Mel Gorman <mgorman@xxxxxxx> wrote:
> page->pfmemalloc can be left set for captured pages so try this but as
> capture is rarely used I'm strongly favouring a partial revert even if
> this works for you. I haven't reproduced this using your workload yet
> but I have found that high-order allocation stress tests for 3.8-rc2 are
> completely screwed. 71% success rates at rest in 3.7 and 6% in 3.8-rc2 so
> I have to chase that down too.
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 9d20c13..c242d21 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -2180,8 +2180,10 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
> current->flags &= ~PF_MEMALLOC;
>
> /* If compaction captured a page, prep and use it */
> - if (page && !prep_new_page(page, order, gfp_mask))
> + if (page && !prep_new_page(page, order, gfp_mask)) {
> + page->pfmemalloc = false;
> goto got_page;
> + }
>
> if (*did_some_progress != COMPACT_SKIPPED) {
> /* Page migration frees to the PCP lists but we want merging */
This (on top of your previous patch) seems to work great after several
hours of testing on both my VM and real machine. I haven't tried your
partial revert, yet. Will try that in a bit on the VM.
--
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>