On Mon 26-09-16 16:47:57, Xishi Qiu wrote:

commit 97a16fc82a7c5b0cfce95c05dfb9561e306ca1b1
(mm, page_alloc: only enforce watermarks for order-0 allocations)
rewrite the high-order check in __zone_watermark_ok(), but I think it
quietly fix a bug. Please see the following.

Before this patch, the high-order check is this:
__zone_watermark_ok()
	...
	for (o = 0; o < order; o++) {
		/* At the next order, this order's pages become unavailable */
		free_pages -= z->free_area[o].nr_free << o;

		/* Require fewer higher order pages to be free */
		min >>= 1;

		if (free_pages <= min)
			return false;
	}
	...

If we have cma memory, and we alloc a high-order movable page, then it's right.

But if we alloc a high-order unmovable page(e.g. alloc kernel stack in dup_task_struct()),
and there are a lot of high-order cma pages, but little high-order unmovable
pages, the it is still return *true*, but we will alloc *failed* finally, because
we cannot fallback from migrate_unmovable to migrate_cma, right?

AFAIR CMA wmark check was always tricky and the above commit has made
the situation at least a bit more clear. Anyway IIRC 

#ifdef CONFIG_CMA
	/* If allocation can't use CMA areas don't use free CMA pages */
	if (!(alloc_flags & ALLOC_CMA))
		free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
#endif

	if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
		return false;

should reduce the prioblem because a lot of CMA pages should just get us
below the wmark + reserve boundary.
-- 
Michal Hocko
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>