diff options
| -rw-r--r-- | src/include/ipxe/tcp.h | 44 | ||||
| -rw-r--r-- | src/net/tcp.c | 162 |
2 files changed, 202 insertions, 4 deletions
diff --git a/src/include/ipxe/tcp.h b/src/include/ipxe/tcp.h index f12b4283f..65fee85e0 100644 --- a/src/include/ipxe/tcp.h +++ b/src/include/ipxe/tcp.h @@ -79,6 +79,48 @@ struct tcp_window_scale_padded_option { */ #define TCP_RX_WINDOW_SCALE 9 +/** TCP selective acknowledgement permitted option */ +struct tcp_sack_permitted_option { + uint8_t kind; + uint8_t length; +} __attribute__ (( packed )); + +/** Padded TCP selective acknowledgement permitted option (used for sending) */ +struct tcp_sack_permitted_padded_option { + uint8_t nop[2]; + struct tcp_sack_permitted_option spopt; +} __attribute__ (( packed )); + +/** Code for the TCP selective acknowledgement permitted option */ +#define TCP_OPTION_SACK_PERMITTED 4 + +/** TCP selective acknowledgement option */ +struct tcp_sack_option { + uint8_t kind; + uint8_t length; +} __attribute__ (( packed )); + +/** TCP selective acknowledgement block */ +struct tcp_sack_block { + uint32_t left; + uint32_t right; +} __attribute__ (( packed )); + +/** Maximum number of selective acknowledgement blocks + * + * This allows for the presence of the TCP timestamp option. + */ +#define TCP_SACK_MAX 3 + +/** Padded TCP selective acknowledgement option (used for sending) */ +struct tcp_sack_padded_option { + uint8_t nop[2]; + struct tcp_sack_option sackopt; +} __attribute__ (( packed )); + +/** Code for the TCP selective acknowledgement option */ +#define TCP_OPTION_SACK 5 + /** TCP timestamp option */ struct tcp_timestamp_option { uint8_t kind; @@ -102,6 +144,8 @@ struct tcp_options { const struct tcp_mss_option *mssopt; /** Window scale option, if present */ const struct tcp_window_scale_option *wsopt; + /** SACK permitted option, if present */ + const struct tcp_sack_permitted_option *spopt; /** Timestamp option, if present */ const struct tcp_timestamp_option *tsopt; }; diff --git a/src/net/tcp.c b/src/net/tcp.c index 719c99360..aeacfd79e 100644 --- a/src/net/tcp.c +++ b/src/net/tcp.c @@ -104,6 +104,9 @@ struct tcp_connection { /** Maximum receive window */ uint32_t max_rcv_win; + /** Selective acknowledgement list (in host-endian order) */ + struct tcp_sack_block sack[TCP_SACK_MAX]; + /** Transmit queue */ struct list_head tx_queue; /** Receive queue */ @@ -129,6 +132,8 @@ enum tcp_flags { TCP_TS_ENABLED = 0x0002, /** TCP acknowledgement is pending */ TCP_ACK_PENDING = 0x0004, + /** TCP selective acknowledgement is enabled */ + TCP_SACK_ENABLED = 0x0008, }; /** TCP internal header @@ -143,6 +148,8 @@ struct tcp_rx_queued_header { * enqueued, and so excludes the SYN, if present. */ uint32_t seq; + /** Next SEQ value, in host-endian order */ + uint32_t nxt; /** Flags * * Only FIN is valid within this flags byte; all other flags @@ -450,6 +457,94 @@ static size_t tcp_xfer_window ( struct tcp_connection *tcp ) { } /** + * Find selective acknowledgement block + * + * @v tcp TCP connection + * @v seq SEQ value in SACK block (in host-endian order) + * @v sack SACK block to fill in (in host-endian order) + * @ret len Length of SACK block + */ +static uint32_t tcp_sack_block ( struct tcp_connection *tcp, uint32_t seq, + struct tcp_sack_block *sack ) { + struct io_buffer *iobuf; + struct tcp_rx_queued_header *tcpqhdr; + uint32_t left = tcp->rcv_ack; + uint32_t right = left; + + /* Find highest block which does not start after SEQ */ + list_for_each_entry ( iobuf, &tcp->rx_queue, list ) { + tcpqhdr = iobuf->data; + if ( tcp_cmp ( tcpqhdr->seq, right ) > 0 ) { + if ( tcp_cmp ( tcpqhdr->seq, seq ) > 0 ) + break; + left = tcpqhdr->seq; + } + if ( tcp_cmp ( tcpqhdr->nxt, right ) > 0 ) + right = tcpqhdr->nxt; + } + + /* Fail if this block does not contain SEQ */ + if ( tcp_cmp ( right, seq ) < 0 ) + return 0; + + /* Populate SACK block */ + sack->left = left; + sack->right = right; + return ( right - left ); +} + +/** + * Update TCP selective acknowledgement list + * + * @v tcp TCP connection + * @v seq SEQ value in first SACK block (in host-endian order) + * @ret count Number of SACK blocks + */ +static unsigned int tcp_sack ( struct tcp_connection *tcp, uint32_t seq ) { + struct tcp_sack_block sack[TCP_SACK_MAX]; + unsigned int old = 0; + unsigned int new = 0; + unsigned int i; + uint32_t len; + + /* Populate first new SACK block */ + len = tcp_sack_block ( tcp, seq, &sack[0] ); + if ( len ) + new++; + + /* Populate remaining new SACK blocks based on old SACK blocks */ + for ( old = 0 ; old < TCP_SACK_MAX ; old++ ) { + + /* Stop if we run out of space in the new list */ + if ( new == TCP_SACK_MAX ) + break; + + /* Skip empty old SACK blocks */ + if ( tcp->sack[old].left == tcp->sack[old].right ) + continue; + + /* Populate new SACK block */ + len = tcp_sack_block ( tcp, tcp->sack[old].left, &sack[new] ); + if ( len == 0 ) + continue; + + /* Eliminate duplicates */ + for ( i = 0 ; i < new ; i++ ) { + if ( sack[i].left == sack[new].left ) { + new--; + break; + } + } + new++; + } + + /* Update SACK list */ + memset ( tcp->sack, 0, sizeof ( tcp->sack ) ); + memcpy ( tcp->sack, sack, ( new * sizeof ( tcp->sack[0] ) ) ); + return new; +} + +/** * Process TCP transmit queue * * @v tcp TCP connection @@ -493,9 +588,10 @@ static size_t tcp_process_tx_queue ( struct tcp_connection *tcp, size_t max_len, } /** - * Transmit any outstanding data + * Transmit any outstanding data (with selective acknowledgement) * * @v tcp TCP connection + * @v sack_seq SEQ for first selective acknowledgement (if any) * * Transmits any outstanding data on the connection. * @@ -503,15 +599,21 @@ static size_t tcp_process_tx_queue ( struct tcp_connection *tcp, size_t max_len, * will have been started if necessary, and so the stack will * eventually attempt to retransmit the failed packet. */ -static void tcp_xmit ( struct tcp_connection *tcp ) { +static void tcp_xmit_sack ( struct tcp_connection *tcp, uint32_t sack_seq ) { struct io_buffer *iobuf; struct tcp_header *tcphdr; struct tcp_mss_option *mssopt; struct tcp_window_scale_padded_option *wsopt; struct tcp_timestamp_padded_option *tsopt; + struct tcp_sack_permitted_padded_option *spopt; + struct tcp_sack_padded_option *sackopt; + struct tcp_sack_block *sack; void *payload; unsigned int flags; + unsigned int sack_count; + unsigned int i; size_t len = 0; + size_t sack_len; uint32_t seq_len; uint32_t app_win; uint32_t max_rcv_win; @@ -590,6 +692,10 @@ static void tcp_xmit ( struct tcp_connection *tcp ) { wsopt->wsopt.kind = TCP_OPTION_WS; wsopt->wsopt.length = sizeof ( wsopt->wsopt ); wsopt->wsopt.scale = TCP_RX_WINDOW_SCALE; + spopt = iob_push ( iobuf, sizeof ( *spopt ) ); + memset ( spopt->nop, TCP_OPTION_NOP, sizeof ( spopt ) ); + spopt->spopt.kind = TCP_OPTION_SACK_PERMITTED; + spopt->spopt.length = sizeof ( spopt->spopt ); } if ( ( flags & TCP_SYN ) || ( tcp->flags & TCP_TS_ENABLED ) ) { tsopt = iob_push ( iobuf, sizeof ( *tsopt ) ); @@ -599,6 +705,21 @@ static void tcp_xmit ( struct tcp_connection *tcp ) { tsopt->tsopt.tsval = htonl ( currticks() ); tsopt->tsopt.tsecr = htonl ( tcp->ts_recent ); } + if ( ( tcp->flags & TCP_SACK_ENABLED ) && + ( ! list_empty ( &tcp->rx_queue ) ) && + ( ( sack_count = tcp_sack ( tcp, sack_seq ) ) != 0 ) ) { + sack_len = ( sack_count * sizeof ( *sack ) ); + sackopt = iob_push ( iobuf, ( sizeof ( *sackopt ) + sack_len )); + memset ( sackopt->nop, TCP_OPTION_NOP, sizeof ( sackopt->nop )); + sackopt->sackopt.kind = TCP_OPTION_SACK; + sackopt->sackopt.length = + ( sizeof ( sackopt->sackopt ) + sack_len ); + sack = ( ( ( void * ) sackopt ) + sizeof ( *sackopt ) ); + for ( i = 0 ; i < sack_count ; i++, sack++ ) { + sack->left = htonl ( tcp->sack[i].left ); + sack->right = htonl ( tcp->sack[i].right ); + } + } if ( len != 0 ) flags |= TCP_PSH; tcphdr = iob_push ( iobuf, sizeof ( *tcphdr ) ); @@ -635,6 +756,17 @@ static void tcp_xmit ( struct tcp_connection *tcp ) { profile_stop ( &tcp_tx_profiler ); } +/** + * Transmit any outstanding data + * + * @v tcp TCP connection + */ +static void tcp_xmit ( struct tcp_connection *tcp ) { + + /* Transmit without an explicit first SACK */ + tcp_xmit_sack ( tcp, tcp->rcv_ack ); +} + /** TCP process descriptor */ static struct process_descriptor tcp_process_desc = PROC_DESC_ONCE ( struct tcp_connection, process, tcp_xmit ); @@ -804,6 +936,12 @@ static void tcp_rx_opts ( struct tcp_connection *tcp, const void *data, case TCP_OPTION_WS: options->wsopt = data; break; + case TCP_OPTION_SACK_PERMITTED: + options->spopt = data; + break; + case TCP_OPTION_SACK: + /* Ignore received SACKs */ + break; case TCP_OPTION_TS: options->tsopt = data; break; @@ -823,6 +961,7 @@ static void tcp_rx_opts ( struct tcp_connection *tcp, const void *data, * @v seq_len Sequence space length to consume */ static void tcp_rx_seq ( struct tcp_connection *tcp, uint32_t seq_len ) { + unsigned int sack; /* Sanity check */ assert ( seq_len > 0 ); @@ -840,6 +979,16 @@ static void tcp_rx_seq ( struct tcp_connection *tcp, uint32_t seq_len ) { /* Update timestamp */ tcp->ts_recent = tcp->ts_val; + /* Update SACK list */ + for ( sack = 0 ; sack < TCP_SACK_MAX ; sack++ ) { + if ( tcp->sack[sack].left == tcp->sack[sack].right ) + continue; + if ( tcp_cmp ( tcp->sack[sack].left, tcp->rcv_ack ) < 0 ) + tcp->sack[sack].left = tcp->rcv_ack; + if ( tcp_cmp ( tcp->sack[sack].right, tcp->rcv_ack ) < 0 ) + tcp->sack[sack].right = tcp->rcv_ack; + } + /* Mark ACK as pending */ tcp->flags |= TCP_ACK_PENDING; } @@ -860,6 +1009,8 @@ static int tcp_rx_syn ( struct tcp_connection *tcp, uint32_t seq, tcp->rcv_ack = seq; if ( options->tsopt ) tcp->flags |= TCP_TS_ENABLED; + if ( options->spopt ) + tcp->flags |= TCP_SACK_ENABLED; if ( options->wsopt ) { tcp->snd_win_scale = options->wsopt->scale; tcp->rcv_win_scale = TCP_RX_WINDOW_SCALE; @@ -1070,6 +1221,7 @@ static void tcp_rx_enqueue ( struct tcp_connection *tcp, uint32_t seq, struct io_buffer *queued; size_t len; uint32_t seq_len; + uint32_t nxt; /* Calculate remaining flags and sequence length. Note that * SYN, if present, has already been processed by this point. @@ -1077,6 +1229,7 @@ static void tcp_rx_enqueue ( struct tcp_connection *tcp, uint32_t seq, flags &= TCP_FIN; len = iob_len ( iobuf ); seq_len = ( len + ( flags ? 1 : 0 ) ); + nxt = ( seq + seq_len ); /* Discard immediately (to save memory) if: * @@ -1087,7 +1240,7 @@ static void tcp_rx_enqueue ( struct tcp_connection *tcp, uint32_t seq, */ if ( ( ! ( tcp->tcp_state & TCP_STATE_RCVD ( TCP_SYN ) ) ) || ( tcp_cmp ( seq, tcp->rcv_ack + tcp->rcv_win ) >= 0 ) || - ( tcp_cmp ( seq + seq_len, tcp->rcv_ack ) < 0 ) || + ( tcp_cmp ( nxt, tcp->rcv_ack ) < 0 ) || ( seq_len == 0 ) ) { free_iob ( iobuf ); return; @@ -1096,6 +1249,7 @@ static void tcp_rx_enqueue ( struct tcp_connection *tcp, uint32_t seq, /* Add internal header */ tcpqhdr = iob_push ( iobuf, sizeof ( *tcpqhdr ) ); tcpqhdr->seq = seq; + tcpqhdr->nxt = nxt; tcpqhdr->flags = flags; /* Add to RX queue */ @@ -1289,7 +1443,7 @@ static int tcp_rx ( struct io_buffer *iobuf, if ( list_empty ( &tcp->rx_queue ) ) { process_add ( &tcp->process ); } else { - tcp_xmit ( tcp ); + tcp_xmit_sack ( tcp, seq ); } /* If this packet was the last we expect to receive, set up |