Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20220401181109.1477354-1-zi.yan@sent.com> <20220401181109.1477354-2-zi.yan@sent.com>
In-Reply-To: <20220401181109.1477354-2-zi.yan@sent.com>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Fri, 1 Apr 2022 11:29:17 -0700
Message-ID: <CAHk-=wifQimj2d6npq-wCi5onYPjzQg4vyO4tFcPJJZr268cRw@mail.gmail.com>
Subject: Re: [PATCH v2 2/2] mm: wrap __find_buddy_pfn() with a necessary buddy
 page validation.
To:     Zi Yan <ziy@nvidia.com>
Cc:     Linux-MM <linux-mm@kvack.org>,
        Steven Rostedt <rostedt@goodmis.org>,
        David Hildenbrand <david@redhat.com>,
        Vlastimil Babka <vbabka@suse.cz>,
        Mel Gorman <mgorman@techsingularity.net>,
        Mike Rapoport <rppt@kernel.org>,
        Oscar Salvador <osalvador@suse.de>,
        Andrew Morton <akpm@linux-foundation.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Fri, Apr 1, 2022 at 11:11 AM Zi Yan <zi.yan@sent.com> wrote:
>
> Whenever the buddy of a page is found from __find_buddy_pfn(),
> page_is_buddy() should be used to check its validity. Add a helper
> function find_buddy_page_pfn() to find the buddy page and do the check
> together.

Well, this certainly looks nicer, except now:

> +extern struct page *find_buddy_page_pfn(struct page *page, unsigned int order);

that 'pfn' no longer makes sense in the name, and

> @@ -1075,11 +1118,11 @@ static inline void __free_one_page(struct page *page,
>                                                                 migratetype);
>                         return;
>                 }
> -               buddy_pfn = __find_buddy_pfn(pfn, order);
> -               buddy = page + (buddy_pfn - pfn);
>
> -               if (!page_is_buddy(page, buddy, order))
> +               buddy = find_buddy_page_pfn(page, order);
> +               if (!buddy)
>                         goto done_merging;
> +               buddy_pfn = page_to_pfn(buddy);

This case now does two "page_to_pfn()" calls (one inside
find_buddy_page_pfn(), and one explicitly on the buddy).

And those page_to_pfn() things can actually be fairly expensive. It
*looks* like just a subtraction, but it's a pointer subtraction that
can end up generating a divide by a non-power-of-two size.

NORMALLY we try very hard to make 'sizeof struct page' be exactly 8
words, but I do not believe this is actually guaranteed.

And yeah, the divide-by-a-constant can be turned into a multiply, but
even that is not necessarily always cheap.

Now, two out of three use-cases didn't actually want the buddy_pfn(),
but this one use-case does look like it might be performance-critical
and a problem.

                 Linus