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;
MIME-Version: 1.0
References: <ae19f8ddc770135572323dd431d0efbe3e419582.camel@linux.ibm.com>
In-Reply-To: <ae19f8ddc770135572323dd431d0efbe3e419582.camel@linux.ibm.com>
From:   Miklos Szeredi <miklos@szeredi.hu>
Date:   Fri, 26 Jul 2019 10:52:31 +0200
Message-ID: <CAJfpegsLKY=M6PSBZjgpKkZTxUYBn+H44BxG2HVLsAVzTzyy_Q@mail.gmail.com>
Subject: Re: Question about vmsplice + SPLICE_F_GIFT
To:     Leonardo Bras <leonardo@linux.ibm.com>
Cc:     linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
        Alexander Viro <viro@zeniv.linux.org.uk>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue, Jul 23, 2019 at 10:33 PM Leonardo Bras <leonardo@linux.ibm.com> wrote:
>
> Hello everybody,
>
> I am not sure if this is the right place to be asking this. If is not,
> I apologize for the inconvenience. Also, please tell me where is a
> better way to as these questions.
>
> I am trying to create a basic C code to test vmsplice + SPLICE_F_GIFT
> for moving memory pages between two processes without copying.
>
> I have followed the man pages and several recipes across the web, but I
> could not reproduce it yet.
>
> Basically, I am doing:
> Sending process:
> - malloc + memcpy for generating pages to transfer
> - vmsplice with SPLICE_F_GIFT sending over named pipe (in a loop)
> Receiving process:
> - Create mmaped file to receive the pages
> - splice with SPLICE_F_MOVE receiving from named pipe (in a loop)

As the splice(2) man page says SPLICE_F_MOVE is currently a no-op.

> I have seen the SPLICE_F_MOVE being used on steal ops from the
> 'pipebuffer', but I couldn't find a way to call it from splice.
>
> Questions:
> It does what I think it does? (reassign memory pages from a process to
> another)

> If so, does page gifting still works?
> If so, is there a basic recipe to test it's workings?

What is the end goal?

It is easy to transfer pages using shared memory (see shm_open(3) and
related API), so why mess with splice?

Thanks,
Miklos