MIME-Version: 1.0
References: <20230210061953.GC2825702@dread.disaster.area> <Y+oCBnz2nLtXrz7O@gondor.apana.org.au>
In-Reply-To: <Y+oCBnz2nLtXrz7O@gondor.apana.org.au>
From:   Andy Lutomirski <luto@kernel.org>
Date:   Mon, 13 Feb 2023 10:01:27 -0800
Message-ID: <CALCETrXKkZw3ojpmTftur1_-dEi6BOo9Q0cems_jgabntNFYig@mail.gmail.com>
Subject: Re: copy on write for splice() from file to pipe?
To:     Herbert Xu <herbert@gondor.apana.org.au>
Cc:     Dave Chinner <david@fromorbit.com>, torvalds@linux-foundation.org,
        metze@samba.org, axboe@kernel.dk, linux-fsdevel@vger.kernel.org,
        linux-api@vger.kernel.org, io-uring@vger.kernel.org,
        linux-kernel@vger.kernel.org, viro@zeniv.linux.org.uk,
        samba-technical@lists.samba.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Feb 13, 2023 at 1:45 AM Herbert Xu <herbert@gondor.apana.org.au> wrote:
>
> Dave Chinner <david@fromorbit.com> wrote:
> >
> > IOWs, the application does not care if the data changes whilst they
> > are in transport attached to the pipe - it only cares that the
> > contents are stable once they have been delivered and are now wholly
> > owned by the network stack IO path so that the OTW encodings
> > (checksum, encryption, whatever) done within the network IO path
> > don't get compromised.
>
> Is this even a real problem? The network stack doesn't care at
> all if you modify the pages while it's being processed.  All the
> things you've mentioned (checksum, encryption, etc.) will be
> self-consistent on the wire.
>
> Even when actual hardware offload is involved it's hard to see how
> things could possibly go wrong unless the hardware was going out of
> its way to do the wrong thing by fetching from memory twice.
>

There's a difference between "kernel speaks TCP (or whatever)
correctly" and "kernel does what the application needs it to do".
When I write programs that send data on the network, I want the kernel
to send the data that I asked it to send.  As a silly but obvious
example, suppose I have two threads, and all I/O is blocking
(O_NONBLOCK is not set, etc):

char buffer[1024] = "A";

Thread A:
send(fd, buffer, 1, 0);

Thread B:
mb();
buffer[0] = 'B';
mb();


Obviously, there are three possible valid outcomes: Thread A can go
first (send returns before B changes the buffer), and 'A' gets sent.
Thread B can go first (the buffer is changed before send() starts),
and 'B' gets sent.  Or both can run concurrently, in which case the
data sent is indeterminate.

But it is not valid for send() to return, then the buffer to change,
and 'B' to get sent, just like:

char foo[] = "A";
send(fd, foo, 1, 0);
foo[0] = 'B';

must send 'A', not 'B'.

The trouble with splice() is that there is no clear point at which the
splice is complete and the data being sent is committed.  I don't
think user applications need the data committed particularly quickly,
but I do think it needs to be committed "eventually* and there needs
to be a point at which the application knows it's been committed.
Right now, if a user program does:

Write 'A' to a file
splice that file to a pipe
splice that pipe to a socket
... wait until when? ...
Write 'B' to a file

There is nothing the user program can wait for to make sure that 'A'
gets sent, but saying that the kernel speaks TCP correctly without
solving this problem doesn't actually solve the problem.