John Fastabend wrote:
> Edward Adam Davis wrote:
> > Syzbot constructed 32 scatterlists, and the data members in struct sk_msg_sg
> > can only store a maximum of MAX_MSG_FRAGS scatterlists.
> > However, the value of MAX_MSG_FRAGS=CONFIG_MAX_SKB_FRAG is less than 32, which
> > leads to the warning reported here.
> >
> > Prevent similar issues from occurring by checking whether sg.end is greater
> > than MAX_MSG_FRAGS.
> >
> > Reported-and-tested-by: syzbot+f2977222e0e95cec15c8@xxxxxxxxxxxxxxxxxxxxxxxxx
> > Signed-off-by: Edward Adam Davis <eadavis@xxxxxx>
> > ---
> > net/tls/tls_sw.c | 2 ++
> > 1 file changed, 2 insertions(+)
> >
> > diff --git a/net/tls/tls_sw.c b/net/tls/tls_sw.c
> > index e37b4d2e2acd..68dbe821f61d 100644
> > --- a/net/tls/tls_sw.c
> > +++ b/net/tls/tls_sw.c
> > @@ -1016,6 +1016,8 @@ static int tls_sw_sendmsg_locked(struct sock *sk, struct msghdr *msg,
> >
> > msg_pl = &rec->msg_plaintext;
> > msg_en = &rec->msg_encrypted;
> > + if (msg_pl->sg.end >= MAX_MSG_FRAGS)
> > + return -EINVAL;
> >
> > orig_size = msg_pl->sg.size;
> > full_record = false;
> > --
> > 2.43.0
> >
>
> I'll test this in a bit, but I suspect this error is because even
> if the msg_pl is full (the sg.end == MAX_MSG_FRAGS) the code is
> missing a full_record=true set to force the loop to do the send
> and abort. My opinion is we should never iterated the loop if the
> msg_pl was full.
>
> I think something like this is actually needed.
>
> diff --git a/net/tls/tls_sw.c b/net/tls/tls_sw.c
> index e37b4d2e2acd..9cfa6f8d51e3 100644
> --- a/net/tls/tls_sw.c
> +++ b/net/tls/tls_sw.c
> @@ -1052,8 +1052,10 @@ static int tls_sw_sendmsg_locked(struct sock *sk, struct msghdr *msg,
> if (ret < 0)
> goto send_end;
> tls_ctx->pending_open_record_frags = true;
> - if (full_record || eor || sk_msg_full(msg_pl))
> + if (full_record || eor || sk_msg_full(msg_pl)) {
> + full_record = true;
> goto copied;
> + }
> continue;
> }
Actually, it needs a bit more than above. That will fix the warning,
but it returns an error on when it should flush the full_record in
some cases. I'll send a fix shortly.
