1 /*-
2 * Copyright (c) 2016-2020 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
26 #ifndef _NETINET_TCP_RACK_H_
27 #define _NETINET_TCP_RACK_H_
28
29 #define RACK_ACKED 0x0001/* The remote endpoint acked this */
30 #define RACK_TO_REXT 0x0002/* A timeout occured on this sendmap entry */
31 #define RACK_DEFERRED 0x0004/* We can't use this for RTT calc - not used */
32 #define RACK_OVERMAX 0x0008/* We have more retran's then we can fit */
33 #define RACK_SACK_PASSED 0x0010/* A sack was done above this block */
34 #define RACK_WAS_SACKPASS 0x0020/* We retransmitted due to SACK pass */
35 #define RACK_HAS_FIN 0x0040/* segment is sent with fin */
36 #define RACK_TLP 0x0080/* segment sent as tail-loss-probe */
37 #define RACK_RWND_COLLAPSED 0x0100/* The peer collapsed the rwnd on the segment */
38 #define RACK_APP_LIMITED 0x0200/* We went app limited after this send */
39 #define RACK_WAS_ACKED 0x0400/* a RTO undid the ack, but it already had a rtt calc done */
40 #define RACK_HAS_SYN 0x0800/* SYN is on this guy */
41 #define RACK_SENT_W_DSACK 0x1000/* Sent with a dsack */
42 #define RACK_SENT_SP 0x2000/* sent in slow path */
43 #define RACK_SENT_FP 0x4000/* sent in fast path */
44 #define RACK_HAD_PUSH 0x8000/* Push was sent on original send */
45 #define RACK_NUM_OF_RETRANS 3
46
47 #define RACK_INITIAL_RTO 1000000 /* 1 second in microseconds */
48
49 #define RACK_REQ_AVG 3 /* Must be less than 256 */
50
51 struct rack_sendmap {
52 TAILQ_ENTRY(rack_sendmap) r_tnext; /* Time of transmit based next */
53 uint32_t r_start; /* Sequence number of the segment */
54 uint32_t r_end; /* End seq, this is 1 beyond actually */
55 uint32_t r_rtr_bytes; /* How many bytes have been retransmitted */
56 uint16_t r_rtr_cnt; /* Retran count, index this -1 to get time
57 * sent */
58 uint16_t r_flags; /* Flags as defined above */
59 struct mbuf *m;
60 uint32_t soff;
61 uint32_t orig_m_len;
62 uint32_t r_nseq_appl; /* If this one is app limited, this is the nxt seq limited */
63 uint8_t r_dupack; /* Dup ack count */
64 uint8_t r_in_tmap; /* Flag to see if its in the r_tnext array */
65 uint8_t r_limit_type; /* is this entry counted against a limit? */
66 uint8_t r_just_ret : 1, /* After sending, the next pkt was just returned, i.e. limited */
67 r_one_out_nr : 1, /* Special case 1 outstanding and not in recovery */
68 r_no_rtt_allowed : 1, /* No rtt measurement allowed */
69 r_hw_tls : 1,
70 r_avail : 4;
71 uint64_t r_tim_lastsent[RACK_NUM_OF_RETRANS];
72 uint64_t r_ack_arrival; /* This is the time of ack-arrival (if SACK'd) */
73 RB_ENTRY(rack_sendmap) r_next; /* RB Tree next */
74 };
75
76 struct deferred_opt_list {
77 TAILQ_ENTRY(deferred_opt_list) next;
78 int optname;
79 uint64_t optval;
80 };
81
82 /*
83 * Timestamps in the rack sendmap are now moving to be
84 * uint64_t's. This means that if you want a uint32_t
85 * usec timestamp (the old usecond timestamp) you simply have
86 * to cast it to uint32_t. The reason we do this is not for
87 * wrap, but we need to get back, at times, to the millisecond
88 * timestamp that is used in the TSTMP option. To do this we
89 * can use the rack_ts_to_msec() inline below which can take
90 * the 64bit ts and make into the correct timestamp millisecond
91 * wise. Thats not possible with the 32bit usecond timestamp since
92 * the seconds wrap too quickly to cover all bases.
93 *
94 * There are quite a few places in rack where I simply cast
95 * back to uint32_t and then end up using the TSTMP_XX()
96 * macros. This is ok, but we could do simple compares if
97 * we ever decided to move all of those variables to 64 bits
98 * as well.
99 */
100
101 inline uint64_t
rack_to_usec_ts(struct timeval * tv)102 rack_to_usec_ts(struct timeval *tv)
103 {
104 return ((tv->tv_sec * HPTS_USEC_IN_SEC) + tv->tv_usec);
105 }
106
107 inline uint32_t
rack_ts_to_msec(uint64_t ts)108 rack_ts_to_msec(uint64_t ts)
109 {
110 return((uint32_t)(ts / HPTS_MSEC_IN_SEC));
111 }
112
113
114 RB_HEAD(rack_rb_tree_head, rack_sendmap);
115 TAILQ_HEAD(rack_head, rack_sendmap);
116 TAILQ_HEAD(def_opt_head, deferred_opt_list);
117
118 /* Map change logging */
119 #define MAP_MERGE 0x01
120 #define MAP_SPLIT 0x02
121 #define MAP_NEW 0x03
122 #define MAP_SACK_M1 0x04
123 #define MAP_SACK_M2 0x05
124 #define MAP_SACK_M3 0x06
125 #define MAP_SACK_M4 0x07
126 #define MAP_SACK_M5 0x08
127 #define MAP_FREE 0x09
128 #define MAP_TRIM_HEAD 0x0a
129
130 #define RACK_LIMIT_TYPE_SPLIT 1
131
132 /*
133 * We use the rate sample structure to
134 * assist in single sack/ack rate and rtt
135 * calculation. In the future we will expand
136 * this in BBR to do forward rate sample
137 * b/w estimation.
138 */
139 #define RACK_RTT_EMPTY 0x00000001 /* Nothing yet stored in RTT's */
140 #define RACK_RTT_VALID 0x00000002 /* We have at least one valid RTT */
141 struct rack_rtt_sample {
142 uint32_t rs_flags;
143 uint32_t rs_rtt_lowest;
144 uint32_t rs_rtt_highest;
145 uint32_t rs_rtt_cnt;
146 uint32_t rs_us_rtt;
147 int32_t confidence;
148 uint64_t rs_rtt_tot;
149 uint16_t rs_us_rtrcnt;
150 };
151
152 #define RACK_LOG_TYPE_ACK 0x01
153 #define RACK_LOG_TYPE_OUT 0x02
154 #define RACK_LOG_TYPE_TO 0x03
155 #define RACK_LOG_TYPE_ALLOC 0x04
156 #define RACK_LOG_TYPE_FREE 0x05
157
158 struct rack_log {
159 union {
160 struct rack_sendmap *rsm; /* For alloc/free */
161 uint64_t sb_acc;/* For out/ack or t-o */
162 };
163 uint32_t th_seq;
164 uint32_t th_ack;
165 uint32_t snd_una;
166 uint32_t snd_nxt; /* th_win for TYPE_ACK */
167 uint32_t snd_max;
168 uint32_t blk_start[4];
169 uint32_t blk_end[4];
170 uint8_t type;
171 uint8_t n_sackblks;
172 uint16_t len; /* Timeout T3=1, TLP=2, RACK=3 */
173 };
174
175 /*
176 * Magic numbers for logging timeout events if the
177 * logging is enabled.
178 */
179 #define RACK_TO_FRM_TMR 1
180 #define RACK_TO_FRM_TLP 2
181 #define RACK_TO_FRM_RACK 3
182 #define RACK_TO_FRM_KEEP 4
183 #define RACK_TO_FRM_PERSIST 5
184 #define RACK_TO_FRM_DELACK 6
185
186 struct rack_opts_stats {
187 uint64_t tcp_rack_tlp_reduce;
188 uint64_t tcp_rack_pace_always;
189 uint64_t tcp_rack_pace_reduce;
190 uint64_t tcp_rack_max_seg;
191 uint64_t tcp_rack_prr_sendalot;
192 uint64_t tcp_rack_min_to;
193 uint64_t tcp_rack_early_seg;
194 uint64_t tcp_rack_reord_thresh;
195 uint64_t tcp_rack_reord_fade;
196 uint64_t tcp_rack_tlp_thresh;
197 uint64_t tcp_rack_pkt_delay;
198 uint64_t tcp_rack_tlp_inc_var;
199 uint64_t tcp_tlp_use;
200 uint64_t tcp_rack_idle_reduce;
201 uint64_t tcp_rack_idle_reduce_high;
202 uint64_t rack_no_timer_in_hpts;
203 uint64_t tcp_rack_min_pace_seg;
204 uint64_t tcp_rack_pace_rate_ca;
205 uint64_t tcp_rack_rr;
206 uint64_t tcp_rack_do_detection;
207 uint64_t tcp_rack_rrr_no_conf_rate;
208 uint64_t tcp_initial_rate;
209 uint64_t tcp_initial_win;
210 uint64_t tcp_hdwr_pacing;
211 uint64_t tcp_gp_inc_ss;
212 uint64_t tcp_gp_inc_ca;
213 uint64_t tcp_gp_inc_rec;
214 uint64_t tcp_rack_force_max_seg;
215 uint64_t tcp_rack_pace_rate_ss;
216 uint64_t tcp_rack_pace_rate_rec;
217 /* Temp counters for dsack */
218 uint64_t tcp_sack_path_1;
219 uint64_t tcp_sack_path_2a;
220 uint64_t tcp_sack_path_2b;
221 uint64_t tcp_sack_path_3;
222 uint64_t tcp_sack_path_4;
223 /* non temp counters */
224 uint64_t tcp_rack_scwnd;
225 uint64_t tcp_rack_noprr;
226 uint64_t tcp_rack_cfg_rate;
227 uint64_t tcp_timely_dyn;
228 uint64_t tcp_rack_mbufq;
229 uint64_t tcp_fillcw;
230 uint64_t tcp_npush;
231 uint64_t tcp_lscwnd;
232 uint64_t tcp_profile;
233 uint64_t tcp_hdwr_rate_cap;
234 uint64_t tcp_pacing_rate_cap;
235 uint64_t tcp_pacing_up_only;
236 uint64_t tcp_use_cmp_acks;
237 uint64_t tcp_rack_abc_val;
238 uint64_t tcp_rec_abc_val;
239 uint64_t tcp_rack_measure_cnt;
240 uint64_t tcp_rack_delayed_ack;
241 uint64_t tcp_rack_rtt_use;
242 uint64_t tcp_data_after_close;
243 uint64_t tcp_defer_opt;
244 uint64_t tcp_rack_fastrsm_hack;
245 uint64_t tcp_rack_beta;
246 uint64_t tcp_rack_beta_ecn;
247 uint64_t tcp_rack_timer_slop;
248 };
249
250 /* RTT shrink reasons */
251 #define RACK_RTTS_INIT 0
252 #define RACK_RTTS_NEWRTT 1
253 #define RACK_RTTS_EXITPROBE 2
254 #define RACK_RTTS_ENTERPROBE 3
255 #define RACK_RTTS_REACHTARGET 4
256 #define RACK_RTTS_SEEHBP 5
257 #define RACK_RTTS_NOBACKOFF 6
258 #define RACK_RTTS_SAFETY 7
259
260 #define RACK_USE_BEG 1
261 #define RACK_USE_END 2
262 #define RACK_USE_END_OR_THACK 3
263
264 #define TLP_USE_ID 1 /* Internet draft behavior */
265 #define TLP_USE_TWO_ONE 2 /* Use 2.1 behavior */
266 #define TLP_USE_TWO_TWO 3 /* Use 2.2 behavior */
267 #define RACK_MIN_BW 8000 /* 64kbps in Bps */
268
269 /* Rack quality indicators for GPUT measurements */
270 #define RACK_QUALITY_NONE 0 /* No quality stated */
271 #define RACK_QUALITY_HIGH 1 /* A normal measurement of a GP RTT */
272 #define RACK_QUALITY_APPLIMITED 2 /* An app limited case that may be of lower quality */
273 #define RACK_QUALITY_PERSIST 3 /* A measurement where we went into persists */
274 #define RACK_QUALITY_PROBERTT 4 /* A measurement where we went into or exited probe RTT */
275 #define RACK_QUALITY_ALLACKED 5 /* All data is now acknowledged */
276
277
278 #define MIN_GP_WIN 6 /* We need at least 6 MSS in a GP measurement */
279 #ifdef _KERNEL
280 #define RACK_OPTS_SIZE (sizeof(struct rack_opts_stats)/sizeof(uint64_t))
281 extern counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
282 #define RACK_OPTS_ADD(name, amm) counter_u64_add(rack_opts_arry[(offsetof(struct rack_opts_stats, name)/sizeof(uint64_t))], (amm))
283 #define RACK_OPTS_INC(name) RACK_OPTS_ADD(name, 1)
284 #endif
285 /*
286 * As we get each SACK we wade through the
287 * rc_map and mark off what is acked.
288 * We also increment rc_sacked as well.
289 *
290 * We also pay attention to missing entries
291 * based on the time and possibly mark them
292 * for retransmit. If we do and we are not already
293 * in recovery we enter recovery. In doing
294 * so we claer prr_delivered/holes_rxt and prr_sent_dur_rec.
295 * We also setup rc_next/rc_snd_nxt/rc_send_end so
296 * we will know where to send from. When not in
297 * recovery rc_next will be NULL and rc_snd_nxt should
298 * equal snd_max.
299 *
300 * Whenever we retransmit from recovery we increment
301 * rc_holes_rxt as we retran a block and mark it as retransmitted
302 * with the time it was sent. During non-recovery sending we
303 * add to our map and note the time down of any send expanding
304 * the rc_map at the tail and moving rc_snd_nxt up with snd_max.
305 *
306 * In recovery during SACK/ACK processing if a chunk has
307 * been retransmitted and it is now acked, we decrement rc_holes_rxt.
308 * When we retransmit from the scoreboard we use
309 * rc_next and rc_snd_nxt/rc_send_end to help us
310 * find what needs to be retran.
311 *
312 * To calculate pipe we simply take (snd_max - snd_una) + rc_holes_rxt
313 * This gets us the effect of RFC6675 pipe, counting twice for
314 * bytes retransmitted.
315 */
316
317 #define TT_RACK_FR_TMR 0x2000
318
319 /*
320 * Locking for the rack control block.
321 * a) Locked by INP_WLOCK
322 * b) Locked by the hpts-mutex
323 *
324 */
325 #define RACK_GP_HIST 4 /* How much goodput history do we maintain? */
326
327 #define RACK_NUM_FSB_DEBUG 16
328 #ifdef _KERNEL
329 struct rack_fast_send_blk {
330 uint32_t left_to_send;
331 uint16_t tcp_ip_hdr_len;
332 uint8_t tcp_flags;
333 uint8_t hoplimit;
334 uint8_t *tcp_ip_hdr;
335 uint32_t recwin;
336 uint32_t off;
337 struct tcphdr *th;
338 struct udphdr *udp;
339 struct mbuf *m;
340 uint32_t o_m_len;
341 uint32_t rfo_apply_push : 1,
342 hw_tls : 1,
343 unused : 30;
344 };
345
346 struct rack_control {
347 /* Second cache line 0x40 from tcp_rack */
348 struct rack_rb_tree_head rc_mtree; /* Tree of all segments Lock(a) */
349 struct rack_head rc_tmap; /* List in transmit order Lock(a) */
350 struct rack_sendmap *rc_tlpsend; /* Remembered place for
351 * tlp_sending Lock(a) */
352 struct rack_sendmap *rc_resend; /* something we have been asked to
353 * resend */
354 struct rack_fast_send_blk fsb; /* The fast-send block */
355 uint32_t timer_slop;
356 uint32_t input_pkt;
357 uint32_t saved_input_pkt;
358 uint32_t rc_hpts_flags;
359 uint32_t rc_fixed_pacing_rate_ca;
360 uint32_t rc_fixed_pacing_rate_rec;
361 uint32_t rc_fixed_pacing_rate_ss;
362 uint32_t cwnd_to_use; /* The cwnd in use */
363 uint32_t rc_timer_exp; /* If a timer ticks of expiry */
364 uint32_t rc_rack_min_rtt; /* lowest RTT seen Lock(a) */
365 uint32_t rc_rack_largest_cwnd; /* Largest CWND we have seen Lock(a) */
366
367 /* Third Cache line 0x80 */
368 struct rack_head rc_free; /* Allocation array */
369 uint64_t last_hw_bw_req;
370 uint64_t crte_prev_rate;
371 uint64_t bw_rate_cap;
372 uint32_t rc_time_last_sent; /* Time we last sent some data and
373 * logged it Lock(a). */
374 uint32_t rc_reorder_ts; /* Last time we saw reordering Lock(a) */
375
376 uint32_t rc_tlp_new_data; /* we need to send new-data on a TLP
377 * Lock(a) */
378 uint32_t rc_prr_out; /* bytes sent during recovery Lock(a) */
379
380 uint32_t rc_prr_recovery_fs; /* recovery fs point Lock(a) */
381
382 uint32_t rc_prr_sndcnt; /* Prr sndcnt Lock(a) */
383
384 uint32_t rc_sacked; /* Tot sacked on scoreboard Lock(a) */
385 uint32_t xxx_rc_last_tlp_seq; /* Last tlp sequence Lock(a) */
386
387 uint32_t rc_prr_delivered; /* during recovery prr var Lock(a) */
388 uint16_t rc_tlp_cnt_out; /* count of times we have sent a TLP without new data */
389 uint16_t xxx_rc_tlp_seg_send_cnt; /* Number of times we have TLP sent
390 * rc_last_tlp_seq Lock(a) */
391
392 uint32_t rc_loss_count; /* How many bytes have been retransmitted
393 * Lock(a) */
394 uint32_t rc_reorder_fade; /* Socket option value Lock(a) */
395
396 /* Forth cache line 0xc0 */
397 /* Times */
398
399 uint32_t rc_rack_tmit_time; /* Rack transmit time Lock(a) */
400 uint32_t rc_holes_rxt; /* Tot retraned from scoreboard Lock(a) */
401
402 /* Variables to track bad retransmits and recover */
403 uint32_t rc_rsm_start; /* RSM seq number we retransmitted Lock(a) */
404 uint32_t rc_cwnd_at; /* cwnd at the retransmit Lock(a) */
405
406 uint32_t rc_ssthresh_at;/* ssthresh at the retransmit Lock(a) */
407 uint32_t rc_num_maps_alloced; /* Number of map blocks (sacks) we
408 * have allocated */
409 uint32_t rc_rcvtime; /* When we last received data */
410 uint32_t rc_num_split_allocs; /* num split map entries allocated */
411
412 uint32_t rc_last_output_to;
413 uint32_t rc_went_idle_time;
414
415 struct rack_sendmap *rc_sacklast; /* sack remembered place
416 * Lock(a) */
417
418 struct rack_sendmap *rc_rsm_at_retran; /* Debug variable kept for
419 * cache line alignment
420 * Lock(a) */
421 struct rack_sendmap *rc_first_appl; /* Pointer to first app limited */
422 struct rack_sendmap *rc_end_appl; /* Pointer to last app limited */
423 /* Cache line split 0x100 */
424 struct sack_filter rack_sf;
425 /* Cache line split 0x140 */
426 /* Flags for various things */
427 uint32_t last_pacing_time;
428 uint32_t rc_pace_max_segs;
429 uint32_t rc_pace_min_segs;
430 uint32_t rc_app_limited_cnt;
431 uint16_t rack_per_of_gp_ss; /* 100 = 100%, so from 65536 = 655 x bw */
432 uint16_t rack_per_of_gp_ca; /* 100 = 100%, so from 65536 = 655 x bw */
433 uint16_t rack_per_of_gp_rec; /* 100 = 100%, so from 65536 = 655 x bw, 0=off */
434 uint16_t rack_per_of_gp_probertt; /* 100 = 100%, so from 65536 = 655 x bw, 0=off */
435 uint32_t rc_high_rwnd;
436 uint32_t ack_count;
437 uint32_t sack_count;
438 uint32_t sack_noextra_move;
439 uint32_t sack_moved_extra;
440 struct rack_rtt_sample rack_rs;
441 const struct tcp_hwrate_limit_table *crte;
442 uint32_t rc_agg_early;
443 uint32_t rc_agg_delayed;
444 uint32_t rc_tlp_rxt_last_time;
445 uint32_t rc_saved_cwnd;
446 uint64_t rc_gp_output_ts; /* chg*/
447 uint64_t rc_gp_cumack_ts; /* chg*/
448 struct timeval act_rcv_time;
449 struct timeval rc_last_time_decay; /* SAD time decay happened here */
450 uint64_t gp_bw;
451 uint64_t init_rate;
452 #ifdef NETFLIX_SHARED_CWND
453 struct shared_cwnd *rc_scw;
454 #endif
455 uint64_t last_gp_comp_bw;
456 uint64_t last_max_bw; /* Our calculated max b/w last */
457 struct time_filter_small rc_gp_min_rtt;
458 struct def_opt_head opt_list;
459 int32_t rc_rtt_diff; /* Timely style rtt diff of our gp_srtt */
460 uint32_t rc_gp_srtt; /* Current GP srtt */
461 uint32_t rc_prev_gp_srtt; /* Previous RTT */
462 uint32_t rc_entry_gp_rtt; /* Entry to PRTT gp-rtt */
463 uint32_t rc_loss_at_start; /* At measurement window where was our lost value */
464
465 uint32_t forced_ack_ts;
466 uint32_t rc_lower_rtt_us_cts; /* Time our GP rtt was last lowered */
467 uint32_t rc_time_probertt_entered;
468 uint32_t rc_time_probertt_starts;
469 uint32_t rc_lowest_us_rtt;
470 uint32_t rc_highest_us_rtt;
471 uint32_t rc_last_us_rtt;
472 uint32_t rc_time_of_last_probertt;
473 uint32_t rc_target_probertt_flight;
474 uint32_t rc_probertt_sndmax_atexit; /* Highest sent to in probe-rtt */
475 uint32_t rc_cwnd_at_erec;
476 uint32_t rc_ssthresh_at_erec;
477 uint32_t dsack_byte_cnt;
478 uint32_t retran_during_recovery;
479 uint32_t rc_gp_lowrtt; /* Lowest rtt seen during GPUT measurement */
480 uint32_t rc_gp_high_rwnd; /* Highest rwnd seen during GPUT measurement */
481 uint32_t rc_snd_max_at_rto; /* For non-sack when the RTO occured what was snd-max */
482 uint32_t rc_out_at_rto;
483 int32_t rc_scw_index;
484 uint32_t rc_tlp_threshold; /* Socket option value Lock(a) */
485 uint32_t rc_last_timeout_snduna;
486 uint32_t challenge_ack_ts;
487 uint32_t challenge_ack_cnt;
488 uint32_t rc_min_to; /* Socket option value Lock(a) */
489 uint32_t rc_pkt_delay; /* Socket option value Lock(a) */
490 struct newreno rc_saved_beta; /*
491 * For newreno cc:
492 * rc_saved_cc are the values we have had
493 * set by the user, if pacing is not happening
494 * (i.e. its early and we have not turned on yet
495 * or it was turned off). The minute pacing
496 * is turned on we pull out the values currently
497 * being used by newreno and replace them with
498 * these values, then save off the old values here,
499 * we also set the flag (if ecn_beta is set) to make
500 * new_reno do less of a backoff for ecn (think abe).
501 */
502 uint16_t rc_early_recovery_segs; /* Socket option value Lock(a) */
503 uint16_t rc_reorder_shift; /* Socket option value Lock(a) */
504 uint8_t rc_no_push_at_mrtt; /* No push when we exceed max rtt */
505 uint8_t num_measurements; /* Number of measurements (up to 0xff, we freeze at 0xff) */
506 uint8_t req_measurements; /* How many measurements are required? */
507 uint8_t rc_tlp_cwnd_reduce; /* Socket option value Lock(a) */
508 uint8_t rc_prr_sendalot;/* Socket option value Lock(a) */
509 uint8_t rc_rate_sample_method;
510 uint8_t rc_gp_hist_idx;
511 };
512 #endif
513
514 #define RACK_TIMELY_CNT_BOOST 5 /* At 5th increase boost */
515 #define RACK_MINRTT_FILTER_TIM 10 /* Seconds */
516
517 #ifdef _KERNEL
518
519 struct tcp_rack {
520 /* First cache line 0x00 */
521 TAILQ_ENTRY(tcp_rack) r_hpts; /* hptsi queue next Lock(b) */
522 int32_t(*r_substate) (struct mbuf *, struct tcphdr *,
523 struct socket *, struct tcpcb *, struct tcpopt *,
524 int32_t, int32_t, uint32_t, int, int, uint8_t); /* Lock(a) */
525 struct tcpcb *rc_tp; /* The tcpcb Lock(a) */
526 struct inpcb *rc_inp; /* The inpcb Lock(a) */
527 uint8_t rc_free_cnt; /* Number of free entries on the rc_free list
528 * Lock(a) */
529 uint8_t client_bufferlvl; /* 0 - 5 normaly, less than or at 2 means its real low */
530 uint8_t no_prr_addback : 1,
531 gp_ready : 1,
532 defer_options: 1,
533 fast_rsm_hack: 1,
534 rc_ack_can_sendout_data: 1, /*
535 * If set it will override pacing restrictions on not sending
536 * data when the pacing timer is running. I.e. you set this
537 * and an ACK will send data. Default is off and its only used
538 * without pacing when we are doing 5G speed up for there
539 * ack filtering.
540 */
541 rc_pacing_cc_set: 1, /*
542 * If we are pacing (pace_always=1) and we have reached the
543 * point where we start pacing (fixed or gp has reached its
544 * magic gp_ready state) this flag indicates we have set in
545 * values to effect CC's backoff's. If pacing is turned off
546 * then we must restore the values saved in rc_saved_beta,
547 * if its going to gp_ready we need to copy the values into
548 * the CC module and set our flags.
549 *
550 * Note this only happens if the cc name is newreno (CCALGONAME_NEWRENO).
551 */
552
553 avail :2;
554 uint8_t avail_bytes;
555 uint32_t rc_rack_rtt; /* RACK-RTT Lock(a) */
556 uint16_t r_mbuf_queue : 1, /* Do we do mbuf queue for non-paced */
557 rtt_limit_mul : 4, /* muliply this by low rtt */
558 r_limit_scw : 1,
559 r_must_retran : 1, /* For non-sack customers we hit an RTO and new data should be resends */
560 r_use_cmp_ack: 1, /* Do we use compressed acks */
561 r_ent_rec_ns: 1, /* We entered recovery and have not sent */
562 r_might_revert: 1, /* Flag to find out if we might need to revert */
563 r_fast_output: 1, /* Fast output is in progress we can skip the bulk of rack_output */
564 r_fsb_inited: 1,
565 r_rack_hw_rate_caps: 1,
566 r_up_only: 1,
567 r_via_fill_cw : 1,
568 r_fill_less_agg : 1;
569
570 uint8_t rc_user_set_max_segs; /* Socket option value Lock(a) */
571 uint8_t rc_labc; /* Appropriate Byte Counting Value */
572 uint16_t forced_ack : 1,
573 rc_gp_incr : 1,
574 rc_gp_bwred : 1,
575 rc_gp_timely_inc_cnt : 3,
576 rc_gp_timely_dec_cnt : 3,
577 r_use_labc_for_rec: 1,
578 rc_highly_buffered: 1, /* The path is highly buffered */
579 rc_dragged_bottom: 1,
580 rc_dack_mode : 1, /* Mac O/S emulation of d-ack */
581 rc_dack_toggle : 1, /* For Mac O/S emulation of d-ack */
582 pacing_longer_than_rtt : 1,
583 rc_gp_filled : 1;
584 uint8_t r_state; /* Current rack state Lock(a) */
585 uint8_t rc_tmr_stopped : 7,
586 t_timers_stopped : 1;
587 uint8_t rc_enobuf : 7, /* count of enobufs on connection provides */
588 rc_on_min_to : 1;
589 uint8_t r_timer_override : 1, /* hpts override Lock(a) */
590 r_is_v6 : 1, /* V6 pcb Lock(a) */
591 rc_in_persist : 1,
592 rc_tlp_in_progress : 1,
593 rc_always_pace : 1, /* Socket option value Lock(a) */
594 rc_pace_to_cwnd : 1,
595 rc_pace_fill_if_rttin_range : 1,
596 rc_srtt_measure_made : 1;
597 uint8_t app_limited_needs_set : 1,
598 use_fixed_rate : 1,
599 rc_has_collapsed : 1,
600 r_rep_attack : 1,
601 r_rep_reverse : 1,
602 rack_hdrw_pacing : 1, /* We are doing Hardware pacing */
603 rack_hdw_pace_ena : 1, /* Is hardware pacing enabled? */
604 rack_attempt_hdwr_pace : 1; /* Did we attempt hdwr pacing (if allowed) */
605 uint8_t rack_tlp_threshold_use : 3, /* only 1, 2 and 3 used so far */
606 rack_rec_nonrxt_use_cr : 1,
607 rack_enable_scwnd : 1,
608 rack_attempted_scwnd : 1,
609 rack_no_prr : 1,
610 rack_scwnd_is_idle : 1;
611 uint8_t rc_allow_data_af_clo: 1,
612 delayed_ack : 1,
613 set_pacing_done_a_iw : 1,
614 use_rack_rr : 1,
615 alloc_limit_reported : 1,
616 sack_attack_disable : 1,
617 do_detection : 1,
618 rc_force_max_seg : 1;
619 uint8_t rack_cwnd_limited : 1,
620 r_early : 1,
621 r_late : 1,
622 r_running_early : 1,
623 r_running_late : 1,
624 r_wanted_output: 1,
625 r_rr_config : 2;
626 uint16_t rc_init_win : 8,
627 rc_gp_rtt_set : 1,
628 rc_gp_dyn_mul : 1,
629 rc_gp_saw_rec : 1,
630 rc_gp_saw_ca : 1,
631 rc_gp_saw_ss : 1,
632 rc_gp_no_rec_chg : 1,
633 in_probe_rtt : 1,
634 measure_saw_probe_rtt : 1;
635 /* Cache line 2 0x40 */
636 struct rack_control r_ctl;
637 } __aligned(CACHE_LINE_SIZE);
638
639 #endif
640 #endif
641