Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Mon, 28 Oct 2019 15:42:22 +0300
From:   "Kirill A. Shutemov" <kirill@shutemov.name>
To:     Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc:     linux-mm@kvack.org, Andrew Morton <akpm@linux-foundation.org>,
        linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Alexander Viro <viro@zeniv.linux.org.uk>,
        Johannes Weiner <hannes@cmpxchg.org>,
        Steven Whitehouse <swhiteho@redhat.com>
Subject: Re: [PATCH] mm/filemap: do not allocate cache pages beyond end of
 file at read
Message-ID: <20191028124222.ld6u3dhhujfqcn7w@box>
References: <157225677483.3442.4227193290486305330.stgit@buzz>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <157225677483.3442.4227193290486305330.stgit@buzz>
User-Agent: NeoMutt/20180716
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, Oct 28, 2019 at 12:59:34PM +0300, Konstantin Khlebnikov wrote:
> Page cache could contain pages beyond end of file during write or
> if read races with truncate. But generic_file_buffered_read() always
> allocates unneeded pages beyond eof if somebody reads here and one
> extra page at the end if file size is page-aligned.
> 
> Function generic_file_buffered_read() calls page_cache_sync_readahead()
> if page not found in cache and then do another lookup. Readahead checks
> file size in __do_page_cache_readahead() before allocating pages.
> After that generic_file_buffered_read() falls back to slow path and
> allocates page for ->readpage() without checking file size.
> 
> This patch checks file size before allocating page for ->readpage().
> 
> Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
> ---
>  mm/filemap.c |    4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/mm/filemap.c b/mm/filemap.c
> index 85b7d087eb45..92abf5f348a9 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -2225,6 +2225,10 @@ static ssize_t generic_file_buffered_read(struct kiocb *iocb,
>  		goto out;
>  
>  no_cached_page:
> +		/* Do not allocate cache pages beyond end of file. */
> +		if (((loff_t)index << PAGE_SHIFT) >= i_size_read(inode))
> +			goto out;
> +
>  		/*
>  		 * Ok, it wasn't cached, so we need to create a new
>  		 * page..
> 
> 

CC Steven.

I've tried something of this sort back in 2013:

http://lore.kernel.org/r/1377099441-2224-1-git-send-email-kirill.shutemov@linux.intel.com

and I've got push back.

Apparently, some filesystems may not have valid i_size before >readpage().
Not sure if it's still the case...

Anyway I don't think it's valid reason for this inefficiency. These
filesystems have to have own implementation of >read_iter() to deal with
this.

-- 
 Kirill A. Shutemov