempted capture not drained already? Not a big deal so
Acked-by: Mel Gorman <mgorman@xxxxxxx>
--
Mel Gorman
SUSE Labs
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>