xref: /freebsd/sys/netinet/tcp_stacks/tcp_rack.h (revision 79ac3c12a714bcd3f2354c52d948aed9575c46d6)
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  * $FreeBSD$
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_REXT	    0x0002/* A timeout occured on this sendmap entry */
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_APP_LIMITED    0x0200/* We went app limited after this send */
41 #define RACK_WAS_ACKED	    0x0400/* a RTO undid the ack, but it already had a rtt calc done */
42 #define RACK_HAS_SYN	    0x0800/* SYN is on this guy */
43 #define RACK_SENT_W_DSACK   0x1000/* Sent with a dsack */
44 #define RACK_SENT_SP	    0x2000/* sent in slow path */
45 #define RACK_SENT_FP        0x4000/* sent in fast path */
46 #define RACK_HAD_PUSH	    0x8000/* Push was sent on original send */
47 #define RACK_NUM_OF_RETRANS 3
48 
49 #define RACK_INITIAL_RTO 1000000 /* 1 second in microseconds */
50 
51 #define RACK_REQ_AVG 3 	/* Must be less than 256 */
52 
53 struct rack_sendmap {
54 	TAILQ_ENTRY(rack_sendmap) r_tnext;	/* Time of transmit based next */
55 	uint32_t r_start;	/* Sequence number of the segment */
56 	uint32_t r_end;		/* End seq, this is 1 beyond actually */
57 	uint32_t r_rtr_bytes;	/* How many bytes have been retransmitted */
58 	uint16_t r_rtr_cnt;	/* Retran count, index this -1 to get time
59 				 * sent */
60 	uint16_t r_flags;	/* Flags as defined above */
61 	struct mbuf *m;
62 	uint32_t soff;
63 	uint32_t orig_m_len;
64 	uint32_t r_nseq_appl;	/* If this one is app limited, this is the nxt seq limited */
65 	uint8_t r_dupack;	/* Dup ack count */
66 	uint8_t r_in_tmap;	/* Flag to see if its in the r_tnext array */
67 	uint8_t r_limit_type;	/* is this entry counted against a limit? */
68 	uint8_t r_just_ret : 1, /* After sending, the next pkt was just returned, i.e. limited  */
69 		r_one_out_nr : 1,	/* Special case 1 outstanding and not in recovery */
70 		r_no_rtt_allowed : 1, /* No rtt measurement allowed */
71 		r_avail : 5;
72 	uint64_t r_tim_lastsent[RACK_NUM_OF_RETRANS];
73 	uint64_t r_ack_arrival;	/* This is the time of ack-arrival (if SACK'd) */
74 	RB_ENTRY(rack_sendmap) r_next;		/* RB Tree next */
75 };
76 
77 struct deferred_opt_list {
78 	TAILQ_ENTRY(deferred_opt_list) next;
79 	int optname;
80 	uint64_t optval;
81 };
82 
83 /*
84  * Timestamps in the rack sendmap are now moving to be
85  * uint64_t's. This means that if you want a uint32_t
86  * usec timestamp (the old usecond timestamp) you simply have
87  * to cast it to uint32_t. The reason we do this is not for
88  * wrap, but we need to get back, at times, to the millisecond
89  * timestamp that is used in the TSTMP option. To do this we
90  * can use the rack_ts_to_msec() inline below which can take
91  * the 64bit ts and make into the correct timestamp millisecond
92  * wise. Thats not possible with the 32bit usecond timestamp since
93  * the seconds wrap too quickly to cover all bases.
94  *
95  * There are quite a few places in rack where I simply cast
96  * back to uint32_t and then end up using the TSTMP_XX()
97  * macros. This is ok, but we could do simple compares if
98  * we ever decided to move all of those variables to 64 bits
99  * as well.
100  */
101 
102 inline uint64_t
103 rack_to_usec_ts(struct timeval *tv)
104 {
105 	return ((tv->tv_sec * HPTS_USEC_IN_SEC) + tv->tv_usec);
106 }
107 
108 inline uint32_t
109 rack_ts_to_msec(uint64_t ts)
110 {
111 	return((uint32_t)(ts / HPTS_MSEC_IN_SEC));
112 }
113 
114 
115 RB_HEAD(rack_rb_tree_head, rack_sendmap);
116 TAILQ_HEAD(rack_head, rack_sendmap);
117 TAILQ_HEAD(def_opt_head, deferred_opt_list);
118 
119 /* Map change logging */
120 #define MAP_MERGE	0x01
121 #define MAP_SPLIT	0x02
122 #define MAP_NEW		0x03
123 #define MAP_SACK_M1	0x04
124 #define MAP_SACK_M2	0x05
125 #define MAP_SACK_M3	0x06
126 #define MAP_SACK_M4	0x07
127 #define MAP_SACK_M5	0x08
128 #define MAP_FREE	0x09
129 #define MAP_TRIM_HEAD	0x0a
130 
131 #define RACK_LIMIT_TYPE_SPLIT	1
132 
133 /*
134  * We use the rate sample structure to
135  * assist in single sack/ack rate and rtt
136  * calculation. In the future we will expand
137  * this in BBR to do forward rate sample
138  * b/w estimation.
139  */
140 #define RACK_RTT_EMPTY 0x00000001	/* Nothing yet stored in RTT's */
141 #define RACK_RTT_VALID 0x00000002	/* We have at least one valid RTT */
142 struct rack_rtt_sample {
143 	uint32_t rs_flags;
144 	uint32_t rs_rtt_lowest;
145 	uint32_t rs_rtt_highest;
146 	uint32_t rs_rtt_cnt;
147 	uint32_t rs_us_rtt;
148 	int32_t  confidence;
149 	uint64_t rs_rtt_tot;
150 	uint16_t rs_us_rtrcnt;
151 };
152 
153 #define RACK_LOG_TYPE_ACK	0x01
154 #define RACK_LOG_TYPE_OUT	0x02
155 #define RACK_LOG_TYPE_TO	0x03
156 #define RACK_LOG_TYPE_ALLOC     0x04
157 #define RACK_LOG_TYPE_FREE      0x05
158 
159 struct rack_log {
160 	union {
161 		struct rack_sendmap *rsm;	/* For alloc/free */
162 		uint64_t sb_acc;/* For out/ack or t-o */
163 	};
164 	uint32_t th_seq;
165 	uint32_t th_ack;
166 	uint32_t snd_una;
167 	uint32_t snd_nxt;	/* th_win for TYPE_ACK */
168 	uint32_t snd_max;
169 	uint32_t blk_start[4];
170 	uint32_t blk_end[4];
171 	uint8_t type;
172 	uint8_t n_sackblks;
173 	uint16_t len;		/* Timeout T3=1, TLP=2, RACK=3 */
174 };
175 
176 /*
177  * Magic numbers for logging timeout events if the
178  * logging is enabled.
179  */
180 #define RACK_TO_FRM_TMR  1
181 #define RACK_TO_FRM_TLP  2
182 #define RACK_TO_FRM_RACK 3
183 #define RACK_TO_FRM_KEEP 4
184 #define RACK_TO_FRM_PERSIST 5
185 #define RACK_TO_FRM_DELACK 6
186 
187 struct rack_opts_stats {
188 	uint64_t tcp_rack_tlp_reduce;
189 	uint64_t tcp_rack_pace_always;
190 	uint64_t tcp_rack_pace_reduce;
191 	uint64_t tcp_rack_max_seg;
192 	uint64_t tcp_rack_prr_sendalot;
193 	uint64_t tcp_rack_min_to;
194 	uint64_t tcp_rack_early_seg;
195 	uint64_t tcp_rack_reord_thresh;
196 	uint64_t tcp_rack_reord_fade;
197 	uint64_t tcp_rack_tlp_thresh;
198 	uint64_t tcp_rack_pkt_delay;
199 	uint64_t tcp_rack_tlp_inc_var;
200 	uint64_t tcp_tlp_use;
201 	uint64_t tcp_rack_idle_reduce;
202 	uint64_t tcp_rack_idle_reduce_high;
203 	uint64_t rack_no_timer_in_hpts;
204 	uint64_t tcp_rack_min_pace_seg;
205 	uint64_t tcp_rack_pace_rate_ca;
206 	uint64_t tcp_rack_rr;
207 	uint64_t tcp_rack_do_detection;
208 	uint64_t tcp_rack_rrr_no_conf_rate;
209 	uint64_t tcp_initial_rate;
210 	uint64_t tcp_initial_win;
211 	uint64_t tcp_hdwr_pacing;
212 	uint64_t tcp_gp_inc_ss;
213 	uint64_t tcp_gp_inc_ca;
214 	uint64_t tcp_gp_inc_rec;
215 	uint64_t tcp_rack_force_max_seg;
216 	uint64_t tcp_rack_pace_rate_ss;
217 	uint64_t tcp_rack_pace_rate_rec;
218 	/* Temp counters for dsack */
219 	uint64_t tcp_sack_path_1;
220 	uint64_t tcp_sack_path_2a;
221 	uint64_t tcp_sack_path_2b;
222 	uint64_t tcp_sack_path_3;
223 	uint64_t tcp_sack_path_4;
224 	/* non temp counters */
225 	uint64_t tcp_rack_scwnd;
226 	uint64_t tcp_rack_noprr;
227 	uint64_t tcp_rack_cfg_rate;
228 	uint64_t tcp_timely_dyn;
229 	uint64_t tcp_rack_mbufq;
230 	uint64_t tcp_fillcw;
231 	uint64_t tcp_npush;
232 	uint64_t tcp_lscwnd;
233 	uint64_t tcp_profile;
234 	uint64_t tcp_hdwr_rate_cap;
235 	uint64_t tcp_pacing_rate_cap;
236 	uint64_t tcp_pacing_up_only;
237 	uint64_t tcp_use_cmp_acks;
238 	uint64_t tcp_rack_abc_val;
239 	uint64_t tcp_rec_abc_val;
240 	uint64_t tcp_rack_measure_cnt;
241 	uint64_t tcp_rack_delayed_ack;
242 	uint64_t tcp_rack_rtt_use;
243 	uint64_t tcp_data_after_close;
244 	uint64_t tcp_defer_opt;
245 	uint64_t tcp_rack_fastrsm_hack;
246 	uint64_t tcp_rack_beta;
247 	uint64_t tcp_rack_beta_ecn;
248 	uint64_t tcp_rack_timer_slop;
249 };
250 
251 /* RTT shrink reasons */
252 #define RACK_RTTS_INIT     0
253 #define RACK_RTTS_NEWRTT   1
254 #define RACK_RTTS_EXITPROBE 2
255 #define RACK_RTTS_ENTERPROBE 3
256 #define RACK_RTTS_REACHTARGET 4
257 #define RACK_RTTS_SEEHBP 5
258 #define RACK_RTTS_NOBACKOFF 6
259 #define RACK_RTTS_SAFETY 7
260 
261 #define RACK_USE_BEG 1
262 #define RACK_USE_END 2
263 #define RACK_USE_END_OR_THACK 3
264 
265 #define TLP_USE_ID	1	/* Internet draft behavior */
266 #define TLP_USE_TWO_ONE 2	/* Use 2.1 behavior */
267 #define TLP_USE_TWO_TWO 3	/* Use 2.2 behavior */
268 #define RACK_MIN_BW 8000	/* 64kbps in Bps */
269 
270 #define MIN_GP_WIN 6	/* We need at least 6 MSS in a GP measurement */
271 #ifdef _KERNEL
272 #define RACK_OPTS_SIZE (sizeof(struct rack_opts_stats)/sizeof(uint64_t))
273 extern counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
274 #define RACK_OPTS_ADD(name, amm) counter_u64_add(rack_opts_arry[(offsetof(struct rack_opts_stats, name)/sizeof(uint64_t))], (amm))
275 #define RACK_OPTS_INC(name) RACK_OPTS_ADD(name, 1)
276 #endif
277 /*
278  * As we get each SACK we wade through the
279  * rc_map and mark off what is acked.
280  * We also increment rc_sacked as well.
281  *
282  * We also pay attention to missing entries
283  * based on the time and possibly mark them
284  * for retransmit. If we do and we are not already
285  * in recovery we enter recovery. In doing
286  * so we claer prr_delivered/holes_rxt and prr_sent_dur_rec.
287  * We also setup rc_next/rc_snd_nxt/rc_send_end so
288  * we will know where to send from. When not in
289  * recovery rc_next will be NULL and rc_snd_nxt should
290  * equal snd_max.
291  *
292  * Whenever we retransmit from recovery we increment
293  * rc_holes_rxt as we retran a block and mark it as retransmitted
294  * with the time it was sent. During non-recovery sending we
295  * add to our map and note the time down of any send expanding
296  * the rc_map at the tail and moving rc_snd_nxt up with snd_max.
297  *
298  * In recovery during SACK/ACK processing if a chunk has
299  * been retransmitted and it is now acked, we decrement rc_holes_rxt.
300  * When we retransmit from the scoreboard we use
301  * rc_next and rc_snd_nxt/rc_send_end to help us
302  * find what needs to be retran.
303  *
304  * To calculate pipe we simply take (snd_max - snd_una) + rc_holes_rxt
305  * This gets us the effect of RFC6675 pipe, counting twice for
306  * bytes retransmitted.
307  */
308 
309 #define TT_RACK_FR_TMR	0x2000
310 
311 /*
312  * Locking for the rack control block.
313  * a) Locked by INP_WLOCK
314  * b) Locked by the hpts-mutex
315  *
316  */
317 #define RACK_GP_HIST 4	/* How much goodput history do we maintain? */
318 
319 #define RACK_NUM_FSB_DEBUG 16
320 struct rack_fast_send_blk {
321 	uint32_t left_to_send;
322 	uint16_t tcp_ip_hdr_len;
323 	uint8_t tcp_flags;
324 	uint8_t hoplimit;
325 	uint8_t *tcp_ip_hdr;
326 	uint32_t recwin;
327 	uint32_t off;
328 	struct tcphdr *th;
329 	struct udphdr *udp;
330 	struct mbuf *m;
331 	uint32_t o_m_len;
332 	uint32_t rfo_apply_push : 1,
333 		unused : 31;
334 };
335 
336 struct rack_control {
337 	/* Second cache line 0x40 from tcp_rack */
338 	struct rack_rb_tree_head rc_mtree; /* Tree of all segments Lock(a) */
339 	struct rack_head rc_tmap;	/* List in transmit order Lock(a) */
340 	struct rack_sendmap *rc_tlpsend;	/* Remembered place for
341 						 * tlp_sending Lock(a) */
342 	struct rack_sendmap *rc_resend;	/* something we have been asked to
343 					 * resend */
344 	struct rack_fast_send_blk fsb;	/* The fast-send block */
345 	uint32_t timer_slop;
346 	uint32_t input_pkt;
347 	uint32_t saved_input_pkt;
348 	uint32_t rc_hpts_flags;
349 	uint32_t rc_fixed_pacing_rate_ca;
350 	uint32_t rc_fixed_pacing_rate_rec;
351 	uint32_t rc_fixed_pacing_rate_ss;
352 	uint32_t cwnd_to_use;	/* The cwnd in use */
353 	uint32_t rc_timer_exp;	/* If a timer ticks of expiry */
354 	uint32_t rc_rack_min_rtt;	/* lowest RTT seen Lock(a) */
355 	uint32_t rc_rack_largest_cwnd;	/* Largest CWND we have seen Lock(a) */
356 
357 	/* Third Cache line 0x80 */
358 	struct rack_head rc_free;	/* Allocation array */
359 	uint64_t last_hw_bw_req;
360 	uint64_t crte_prev_rate;
361 	uint64_t bw_rate_cap;
362 	uint32_t rc_time_last_sent;	/* Time we last sent some data and
363 					 * logged it Lock(a). */
364 	uint32_t rc_reorder_ts;	/* Last time we saw reordering Lock(a) */
365 
366 	uint32_t rc_tlp_new_data;	/* we need to send new-data on a TLP
367 					 * Lock(a) */
368 	uint32_t rc_prr_out;	/* bytes sent during recovery Lock(a) */
369 
370 	uint32_t rc_prr_recovery_fs;	/* recovery fs point Lock(a) */
371 
372 	uint32_t rc_prr_sndcnt;	/* Prr sndcnt Lock(a) */
373 
374 	uint32_t rc_sacked;	/* Tot sacked on scoreboard Lock(a) */
375 	uint32_t xxx_rc_last_tlp_seq;	/* Last tlp sequence Lock(a) */
376 
377 	uint32_t rc_prr_delivered;	/* during recovery prr var Lock(a) */
378 	uint16_t rc_tlp_cnt_out;	/* count of times we have sent a TLP without new data */
379 	uint16_t xxx_rc_tlp_seg_send_cnt;	/* Number of times we have TLP sent
380 					 * rc_last_tlp_seq Lock(a) */
381 
382 	uint32_t rc_loss_count;	/* How many bytes have been retransmitted
383 				 * Lock(a) */
384 	uint32_t rc_reorder_fade;	/* Socket option value Lock(a) */
385 
386 	/* Forth cache line 0xc0  */
387 	/* Times */
388 
389 	uint32_t rc_rack_tmit_time;	/* Rack transmit time Lock(a) */
390 	uint32_t rc_holes_rxt;	/* Tot retraned from scoreboard Lock(a) */
391 
392 	/* Variables to track bad retransmits and recover */
393 	uint32_t rc_rsm_start;	/* RSM seq number we retransmitted Lock(a) */
394 	uint32_t rc_cwnd_at;	/* cwnd at the retransmit Lock(a) */
395 
396 	uint32_t rc_ssthresh_at;/* ssthresh at the retransmit Lock(a) */
397 	uint32_t rc_num_maps_alloced;	/* Number of map blocks (sacks) we
398 					 * have allocated */
399 	uint32_t rc_rcvtime;	/* When we last received data */
400 	uint32_t rc_num_split_allocs;	/* num split map entries allocated */
401 
402 	uint32_t rc_last_output_to;
403 	uint32_t rc_went_idle_time;
404 
405 	struct rack_sendmap *rc_sacklast;	/* sack remembered place
406 						 * Lock(a) */
407 
408 	struct rack_sendmap *rc_rsm_at_retran;	/* Debug variable kept for
409 						 * cache line alignment
410 						 * Lock(a) */
411 	struct rack_sendmap *rc_first_appl;	/* Pointer to first app limited */
412 	struct rack_sendmap *rc_end_appl;	/* Pointer to last app limited */
413 	/* Cache line split 0x100 */
414 	struct sack_filter rack_sf;
415 	/* Cache line split 0x140 */
416 	/* Flags for various things */
417 	uint32_t last_pacing_time;
418 	uint32_t rc_pace_max_segs;
419 	uint32_t rc_pace_min_segs;
420 	uint32_t rc_app_limited_cnt;
421 	uint16_t rack_per_of_gp_ss; /* 100 = 100%, so from 65536 = 655 x bw  */
422 	uint16_t rack_per_of_gp_ca; /* 100 = 100%, so from 65536 = 655 x bw  */
423 	uint16_t rack_per_of_gp_rec; /* 100 = 100%, so from 65536 = 655 x bw, 0=off */
424 	uint16_t rack_per_of_gp_probertt; /* 100 = 100%, so from 65536 = 655 x bw, 0=off */
425 	uint32_t rc_high_rwnd;
426 	uint32_t ack_count;
427 	uint32_t sack_count;
428 	uint32_t sack_noextra_move;
429 	uint32_t sack_moved_extra;
430 	struct rack_rtt_sample rack_rs;
431 	const struct tcp_hwrate_limit_table *crte;
432 	uint32_t rc_agg_early;
433 	uint32_t rc_agg_delayed;
434 	uint32_t rc_tlp_rxt_last_time;
435 	uint32_t rc_saved_cwnd;
436 	uint64_t rc_gp_output_ts; /* chg*/
437 	uint64_t rc_gp_cumack_ts; /* chg*/
438 	struct timeval act_rcv_time;
439 	struct timeval rc_last_time_decay;	/* SAD time decay happened here */
440 	uint64_t gp_bw;
441 	uint64_t init_rate;
442 #ifdef NETFLIX_SHARED_CWND
443 	struct shared_cwnd *rc_scw;
444 #endif
445 	uint64_t last_gp_comp_bw;
446 	uint64_t last_max_bw;	/* Our calculated max b/w last */
447 	struct time_filter_small rc_gp_min_rtt;
448 	struct def_opt_head opt_list;
449 	int32_t rc_rtt_diff;		/* Timely style rtt diff of our gp_srtt */
450 	uint32_t rc_gp_srtt;		/* Current GP srtt */
451 	uint32_t rc_prev_gp_srtt;	/* Previous RTT */
452 	uint32_t rc_entry_gp_rtt;	/* Entry to PRTT gp-rtt */
453 	uint32_t rc_loss_at_start;	/* At measurement window where was our lost value */
454 
455 	uint32_t forced_ack_ts;
456 	uint32_t rc_lower_rtt_us_cts;	/* Time our GP rtt was last lowered */
457 	uint32_t rc_time_probertt_entered;
458 	uint32_t rc_time_probertt_starts;
459 	uint32_t rc_lowest_us_rtt;
460 	uint32_t rc_highest_us_rtt;
461 	uint32_t rc_last_us_rtt;
462 	uint32_t rc_time_of_last_probertt;
463 	uint32_t rc_target_probertt_flight;
464 	uint32_t rc_probertt_sndmax_atexit;	/* Highest sent to in probe-rtt */
465 	uint32_t rc_cwnd_at_erec;
466 	uint32_t rc_ssthresh_at_erec;
467 	uint32_t dsack_byte_cnt;
468 	uint32_t retran_during_recovery;
469 	uint32_t rc_gp_lowrtt;			/* Lowest rtt seen during GPUT measurement */
470 	uint32_t rc_gp_high_rwnd;		/* Highest rwnd seen during GPUT measurement */
471 	uint32_t rc_snd_max_at_rto;	/* For non-sack when the RTO occured what was snd-max */
472 	uint32_t rc_out_at_rto;
473 	int32_t rc_scw_index;
474 	uint32_t rc_tlp_threshold;	/* Socket option value Lock(a) */
475 	uint32_t rc_last_timeout_snduna;
476 	uint32_t challenge_ack_ts;
477 	uint32_t challenge_ack_cnt;
478 	uint32_t rc_min_to;	/* Socket option value Lock(a) */
479 	uint32_t rc_pkt_delay;	/* Socket option value Lock(a) */
480 	struct newreno rc_saved_beta;	/*
481 					 * For newreno cc:
482 					 * rc_saved_cc are the values we have had
483 					 * set by the user, if pacing is not happening
484 					 * (i.e. its early and we have not turned on yet
485 					 *  or it was turned off). The minute pacing
486 					 * is turned on we pull out the values currently
487 					 * being used by newreno and replace them with
488 					 * these values, then save off the old values here,
489 					 * we also set the flag (if ecn_beta is set) to make
490 					 * new_reno do less of a backoff for ecn (think abe).
491 					 */
492 	uint16_t rc_early_recovery_segs;	/* Socket option value Lock(a) */
493 	uint16_t rc_reorder_shift;	/* Socket option value Lock(a) */
494 	uint8_t rc_no_push_at_mrtt;	/* No push when we exceed max rtt */
495 	uint8_t num_measurements;	/* Number of measurements (up to 0xff, we freeze at 0xff)  */
496 	uint8_t req_measurements;	/* How many measurements are required? */
497 	uint8_t rc_tlp_cwnd_reduce;	/* Socket option value Lock(a) */
498 	uint8_t rc_prr_sendalot;/* Socket option value Lock(a) */
499 	uint8_t rc_rate_sample_method;
500 	uint8_t rc_gp_hist_idx;
501 };
502 
503 #define RACK_TIMELY_CNT_BOOST 5	/* At 5th increase boost */
504 #define RACK_MINRTT_FILTER_TIM 10 /* Seconds */
505 
506 #ifdef _KERNEL
507 
508 struct tcp_rack {
509 	/* First cache line 0x00 */
510 	TAILQ_ENTRY(tcp_rack) r_hpts;	/* hptsi queue next Lock(b) */
511 	int32_t(*r_substate) (struct mbuf *, struct tcphdr *,
512 	    struct socket *, struct tcpcb *, struct tcpopt *,
513 	    int32_t, int32_t, uint32_t, int, int, uint8_t);	/* Lock(a) */
514 	struct tcpcb *rc_tp;	/* The tcpcb Lock(a) */
515 	struct inpcb *rc_inp;	/* The inpcb Lock(a) */
516 	uint8_t rc_free_cnt;	/* Number of free entries on the rc_free list
517 				 * Lock(a) */
518 	uint8_t client_bufferlvl;	/* 0 - 5 normaly, less than or at 2 means its real low */
519 	uint8_t no_prr_addback : 1,
520 		gp_ready : 1,
521 		defer_options: 1,
522 		fast_rsm_hack: 1,
523 		rc_ack_can_sendout_data: 1, /*
524 					     * If set it will override pacing restrictions on not sending
525 					     * data when the pacing timer is running. I.e. you set this
526 					     * and an ACK will send data. Default is off and its only used
527 					     * without pacing when we are doing 5G speed up for there
528 					     * ack filtering.
529 					     */
530 		rc_pacing_cc_set: 1,	     /*
531 					      * If we are pacing (pace_always=1) and we have reached the
532 					      * point where we start pacing (fixed or gp has reached its
533 					      * magic gp_ready state) this flag indicates we have set in
534 					      * values to effect CC's backoff's. If pacing is turned off
535 					      * then we must restore the values saved in rc_saved_beta,
536 					      * if its going to gp_ready we need to copy the values into
537 					      * the CC module and set our flags.
538 					      *
539 					      * Note this only happens if the cc name is newreno (CCALGONAME_NEWRENO).
540 					      */
541 
542 		avail :2;
543 	uint8_t avail_bytes;
544 	uint32_t rc_rack_rtt;	/* RACK-RTT Lock(a) */
545 	uint16_t r_mbuf_queue : 1,	/* Do we do mbuf queue for non-paced */
546 		 rtt_limit_mul : 4,	/* muliply this by low rtt */
547 		 r_limit_scw : 1,
548 		 r_must_retran : 1,	/* For non-sack customers we hit an RTO and new data should be resends */
549 		 r_use_cmp_ack: 1,	/* Do we use compressed acks */
550 		 r_ent_rec_ns: 1,	/* We entered recovery and have not sent */
551 		 r_might_revert: 1,	/* Flag to find out if we might need to revert */
552 		 r_fast_output: 1, 	/* Fast output is in progress we can skip the bulk of rack_output */
553 		 r_fsb_inited: 1,
554 		 r_rack_hw_rate_caps: 1,
555 		 r_up_only: 1,
556 		 r_via_fill_cw : 1,
557 		 r_fill_less_agg : 1;
558 
559 	uint8_t rc_user_set_max_segs;	/* Socket option value Lock(a) */
560 	uint8_t rc_labc;		/* Appropriate Byte Counting Value */
561 	uint16_t forced_ack : 1,
562 		rc_gp_incr : 1,
563 		rc_gp_bwred : 1,
564 		rc_gp_timely_inc_cnt : 3,
565 		rc_gp_timely_dec_cnt : 3,
566 		r_use_labc_for_rec: 1,
567 		rc_highly_buffered: 1,		/* The path is highly buffered */
568 		rc_dragged_bottom: 1,
569 		rc_dack_mode : 1,		/* Mac O/S emulation of d-ack */
570 		rc_dack_toggle : 1,		/* For Mac O/S emulation of d-ack */
571 		pacing_longer_than_rtt : 1,
572 		rc_gp_filled : 1;
573 	uint8_t r_state;	/* Current rack state Lock(a) */
574 	uint8_t rc_tmr_stopped : 7,
575 		t_timers_stopped : 1;
576 	uint8_t rc_enobuf : 7,	/* count of enobufs on connection provides */
577 		rc_on_min_to : 1;
578 	uint8_t r_timer_override : 1,	/* hpts override Lock(a) */
579 		r_is_v6 : 1,	/* V6 pcb Lock(a)  */
580 		rc_in_persist : 1,
581 		rc_tlp_in_progress : 1,
582 		rc_always_pace : 1,	/* Socket option value Lock(a) */
583 		rc_pace_to_cwnd : 1,
584 		rc_pace_fill_if_rttin_range : 1,
585 		rc_srtt_measure_made : 1;
586 	uint8_t app_limited_needs_set : 1,
587 		use_fixed_rate : 1,
588 		rc_has_collapsed : 1,
589 		r_rep_attack : 1,
590 		r_rep_reverse : 1,
591 		rack_hdrw_pacing : 1,  /* We are doing Hardware pacing */
592 		rack_hdw_pace_ena : 1, /* Is hardware pacing enabled? */
593 		rack_attempt_hdwr_pace : 1; /* Did we attempt hdwr pacing (if allowed) */
594 	uint8_t rack_tlp_threshold_use : 3,	/* only 1, 2 and 3 used so far */
595 		rack_rec_nonrxt_use_cr : 1,
596 		rack_enable_scwnd : 1,
597 		rack_attempted_scwnd : 1,
598 		rack_no_prr : 1,
599 		rack_scwnd_is_idle : 1;
600 	uint8_t rc_allow_data_af_clo: 1,
601 		delayed_ack : 1,
602 		set_pacing_done_a_iw : 1,
603 		use_rack_rr : 1,
604 		alloc_limit_reported : 1,
605 		sack_attack_disable : 1,
606 		do_detection : 1,
607 		rc_force_max_seg : 1;
608 	uint8_t rack_cwnd_limited : 1,
609 		r_early : 1,
610 		r_late : 1,
611 		r_running_early : 1,
612 		r_running_late : 1,
613 		r_wanted_output: 1,
614 		r_rr_config : 2;
615 	uint16_t rc_init_win : 8,
616 		rc_gp_rtt_set : 1,
617 		rc_gp_dyn_mul : 1,
618 		rc_gp_saw_rec : 1,
619 		rc_gp_saw_ca : 1,
620 		rc_gp_saw_ss : 1,
621 		rc_gp_no_rec_chg : 1,
622 		in_probe_rtt : 1,
623 		measure_saw_probe_rtt : 1;
624 	/* Cache line 2 0x40 */
625 	struct rack_control r_ctl;
626 }        __aligned(CACHE_LINE_SIZE);
627 
628 #endif
629 #endif
630