Re: [PATCHv6 32/36] thp: reintroduce split_huge_page()

[Vlastimil Babka]

On 06/03/2015 07:06 PM, Kirill A. Shutemov wrote:
> This patch adds implementation of split_huge_page() for new
> refcountings.
>
> Unlike previous implementation, new split_huge_page() can fail if
> somebody holds GUP pin on the page. It also means that pin on page
> would prevent it from bening split under you. It makes situation in
> many places much cleaner.
>
> The basic scheme of split_huge_page():
>
>    - Check that sum of mapcounts of all subpage is equal to page_count()
>      plus one (caller pin). Foll off with -EBUSY. This way we can avoid
>      useless PMD-splits.
>
>    - Freeze the page counters by splitting all PMD and setup migration
>      PTEs.
>
>    - Re-check sum of mapcounts against page_count(). Page's counts are
>      stable now. -EBUSY if page is pinned.
>
>    - Split compound page.
>
>    - Unfreeze the page by removing migration entries.
>
> Signed-off-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> Tested-by: Sasha Levin <sasha.levin@xxxxxxxxxx>

[...]

> +
> +static int __split_huge_page_tail(struct page *head, int tail,
> +		struct lruvec *lruvec, struct list_head *list)
> +{
> +	int mapcount;
> +	struct page *page_tail = head + tail;
> +
> +	mapcount = page_mapcount(page_tail);

Isn't page_mapcount() unnecessarily heavyweight here? When you are 
splitting a page, it already should have zero compound_mapcount() and 
shouldn't be PageDoubleMap(), no? So you should care about 
page->_mapcount only? Sure, splitting THP is not a hotpath, but when 
done 512 times per split, it could make some difference in the split's 
latency.

> +	VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
> +
> +	/*
> +	 * tail_page->_count is zero and not changing from under us. But
> +	 * get_page_unless_zero() may be running from under us on the
> +	 * tail_page. If we used atomic_set() below instead of atomic_add(), we
> +	 * would then run atomic_set() concurrently with
> +	 * get_page_unless_zero(), and atomic_set() is implemented in C not
> +	 * using locked ops. spin_unlock on x86 sometime uses locked ops
> +	 * because of PPro errata 66, 92, so unless somebody can guarantee
> +	 * atomic_set() here would be safe on all archs (and not only on x86),
> +	 * it's safer to use atomic_add().

I would be surprised if this was the first place to use atomic_set() 
with potential concurrent atomic_add(). Shouldn't atomic_*() API 
guarantee that this works?

> +	 */
> +	atomic_add(page_mapcount(page_tail) + 1, &page_tail->_count);

You already have the value in mapcount variable, so why read it again.


--
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/