1 /*- 2 * Copyright (c) 2016-9 Netflix, Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 * $FreeBSD$ 26 */ 27 28 #ifndef _NETINET_TCP_RACK_H_ 29 #define _NETINET_TCP_RACK_H_ 30 31 #define RACK_ACKED 0x0001/* The remote endpoint acked this */ 32 #define RACK_TO_MIXED 0x0002/* A timeout occured that mixed the send order - not used */ 33 #define RACK_DEFERRED 0x0004/* We can't use this for RTT calc - not used */ 34 #define RACK_OVERMAX 0x0008/* We have more retran's then we can fit */ 35 #define RACK_SACK_PASSED 0x0010/* A sack was done above this block */ 36 #define RACK_WAS_SACKPASS 0x0020/* We retransmitted due to SACK pass */ 37 #define RACK_HAS_FIN 0x0040/* segment is sent with fin */ 38 #define RACK_TLP 0x0080/* segment sent as tail-loss-probe */ 39 #define RACK_RWND_COLLAPSED 0x0100/* The peer collapsed the rwnd on the segment */ 40 #define RACK_NUM_OF_RETRANS 3 41 42 #define RACK_INITIAL_RTO 1000 /* 1 second in milli seconds */ 43 44 struct rack_sendmap { 45 uint32_t r_start; /* Sequence number of the segment */ 46 uint32_t r_end; /* End seq, this is 1 beyond actually */ 47 TAILQ_ENTRY(rack_sendmap) r_tnext; /* Time of transmit based next */ 48 RB_ENTRY(rack_sendmap) r_next; /* RB Tree next */ 49 uint32_t r_rtr_bytes; /* How many bytes have been retransmitted */ 50 uint16_t r_rtr_cnt; /* Retran count, index this -1 to get time 51 * sent */ 52 uint16_t r_flags; /* Flags as defined above */ 53 uint32_t r_tim_lastsent[RACK_NUM_OF_RETRANS]; 54 uint8_t r_dupack; /* Dup ack count */ 55 uint8_t r_in_tmap; /* Flag to see if its in the r_tnext array */ 56 uint8_t r_limit_type; /* is this entry counted against a limit? */ 57 uint8_t r_resv[49]; 58 }; 59 60 RB_HEAD(rack_rb_tree_head, rack_sendmap); 61 TAILQ_HEAD(rack_head, rack_sendmap); 62 63 #define RACK_LIMIT_TYPE_SPLIT 1 64 65 /* 66 * We use the rate sample structure to 67 * assist in single sack/ack rate and rtt 68 * calculation. In the future we will expand 69 * this in BBR to do forward rate sample 70 * b/w estimation. 71 */ 72 #define RACK_RTT_EMPTY 0x00000001 /* Nothing yet stored in RTT's */ 73 #define RACK_RTT_VALID 0x00000002 /* We have at least one valid RTT */ 74 struct rack_rtt_sample { 75 uint32_t rs_flags; 76 uint32_t rs_rtt_lowest; 77 uint32_t rs_rtt_highest; 78 uint32_t rs_rtt_cnt; 79 uint64_t rs_rtt_tot; 80 }; 81 82 #define RACK_LOG_TYPE_ACK 0x01 83 #define RACK_LOG_TYPE_OUT 0x02 84 #define RACK_LOG_TYPE_TO 0x03 85 #define RACK_LOG_TYPE_ALLOC 0x04 86 #define RACK_LOG_TYPE_FREE 0x05 87 88 89 struct rack_log { 90 union { 91 struct rack_sendmap *rsm; /* For alloc/free */ 92 uint64_t sb_acc;/* For out/ack or t-o */ 93 }; 94 uint32_t th_seq; 95 uint32_t th_ack; 96 uint32_t snd_una; 97 uint32_t snd_nxt; /* th_win for TYPE_ACK */ 98 uint32_t snd_max; 99 uint32_t blk_start[4]; 100 uint32_t blk_end[4]; 101 uint8_t type; 102 uint8_t n_sackblks; 103 uint16_t len; /* Timeout T3=1, TLP=2, RACK=3 */ 104 }; 105 106 /* 107 * Magic numbers for logging timeout events if the 108 * logging is enabled. 109 */ 110 #define RACK_TO_FRM_TMR 1 111 #define RACK_TO_FRM_TLP 2 112 #define RACK_TO_FRM_RACK 3 113 #define RACK_TO_FRM_KEEP 4 114 #define RACK_TO_FRM_PERSIST 5 115 #define RACK_TO_FRM_DELACK 6 116 117 struct rack_opts_stats { 118 uint64_t tcp_rack_prop_rate; 119 uint64_t tcp_rack_prop; 120 uint64_t tcp_rack_tlp_reduce; 121 uint64_t tcp_rack_early_recov; 122 uint64_t tcp_rack_pace_always; 123 uint64_t tcp_rack_pace_reduce; 124 uint64_t tcp_rack_max_seg; 125 uint64_t tcp_rack_prr_sendalot; 126 uint64_t tcp_rack_min_to; 127 uint64_t tcp_rack_early_seg; 128 uint64_t tcp_rack_reord_thresh; 129 uint64_t tcp_rack_reord_fade; 130 uint64_t tcp_rack_tlp_thresh; 131 uint64_t tcp_rack_pkt_delay; 132 uint64_t tcp_rack_tlp_inc_var; 133 uint64_t tcp_tlp_use; 134 uint64_t tcp_rack_idle_reduce; 135 uint64_t tcp_rack_idle_reduce_high; 136 uint64_t rack_no_timer_in_hpts; 137 uint64_t tcp_rack_min_pace_seg; 138 uint64_t tcp_rack_min_pace; 139 uint64_t tcp_rack_cheat; 140 uint64_t tcp_rack_no_sack; 141 }; 142 143 #define TLP_USE_ID 1 /* Internet draft behavior */ 144 #define TLP_USE_TWO_ONE 2 /* Use 2.1 behavior */ 145 #define TLP_USE_TWO_TWO 3 /* Use 2.2 behavior */ 146 147 #ifdef _KERNEL 148 #define RACK_OPTS_SIZE (sizeof(struct rack_opts_stats)/sizeof(uint64_t)) 149 extern counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; 150 #define RACK_OPTS_ADD(name, amm) counter_u64_add(rack_opts_arry[(offsetof(struct rack_opts_stats, name)/sizeof(uint64_t))], (amm)) 151 #define RACK_OPTS_INC(name) RACK_OPTS_ADD(name, 1) 152 #endif 153 /* 154 * As we get each SACK we wade through the 155 * rc_map and mark off what is acked. 156 * We also increment rc_sacked as well. 157 * 158 * We also pay attention to missing entries 159 * based on the time and possibly mark them 160 * for retransmit. If we do and we are not already 161 * in recovery we enter recovery. In doing 162 * so we claer prr_delivered/holes_rxt and prr_sent_dur_rec. 163 * We also setup rc_next/rc_snd_nxt/rc_send_end so 164 * we will know where to send from. When not in 165 * recovery rc_next will be NULL and rc_snd_nxt should 166 * equal snd_max. 167 * 168 * Whenever we retransmit from recovery we increment 169 * rc_holes_rxt as we retran a block and mark it as retransmitted 170 * with the time it was sent. During non-recovery sending we 171 * add to our map and note the time down of any send expanding 172 * the rc_map at the tail and moving rc_snd_nxt up with snd_max. 173 * 174 * In recovery during SACK/ACK processing if a chunk has 175 * been retransmitted and it is now acked, we decrement rc_holes_rxt. 176 * When we retransmit from the scoreboard we use 177 * rc_next and rc_snd_nxt/rc_send_end to help us 178 * find what needs to be retran. 179 * 180 * To calculate pipe we simply take (snd_max - snd_una) + rc_holes_rxt 181 * This gets us the effect of RFC6675 pipe, counting twice for 182 * bytes retransmitted. 183 */ 184 185 #define TT_RACK_FR_TMR 0x2000 186 187 /* 188 * Locking for the rack control block. 189 * a) Locked by INP_WLOCK 190 * b) Locked by the hpts-mutex 191 * 192 */ 193 #define RACK_GP_HIST 4 /* How much goodput history do we maintain? */ 194 195 struct rack_control { 196 /* Second cache line 0x40 from tcp_rack */ 197 struct rack_rb_tree_head rc_mtree; /* Tree of all segments Lock(a) */ 198 struct rack_head rc_tmap; /* List in transmit order Lock(a) */ 199 struct rack_sendmap *rc_tlpsend; /* Remembered place for 200 * tlp_sending Lock(a) */ 201 struct rack_sendmap *rc_resend; /* something we have been asked to 202 * resend */ 203 struct timeval rc_last_time_decay; /* SAD time decay happened here */ 204 uint32_t input_pkt; 205 uint32_t saved_input_pkt; 206 uint32_t rc_hpts_flags; 207 uint32_t rc_timer_exp; /* If a timer ticks of expiry */ 208 uint32_t rc_rack_min_rtt; /* lowest RTT seen Lock(a) */ 209 uint32_t rc_rack_largest_cwnd; /* Largest CWND we have seen Lock(a) */ 210 211 /* Third Cache line 0x80 */ 212 struct rack_head rc_free; /* Allocation array */ 213 uint32_t rc_time_last_sent; /* Time we last sent some data and 214 * logged it Lock(a). */ 215 uint32_t rc_reorder_ts; /* Last time we saw reordering Lock(a) */ 216 217 uint32_t rc_tlp_new_data; /* we need to send new-data on a TLP 218 * Lock(a) */ 219 uint32_t rc_prr_out; /* bytes sent during recovery Lock(a) */ 220 221 uint32_t rc_prr_recovery_fs; /* recovery fs point Lock(a) */ 222 223 uint32_t rc_prr_sndcnt; /* Prr sndcnt Lock(a) */ 224 225 uint32_t rc_sacked; /* Tot sacked on scoreboard Lock(a) */ 226 uint32_t rc_last_tlp_seq; /* Last tlp sequence Lock(a) */ 227 228 uint32_t rc_prr_delivered; /* during recovery prr var Lock(a) */ 229 uint16_t rc_tlp_send_cnt; /* Number of TLP sends we have done 230 * since peer spoke to us Lock(a) */ 231 uint16_t rc_tlp_seg_send_cnt; /* Number of times we have TLP sent 232 * rc_last_tlp_seq Lock(a) */ 233 234 uint32_t rc_loss_count; /* During recovery how many segments were lost 235 * Lock(a) */ 236 uint32_t rc_reorder_fade; /* Socket option value Lock(a) */ 237 238 /* Forth cache line 0xc0 */ 239 /* Times */ 240 241 uint32_t rc_rack_tmit_time; /* Rack transmit time Lock(a) */ 242 uint32_t rc_holes_rxt; /* Tot retraned from scoreboard Lock(a) */ 243 244 /* Variables to track bad retransmits and recover */ 245 uint32_t rc_rsm_start; /* RSM seq number we retransmitted Lock(a) */ 246 uint32_t rc_cwnd_at; /* cwnd at the retransmit Lock(a) */ 247 248 uint32_t rc_ssthresh_at;/* ssthresh at the retransmit Lock(a) */ 249 uint32_t rc_num_maps_alloced; /* Number of map blocks (sacks) we 250 * have allocated */ 251 uint32_t rc_rcvtime; /* When we last received data */ 252 uint32_t rc_num_split_allocs; /* num split map entries allocated */ 253 254 uint32_t rc_last_output_to; 255 uint32_t rc_went_idle_time; 256 257 struct rack_sendmap *rc_sacklast; /* sack remembered place 258 * Lock(a) */ 259 260 struct rack_sendmap *rc_rsm_at_retran; /* Debug variable kept for 261 * cache line alignment 262 * Lock(a) */ 263 struct timeval rc_last_ack; 264 /* Cache line split 0x100 */ 265 struct sack_filter rack_sf; 266 /* Cache line split 0x140 */ 267 /* Flags for various things */ 268 uint32_t rc_pace_max_segs; 269 uint32_t rc_pace_min_segs; 270 uint32_t rc_high_rwnd; 271 uint32_t ack_count; 272 uint32_t sack_count; 273 uint32_t sack_noextra_move; 274 uint32_t sack_moved_extra; 275 struct rack_rtt_sample rack_rs; 276 uint32_t rc_tlp_rxt_last_time; 277 uint32_t rc_saved_cwnd; 278 uint32_t rc_gp_history[RACK_GP_HIST]; 279 uint32_t rc_tlp_threshold; /* Socket option value Lock(a) */ 280 uint16_t rc_early_recovery_segs; /* Socket option value Lock(a) */ 281 uint16_t rc_reorder_shift; /* Socket option value Lock(a) */ 282 uint16_t rc_pkt_delay; /* Socket option value Lock(a) */ 283 uint8_t rc_prop_rate; /* Socket option value Lock(a) */ 284 uint8_t rc_prop_reduce; /* Socket option value Lock(a) */ 285 uint8_t rc_tlp_cwnd_reduce; /* Socket option value Lock(a) */ 286 uint8_t rc_early_recovery; /* Socket option value Lock(a) */ 287 uint8_t rc_prr_sendalot;/* Socket option value Lock(a) */ 288 uint8_t rc_min_to; /* Socket option value Lock(a) */ 289 uint8_t rc_tlp_rtx_out; /* This is TLPRtxOut in the draft */ 290 uint8_t rc_rate_sample_method; 291 uint8_t rc_gp_hist_idx: 7, 292 rc_gp_hist_filled: 1; 293 294 }; 295 296 #ifdef _KERNEL 297 298 struct tcp_rack { 299 /* First cache line 0x00 */ 300 TAILQ_ENTRY(tcp_rack) r_hpts; /* hptsi queue next Lock(b) */ 301 int32_t(*r_substate) (struct mbuf *, struct tcphdr *, 302 struct socket *, struct tcpcb *, struct tcpopt *, 303 int32_t, int32_t, uint32_t, int, int); /* Lock(a) */ 304 struct tcpcb *rc_tp; /* The tcpcb Lock(a) */ 305 struct inpcb *rc_inp; /* The inpcb Lock(a) */ 306 uint32_t rc_free_cnt; /* Number of free entries on the rc_free list 307 * Lock(a) */ 308 uint32_t rc_rack_rtt; /* RACK-RTT Lock(a) */ 309 uint16_t r_wanted_output; /* Output routine wanted to be called */ 310 uint16_t r_cpu; /* CPU that the INP is running on Lock(a) */ 311 uint16_t rc_pace_max_segs; /* Socket option value Lock(a) */ 312 uint16_t rc_pace_reduce;/* Socket option value Lock(a) */ 313 314 uint8_t r_state; /* Current rack state Lock(a) */ 315 uint8_t rc_tmr_stopped : 7, 316 t_timers_stopped : 1; 317 uint8_t rc_enobuf; /* count of enobufs on connection provides 318 * backoff Lock(a) */ 319 uint8_t r_timer_override : 1, /* hpts override Lock(a) */ 320 r_tlp_running : 1, /* Running from a TLP timeout Lock(a) */ 321 r_is_v6 : 1, /* V6 pcb Lock(a) */ 322 rc_in_persist : 1, 323 rc_last_pto_set : 1, /* XXX not used */ 324 rc_tlp_in_progress : 1, 325 rc_always_pace : 1, /* Socket option value Lock(a) */ 326 tlp_timer_up : 1; /* The tlp timer is up flag Lock(a) */ 327 uint8_t r_enforce_min_pace : 2, 328 rc_has_collapsed : 1, 329 r_rep_attack : 1, 330 r_rep_reverse : 1, 331 r_xxx_min_pace_seg_thresh : 3; 332 uint8_t rack_tlp_threshold_use; 333 uint8_t rc_allow_data_af_clo: 1, 334 delayed_ack : 1, 335 set_pacing_done_a_iw : 1, 336 use_rack_cheat : 1, 337 alloc_limit_reported : 1, 338 sack_attack_disable : 1, 339 do_detection : 1, 340 rc_avail : 1; 341 uint16_t rack_per_of_gp; 342 /* Cache line 2 0x40 */ 343 struct rack_control r_ctl; 344 } __aligned(CACHE_LINE_SIZE); 345 346 #endif 347 #endif 348