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;
Date:   Sun, 16 Apr 2023 12:44:53 -0700 (PDT)
From:   Hugh Dickins <hughd@google.com>
To:     Zi Yan <ziy@nvidia.com>
cc:     "Matthew Wilcox (Oracle)" <willy@infradead.org>,
        Yang Shi <shy828301@gmail.com>, Yu Zhao <yuzhao@google.com>,
        linux-mm@kvack.org,
        "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>,
        Ryan Roberts <ryan.roberts@arm.com>,
        =?ISO-8859-15?Q?Michal_Koutn=FD?= <mkoutny@suse.com>,
        Roman Gushchin <roman.gushchin@linux.dev>,
        Zach O'Keefe <zokeefe@google.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        linux-kernel@vger.kernel.org, cgroups@vger.kernel.org,
        linux-fsdevel@vger.kernel.org, linux-kselftest@vger.kernel.org
Subject: Re: [PATCH v3 6/7] mm: truncate: split huge page cache page to a
 non-zero order if possible.
In-Reply-To: <20230403201839.4097845-7-zi.yan@sent.com>
Message-ID: <9dd96da-efa2-5123-20d4-4992136ef3ad@google.com>
References: <20230403201839.4097845-1-zi.yan@sent.com> <20230403201839.4097845-7-zi.yan@sent.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Precedence: bulk

On Mon, 3 Apr 2023, Zi Yan wrote:

> From: Zi Yan <ziy@nvidia.com>
> 
> To minimize the number of pages after a huge page truncation, we do not
> need to split it all the way down to order-0. The huge page has at most
> three parts, the part before offset, the part to be truncated, the part
> remaining at the end. Find the greatest common divisor of them to
> calculate the new page order from it, so we can split the huge
> page to this order and keep the remaining pages as large and as few as
> possible.
> 
> Signed-off-by: Zi Yan <ziy@nvidia.com>
> ---
>  mm/truncate.c | 21 +++++++++++++++++++--
>  1 file changed, 19 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/truncate.c b/mm/truncate.c
> index 86de31ed4d32..817efd5e94b4 100644
> --- a/mm/truncate.c
> +++ b/mm/truncate.c
> @@ -22,6 +22,7 @@
>  #include <linux/buffer_head.h>	/* grr. try_to_release_page */
>  #include <linux/shmem_fs.h>
>  #include <linux/rmap.h>
> +#include <linux/gcd.h>

Really?

>  #include "internal.h"
>  
>  /*
> @@ -211,7 +212,8 @@ int truncate_inode_folio(struct address_space *mapping, struct folio *folio)
>  bool truncate_inode_partial_folio(struct folio *folio, loff_t start, loff_t end)
>  {
>  	loff_t pos = folio_pos(folio);
> -	unsigned int offset, length;
> +	unsigned int offset, length, remaining;
> +	unsigned int new_order = folio_order(folio);
>  
>  	if (pos < start)
>  		offset = start - pos;
> @@ -222,6 +224,7 @@ bool truncate_inode_partial_folio(struct folio *folio, loff_t start, loff_t end)
>  		length = length - offset;
>  	else
>  		length = end + 1 - pos - offset;
> +	remaining = folio_size(folio) - offset - length;
>  
>  	folio_wait_writeback(folio);
>  	if (length == folio_size(folio)) {
> @@ -236,11 +239,25 @@ bool truncate_inode_partial_folio(struct folio *folio, loff_t start, loff_t end)
>  	 */
>  	folio_zero_range(folio, offset, length);
>  
> +	/*
> +	 * Use the greatest common divisor of offset, length, and remaining
> +	 * as the smallest page size and compute the new order from it. So we
> +	 * can truncate a subpage as large as possible. Round up gcd to
> +	 * PAGE_SIZE, otherwise ilog2 can give -1 when gcd/PAGE_SIZE is 0.
> +	 */
> +	new_order = ilog2(round_up(gcd(gcd(offset, length), remaining),
> +				   PAGE_SIZE) / PAGE_SIZE);

Gosh.  In mm/readahead.c I can see "order = __ffs(index)",
and I think something along those lines would be more appropriate here.

But, if there's any value at all to choosing intermediate orders here in
truncation, I don't think choosing a single order is the right approach -
more easily implemented, yes, but is it worth doing?

What you'd actually want (if anything) is to choose the largest orders
possible, with smaller and smaller orders filling in the rest (I expect
there's a technical name for this, but I don't remember - bin packing
is something else, I think).

As this code stands, truncate a 2M huge page at 1M and you get two 1M
pieces (one then discarded) - nice; but truncate it at 1M+1 and you get
lots of order 2 (forced up from 1) pieces.  Seems weird, and not worth
the effort.

Hugh

> +
> +	/* order-1 THP not supported, downgrade to order-0 */
> +	if (new_order == 1)
> +		new_order = 0;
> +
> +
>  	if (folio_has_private(folio))
>  		folio_invalidate(folio, offset, length);
>  	if (!folio_test_large(folio))
>  		return true;
> -	if (split_folio(folio) == 0)
> +	if (split_huge_page_to_list_to_order(&folio->page, NULL, new_order) == 0)
>  		return true;
>  	if (folio_test_dirty(folio))
>  		return false;
> -- 
> 2.39.2