Re: [PATCH 01/16] mm: delayed page activation

[Wu Fengguang]

On Sun, Dec 04, 2005 at 03:11:28PM +0300, Nikita Danilov wrote:
> Wu Fengguang writes:
>  > When a page is referenced the second time in inactive_list, mark it with
>  > PG_activate instead of moving it into active_list immediately. The actual
>  > moving work is delayed to vmscan time.
>  > 
>  > This implies two essential changes:
>  > - keeps the adjecency of pages in lru;
> 
> But this change destroys LRU ordering: at the time when shrink_list()
> inspects PG_activate bit, information about order in which
> mark_page_accessed() was called against pages is lost. E.g., suppose

Thanks.
But this order of re-access time may be pointless. In fact the original
mark_page_accessed() is doing another inversion: inversion of page lifetime.
In the word of CLOCK-Pro, a page first being re-accessed has lower
inter-reference distance, and therefore should be better protected(if ignore
possible read-ahead effects). If we move re-accessed pages immediately into
active_list, we are pushing them closer to danger of eviction.

btw, the current vmscan code clears PG_referenced flag when moving pages to
active_list. I followed the convention by doing this in the patch:

--- linux-2.6.15-rc2-mm1.orig/mm/vmscan.c
+++ linux-2.6.15-rc2-mm1/mm/vmscan.c
@@ -454,6 +454,12 @@ static int shrink_list(struct list_head
                if (PageWriteback(page))
                        goto keep_locked;

+               if (PageActivate(page)) {
+                       ClearPageActivate(page);
+                       ClearPageReferenced(page);
+                       goto activate_locked;
+               }
+
                referenced = page_referenced(page, 1, sc->priority <= 0);
                /* In active use or really unfreeable?  Activate it. */
                if (referenced && page_mapping_inuse(page))

Though I have a strong feeling that with the extra PG_activate bit, the
+                       ClearPageReferenced(page);
line should be removed. That is, let the extra reference record live through it.
The point is to smooth out the inter-reference distance. Imagine the following
situation:

-      +            -   +           +   -                   -   +              -
1                   2                   3                   4                  5
        +: reference time
        -: shrink_list time

One page have an average inter-reference distance that is smaller than the
inter-scan distance. But the distances vary a bit. Here we'd better let the
reference count accumulate, or at the 3rd shrink_list time it will be evicted.
Though it has a side effect of favoriting non-mmaped file a bit more than
before, and I was not quite sure about it.

> inactive list initially contained pages
> 
>      /* head */ (P1, P2, P3) /* tail */
> 
> all of them referenced. Then mark_page_accessed(), is called against P1,
> P2, and P3 (in that order). With the old code active list would end up 
> 
>      /* head */ (P3, P2, P1) /* tail */
> 
> which corresponds to LRU. With delayed page activation, pages are moved
> to head of the active list in the order they are analyzed by
> shrink_list(), which gives
> 
>      /* head */ (P1, P2, P3) /* tail */
> 
> on the active list, that is _inverse_ LRU order.

Thanks,
Wu
-
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/