MIME-Version: 1.0
References: <1677529301-19530-1-git-send-email-george.kennedy@oracle.com> <face2b1b-0f2e-4a79-a71b-79681fc56273@t-8ch.de>
In-Reply-To: <face2b1b-0f2e-4a79-a71b-79681fc56273@t-8ch.de>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Mon, 27 Feb 2023 17:58:12 -0800
Message-ID: <CAHk-=whxC0ob3f--B-igBp34ujA5PZQMcTnDQ8Zh8HA7+ZbeWg@mail.gmail.com>
Subject: Re: [PATCH v2] vc_screen: modify vcs_size() handling in vcs_read()
To:     linux@weissschuh.net
Cc:     George Kennedy <george.kennedy@oracle.com>, jirislaby@kernel.org,
        gregkh@linuxfoundation.org, sfr@canb.auug.org.au,
        akpm@linux-foundation.org, linux-kernel@vger.kernel.org,
        linux-serial@vger.kernel.org, regressions@lists.linux.dev
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Mon, Feb 27, 2023 at 5:46=E2=80=AFPM <linux@weissschuh.net> wrote:
>
> This still seems to be broken for me.

Looks that way.

> I still need the patch from
>
> https://lore.kernel.org/lkml/20230220064612.1783-1-linux@weissschuh.net/

.. and that has the same problem with "what if the error happens
during an iteration that wasn't the first, and we already succeeded
partially".

The "goto unlock_out" is bogus, since it jumps over all the "update
pos and check if we read something".

It was the correct thing to do *above* the loop, but not inside the loop.

IOW, I think the proper patch is to also turn the "goto unlock_out"
into a "break". Mind testing something like this (whitespace-damaged,
but you get the idea):

    --- a/drivers/tty/vt/vc_screen.c
    +++ b/drivers/tty/vt/vc_screen.c
    @@ -403,10 +403,11 @@ vcs_read(struct file *file, char __user
*buf, size_t count, loff_t *ppos)
                unsigned int this_round, skip =3D 0;
                int size;

    -           ret =3D -ENXIO;
                vc =3D vcs_vc(inode, &viewed);
    -           if (!vc)
    -                   goto unlock_out;
    +           if (!vc) {
    +                   ret =3D -ENXIO;
    +                   break;
    +           }

                /* Check whether we are above size each round,
                 * as copy_to_user at the end of this loop

which hopefully really fixes this (at least I don't see any other
"goto unlock_out" cases).

              Linus