Re: [patch v2 4/4] mm, thp: do not perform sync compaction on pagefault

[Mel Gorman]

On Thu, May 01, 2014 at 02:35:48PM -0700, David Rientjes wrote:
> Synchronous memory compaction can be very expensive: it can iterate an enormous 
> amount of memory without aborting, constantly rescheduling, waiting on page
> locks and lru_lock, etc, if a pageblock cannot be defragmented.
> 
> Unfortunately, it's too expensive for pagefault for transparent hugepages and 
> it's much better to simply fallback to pages.  On 128GB machines, we find that 
> synchronous memory compaction can take O(seconds) for a single thp fault.
> 
> Now that async compaction remembers where it left off without strictly relying
> on sync compaction, this makes thp allocations best-effort without causing
> egregious latency during pagefault.
> 
> Signed-off-by: David Rientjes <rientjes@xxxxxxxxxx>

Compaction uses MIGRATE_SYNC_LIGHT which in the current implementation
avoids calling ->writepage to clean dirty page in the fallback migrate
case

static int fallback_migrate_page(struct address_space *mapping,
        struct page *newpage, struct page *page, enum migrate_mode mode)
{
        if (PageDirty(page)) {
                /* Only writeback pages in full synchronous migration */
                if (mode != MIGRATE_SYNC)
                        return -EBUSY;
                return writeout(mapping, page);
        }

or waiting on page writeback in other cases

                /*
                 * Only in the case of a full synchronous migration is it
                 * necessary to wait for PageWriteback. In the async case,
                 * the retry loop is too short and in the sync-light case,
                 * the overhead of stalling is too much
                 */
                if (mode != MIGRATE_SYNC) {
                        rc = -EBUSY;
                        goto uncharge;
                }

or on acquiring the page lock for unforced migrations.

However, buffers still get locked in the SYNC_LIGHT case causing stalls
in buffer_migrate_lock_buffers which may be undesirable or maybe you are
hitting some other case. It would be preferable to identify what is
getting stalled in SYNC_LIGHT compaction and fix that rather than
disabling it entirely. You may also want to distinguish between a direct
compaction by a process and collapsing huge pages as done by khugepaged.

-- 
Mel Gorman
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/