On Wed, May 15, 2024 at 11:34:27AM -0400, Jon Maloy wrote:
...
  commit a469fc393fa1 ("tcp, tap: Don't
increase tap-side sequence counter for dropped frames")
 delayed update of conn->seq_to_tap until the moment the corresponding
 frame has been successfully pushed out. This has the advantage that we
 immediately can make a new attempt to transmit a frame after a failed
 trasnmit, rather than waiting for the peer to later discover a gap and
 trigger the fast retransmit mechanism to solve the problem.
 
 This approach has turned out to cause a problem with spurious sequence
 number updates during peer-initiated retransmits, and we have realized
 it may not be the best way to solve the above issue.
 
 We now restore the previous method, by updating the said field at the
 moment a frame is added to the outqueue. To retain the advantage of
 having a quick re-attempt based on local failure detection, we now scan
 through the part of the outqueue that had do be dropped, and restore the
 sequence counter for each affected connection to the most appropriate
 value.
 
 Signed-off-by: Jon Maloy <jmaloy(a)redhat.com>
 
 ---
 v2: - Re-spun loop in tcp_revert_seq() and some other changes based on
       feedback from Stefano Brivio.
     - Added paranoid test to avoid that seq_to_tap becomes lower than
       seq_ack_from_tap.
 
 v3: - Identical to v2. Called v3 because it was embedded in a series
       with that version.
 
 v4: - In tcp_revert_seq(), we read the sequence number from the TCP
       header instead of keeping a copy in struct tcp_buf_seq_update.
     - Since the only remaining field in struct tcp_buf_seq_update is
       a pointer to struct tcp_tap_conn, we eliminate the struct
       altogether, and make the tcp6/tcp3_buf_seq_update arrays into
       arrays of said pointer.
     - Removed 'paranoid' test in tcp_revert_seq. If it happens, it
       is not fatal, and will be caught by other code anyway.
     - Separated from the series again.
 ---
  tcp.c | 59 +++++++++++++++++++++++++++++++++++++----------------------
  1 file changed, 37 insertions(+), 22 deletions(-)
 
 diff --git a/tcp.c b/tcp.c
 index 21d0af0..976dba8 100644
 --- a/tcp.c
 +++ b/tcp.c
 @@ -410,16 +410,6 @@ 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 */
  /**
   * struct tcp_payload_t - TCP header and data to send segments with payload
 @@ -461,7 +451,8 @@ static struct tcp_payload_t	tcp4_payload[TCP_FRAMES_MEM];
  
  static_assert(MSS4 <= sizeof(tcp4_payload[0].data), "MSS4 is greater than
65516");
  
 -static struct tcp_buf_seq_update tcp4_seq_update[TCP_FRAMES_MEM];
 +/* References tracking the owner connection of frames in the tap outqueue */
 +static struct tcp_tap_conn *tcp4_frame_conns[TCP_FRAMES_MEM];
  static unsigned int tcp4_payload_used;
  
  static struct tap_hdr		tcp4_flags_tap_hdr[TCP_FRAMES_MEM];
 @@ -483,7 +474,8 @@ static struct tcp_payload_t	tcp6_payload[TCP_FRAMES_MEM];
  
  static_assert(MSS6 <= sizeof(tcp6_payload[0].data), "MSS6 is greater than
65516");
  
 -static struct tcp_buf_seq_update tcp6_seq_update[TCP_FRAMES_MEM];
 +/* References tracking the owner connection of frames in the tap outqueue */
 +static struct tcp_tap_conn *tcp6_frame_conns[TCP_FRAMES_MEM];
  static unsigned int tcp6_payload_used;
  
  static struct tap_hdr		tcp6_flags_tap_hdr[TCP_FRAMES_MEM];
 @@ -1261,25 +1253,49 @@ static void tcp_flags_flush(const struct ctx *c)
  	tcp4_flags_used = 0;
  }
  
 +/**
 + * tcp_revert_seq() - Revert affected conn->seq_to_tap after failed transmission
 + * @conns:       Array of connection pointers corresponding to queued frames
 + * @frames:      Two-dimensional array containing queued frames with sub-iovs 
You can make the 2d array explicit in the type as:
	struct iovec (*frames)[TCP_NUM_IOVS];
See, for example the 'tap_iov' local in udp_tap_send().   (I recommend
the command line tool 'cdecl', also available online at cdecl.org for
working out confusing pointer-to-array types).

...
  + * @num_frames:  Number of entries in the two arrays
to be compared
 + */
 +static void tcp_revert_seq(struct tcp_tap_conn **conns, struct iovec *frames,
 +			   int num_frames)
 +{
 +	int c, f;
 +
 +	for (c = 0, f = 0; c < num_frames; c++, f += TCP_NUM_IOVS) { 
Nit: I find having the two parallel counters kind of confusing.  It
naturally goes away with the type change suggested above, but even
without that I'd prefer an explicit multiply in the body.  I strongly
suspect the compiler will be better at working out if the strength
reduction is worth it.

...
  +		struct tcp_tap_conn *conn = conns[c];
 +		struct tcphdr *th = frames[f + TCP_IOV_PAYLOAD].iov_base;
 +		uint32_t seq = ntohl(th->seq);
 +
 +		if (SEQ_LE(conn->seq_to_tap, seq)) 
Isn't this test inverted?  We want to rewind seq_to_tap if seq is less
than it, rather than the other way aruond.

...
  +			continue;
 +
 +		conn->seq_to_tap = seq;
 +	}
 +}
 +
  /**
   * tcp_payload_flush() - Send out buffers for segments with data
   * @c:		Execution context
   */
  static void tcp_payload_flush(const struct ctx *c)
  {
 -	unsigned i;
  	size_t m;
  
  	m = tap_send_frames(c, &tcp6_l2_iov[0][0], TCP_NUM_IOVS,
  			    tcp6_payload_used);
 -	for (i = 0; i < m; i++)
 -		*tcp6_seq_update[i].seq += tcp6_seq_update[i].len;
 +	if (m != tcp6_payload_used)
 +		tcp_revert_seq(tcp6_frame_conns, &tcp6_l2_iov[m][0], 
With the type change above this would become just &tcp_l2_iov[m].

...
  +			       tcp6_payload_used - m);
  	tcp6_payload_used = 0;
  
  	m = tap_send_frames(c, &tcp4_l2_iov[0][0], TCP_NUM_IOVS,
  			    tcp4_payload_used);
 -	for (i = 0; i < m; i++)
 -		*tcp4_seq_update[i].seq += tcp4_seq_update[i].len;
 +	if (m != tcp4_payload_used)
 +		tcp_revert_seq(tcp4_frame_conns, &tcp4_l2_iov[m][0],
 +			       tcp4_payload_used - m);
  	tcp4_payload_used = 0;
  }
  
 @@ -2129,10 +2145,11 @@ static int tcp_sock_consume(const struct tcp_tap_conn *conn,
uint32_t ack_seq)
  static void tcp_data_to_tap(const struct ctx *c, struct tcp_tap_conn *conn,
  			    ssize_t dlen, int no_csum, uint32_t seq)
  {
 -	uint32_t *seq_update = &conn->seq_to_tap;
  	struct iovec *iov;
  	size_t l4len;
  
 +	conn->seq_to_tap = seq + dlen;
 +
  	if (CONN_V4(conn)) {
  		struct iovec *iov_prev = tcp4_l2_iov[tcp4_payload_used - 1];
  		const uint16_t *check = NULL;
 @@ -2142,8 +2159,7 @@ static void tcp_data_to_tap(const struct ctx *c, struct
tcp_tap_conn *conn,
  			check = &iph->check;
  		}
  
 -		tcp4_seq_update[tcp4_payload_used].seq = seq_update;
 -		tcp4_seq_update[tcp4_payload_used].len = dlen;
 +		tcp4_frame_conns[tcp4_payload_used] = conn;
  
  		iov = tcp4_l2_iov[tcp4_payload_used++];
  		l4len = tcp_l2_buf_fill_headers(c, conn, iov, dlen, check, seq);
 @@ -2151,8 +2167,7 @@ static void tcp_data_to_tap(const struct ctx *c, struct
tcp_tap_conn *conn,
  		if (tcp4_payload_used > TCP_FRAMES_MEM - 1)
  			tcp_payload_flush(c);
  	} else if (CONN_V6(conn)) {
 -		tcp6_seq_update[tcp6_payload_used].seq = seq_update;
 -		tcp6_seq_update[tcp6_payload_used].len = dlen;
 +		tcp6_frame_conns[tcp6_payload_used] = conn;
  
  		iov = tcp6_l2_iov[tcp6_payload_used++];
  		l4len = tcp_l2_buf_fill_headers(c, conn, iov, dlen, NULL, seq); 
-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

On Wed, May 15, 2024 at 10:57:06PM -0400, Jon Maloy wrote:
...
  
 
 On 2024-05-15 22:24, David Gibson wrote:
  On Wed, May 15, 2024 at 11:34:27AM -0400, Jon
Maloy wrote:
  commit a469fc393fa1 ("tcp, tap: Don't
increase tap-side sequence counter for dropped frames")
 delayed update of conn->seq_to_tap until the moment the corresponding
 frame has been successfully pushed out. This has the advantage that we
 immediately can make a new attempt to transmit a frame after a failed
 trasnmit, rather than waiting for the peer to later discover a gap and
 trigger the fast retransmit mechanism to solve the problem.
 
 This approach has turned out to cause a problem with spurious sequence
 number updates during peer-initiated retransmits, and we have realized
 it may not be the best way to solve the above issue.
 
 We now restore the previous method, by updating the said field at the
 moment a frame is added to the outqueue. To retain the advantage of
 having a quick re-attempt based on local failure detection, we now scan
 through the part of the outqueue that had do be dropped, and restore the
 sequence counter for each affected connection to the most appropriate
 value.
 
 Signed-off-by: Jon Maloy <jmaloy(a)redhat.com>
 
 ---
 v2: - Re-spun loop in tcp_revert_seq() and some other changes based on
        feedback from Stefano Brivio.
      - Added paranoid test to avoid that seq_to_tap becomes lower than
        seq_ack_from_tap.
 
 v3: - Identical to v2. Called v3 because it was embedded in a series
        with that version.
 
 v4: - In tcp_revert_seq(), we read the sequence number from the TCP
        header instead of keeping a copy in struct tcp_buf_seq_update.
      - Since the only remaining field in struct tcp_buf_seq_update is
        a pointer to struct tcp_tap_conn, we eliminate the struct
        altogether, and make the tcp6/tcp3_buf_seq_update arrays into
        arrays of said pointer.
      - Removed 'paranoid' test in tcp_revert_seq. If it happens, it
        is not fatal, and will be caught by other code anyway.
      - Separated from the series again.
 ---
   tcp.c | 59 +++++++++++++++++++++++++++++++++++++----------------------
   1 file changed, 37 insertions(+), 22 deletions(-)
 
 diff --git a/tcp.c b/tcp.c
 index 21d0af0..976dba8 100644
 --- a/tcp.c
 +++ b/tcp.c
 @@ -410,16 +410,6 @@ 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 */
   /**
    * struct tcp_payload_t - TCP header and data to send segments with payload
 @@ -461,7 +451,8 @@ static struct tcp_payload_t	tcp4_payload[TCP_FRAMES_MEM];
   static_assert(MSS4 <= sizeof(tcp4_payload[0].data), "MSS4 is greater than
65516");
 -static struct tcp_buf_seq_update tcp4_seq_update[TCP_FRAMES_MEM];
 +/* References tracking the owner connection of frames in the tap outqueue */
 +static struct tcp_tap_conn *tcp4_frame_conns[TCP_FRAMES_MEM];
   static unsigned int tcp4_payload_used;
   static struct tap_hdr		tcp4_flags_tap_hdr[TCP_FRAMES_MEM];
 @@ -483,7 +474,8 @@ static struct tcp_payload_t	tcp6_payload[TCP_FRAMES_MEM];
   static_assert(MSS6 <= sizeof(tcp6_payload[0].data), "MSS6 is greater than
65516");
 -static struct tcp_buf_seq_update tcp6_seq_update[TCP_FRAMES_MEM];
 +/* References tracking the owner connection of frames in the tap outqueue */
 +static struct tcp_tap_conn *tcp6_frame_conns[TCP_FRAMES_MEM];
   static unsigned int tcp6_payload_used;
   static struct tap_hdr		tcp6_flags_tap_hdr[TCP_FRAMES_MEM];
 @@ -1261,25 +1253,49 @@ static void tcp_flags_flush(const struct ctx *c)
   	tcp4_flags_used = 0;
   }
 +/**
 + * tcp_revert_seq() - Revert affected conn->seq_to_tap after failed transmission
 + * @conns:       Array of connection pointers corresponding to queued frames
 + * @frames:      Two-dimensional array containing queued frames with sub-iovs 
 You
can make the 2d array explicit in the type as:
 	struct iovec (*frames)[TCP_NUM_IOVS];
 See, for example the 'tap_iov' local in udp_tap_send().   (I recommend
 the command line tool 'cdecl', also available online at cdecl.org for
 working out confusing pointer-to-array types). 
 Nice. I  wasn't quite happy
with this.
  
  + * @num_frames:  Number of entries in the two
arrays to be compared
 + */
 +static void tcp_revert_seq(struct tcp_tap_conn **conns, struct iovec *frames,
 +			   int num_frames)
 +{
 +	int c, f;
 +
 +	for (c = 0, f = 0; c < num_frames; c++, f += TCP_NUM_IOVS) { 
 Nit: I find
having the two parallel counters kind of confusing.  It
 naturally goes away with the type change suggested above, but even
 without that I'd prefer an explicit multiply in the body.  I strongly
 suspect the compiler will be better at working out if the strength
 reduction is worth it.
 
  +		struct tcp_tap_conn *conn = conns[c];
 +		struct tcphdr *th = frames[f + TCP_IOV_PAYLOAD].iov_base;
 +		uint32_t seq = ntohl(th->seq);
 +
 +		if (SEQ_LE(conn->seq_to_tap, seq)) 
 Isn't this test inverted?  We want
to rewind seq_to_tap if seq is less
 than it, rather than the other way aruond. 
 No. We do 'continue', i.e.,
nothing, if this condition is fulfilled.
 This may look a little non-intuitive here, but makes sense when I add the
 next patch. 
Oh, of course, my mistake.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

Hi David,
Found it (not your missing comment, but the bug), and it fixed the problem.
I'll post this patch separately shortly.

///jon


On 2024-06-04 13:36, Jon Maloy wrote:
...
  Hi David,
 This is the last comment I received from you regarding this patch.
 See below for further comment.

 On 2024-05-16 00:16, David Gibson wrote:
  On Wed, May 15, 2024 at 10:57:06PM -0400, Jon
Maloy wrote:
 
 On 2024-05-15 22:24, David Gibson wrote:
  On Wed, May 15, 2024 at 11:34:27AM -0400, Jon
Maloy wrote:
> commit a469fc393fa1 ("tcp, tap: Don't increase tap-side sequence 
> counter for dropped frames")
> delayed update of conn->seq_to_tap until the moment the corresponding
> frame has been successfully pushed out. This has the advantage 
> that we
> immediately can make a new attempt to transmit a frame after a failed
> trasnmit, rather than waiting for the peer to later discover a gap 
> and
> trigger the fast retransmit mechanism to solve the problem.
>
> This approach has turned out to cause a problem with spurious 
> sequence
> number updates during peer-initiated retransmits, and we have 
> realized
> it may not be the best way to solve the above issue.
>
> We now restore the previous method, by updating the said field at the
> moment a frame is added to the outqueue. To retain the advantage of
> having a quick re-attempt based on local failure detection, we now 
> scan
> through the part of the outqueue that had do be dropped, and 
> restore the
> sequence counter for each affected connection to the most appropriate
> value.
>
> Signed-off-by: Jon Maloy <jmaloy(a)redhat.com>
>
> ---
> v2: - Re-spun loop in tcp_revert_seq() and some other changes 
> based on
>         feedback from Stefano Brivio.
>       - Added paranoid test to avoid that seq_to_tap becomes lower 
> than
>         seq_ack_from_tap.
>
> v3: - Identical to v2. Called v3 because it was embedded in a series
>         with that version.
>
> v4: - In tcp_revert_seq(), we read the sequence number from the TCP
>         header instead of keeping a copy in struct 
> tcp_buf_seq_update.
>       - Since the only remaining field in struct 
> tcp_buf_seq_update is
>         a pointer to struct tcp_tap_conn, we eliminate the struct
>         altogether, and make the tcp6/tcp3_buf_seq_update arrays into
>         arrays of said pointer.
>       - Removed 'paranoid' test in tcp_revert_seq. If it happens, it
>         is not fatal, and will be caught by other code anyway.
>       - Separated from the series again.
> ---
>    tcp.c | 59 
> +++++++++++++++++++++++++++++++++++++----------------------
>    1 file changed, 37 insertions(+), 22 deletions(-)
>
> diff --git a/tcp.c b/tcp.c
> index 21d0af0..976dba8 100644
> --- a/tcp.c
> +++ b/tcp.c
> @@ -410,16 +410,6 @@ 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 */
>    /**
>     * struct tcp_payload_t - TCP header and data to send segments 
> with payload
> @@ -461,7 +451,8 @@ static struct tcp_payload_t 
> tcp4_payload[TCP_FRAMES_MEM];
>    static_assert(MSS4 <= sizeof(tcp4_payload[0].data), "MSS4 is 
> greater than 65516");
> -static struct tcp_buf_seq_update tcp4_seq_update[TCP_FRAMES_MEM];
> +/* References tracking the owner connection of frames in the tap 
> outqueue */
> +static struct tcp_tap_conn *tcp4_frame_conns[TCP_FRAMES_MEM];
>    static unsigned int tcp4_payload_used;
>    static struct tap_hdr tcp4_flags_tap_hdr[TCP_FRAMES_MEM];
> @@ -483,7 +474,8 @@ static struct tcp_payload_t 
> tcp6_payload[TCP_FRAMES_MEM];
>    static_assert(MSS6 <= sizeof(tcp6_payload[0].data), "MSS6 is 
> greater than 65516");
> -static struct tcp_buf_seq_update tcp6_seq_update[TCP_FRAMES_MEM];
> +/* References tracking the owner connection of frames in the tap 
> outqueue */
> +static struct tcp_tap_conn *tcp6_frame_conns[TCP_FRAMES_MEM];
>    static unsigned int tcp6_payload_used;
>    static struct tap_hdr tcp6_flags_tap_hdr[TCP_FRAMES_MEM];
> @@ -1261,25 +1253,49 @@ static void tcp_flags_flush(const struct 
> ctx *c)
>        tcp4_flags_used = 0;
>    }
> +/**
> + * tcp_revert_seq() - Revert affected conn->seq_to_tap after 
> failed transmission
> + * @conns:       Array of connection pointers corresponding to 
> queued frames
> + * @frames:      Two-dimensional array containing queued frames 
> with sub-iovs
 You can make the 2d array explicit in the type as:
     struct iovec (*frames)[TCP_NUM_IOVS];
 See, for example the 'tap_iov' local in udp_tap_send().   (I recommend
 the command line tool 'cdecl', also available online at cdecl.org for
 working out confusing pointer-to-array types). 
 Nice. I  wasn't quite happy
with this.
 > + * @num_frames:  Number of entries in the
two arrays to be compared
> + */
> +static void tcp_revert_seq(struct tcp_tap_conn **conns, struct 
> iovec *frames,
> +               int num_frames)
> +{
> +    int c, f;
> +
> +    for (c = 0, f = 0; c < num_frames; c++, f += TCP_NUM_IOVS) {
 Nit: I find having the two parallel counters kind of confusing.  It
 naturally goes away with the type change suggested above, but even
 without that I'd prefer an explicit multiply in the body.  I strongly
 suspect the compiler will be better at working out if the strength
 reduction is worth it.

> +        struct tcp_tap_conn *conn = conns[c];
> +        struct tcphdr *th = frames[f + TCP_IOV_PAYLOAD].iov_base;
> +        uint32_t seq = ntohl(th->seq);
> +
> +        if (SEQ_LE(conn->seq_to_tap, seq))
 Isn't this test inverted?  We want to rewind seq_to_tap if seq is less
 than it, rather than the other way aruond. 
 No. We do 'continue', i.e.,
nothing, if this condition is fulfilled.
 This may look a little non-intuitive here, but makes sense when I 
 add the
 next patch. 
 Oh, of course, my mistake.
 
 The code now (v7) looks as follows:

 /**
  * tcp_revert_seq() - Revert affected conn->seq_to_tap after failed 
 transmission
  * @conns:       Array of connection pointers corresponding to queued 
 frames
  * @frames:      Two-dimensional array containing queued frames with 
 sub-iovs
  * @num_frames:  Number of entries in the two arrays to be compared
  */
 static void tcp_revert_seq(struct tcp_tap_conn **conns, struct iovec 
 (*frames)[TCP_NUM_IOVS],
                            int num_frames)
 {
         int i;

         for (i = 0; i < num_frames; i++) {
                 struct tcp_tap_conn *conn = conns[i];
                 struct tcphdr *th = frames[i][TCP_IOV_PAYLOAD].iov_base;
                 uint32_t seq = ntohl(th->seq);

                 if (SEQ_LE(conn->seq_to_tap, seq))
                         continue;

                 conn->seq_to_tap = seq;
                 tcp_set_peek_offset(conn->sock, seq - 
 conn->seq_ack_from_tap);
         }
 }

 /**
  * tcp_payload_flush() - Send out buffers for segments with data
  * @c:          Execution context
  */
 static void tcp_payload_flush(const struct ctx *c)
 {
         size_t m;

         m = tap_send_frames(c, &tcp6_l2_iov[0][0], TCP_NUM_IOVS,
                             tcp6_payload_used);
         if (m != tcp6_payload_used) {
                 tcp_revert_seq(tcp6_frame_conns, &tcp6_l2_iov[m],
                                tcp6_payload_used - m);
         }
         tcp6_payload_used = 0;

         m = tap_send_frames(c, &tcp4_l2_iov[0][0], TCP_NUM_IOVS,
                             tcp4_payload_used);
         if (m != tcp4_payload_used) {
                 tcp_revert_seq(tcp4_frame_conns, &tcp4_l2_iov[m],
                                tcp4_payload_used - m);
         }
         tcp4_payload_used = 0;
 }

 Was this the version you were talking about on Monday morning?
 Did you spot some bug here which I am missing?

 Thanks
 ///jon
 

On Wed, May 15, 2024 at 11:34:28AM -0400, Jon Maloy wrote:
...
 
 From
linux-6.9.0 the kernel will contain 
 commit 05ea491641d3 ("tcp: add support for
SO_PEEK_OFF socket option").
 
 This new feature makes is possible to call recv_msg(MSG_PEEK) and make
 it start reading data from a given offset set by the SO_PEEK_OFF socket
 option. This way, we can avoid repeated reading of already read bytes of
 a received message, hence saving read cycles when forwarding TCP
 messages in the host->name space direction.
 
 In this commit, we add functionality to leverage this feature when
 available, while we fall back to the previous behavior when not.
 
 Measurements with iperf3 shows that throughput increases with 15-20
 percent in the host->namespace direction when this feature is used.
 
 Signed-off-by: Jon Maloy <jmaloy(a)redhat.com> 
I'm pretty sure this needs one more call to tcp_set_peek_offset()
inside the revert function introduced in the previous patch.
Otherwise looks good.

...
  
 ---
 v2: - Some smaller changes as suggested by David Gibson and Stefano Brivio.
     - Moved initial set_peek_offset(0) to only the locations where the socket is set
       to ESTABLISHED.
     - Removed the per-packet synchronization between sk_peek_off and
       already_sent. Instead only doing it in retransmit situations.
     - The problem I found when trouble shooting the occasionally occurring
       out of synch values between 'already_sent' and 'sk_peek_offset'
may
       have deeper implications that we may need to be investigate.
 
 v3: - Rebased to most recent version of tcp.c, plus the previous
       patch in this series.
     - Some changes based on feedback from PASST team
 
 v4: - Some small changes based on feedback from Stefan/David.
 ---
  tcp.c | 58 ++++++++++++++++++++++++++++++++++++++++++++++++++--------
  1 file changed, 50 insertions(+), 8 deletions(-)
 
 diff --git a/tcp.c b/tcp.c
 index 976dba8..4163bf9 100644
 --- a/tcp.c
 +++ b/tcp.c
 @@ -511,6 +511,9 @@ static struct iovec	tcp6_l2_iov		[TCP_FRAMES_MEM][TCP_NUM_IOVS];
  static struct iovec	tcp4_l2_flags_iov	[TCP_FRAMES_MEM][TCP_NUM_IOVS];
  static struct iovec	tcp6_l2_flags_iov	[TCP_FRAMES_MEM][TCP_NUM_IOVS];
  
 +/* Does the kernel support TCP_PEEK_OFF? */
 +static bool peek_offset_cap;
 +
  /* sendmsg() to socket */
  static struct iovec	tcp_iov			[UIO_MAXIOV];
  
 @@ -526,6 +529,20 @@ static_assert(ARRAY_SIZE(tc_hash) >= FLOW_MAX,
  int init_sock_pool4		[TCP_SOCK_POOL_SIZE];
  int init_sock_pool6		[TCP_SOCK_POOL_SIZE];
  
 +/**
 + * tcp_set_peek_offset() - Set SO_PEEK_OFF offset on a socket if supported
 + * @s:          Socket to update
 + * @offset:     Offset in bytes
 + */
 +static void tcp_set_peek_offset(int s, int offset)
 +{
 +	if (!peek_offset_cap)
 +		return;
 +
 +	if (setsockopt(s, SOL_SOCKET, SO_PEEK_OFF, &offset, sizeof(offset)))
 +		err("Failed to set SO_PEEK_OFF to %u in socket %i", offset, s);
 +}
 +
  /**
   * tcp_conn_epoll_events() - epoll events mask for given connection state
   * @events:	Current connection events
 @@ -2197,14 +2214,15 @@ static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn
*conn)
  	uint32_t already_sent, seq;
  	struct iovec *iov;
  
 +	/* How much have we read/sent since last received ack ? */
  	already_sent = conn->seq_to_tap - conn->seq_ack_from_tap;
 -
  	if (SEQ_LT(already_sent, 0)) {
  		/* RFC 761, section 2.1. */
  		flow_trace(conn, "ACK sequence gap: ACK for %u, sent: %u",
  			   conn->seq_ack_from_tap, conn->seq_to_tap);
  		conn->seq_to_tap = conn->seq_ack_from_tap;
  		already_sent = 0;
 +		tcp_set_peek_offset(s, 0);
  	}
  
  	if (!wnd_scaled || already_sent >= wnd_scaled) {
 @@ -2222,11 +2240,16 @@ static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn
*conn)
  		iov_rem = (wnd_scaled - already_sent) % mss;
  	}
  
 -	mh_sock.msg_iov = iov_sock;
 -	mh_sock.msg_iovlen = fill_bufs + 1;
 -
 -	iov_sock[0].iov_base = tcp_buf_discard;
 -	iov_sock[0].iov_len = already_sent;
 +	/* Prepare iov according to kernel capability */
 +	if (!peek_offset_cap) {
 +		mh_sock.msg_iov = iov_sock;
 +		iov_sock[0].iov_base = tcp_buf_discard;
 +		iov_sock[0].iov_len = already_sent;
 +		mh_sock.msg_iovlen = fill_bufs + 1;
 +	} else {
 +		mh_sock.msg_iov = &iov_sock[1];
 +		mh_sock.msg_iovlen = fill_bufs;
 +	}
  
  	if (( v4 && tcp4_payload_used + fill_bufs > TCP_FRAMES_MEM) ||
  	    (!v4 && tcp6_payload_used + fill_bufs > TCP_FRAMES_MEM)) {
 @@ -2267,7 +2290,10 @@ static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn
*conn)
  		return 0;
  	}
  
 -	sendlen = len - already_sent;
 +	sendlen = len;
 +	if (!peek_offset_cap)
 +		sendlen -= already_sent;
 +
  	if (sendlen <= 0) {
  		conn_flag(c, conn, STALLED);
  		return 0;
 @@ -2438,6 +2464,7 @@ static int tcp_data_from_tap(struct ctx *c, struct tcp_tap_conn
*conn,
  			   "fast re-transmit, ACK: %u, previous sequence: %u",
  			   max_ack_seq, conn->seq_to_tap);
  		conn->seq_to_tap = max_ack_seq;
 +		tcp_set_peek_offset(conn->sock, 0);
  		tcp_data_from_sock(c, conn);
  	}
  
 @@ -2530,6 +2557,7 @@ static void tcp_conn_from_sock_finish(struct ctx *c, struct
tcp_tap_conn *conn,
  	conn->seq_ack_to_tap = conn->seq_from_tap;
  
  	conn_event(c, conn, ESTABLISHED);
 +	tcp_set_peek_offset(conn->sock, 0);
  
  	/* The client might have sent data already, which we didn't
  	 * dequeue waiting for SYN,ACK from tap -- check now.
 @@ -2610,6 +2638,7 @@ int tcp_tap_handler(struct ctx *c, uint8_t pif, sa_family_t af,
  			goto reset;
  
  		conn_event(c, conn, ESTABLISHED);
 +		tcp_set_peek_offset(conn->sock, 0);
  
  		if (th->fin) {
  			conn->seq_from_tap++;
 @@ -2863,6 +2892,7 @@ void tcp_timer_handler(struct ctx *c, union epoll_ref ref)
  			flow_dbg(conn, "ACK timeout, retry");
  			conn->retrans++;
  			conn->seq_to_tap = conn->seq_ack_from_tap;
 +			tcp_set_peek_offset(conn->sock, 0);
  			tcp_data_from_sock(c, conn);
  			tcp_timer_ctl(c, conn);
  		}
 @@ -3154,7 +3184,8 @@ static void tcp_sock_refill_init(const struct ctx *c)
   */
  int tcp_init(struct ctx *c)
  {
 -	unsigned b;
 +	unsigned int b, optv = 0;
 +	int s;
  
  	for (b = 0; b < TCP_HASH_TABLE_SIZE; b++)
  		tc_hash[b] = FLOW_SIDX_NONE;
 @@ -3178,6 +3209,17 @@ int tcp_init(struct ctx *c)
  		NS_CALL(tcp_ns_socks_init, c);
  	}
  
 +	/* Probe for SO_PEEK_OFF support */
 +	s = socket(AF_INET, SOCK_STREAM | SOCK_CLOEXEC, IPPROTO_TCP);
 +	if (s < 0) {
 +		warn("Temporary TCP socket creation failed");
 +	} else {
 +		if (!setsockopt(s, SOL_SOCKET, SO_PEEK_OFF, &optv, sizeof(int)))
 +			peek_offset_cap = true;
 +		close(s);
 +	}
 +	info("SO_PEEK_OFF%ssupported", peek_offset_cap ? " " : " not
");
 +
  	return 0;
  }
   
-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

On 2024-05-15 16:24, Stefano Brivio wrote:
...
  On Wed, 15 May 2024 11:34:29 -0400
 Jon Maloy <jmaloy(a)redhat.com> wrote:

  A bug in kernel TCP may lead to a deadlock where
a zero window is sent
 from the peer, while it is unable to send out window updates even after
 reads have freed up enough buffer space to permit a larger window.
 In this situation, new window advertisemnts from the peer can only be
 triggered by packets arriving from this side.

 However, such packets are never sent, because the zero-window condition
 currently prevents this side from sending out any packets whatsoever
 to the peer.

 We notice that the above bug is triggered *only* after the peer has
 dropped an arriving packet because of severe memory squeeze, and that we
 hence always enter a retransmission situation when this occurs. This
 also means that it goes against the RFC 9293 recommendation that a
 previously advertised window never should shrink.

 RFC 9293 gives the solution to this situation. In chapter 3.6.1 we find
 the following statement:
 "A TCP receiver SHOULD NOT shrink the window, i.e., move the right
 window edge to the left (SHLD-14). However, a sending TCP peer MUST
 be robust against window shrinking, which may cause the
 "usable window" (see Section 3.8.6.2.1) to become negative (MUST-34).

 If this happens, the sender SHOULD NOT send new data (SHLD-15), but
 SHOULD retransmit normally the old unacknowledged data between SND.UNA
 and SND.UNA+SND.WND (SHLD-16). The sender MAY also retransmit old data
 beyond SND.UNA+SND.WND (MAY-7)"

 We never see the window become negative, but we interpret this as a
 recommendation to use the previously available window during
 retransmission even when the currently advertised window is zero.

 We use the above mechanism only at timer-induced retransmits.
 In the case we receive duplicate ack and a zero window, but
 still know we have outstanding data acks waiting, we send out an
 empty "fast probe" instead of doing fast retransmit. This averts
 the risk of overwhelming a memory squeezed peer with retransmits,
 while still forcing it to send out a new window update when the
 probe is received. This entails a theoretical risk of redundant
 retransmits from the peer, but that is a risk worth taking.

 In case of a zero-window non-retransmission situation where there
 is no new data to be sent, we also add a simple zero-window probing
 feature. By sending an empty packet at regular timeout events we
 resolve the situation described above, since the peer receives the
 necessary trigger to advertise its window once it becomes non-zero
 again.

 It should be noted that although this solves the problem we have at
 hand, it is not a genuine solution to the kernel bug. There may well
 be TCP stacks around in other OS-es which don't do this, nor have
 keep-alive probing as an alternatve way to solve the situation.

 Signed-off-by: Jon Maloy <jmaloy(a)redhat.com>

 ---
 v2: - Using previously advertised window during retransmission, instead
        highest send sequencece number in the cycle.
 v3: - Rebased to newest code
      - Changes based on feedback from PASST team
      - Sending out empty probe message at timer expiration when
        we are not in retransmit situation.
 v4: - Some small changes based on feedback from PASST team.
      - Replaced fast retransmit with a one-time 'fast probe' when
        window is zero.
 ---
   tcp.c      | 32 +++++++++++++++++++++++++++-----
   tcp_conn.h |  2 ++
   2 files changed, 29 insertions(+), 5 deletions(-)

 diff --git a/tcp.c b/tcp.c
 index 4163bf9..a33f494 100644
 --- a/tcp.c
 +++ b/tcp.c
 @@ -1761,9 +1761,15 @@ static void tcp_get_tap_ws(struct tcp_tap_conn *conn,
    */
   static void tcp_tap_window_update(struct tcp_tap_conn *conn, unsigned wnd)
   {
 +	uint32_t wnd_edge;
 +
   	wnd = MIN(MAX_WINDOW, wnd << conn->ws_from_tap);
   	conn->wnd_from_tap = MIN(wnd >> conn->ws_from_tap, USHRT_MAX);
   
 +	wnd_edge = conn->seq_ack_from_tap + wnd;
 +	if (wnd && SEQ_GT(wnd_edge, conn->seq_wnd_edge_from_tap)) 
 Here,
cppcheck ('make cppcheck') says:

 tcp.c:1770:6: style: Condition 'wnd' is always true [knownConditionTrueFalse]
   if (wnd && SEQ_GT(wnd_edge, conn->seq_wnd_edge_from_tap))
       ^
 tcp.c:1766:8: note: Assignment
'wnd=((1<<(16+8))<(wnd<<conn->ws_from_tap))?(1<<(16+8)):(wnd<<conn->ws_from_tap)',
assigned value is less than 1
   wnd = MIN(MAX_WINDOW, wnd << conn->ws_from_tap);
         ^
 tcp.c:1770:6: note: Condition 'wnd' is always true
   if (wnd && SEQ_GT(wnd_edge, conn->seq_wnd_edge_from_tap))
       ^

 See the comment in tcp_update_seqack_wnd() and related suppression.

 It's clearly a false positive (if you omit the MIN() macro, it goes
 away), so we need that same suppression here. 
Ok. I'll change it. Still a
little annoying when our tools are causing 
us extra job because they aren't up to the task.
...
 
  +		conn->seq_wnd_edge_from_tap = wnd_edge;
 +
   	/* FIXME: reflect the tap-side receiver's window back to the sock-side
   	 * sender by adjusting SO_RCVBUF? */
   }
 @@ -1796,6 +1802,7 @@ static void tcp_seq_init(const struct ctx *c, struct tcp_tap_conn
*conn,
   	ns = (now->tv_sec * 1000000000 + now->tv_nsec) >> 5;
   
   	conn->seq_to_tap = ((uint32_t)(hash >> 32) ^ (uint32_t)hash) + ns;
 +	conn->seq_wnd_edge_from_tap = conn->seq_to_tap;
   }
   
   /**
 @@ -2205,13 +2212,12 @@ static void tcp_data_to_tap(const struct ctx *c, struct
tcp_tap_conn *conn,
    */
   static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn *conn)
   {
 -	uint32_t wnd_scaled = conn->wnd_from_tap << conn->ws_from_tap;
   	int fill_bufs, send_bufs = 0, last_len, iov_rem = 0;
   	int sendlen, len, dlen, v4 = CONN_V4(conn);
 +	uint32_t already_sent, max_send, seq;
   	int s = conn->sock, i, ret = 0;
   	struct msghdr mh_sock = { 0 };
   	uint16_t mss = MSS_GET(conn);
 -	uint32_t already_sent, seq;
   	struct iovec *iov;
   
   	/* How much have we read/sent since last received ack ? */
 @@ -2225,19 +2231,24 @@ static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn
*conn)
   		tcp_set_peek_offset(s, 0);
   	}
   
 -	if (!wnd_scaled || already_sent >= wnd_scaled) {
 +	/* How much are we still allowed to send within current window ? */
 +	max_send = conn->seq_wnd_edge_from_tap - conn->seq_to_tap;
 +	if (SEQ_LE(max_send, 0)) {
 +		flow_trace(conn, "Empty window: win_upper: %u, sent: %u", 
 This is not
win_upper anymore, and the window is actually full rather
 than empty (it's... empty of space). Maybe:

 		flow_trace(conn, "Window full: right edge: %u, sent: %u" 
yes.
...
 
  +			   conn->seq_wnd_edge_from_tap,
conn->seq_to_tap);
 +		conn->seq_wnd_edge_from_tap = conn->seq_to_tap;
   		conn_flag(c, conn, STALLED);
   		conn_flag(c, conn, ACK_FROM_TAP_DUE);
   		return 0;
   	}
   
   	/* Set up buffer descriptors we'll fill completely and partially. */
 -	fill_bufs = DIV_ROUND_UP(wnd_scaled - already_sent, mss);
 +	fill_bufs = DIV_ROUND_UP(max_send,  mss);
   	if (fill_bufs > TCP_FRAMES) {
   		fill_bufs = TCP_FRAMES;
   		iov_rem = 0;
   	} else {
 -		iov_rem = (wnd_scaled - already_sent) % mss;
 +		iov_rem = max_send % mss;
   	}
   
   	/* Prepare iov according to kernel capability */
 @@ -2466,6 +2477,13 @@ static int tcp_data_from_tap(struct ctx *c, struct tcp_tap_conn
*conn,
   		conn->seq_to_tap = max_ack_seq;
   		tcp_set_peek_offset(conn->sock, 0);
   		tcp_data_from_sock(c, conn);
 +	} else if (!max_ack_seq_wnd && SEQ_GT(conn->seq_to_tap, max_ack_seq)) {
 +		/* Force peer to send new advertisement now, but only once */ 
 Two questions:

 - which advertisement? We're sending a zero-window probe, not forcing
    the peer to do much really. I would rather just state that we're
    sending a probe 
Actually, it is only clear if you know the code of the (linux)
peer.
I realized this was maybe a too strong statement, but this is really
what happens.

...
 
 - what guarantees it only happens once? If we get more data from the
    socket, we'll get again SEQ_GT(conn->seq_to_tap, max_ack_seq) in a
    bit, and send another ACK (duplicate) to the peer, without the peer
    necessarily ever advertising a non-zero window meanwhile 
Right. I need to give
this  a new spin, but I think I am on the right 
track.
I *really* want this to be solved here, and not need to fall back to
  the timeout except under exceptional conditions.
This happens really often in my tests, and with this fix it actually
works like a charm.

...
 
    I'm struggling a bit to understand how this can work "cleanly", a
    packet capture of this mechanism in action would certainly help. 
Ok. In v5.
...
 
  +		flow_trace(conn, "fast probe, ACK: %u,
previous sequence: %u",
 +			   max_ack_seq, conn->seq_to_tap);
 +		tcp_send_flag(c, conn, ACK);
 +		conn->seq_to_tap = max_ack_seq;
 +		tcp_set_peek_offset(conn->sock, 0);
   	}
   
   	if (!iov_i)
 @@ -2911,6 +2929,10 @@ void tcp_timer_handler(struct ctx *c, union epoll_ref ref)
   			flow_dbg(conn, "activity timeout");
   			tcp_rst(c, conn);
   		}
 +		/* No data exchanged recently? Keep connection alive. */ 
 ...I just spotted this
from v3: this is not the reason why we're
 sending a keep-alive. We're sending a keep-alive segment because the
 peer advertised its window as zero.

 I also realised that this is not scheduled additionally, so it will
 just trigger on an activity timeout, I suppose. We should reschedule
 this after ACK_TIMEOUT, instead (that was my earlier suggestion, I
 didn't check anymore) when the peer advertises a zero window.

  +		if (conn->seq_to_tap ==
conn->seq_ack_from_tap && 
 ...this part will only work if we reset
seq_to_tap to seq_ack_from_tap
 earlier, and we have no pending data to send, which is not necessarily
 the case if we want to send a zero-window probe.

  +		    conn->seq_from_tap ==
conn->seq_ack_to_tap)
 +			tcp_send_flag(c, conn, ACK); 
 I think the conditions should simply be:

 - the window currently advertised by the peer is zero

 - we don't have pending data to acknowledge (otherwise the peer can
    interpret our keep-alive as a duplicate ACK) 
That is ok.
The deadlock situation we have ended up in will anyway be resolved by the
real retransmit that will happen before we get here.
I am now wondering if this probe it serves any purpose at all?
...
 
>   	}
>   }
>   
> diff --git a/tcp_conn.h b/tcp_conn.h
> index d280b22..5cbad2a 100644
> --- a/tcp_conn.h
> +++ b/tcp_conn.h
> @@ -30,6 +30,7 @@
>    * @wnd_to_tap:		Sending window advertised to tap, unscaled (as sent)
>    * @seq_to_tap:		Next sequence for packets to tap
>    * @seq_ack_from_tap:	Last ACK number received from tap
> + * @seq_wnd_edge_from_tap: Right edge of last non-zero window from tap
>    * @seq_from_tap:	Next sequence for packets from tap (not actually sent)
>    * @seq_ack_to_tap:	Last ACK number sent to tap
>    * @seq_init_from_tap:	Initial sequence number from tap
> @@ -101,6 +102,7 @@ struct tcp_tap_conn {
>   
>   	uint32_t	seq_to_tap;
>   	uint32_t	seq_ack_from_tap;
> +	uint32_t	seq_wnd_edge_from_tap;
>   	uint32_t	seq_from_tap;
>   	uint32_t	seq_ack_to_tap;
>   	uint32_t	seq_init_from_tap; 
I'll re-spin the two first ones asap, so you can apply them, and then I 
will try to improve
this one further.

///jon

On Wed, 15 May 2024 19:10:49 -0400
Jon Maloy <jmaloy(a)redhat.com> wrote:

...
  On 2024-05-15 16:24, Stefano Brivio wrote:
  On Wed, 15 May 2024 11:34:29 -0400
 Jon Maloy <jmaloy(a)redhat.com> wrote:
  
  A bug in kernel TCP may lead to a deadlock where
a zero window is sent
 from the peer, while it is unable to send out window updates even after
 reads have freed up enough buffer space to permit a larger window.
 In this situation, new window advertisemnts from the peer can only be
 triggered by packets arriving from this side.

 However, such packets are never sent, because the zero-window condition
 currently prevents this side from sending out any packets whatsoever
 to the peer.

 We notice that the above bug is triggered *only* after the peer has
 dropped an arriving packet because of severe memory squeeze, and that we
 hence always enter a retransmission situation when this occurs. This
 also means that it goes against the RFC 9293 recommendation that a
 previously advertised window never should shrink.

 RFC 9293 gives the solution to this situation. In chapter 3.6.1 we find
 the following statement:
 "A TCP receiver SHOULD NOT shrink the window, i.e., move the right
 window edge to the left (SHLD-14). However, a sending TCP peer MUST
 be robust against window shrinking, which may cause the
 "usable window" (see Section 3.8.6.2.1) to become negative (MUST-34).

 If this happens, the sender SHOULD NOT send new data (SHLD-15), but
 SHOULD retransmit normally the old unacknowledged data between SND.UNA
 and SND.UNA+SND.WND (SHLD-16). The sender MAY also retransmit old data
 beyond SND.UNA+SND.WND (MAY-7)"

 We never see the window become negative, but we interpret this as a
 recommendation to use the previously available window during
 retransmission even when the currently advertised window is zero.

 We use the above mechanism only at timer-induced retransmits.
 In the case we receive duplicate ack and a zero window, but
 still know we have outstanding data acks waiting, we send out an
 empty "fast probe" instead of doing fast retransmit. This averts
 the risk of overwhelming a memory squeezed peer with retransmits,
 while still forcing it to send out a new window update when the
 probe is received. This entails a theoretical risk of redundant
 retransmits from the peer, but that is a risk worth taking.

 In case of a zero-window non-retransmission situation where there
 is no new data to be sent, we also add a simple zero-window probing
 feature. By sending an empty packet at regular timeout events we
 resolve the situation described above, since the peer receives the
 necessary trigger to advertise its window once it becomes non-zero
 again.

 It should be noted that although this solves the problem we have at
 hand, it is not a genuine solution to the kernel bug. There may well
 be TCP stacks around in other OS-es which don't do this, nor have
 keep-alive probing as an alternatve way to solve the situation.

 Signed-off-by: Jon Maloy <jmaloy(a)redhat.com>

 ---
 v2: - Using previously advertised window during retransmission, instead
        highest send sequencece number in the cycle.
 v3: - Rebased to newest code
      - Changes based on feedback from PASST team
      - Sending out empty probe message at timer expiration when
        we are not in retransmit situation.
 v4: - Some small changes based on feedback from PASST team.
      - Replaced fast retransmit with a one-time 'fast probe' when
        window is zero.
 ---
   tcp.c      | 32 +++++++++++++++++++++++++++-----
   tcp_conn.h |  2 ++
   2 files changed, 29 insertions(+), 5 deletions(-)

 diff --git a/tcp.c b/tcp.c
 index 4163bf9..a33f494 100644
 --- a/tcp.c
 +++ b/tcp.c
 @@ -1761,9 +1761,15 @@ static void tcp_get_tap_ws(struct tcp_tap_conn *conn,
    */
   static void tcp_tap_window_update(struct tcp_tap_conn *conn, unsigned wnd)
   {
 +	uint32_t wnd_edge;
 +
   	wnd = MIN(MAX_WINDOW, wnd << conn->ws_from_tap);
   	conn->wnd_from_tap = MIN(wnd >> conn->ws_from_tap, USHRT_MAX);
   
 +	wnd_edge = conn->seq_ack_from_tap + wnd;
 +	if (wnd && SEQ_GT(wnd_edge, conn->seq_wnd_edge_from_tap))   
 Here,
cppcheck ('make cppcheck') says:

 tcp.c:1770:6: style: Condition 'wnd' is always true [knownConditionTrueFalse]
   if (wnd && SEQ_GT(wnd_edge, conn->seq_wnd_edge_from_tap))
       ^
 tcp.c:1766:8: note: Assignment
'wnd=((1<<(16+8))<(wnd<<conn->ws_from_tap))?(1<<(16+8)):(wnd<<conn->ws_from_tap)',
assigned value is less than 1
   wnd = MIN(MAX_WINDOW, wnd << conn->ws_from_tap);
         ^
 tcp.c:1770:6: note: Condition 'wnd' is always true
   if (wnd && SEQ_GT(wnd_edge, conn->seq_wnd_edge_from_tap))
       ^

 See the comment in tcp_update_seqack_wnd() and related suppression.

 It's clearly a false positive (if you omit the MIN() macro, it goes
 away), so we need that same suppression here.   
 Ok. I'll change it. Still a
little annoying when our tools are causing 
 us extra job because they aren't up to the task. 
Just like in David's experience, cppcheck probably saved me a
substantial number of embarrassing situations.

I would say it's very much up to the task to be honest, a few false
positives are something we can definitely expect.

...
 
   
  +		conn->seq_wnd_edge_from_tap = wnd_edge;
 +
   	/* FIXME: reflect the tap-side receiver's window back to the sock-side
   	 * sender by adjusting SO_RCVBUF? */
   }
 @@ -1796,6 +1802,7 @@ static void tcp_seq_init(const struct ctx *c, struct tcp_tap_conn
*conn,
   	ns = (now->tv_sec * 1000000000 + now->tv_nsec) >> 5;
   
   	conn->seq_to_tap = ((uint32_t)(hash >> 32) ^ (uint32_t)hash) + ns;
 +	conn->seq_wnd_edge_from_tap = conn->seq_to_tap;
   }
   
   /**
 @@ -2205,13 +2212,12 @@ static void tcp_data_to_tap(const struct ctx *c, struct
tcp_tap_conn *conn,
    */
   static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn *conn)
   {
 -	uint32_t wnd_scaled = conn->wnd_from_tap << conn->ws_from_tap;
   	int fill_bufs, send_bufs = 0, last_len, iov_rem = 0;
   	int sendlen, len, dlen, v4 = CONN_V4(conn);
 +	uint32_t already_sent, max_send, seq;
   	int s = conn->sock, i, ret = 0;
   	struct msghdr mh_sock = { 0 };
   	uint16_t mss = MSS_GET(conn);
 -	uint32_t already_sent, seq;
   	struct iovec *iov;
   
   	/* How much have we read/sent since last received ack ? */
 @@ -2225,19 +2231,24 @@ static int tcp_data_from_sock(struct ctx *c, struct tcp_tap_conn
*conn)
   		tcp_set_peek_offset(s, 0);
   	}
   
 -	if (!wnd_scaled || already_sent >= wnd_scaled) {
 +	/* How much are we still allowed to send within current window ? */
 +	max_send = conn->seq_wnd_edge_from_tap - conn->seq_to_tap;
 +	if (SEQ_LE(max_send, 0)) {
 +		flow_trace(conn, "Empty window: win_upper: %u, sent: %u",   
 This is
not win_upper anymore, and the window is actually full rather
 than empty (it's... empty of space). Maybe:

 		flow_trace(conn, "Window full: right edge: %u, sent: %u"   
 yes.
   
  +			   conn->seq_wnd_edge_from_tap,
conn->seq_to_tap);
 +		conn->seq_wnd_edge_from_tap = conn->seq_to_tap;
   		conn_flag(c, conn, STALLED);
   		conn_flag(c, conn, ACK_FROM_TAP_DUE);
   		return 0;
   	}
   
   	/* Set up buffer descriptors we'll fill completely and partially. */
 -	fill_bufs = DIV_ROUND_UP(wnd_scaled - already_sent, mss);
 +	fill_bufs = DIV_ROUND_UP(max_send,  mss);
   	if (fill_bufs > TCP_FRAMES) {
   		fill_bufs = TCP_FRAMES;
   		iov_rem = 0;
   	} else {
 -		iov_rem = (wnd_scaled - already_sent) % mss;
 +		iov_rem = max_send % mss;
   	}
   
   	/* Prepare iov according to kernel capability */
 @@ -2466,6 +2477,13 @@ static int tcp_data_from_tap(struct ctx *c, struct tcp_tap_conn
*conn,
   		conn->seq_to_tap = max_ack_seq;
   		tcp_set_peek_offset(conn->sock, 0);
   		tcp_data_from_sock(c, conn);
 +	} else if (!max_ack_seq_wnd && SEQ_GT(conn->seq_to_tap, max_ack_seq)) {
 +		/* Force peer to send new advertisement now, but only once */   
 Two questions:

 - which advertisement? We're sending a zero-window probe, not forcing
    the peer to do much really. I would rather just state that we're
    sending a probe   
 Actually, it is only clear if you know the code of the
(linux) peer.
 I realized this was maybe a too strong statement, but this is really
 what happens. 
Our peer is not necessarily the Linux kernel, though, and, especially,
not a specific (broken) version of it.

...
 
  - what
guarantees it only happens once? If we get more data from the
    socket, we'll get again SEQ_GT(conn->seq_to_tap, max_ack_seq) in a
    bit, and send another ACK (duplicate) to the peer, without the peer
    necessarily ever advertising a non-zero window meanwhile   
 Right. I need to
give this  a new spin, but I think I am on the right 
 track.
 I *really* want this to be solved here, and not need to fall back to
   the timeout except under exceptional conditions.
 This happens really often in my tests, and with this fix it actually
 works like a charm.
 
 
    I'm struggling a bit to understand how this can work "cleanly", a
    packet capture of this mechanism in action would certainly help.   
 Ok. In v5.
   
  +		flow_trace(conn, "fast probe, ACK: %u,
previous sequence: %u",
 +			   max_ack_seq, conn->seq_to_tap);
 +		tcp_send_flag(c, conn, ACK);
 +		conn->seq_to_tap = max_ack_seq;
 +		tcp_set_peek_offset(conn->sock, 0);
   	}
   
   	if (!iov_i)
 @@ -2911,6 +2929,10 @@ void tcp_timer_handler(struct ctx *c, union epoll_ref ref)
   			flow_dbg(conn, "activity timeout");
   			tcp_rst(c, conn);
   		}
 +		/* No data exchanged recently? Keep connection alive. */   
 ...I just spotted
this from v3: this is not the reason why we're
 sending a keep-alive. We're sending a keep-alive segment because the
 peer advertised its window as zero.

 I also realised that this is not scheduled additionally, so it will
 just trigger on an activity timeout, I suppose. We should reschedule
 this after ACK_TIMEOUT, instead (that was my earlier suggestion, I
 didn't check anymore) when the peer advertises a zero window.
  
  +		if (conn->seq_to_tap ==
conn->seq_ack_from_tap &&   
 ...this part will only work if we reset
seq_to_tap to seq_ack_from_tap
 earlier, and we have no pending data to send, which is not necessarily
 the case if we want to send a zero-window probe.
  
  +		    conn->seq_from_tap ==
conn->seq_ack_to_tap)
 +			tcp_send_flag(c, conn, ACK);   
 I think the conditions should simply be:

 - the window currently advertised by the peer is zero

 - we don't have pending data to acknowledge (otherwise the peer can
    interpret our keep-alive as a duplicate ACK)   
 That is ok.
 The deadlock situation we have ended up in will anyway be resolved by the
 real retransmit that will happen before we get here.
 I am now wondering if this probe it serves any purpose at all? 
Well, RFC 9293 says we must implement it, so it's a good idea to do
that, I guess.

Let's say your first probe is lost for whatever reason (if the system
is low on memory, that is actually likely to happen with tap devices).

...
 
   
>   	}
>   }
>   
> diff --git a/tcp_conn.h b/tcp_conn.h
> index d280b22..5cbad2a 100644
> --- a/tcp_conn.h
> +++ b/tcp_conn.h
> @@ -30,6 +30,7 @@
>    * @wnd_to_tap:		Sending window advertised to tap, unscaled (as sent)
>    * @seq_to_tap:		Next sequence for packets to tap
>    * @seq_ack_from_tap:	Last ACK number received from tap
> + * @seq_wnd_edge_from_tap: Right edge of last non-zero window from tap
>    * @seq_from_tap:	Next sequence for packets from tap (not actually sent)
>    * @seq_ack_to_tap:	Last ACK number sent to tap
>    * @seq_init_from_tap:	Initial sequence number from tap
> @@ -101,6 +102,7 @@ struct tcp_tap_conn {
>   
>   	uint32_t	seq_to_tap;
>   	uint32_t	seq_ack_from_tap;
> +	uint32_t	seq_wnd_edge_from_tap;
>   	uint32_t	seq_from_tap;
>   	uint32_t	seq_ack_to_tap;
>   	uint32_t	seq_init_from_tap;   
 
 I'll re-spin the two first ones asap, so you can apply them, and then I 
 will try to improve
 this one further. 
Just mind that if 2/3 makes the issue you're working around here more
likely to happen (me, I've never seen it), we shouldn't go ahead with
2/3 without this, I guess.

-- 
Stefano