From mboxrd@z Thu Jan  1 00:00:00 1970
From: Jeff Garzik <jeff@garzik.org>
Subject: Re: [PATCH 1/3 v2] chunkd: remove sendfile(2) zero-copy support
Date: Sun, 18 Jul 2010 19:10:09 -0400
Message-ID: <4C4389D1.6000201@garzik.org>
References: <20100717054616.GA9993@havoc.gtf.org> <20100717054651.GA10037@havoc.gtf.org> <4C4278EA.8000605@redhat.com>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit
Return-path: <hail-devel-owner@vger.kernel.org>
DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed;
        d=gmail.com; s=gamma;
        h=domainkey-signature:received:received:sender:message-id:date:from
         :user-agent:mime-version:to:cc:subject:references:in-reply-to
         :content-type:content-transfer-encoding;
        bh=Ejd6MR4wYHvbLjx34O9gOt98KrRc3yqBDxX/4mltmgA=;
        b=PkQVfozuhD52ZKrEUH2jbn9mFGZVbUvpQM7xroDuYAFWNVShUer8Ctpeg/9h4YnSla
         pRiaijr3wFhwwhR6WWWLxUTt71s9V/iOTwqAv5yIJFcVGsiYBEkGxq0Jn737ywjr0OB4
         RqTdr9j1GJHY4iqmFBepSI6IfE1ftz+VoYek0=
In-Reply-To: <4C4278EA.8000605@redhat.com>
Sender: hail-devel-owner@vger.kernel.org
List-ID: <hail-devel.vger.kernel.org>
Content-Type: text/plain; charset="us-ascii"; format="flowed"
To: Steven Dake <sdake@redhat.com>
Cc: hail-devel@vger.kernel.org

On 07/17/2010 11:45 PM, Steven Dake wrote:
> On 07/16/2010 10:46 PM, Jeff Garzik wrote:
>> chunkd: remove sendfile(2) zero-copy support
>>
>> chunkd will be soon checksumming data in main memory. That removes
>> the utility of a zero-copy interface which bypasses the on-heap
>> data requirement.
>>
>> Signed-off-by: Jeff Garzik<jgarzik@redhat.com>

> May be able to use vmsplice with sendfile (if linux is only target
> platform). Haven't tried it myself, but the operations look interesting
> at achieving zero copy with sockets from memory addresses.

Even though the man pages say "only for pipes", this syscall definitely 
works with TCP.  The big question:  is it actually faster than 
read()+write() ?

Years ago, I experimented with using some fancy new Linux-specific 
syscalls in a from-scratch implementation of cp(1).  It turned out that 
read()+write() was faster than other methods.

That was file->file copying.  It's probably worth investigating 
vmsplice() for our file->checksum->TCP case, definitely.

	Jeff