Re: [PATCH RFT 4/5] tcp, tap: Don't increase tap-side sequence counter for dropped frames

[Stefano Brivio <sbrivio@redhat.com>]

On Thu, 28 Sep 2023 11:58:45 +1000
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Wed, Sep 27, 2023 at 07:06:03PM +0200, Stefano Brivio wrote:
> > On Mon, 25 Sep 2023 14:47:52 +1000
> > David Gibson <david@gibson.dropbear.id.au> wrote:
> >   
> > > On Sat, Sep 23, 2023 at 12:06:09AM +0200, Stefano Brivio wrote:  
> > > > ...so that we'll retry sending them, instead of more-or-less silently
> > > > dropping them. This happens quite frequently if our sending buffer on
> > > > the UNIX domain socket is heavily constrained (for instance, by the
> > > > 208 KiB default memory limit).
> > > > 
> > > > It might be argued that dropping frames is part of the expected TCP
> > > > flow: we don't dequeue those from the socket anyway, so we'll
> > > > eventually retransmit them.
> > > > 
> > > > But we don't need the receiver to tell us (by the way of duplicate or
> > > > missing ACKs) that we couldn't send them: we already know as
> > > > sendmsg() reports that. This seems to considerably increase
> > > > throughput stability and throughput itself for TCP connections with
> > > > default wmem_max values.
> > > > 
> > > > Unfortunately, the 16 bits left as padding in the frame descriptors    
> > > 
> > > I assume you're referring to the 'pad' fields in tcp[46]_l2_buf_t,
> > > yes?  
> > 
> > Right, that.
> >   
> > > For AVX2 we have substantially more space here.  Couldn't we put
> > > a conn (or seq) pointer in here at the cost of a few bytes MSS for
> > > non-AVX2 and zero cost for AVX2 (which is probably the majority case)?  
> > 
> > Yes, true. On the other hand, having this parallel array only affects
> > readability I guess, whereas inserting pointers and lengths in
> > tcp[46]_l2_buf_t actually decreases the usable MSS (not just on
> > non-AVX2 x86, but also on other architectures). So I'd rather stick to
> > this.  
> 
> Yeah, I guess so.
> 
> Actually.. I did just think of one other option.  It avoids both any
> extra padding and a parallel array, but at the cost of additional work
> when frames are dropped.  We could use that 16-bits of padding to
> store the TCP payload length.  Then when we don't manage to send all
> our frames, we do another loop through and add up how many stream
> bytes we actually sent to update the seq pointer.

Hmm, yes. It's slightly more memory efficient, but the complexity seems
a bit overkill to me.

> > > > we use internally aren't enough to uniquely identify for which
> > > > connection we should update sequence numbers: create a parallel
> > > > array of pointers to sequence numbers and L4 lengths, of
> > > > TCP_FRAMES_MEM size, and go through it after calling sendmsg().
> > > > 
> > > > Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
> > > > ---
> > > >  tap.c | 10 +++++++---
> > > >  tap.h |  2 +-
> > > >  tcp.c | 43 ++++++++++++++++++++++++++++++++++++-------
> > > >  3 files changed, 44 insertions(+), 11 deletions(-)
> > > > 
> > > > diff --git a/tap.c b/tap.c
> > > > index 93db989..b30ff81 100644
> > > > --- a/tap.c
> > > > +++ b/tap.c
> > > > @@ -413,13 +413,15 @@ static size_t tap_send_frames_passt(const struct ctx *c,
> > > >   * @c:		Execution context
> > > >   * @iov:	Array of buffers, each containing one frame (with L2 headers)
> > > >   * @n:		Number of buffers/frames in @iov
> > > > + *
> > > > + * Return: number of frames actually sent
> > > >   */
> > > > -void tap_send_frames(struct ctx *c, const struct iovec *iov, size_t n)
> > > > +size_t tap_send_frames(struct ctx *c, const struct iovec *iov, size_t n)
> > > >  {
> > > >  	size_t m;
> > > >  
> > > >  	if (!n)
> > > > -		return;
> > > > +		return 0;
> > > >  
> > > >  	if (c->mode == MODE_PASST)
> > > >  		m = tap_send_frames_passt(c, iov, n);
> > > > @@ -427,9 +429,11 @@ void tap_send_frames(struct ctx *c, const struct iovec *iov, size_t n)
> > > >  		m = tap_send_frames_pasta(c, iov, n);
> > > >  
> > > >  	if (m < n)
> > > > -		debug("tap: dropped %lu frames of %lu due to short send", n - m, n);
> > > > +		debug("tap: failed to send %lu frames of %lu", n - m, n);
> > > >  
> > > >  	pcap_multiple(iov, m, c->mode == MODE_PASST ? sizeof(uint32_t) : 0);
> > > > +
> > > > +	return m;
> > > >  }
> > > >  
> > > >  /**
> > > > diff --git a/tap.h b/tap.h
> > > > index 021fb7c..952fafc 100644
> > > > --- a/tap.h
> > > > +++ b/tap.h
> > > > @@ -73,7 +73,7 @@ void tap_icmp6_send(const struct ctx *c,
> > > >  		    const struct in6_addr *src, const struct in6_addr *dst,
> > > >  		    void *in, size_t len);
> > > >  int tap_send(const struct ctx *c, const void *data, size_t len);
> > > > -void tap_send_frames(struct ctx *c, const struct iovec *iov, size_t n);
> > > > +size_t tap_send_frames(struct ctx *c, const struct iovec *iov, size_t n);
> > > >  void tap_update_mac(struct tap_hdr *taph,
> > > >  		    const unsigned char *eth_d, const unsigned char *eth_s);
> > > >  void tap_listen_handler(struct ctx *c, uint32_t events);
> > > > diff --git a/tcp.c b/tcp.c
> > > > index 4606f17..76b7b8d 100644
> > > > --- a/tcp.c
> > > > +++ b/tcp.c
> > > > @@ -434,6 +434,16 @@ static int tcp_sock_ns		[NUM_PORTS][IP_VERSIONS];
> > > >   */
> > > >  static union inany_addr low_rtt_dst[LOW_RTT_TABLE_SIZE];
> > > >  
> > > > +/**
> > > > + * tcp_buf_seq_update - Sequences to update with length of frames once sent
> > > > + * @seq:	Pointer to sequence number sent to tap-side, to be updated
> > > > + * @len:	TCP payload length
> > > > + */
> > > > +struct tcp_buf_seq_update {
> > > > +	uint32_t *seq;
> > > > +	uint16_t len;
> > > > +};
> > > > +
> > > >  /* Static buffers */
> > > >  
> > > >  /**
> > > > @@ -462,6 +472,8 @@ static struct tcp4_l2_buf_t {
> > > >  #endif
> > > >  tcp4_l2_buf[TCP_FRAMES_MEM];
> > > >  
> > > > +static struct tcp_buf_seq_update tcp4_l2_buf_seq_update[TCP_FRAMES_MEM];
> > > > +
> > > >  static unsigned int tcp4_l2_buf_used;
> > > >  
> > > >  /**
> > > > @@ -490,6 +502,8 @@ struct tcp6_l2_buf_t {
> > > >  #endif
> > > >  tcp6_l2_buf[TCP_FRAMES_MEM];
> > > >  
> > > > +static struct tcp_buf_seq_update tcp6_l2_buf_seq_update[TCP_FRAMES_MEM];
> > > > +
> > > >  static unsigned int tcp6_l2_buf_used;
> > > >  
> > > >  /* recvmsg()/sendmsg() data for tap */
> > > > @@ -1369,10 +1383,17 @@ static void tcp_l2_flags_buf_flush(struct ctx *c)
> > > >   */
> > > >  static void tcp_l2_data_buf_flush(struct ctx *c)
> > > >  {
> > > > -	tap_send_frames(c, tcp6_l2_iov, tcp6_l2_buf_used);
> > > > +	unsigned i;
> > > > +	size_t m;
> > > > +
> > > > +	m = tap_send_frames(c, tcp6_l2_iov, tcp6_l2_buf_used);
> > > > +	for (i = 0; i < m; i++)
> > > > +		*tcp6_l2_buf_seq_update[i].seq += tcp6_l2_buf_seq_update[i].len;
> > > >  	tcp6_l2_buf_used = 0;
> > > >  
> > > > -	tap_send_frames(c, tcp4_l2_iov, tcp4_l2_buf_used);
> > > > +	m = tap_send_frames(c, tcp4_l2_iov, tcp4_l2_buf_used);
> > > > +	for (i = 0; i < m; i++)
> > > > +		*tcp4_l2_buf_seq_update[i].seq += tcp4_l2_buf_seq_update[i].len;
> > > >  	tcp4_l2_buf_used = 0;
> > > >  }
> > > >  
> > > > @@ -2149,10 +2170,11 @@ static int tcp_sock_consume(struct tcp_tap_conn *conn, uint32_t ack_seq)
> > > >   * @plen:	Payload length at L4
> > > >   * @no_csum:	Don't compute IPv4 checksum, use the one from previous buffer
> > > >   * @seq:	Sequence number to be sent
> > > > - * @now:	Current timestamp
> > > > + * @seq_update:	Pointer to sequence number to update on successful send
> > > >   */
> > > >  static void tcp_data_to_tap(struct ctx *c, struct tcp_tap_conn *conn,
> > > > -			    ssize_t plen, int no_csum, uint32_t seq)
> > > > +			    ssize_t plen, int no_csum, uint32_t seq,
> > > > +			    uint32_t *seq_update)    
> > > 
> > > seq_update is always &conn->seq_to_tap, so there's no need for an
> > > additional parameter.  
> > 
> > Oh, right, I'll drop that.
> >   
> > > >  {
> > > >  	struct iovec *iov;
> > > >  
> > > > @@ -2160,6 +2182,9 @@ static void tcp_data_to_tap(struct ctx *c, struct tcp_tap_conn *conn,
> > > >  		struct tcp4_l2_buf_t *b = &tcp4_l2_buf[tcp4_l2_buf_used];
> > > >  		uint16_t *check = no_csum ? &(b - 1)->iph.check : NULL;
> > > >  
> > > > +		tcp4_l2_buf_seq_update[tcp4_l2_buf_used].seq = seq_update;
> > > > +		tcp4_l2_buf_seq_update[tcp4_l2_buf_used].len = plen;
> > > > +
> > > >  		iov = tcp4_l2_iov + tcp4_l2_buf_used++;
> > > >  		iov->iov_len = tcp_l2_buf_fill_headers(c, conn, b, plen,
> > > >  						       check, seq);
> > > > @@ -2168,6 +2193,9 @@ static void tcp_data_to_tap(struct ctx *c, struct tcp_tap_conn *conn,
> > > >  	} else if (CONN_V6(conn)) {
> > > >  		struct tcp6_l2_buf_t *b = &tcp6_l2_buf[tcp6_l2_buf_used];
> > > >  
> > > > +		tcp6_l2_buf_seq_update[tcp6_l2_buf_used].seq = seq_update;
> > > > +		tcp6_l2_buf_seq_update[tcp6_l2_buf_used].len = plen;
> > > > +
> > > >  		iov = tcp6_l2_iov + tcp6_l2_buf_used++;
> > > >  		iov->iov_len = tcp_l2_buf_fill_headers(c, conn, b, plen,
> > > >  						       NULL, seq);
> > > > @@ -2193,7 +2221,7 @@ static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn *conn)
> > > >  	int s = conn->sock, i, ret = 0;
> > > >  	struct msghdr mh_sock = { 0 };
> > > >  	uint16_t mss = MSS_GET(conn);
> > > > -	uint32_t already_sent;
> > > > +	uint32_t already_sent, seq;
> > > >  	struct iovec *iov;
> > > >  
> > > >  	already_sent = conn->seq_to_tap - conn->seq_ack_from_tap;
> > > > @@ -2282,14 +2310,15 @@ static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn *conn)
> > > >  
> > > >  	/* Finally, queue to tap */
> > > >  	plen = mss;
> > > > +	seq = conn->seq_to_tap;    
> > > 
> > > This will only be correct if tcp_l2_data_buf_flush() is *always*
> > > called between tcp_data_from_sock() calls for the same socket.  That
> > > should be true for the normal course of things.  However, couldn't it
> > > happen that we get a normal socket EPOLLIN event for a particular
> > > connection - calling tcp_data_from_sock() - but in the same epoll()
> > > round we also get a tap ack for the same connection which causes
> > > another call to tcp_data_from_sock() (with the change from patch
> > > 2/5).  IIRC those would both happen before the deferred handling and
> > > therefore the data_buf_flush().  
> > 
> > Ah, yes, I actually wrote this before 2/5 and concluded it was okay :/
> > but with that change, it's not. Unless we drop that change from 2/5.  
> 
> Even if we drop the change, it's a worryingly subtle constraint.

Another option to avoid this...

> > > Not sure how to deal with that short of separate 'seq_queued' and
> > > 'seq_sent' counters in the connection structure, which is a bit
> > > unfortunate.  
> > 
> > I wonder how bad it is if we call tcp_l2_data_buf_flush()
> > unconditionally before calling tcp_data_from_sock() from
> > tcp_tap_handler(). But again, maybe this is not needed at all, we
> > should check that epoll detail from 2/5 first...

other than this one, would be to use that external table to update
sequence numbers *in the frames* as we send stuff out.

-- 
Stefano