Re: [PATCH v4] nvme-tcp: fix connect failure on receiving partial ICResp PDU

[Sagi Grimberg]

On 29/01/2025 0:02, Caleb Sander wrote:
> On Tue, Jan 28, 2025 at 12:28 AM Hannes Reinecke <hare@xxxxxxx> wrote:
> > On 1/27/25 18:38, Caleb Sander wrote:
> > > On Sun, Jan 26, 2025 at 11:37 PM Hannes Reinecke <hare@xxxxxxx> wrote:
> > > > On 1/24/25 19:43, Caleb Sander Mateos wrote:
> > > > > nvme_tcp_init_connection() attempts to receive an ICResp PDU but only
> > > > > checks that the return value from recvmsg() is non-negative. If the
> > > > > sender closes the TCP connection or sends fewer than 128 bytes, this
> > > > > check will pass even though the full PDU wasn't received.
> > > > >
> > > > > Ensure the full ICResp PDU is received by checking that recvmsg()
> > > > > returns the expected 128 bytes.
> > > > >
> > > > > Additionally set the MSG_WAITALL flag for recvmsg(), as a sender could
> > > > > split the ICResp over multiple TCP frames. Without MSG_WAITALL,
> > > > > recvmsg() could return prematurely with only part of the PDU.
> > > > >
> > > > > Signed-off-by: Caleb Sander Mateos <csander@xxxxxxxxxxxxxxx>
> > > > > Fixes: 3f2304f8c6d6 ("nvme-tcp: add NVMe over TCP host driver")
> > > > > ---
> > > > > v4: keep recvmsg() error return value
> > > > > v3: fix return value to indicate error
> > > > > v2: add Fixes tag
> > > > >
> > > > >     drivers/nvme/host/tcp.c | 5 ++++-
> > > > >     1 file changed, 4 insertions(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
> > > > > index e9ff6babc540..56679eb8c0d6 100644
> > > > > --- a/drivers/nvme/host/tcp.c
> > > > > +++ b/drivers/nvme/host/tcp.c
> > > > > @@ -1446,15 +1446,18 @@ static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
> > > > >         iov.iov_len = sizeof(*icresp);
> > > > >         if (nvme_tcp_queue_tls(queue)) {
> > > > >                 msg.msg_control = cbuf;
> > > > >                 msg.msg_controllen = sizeof(cbuf);
> > > > >         }
> > > > > +     msg.msg_flags = MSG_WAITALL;
> > > > >         ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
> > > > >                         iov.iov_len, msg.msg_flags);
> > > > But won't we have to wait for a TCP timeout now if the sender sends less
> > > > than 128 bytes? With this patch we always wait for 128 bytes, and
> > > > possibly wait for TCP timeout if not.
> > > Yes, if the NVMe/TCP controller sends less than 128 bytes, we need to
> > > wait for it to send the remainder of the ICResp PDU. That's just how
> > > the NVMe/TCP protocol works. If we want to protect against
> > > buggy/malicious controllers that don't send a full ICResp, we need a
> > > timeout mechanism. That's the purpose of the existing
> > > `queue->sock->sk->sk_rcvtimeo = 10 * HZ;` in nvme_tcp_alloc_queue().
> > > Note that recvmsg() timing out was already possible in the original
> > > code if the controller didn't send anything on the TCP connection
> > > after accepting it.
> > Hmm. With checking the code 'rcvtimeo' is only evaluated if MSG_WAITALL
> > is _not_ set. Makes me wonder why we do set it...
> > But that's beside the point.
> I am not seeing where sk_rcvtimeo is ignored when MSG_WAITALL is used,
> can you point me to the code?
> It is certainly ignored for *non-blocking* recvmsg() (MSG_DONTWAIT),
> but I don't see why it would be ignored for MSG_WAITALL. man 7 socket
> also suggests it applies to all blocking socket I/O.
>
> static inline long sock_rcvtimeo(const struct sock *sk, bool noblock)
> {
>          return noblock ? 0 : sk->sk_rcvtimeo;
> }
>
> In tcp_recvmsg_locked():
> timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
>
> > > > Testcase for this would be nice ...
> > > >
> > > > And I need to check if secure concatenation is affected here; with
> > > > secure concatenation we need to peek at the first packet to check
> > > > if it's an ICRESP or a TLS negotiation.
> > > Are you saying that with secure concatenation we don't know in advance
> > > whether the connection is using TLS between the TCP and NVMe/TCP
> > > protocol layers? Wouldn't the host already need to know that when it
> > > sent its ICReq PDU?
> > No, the host doesn't need to know. TLS is enabled by the lower
> > layers.
> >
> > But upon further checking, I guess it'll be okay with secure
> > concatenation.
> >
> > Nevertheless, I would vastly prefer to have a receive loop here
> > instead of waiting to receive the full amount as per MSG_WAITALL.
> > The entire tcp code is using nonblocking calls, so I'd rather
> > keep it that way and implement a receive loop here.
> The driver uses non-blocking socket I/O in the I/O path, but not the
> connect path. nvme_tcp_init_connection() is already using blocking
> socket I/O to send the ICReq PDU and receive the ICResp. I suppose we
> could change the connect path to register poll interest and use
> non-blocking operations, but that seems like a more involved code
> change and orthogonal to the issue of receiving the full ICResp.

I agree. the driver blocks on ICResp rcvmsg, there is no reason to
use async interface.

Caleb, can you please make sure to test this patch with TLS?
Do you have a reliable way to reproduce this?