tas: Implement TCP Window Scale option

include/packet_defs.h

@@ -169,13 +169,19 @@ struct tcp_hdr {
 #define TCP_OPT_END_OF_OPTIONS 0
 #define TCP_OPT_NO_OP 1
 #define TCP_OPT_MSS 2
+#define TCP_OPT_WS 3
 #define TCP_OPT_TIMESTAMP 8
 struct tcp_mss_opt {
   uint8_t kind;
   uint8_t length;
   beui16_t mss;
 } __attribute__((packed));
 
+struct tcp_ws_opt {
+  uint8_t kind;
+  uint8_t length;
+  uint8_t scale;
+} __attribute__((packed));
 
 struct tcp_timestamp_opt {
   uint8_t kind;

include/tas_memif.h

@@ -262,13 +262,17 @@ struct flextcp_pl_flowst {
   /** Sequence number of queue pointer bumps */
   uint16_t bump_seq;
 
-  // 56
+  /** TX Window scale factor */
+  uint8_t tx_window_scale;
+  /** RX Window scale factor */
+  uint8_t rx_window_scale;
+  // 58
 
   /********************************************************/
   /* read-write fields */
 
-  /** spin lock */
-  volatile uint32_t lock;
+  /** Duplicate ack count */
+  uint16_t rx_dupack_cnt;
 
   /** Bytes available for received segments at next position */
   uint32_t rx_avail;
@@ -279,8 +283,9 @@ struct flextcp_pl_flowst {
   uint32_t rx_next_seq;
   /** Bytes available in remote end for received segments */
   uint32_t rx_remote_avail;
-  /** Duplicate ack count */
-  uint32_t rx_dupack_cnt;
+
+  /** spin lock */
+  volatile uint32_t lock;
 
 #ifdef FLEXNIC_PL_OOO_RECV
   /* Start of interval of out-of-order received data */

tas/config.c

@@ -48,6 +48,7 @@ enum cfg_params {
   CP_TCP_TXBUF_LEN,
   CP_TCP_HANDSHAKE_TO,
   CP_TCP_HANDSHAKE_RETRIES,
+  CP_TCP_WINDOW_SCALE,
   CP_CC,
   CP_CC_CONTROL_GRANULARITY,
   CP_CC_CONTROL_INTERVAL,
@@ -123,6 +124,9 @@ static struct option opts[] = {
     { .name = "tcp-handshake-retries",
       .has_arg = required_argument,
       .val = CP_TCP_HANDSHAKE_RETRIES },
+    { .name = "tcp-window-scale",
+      .has_arg = required_argument,
+      .val = CP_TCP_WINDOW_SCALE },
     { .name = "cc",
       .has_arg = required_argument,
       .val = CP_CC },
@@ -327,6 +331,12 @@ int config_parse(struct configuration *c, int argc, char *argv[])
           goto failed;
         }
         break;
+      case CP_TCP_WINDOW_SCALE:
+        if (parse_int8(optarg, &c->tcp_window_scale) != 0) {
+          fprintf(stderr, "tcp window scale parsing failed\n");
+          goto failed;
+        }
+        break;
       case CP_CC:
         if (!strcmp(optarg, "dctcp-win")) {
           c->cc_algorithm = CONFIG_CC_DCTCP_WIN;
@@ -565,6 +575,7 @@ static int config_defaults(struct configuration *c, char *progname)
   c->tcp_txbuf_len = 8192;
   c->tcp_handshake_to = 10000;
   c->tcp_handshake_retries = 10;
+  c->tcp_window_scale = 0;
   c->cc_algorithm = CONFIG_CC_DCTCP_RATE;
   c->cc_control_granularity = 50;
   c->cc_control_interval = 2;
@@ -635,6 +646,8 @@ static void print_usage(struct configuration *c, char *progname)
           "[default: %"PRIu32"]\n"
       "  --tcp-handshake-retries=RETRIES  Handshake retries "
           "[default: %"PRIu32"]\n"
+      "  --tcp-window-scale=SCALE    Window scale factor "
+          "[default: %"PRIu8"]\n"
       "\n"
       "Congestion control parameters:\n"
       "  --cc=ALGORITHM              Congestion-control algorithm "
@@ -711,7 +724,7 @@ static void print_usage(struct configuration *c, char *progname)
       progname, c->shm_len,
       c->nic_rx_len, c->nic_tx_len, c->app_kin_len, c->app_kout_len,
       c->tcp_rtt_init, c->tcp_link_bw, c->tcp_rxbuf_len, c->tcp_txbuf_len,
-      c->tcp_handshake_to, c->tcp_handshake_retries,
+      c->tcp_handshake_to, c->tcp_handshake_retries, c->tcp_window_scale,
       c->cc_control_granularity, c->cc_control_interval, c->cc_rexmit_ints,
       (double) c->cc_dctcp_weight / UINT32_MAX, c->cc_dctcp_min,
       c->cc_const_rate, c->cc_timely_tlow, c->cc_timely_thigh,

tas/fast/fast_flows.c

@@ -66,10 +66,11 @@ static void flow_rx_seq_write(struct flextcp_pl_flowst *fs, uint32_t seq,
 static void flow_tx_segment(struct dataplane_context *ctx,
     struct network_buf_handle *nbh, struct flextcp_pl_flowst *fs,
     uint32_t seq, uint32_t ack, uint32_t rxwnd, uint16_t payload,
-    uint32_t payload_pos, uint32_t ts_echo, uint32_t ts_my, uint8_t fin);
+    uint32_t payload_pos, uint32_t ts_echo, uint32_t ts_my,
+    uint8_t window_scale, uint8_t fin);
 static void flow_tx_ack(struct dataplane_context *ctx, uint32_t seq,
     uint32_t ack, uint32_t rxwnd, uint32_t echo_ts, uint32_t my_ts,
-    struct network_buf_handle *nbh, struct tcp_timestamp_opt *ts_opt);
+    uint8_t window_scale, struct network_buf_handle *nbh, struct tcp_timestamp_opt *ts_opt);
 static void flow_reset_retransmit(struct flextcp_pl_flowst *fs);
 
 static inline void tcp_checksums(struct network_buf_handle *nbh,
@@ -195,7 +196,7 @@ int fast_flows_qman(struct dataplane_context *ctx, uint32_t queue,
 
   /* send out segment */
   flow_tx_segment(ctx, nbh, fs, tx_seq, ack, rx_wnd, len, tx_pos,
-      fs->tx_next_ts, ts, fin);
+      fs->tx_next_ts, ts, fs->rx_window_scale, fin);
 unlock:
   fs_unlock(fs);
   return ret;
@@ -499,7 +500,7 @@ int fast_flows_packet(struct dataplane_context *ctx,
     }
   }
 
-  fs->rx_remote_avail = f_beui16(p->tcp.wnd);
+  fs->rx_remote_avail = f_beui16(p->tcp.wnd) << fs->tx_window_scale;
 
   /* make sure we don't receive anymore payload after FIN */
   if ((fs->rx_base_sp & FLEXNIC_PL_FLOWST_RXFIN) == FLEXNIC_PL_FLOWST_RXFIN &&
@@ -631,7 +632,7 @@ int fast_flows_packet(struct dataplane_context *ctx,
   /* if we need to send an ack, also send packet to TX pipeline to do so */
   if (trigger_ack) {
     flow_tx_ack(ctx, fs->tx_next_seq, fs->rx_next_seq, fs->rx_avail,
-        fs->tx_next_ts, ts, nbh, opts->ts);
+        fs->tx_next_ts, ts, fs->rx_window_scale, nbh, opts->ts);
   }
 
   fs_unlock(fs);
@@ -758,7 +759,7 @@ int fast_flows_bump(struct dataplane_context *ctx, uint32_t flow_id,
    * we're not sending anyways. */
   if (new_avail == 0 && rx_avail_prev == 0 && fs->rx_avail != 0) {
     flow_tx_segment(ctx, nbh, fs, fs->tx_next_seq, fs->rx_next_seq,
-        fs->rx_avail, 0, 0, fs->tx_next_ts, ts, 0);
+        fs->rx_avail, 0, 0, fs->tx_next_ts, ts, fs->rx_window_scale, 0);
     ret = 0;
   }
 
@@ -877,7 +878,8 @@ static void flow_rx_seq_write(struct flextcp_pl_flowst *fs, uint32_t seq,
 static void flow_tx_segment(struct dataplane_context *ctx,
     struct network_buf_handle *nbh, struct flextcp_pl_flowst *fs,
     uint32_t seq, uint32_t ack, uint32_t rxwnd, uint16_t payload,
-    uint32_t payload_pos, uint32_t ts_echo, uint32_t ts_my, uint8_t fin)
+    uint32_t payload_pos, uint32_t ts_echo, uint32_t ts_my,
+    uint8_t window_scale, uint8_t fin)
 {
   uint16_t hdrs_len, optlen, fin_fl;
   struct pkt_tcp *p = network_buf_buf(nbh);
@@ -915,7 +917,7 @@ static void flow_tx_segment(struct dataplane_context *ctx,
   p->tcp.seqno = t_beui32(seq);
   p->tcp.ackno = t_beui32(ack);
   TCPH_HDRLEN_FLAGS_SET(&p->tcp, 5 + optlen / 4, TCP_PSH | TCP_ACK | fin_fl);
-  p->tcp.wnd = t_beui16(MIN(0xFFFF, rxwnd));
+  p->tcp.wnd = t_beui16(MIN(0xFFFF, rxwnd >> window_scale));
   p->tcp.chksum = 0;
   p->tcp.urgp = t_beui16(0);
 
@@ -955,7 +957,7 @@ static void flow_tx_segment(struct dataplane_context *ctx,
 }
 
 static void flow_tx_ack(struct dataplane_context *ctx, uint32_t seq,
-    uint32_t ack, uint32_t rxwnd, uint32_t echots, uint32_t myts,
+    uint32_t ack, uint32_t rxwnd, uint32_t echots, uint32_t myts, uint8_t window_scale,
     struct network_buf_handle *nbh, struct tcp_timestamp_opt *ts_opt)
 {
   struct pkt_tcp *p;
@@ -998,7 +1000,7 @@ static void flow_tx_ack(struct dataplane_context *ctx, uint32_t seq,
   p->tcp.seqno = t_beui32(seq);
   p->tcp.ackno = t_beui32(ack);
   TCPH_HDRLEN_FLAGS_SET(&p->tcp, TCPH_HDRLEN(&p->tcp), TCP_ACK | ecn_flags);
-  p->tcp.wnd = t_beui16(MIN(0xFFFF, rxwnd));
+  p->tcp.wnd = t_beui16(MIN(0xFFFF, rxwnd >> window_scale));
   p->tcp.urgp = t_beui16(0);
 
   /* fill in timestamp option */

tas/include/config.h

@@ -63,6 +63,8 @@ struct configuration {
   uint32_t tcp_handshake_to;
   /** # of retries for dropped handshake packets */
   uint32_t tcp_handshake_retries;
+  /** Window scale configuration */
+  uint8_t tcp_window_scale;
   /** IP address for this host */
   uint32_t ip;
   /** IP prefix length for this host */

tas/slow/internal.h

@@ -117,32 +117,35 @@ enum nicif_connection_flags {
 /**
  * Register flow (must be called from poll thread).
  *
- * @param db          Doorbell ID
- * @param mac_remote  MAC address of the remote host
- * @param ip_local    Local IP address
- * @param port_local  Local port number
- * @param ip_remote   Remote IP address
- * @param port_remote Remote port number
- * @param rx_base     Base address of circular receive buffer
- * @param rx_len      Length of circular receive buffer
- * @param tx_base     Base address of circular transmit buffer
- * @param tx_len      Length of circular transmit buffer
- * @param remote_seq  Next sequence number expected from remote host
- * @param local_seq   Next sequence number for transmission
- * @param app_opaque  Opaque value to pass in notificaitions
- * @param flags       See #nicif_connection_flags.
- * @param rate        Congestion rate to set [Kbps]
- * @param fn_core     FlexNIC emulator core for the connection
- * @param flow_group  Flow group
- * @param pf_id       Pointer to location where flow id should be stored
+ * @param db                Doorbell ID
+ * @param mac_remote        MAC address of the remote host
+ * @param ip_local          Local IP address
+ * @param port_local        Local port number
+ * @param ip_remote         Remote IP address
+ * @param port_remote       Remote port number
+ * @param rx_base           Base address of circular receive buffer
+ * @param rx_len            Length of circular receive buffer
+ * @param tx_base           Base address of circular transmit buffer
+ * @param tx_len            Length of circular transmit buffer
+ * @param remote_seq        Next sequence number expected from remote host
+ * @param local_seq         Next sequence number for transmission
+ * @param app_opaque        Opaque value to pass in notificaitions
+ * @param flags             See #nicif_connection_flags.
+ * @param rate              Congestion rate to set [Kbps]
+ * @param rx_window_scale   Window scale factor for RX window
+ * @param tx_window_scale   Window scale factor for TX window
+ * @param fn_core           FlexNIC emulator core for the connection
+ * @param flow_group        Flow group
+ * @param pf_id             Pointer to location where flow id should be stored
  *
  * @return 0 on success, <0 else
  */
 int nicif_connection_add(uint32_t db, uint64_t mac_remote, uint32_t ip_local,
     uint16_t port_local, uint32_t ip_remote, uint16_t port_remote,
     uint64_t rx_base, uint32_t rx_len, uint64_t tx_base, uint32_t tx_len,
     uint32_t remote_seq, uint32_t local_seq, uint64_t app_opaque,
-    uint32_t flags, uint32_t rate, uint32_t fn_core, uint16_t flow_group,
+    uint32_t flags, uint32_t rate, uint8_t rx_window_scale, uint8_t tx_window_scale,
+    uint32_t fn_core, uint16_t flow_group,
     uint32_t *pf_id);
 
 /**
@@ -476,6 +479,10 @@ struct connection {
     uint32_t local_seq;
     /** Timestamp received with SYN/SYN-ACK packet */
     uint32_t syn_ts;
+    /** TX Window scale factor */
+    uint8_t tx_window_scale;
+    /** RX Window scale factor */
+    uint8_t rx_window_scale;
   /**@}*/
 
   /**

tas/slow/nicif.c

@@ -176,7 +176,8 @@ int nicif_connection_add(uint32_t db, uint64_t mac_remote, uint32_t ip_local,
     uint16_t port_local, uint32_t ip_remote, uint16_t port_remote,
     uint64_t rx_base, uint32_t rx_len, uint64_t tx_base, uint32_t tx_len,
     uint32_t remote_seq, uint32_t local_seq, uint64_t app_opaque,
-    uint32_t flags, uint32_t rate, uint32_t fn_core, uint16_t flow_group,
+    uint32_t flags, uint32_t rate, uint8_t rx_window_scale, uint8_t tx_window_scale,
+    uint32_t fn_core, uint16_t flow_group,
     uint32_t *pf_id)
 {
   struct flextcp_pl_flowst *fs;
@@ -222,6 +223,8 @@ int nicif_connection_add(uint32_t db, uint64_t mac_remote, uint32_t ip_local,
   fs->flow_group = flow_group;
   fs->lock = 0;
   fs->bump_seq = 0;
+  fs->tx_window_scale = tx_window_scale;
+  fs->rx_window_scale = rx_window_scale;
 
   fs->rx_avail = rx_len;
   fs->rx_next_pos = 0;