Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 7 Oct 2021 21:55:21 +0100
From:   Matthew Wilcox <willy@infradead.org>
To:     Andrew Morton <akpm@linux-foundation.org>
Cc:     Mel Gorman <mgorman@techsingularity.net>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH v2] mm: Optimise put_pages_list()
Message-ID: <YV9eueky+lBfSWA3@casper.infradead.org>
References: <20211007192138.561673-1-willy@infradead.org>
 <20211007123109.6a49c7c625e414acf7546c89@linux-foundation.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20211007123109.6a49c7c625e414acf7546c89@linux-foundation.org>
Precedence: bulk

On Thu, Oct 07, 2021 at 12:31:09PM -0700, Andrew Morton wrote:
> On Thu,  7 Oct 2021 20:21:37 +0100 "Matthew Wilcox (Oracle)" <willy@infradead.org> wrote:
> 
> > Instead of calling put_page() one page at a time, pop pages off
> > the list if their refcount was too high and pass the remainder to
> > put_unref_page_list().  This should be a speed improvement, but I have
> > no measurements to support that.  Current callers do not care about
> > performance, but I hope to add some which do.
> 
> Don't you think it would actually be slower to take an additional pass
> across the list?  If the list is long enough to cause cache thrashing. 
> Maybe it's faster for small lists.

My first response is an appeal to authority -- release_pages() does
this same thing.  Only it takes an array, constructs a list and passes
that to put_unref_page_list().  So if that's slower (and lists _are_
slower than arrays), we should have a put_unref_page_array().

Second, we can follow through the code paths and reason about it.

Before:

while (!list_empty(pages)) {
	put_page(victim);
		page = compound_head(page);
		if (put_page_testzero(page))
			__put_page(page);
				__put_single_page(page)
					__page_cache_release(page);
					mem_cgroup_uncharge(page);
					<---
free_unref_page(page, 0);
	free_unref_page_prepare()
        local_lock_irqsave(&pagesets.lock, flags);
        free_unref_page_commit(page, pfn, migratetype, order);
        local_unlock_irqrestore(&pagesets.lock, flags);

After:

free_unref_page_list(pages);
        list_for_each_entry_safe(page, next, list, lru) {
                if (!free_unref_page_prepare(page, pfn, 0)) {
        }

        local_lock_irqsave(&pagesets.lock, flags);
        list_for_each_entry_safe(page, next, list, lru) {
		free_unref_page_commit()
	}
        local_unlock_irqrestore(&pagesets.lock, flags);

So the major win here is that we disable/enable interrupts once per
batch rather than once per page.