xref: /freebsd/sys/netinet/tcp_ratelimit.c (revision 43a5ec4eb41567cc92586503212743d89686d78f)
1 /*-
2  *
3  * SPDX-License-Identifier: BSD-3-Clause
4  *
5  * Copyright (c) 2018-2020
6  *	Netflix Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  */
30 /**
31  * Author: Randall Stewart <rrs@netflix.com>
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_tcpdebug.h"
40 #include "opt_ratelimit.h"
41 #include <sys/param.h>
42 #include <sys/kernel.h>
43 #include <sys/malloc.h>
44 #include <sys/mbuf.h>
45 #include <sys/socket.h>
46 #include <sys/socketvar.h>
47 #include <sys/sysctl.h>
48 #include <sys/eventhandler.h>
49 #include <sys/mutex.h>
50 #include <sys/ck.h>
51 #include <net/if.h>
52 #include <net/if_var.h>
53 #include <netinet/in.h>
54 #include <netinet/in_pcb.h>
55 #define TCPSTATES		/* for logging */
56 #include <netinet/tcp_var.h>
57 #ifdef INET6
58 #include <netinet6/tcp6_var.h>
59 #endif
60 #include <netinet/tcp_hpts.h>
61 #include <netinet/tcp_log_buf.h>
62 #include <netinet/tcp_ratelimit.h>
63 #ifndef USECS_IN_SECOND
64 #define USECS_IN_SECOND 1000000
65 #endif
66 /*
67  * For the purposes of each send, what is the size
68  * of an ethernet frame.
69  */
70 MALLOC_DEFINE(M_TCPPACE, "tcp_hwpace", "TCP Hardware pacing memory");
71 #ifdef RATELIMIT
72 
73 /*
74  * The following preferred table will seem weird to
75  * the casual viewer. Why do we not have any rates below
76  * 1Mbps? Why do we have a rate at 1.44Mbps called common?
77  * Why do the rates cluster in the 1-100Mbps range more
78  * than others? Why does the table jump around at the beginnign
79  * and then be more consistently raising?
80  *
81  * Let me try to answer those questions. A lot of
82  * this is dependant on the hardware. We have three basic
83  * supporters of rate limiting
84  *
85  * Chelsio - Supporting 16 configurable rates.
86  * Mlx  - c4 supporting 13 fixed rates.
87  * Mlx  - c5 & c6 supporting 127 configurable rates.
88  *
89  * The c4 is why we have a common rate that is available
90  * in all rate tables. This is a selected rate from the
91  * c4 table and we assure its available in all ratelimit
92  * tables. This way the tcp_ratelimit code has an assured
93  * rate it should always be able to get. This answers a
94  * couple of the questions above.
95  *
96  * So what about the rest, well the table is built to
97  * try to get the most out of a joint hardware/software
98  * pacing system.  The software pacer will always pick
99  * a rate higher than the b/w that it is estimating
100  *
101  * on the path. This is done for two reasons.
102  * a) So we can discover more b/w
103  * and
104  * b) So we can send a block of MSS's down and then
105  *    have the software timer go off after the previous
106  *    send is completely out of the hardware.
107  *
108  * But when we do <b> we don't want to have the delay
109  * between the last packet sent by the hardware be
110  * excessively long (to reach our desired rate).
111  *
112  * So let me give an example for clarity.
113  *
114  * Lets assume that the tcp stack sees that 29,110,000 bps is
115  * what the bw of the path is. The stack would select the
116  * rate 31Mbps. 31Mbps means that each send that is done
117  * by the hardware will cause a 390 micro-second gap between
118  * the packets sent at that rate. For 29,110,000 bps we
119  * would need 416 micro-seconds gap between each send.
120  *
121  * Note that are calculating a complete time for pacing
122  * which includes the ethernet, IP and TCP overhead. So
123  * a full 1514 bytes is used for the above calculations.
124  * My testing has shown that both cards are also using this
125  * as their basis i.e. full payload size of the ethernet frame.
126  * The TCP stack caller needs to be aware of this and make the
127  * appropriate overhead calculations be included in its choices.
128  *
129  * Now, continuing our example, we pick a MSS size based on the
130  * delta between the two rates (416 - 390) divided into the rate
131  * we really wish to send at rounded up.  That results in a MSS
132  * send of 17 mss's at once. The hardware then will
133  * run out of data in a single 17MSS send in 6,630 micro-seconds.
134  *
135  * On the other hand the software pacer will send more data
136  * in 7,072 micro-seconds. This means that we will refill
137  * the hardware 52 microseconds after it would have sent
138  * next if it had not ran out of data. This is a win since we are
139  * only sending every 7ms or so and yet all the packets are spaced on
140  * the wire with 94% of what they should be and only
141  * the last packet is delayed extra to make up for the
142  * difference.
143  *
144  * Note that the above formula has two important caveat.
145  * If we are above (b/w wise) over 100Mbps we double the result
146  * of the MSS calculation. The second caveat is if we are 500Mbps
147  * or more we just send the maximum MSS at once i.e. 45MSS. At
148  * the higher b/w's even the cards have limits to what times (timer granularity)
149  * they can insert between packets and start to send more than one
150  * packet at a time on the wire.
151  *
152  */
153 #define COMMON_RATE 180500
154 const uint64_t desired_rates[] = {
155 	122500,			/* 1Mbps  - rate 1 */
156 	180500,			/* 1.44Mpbs - rate 2  common rate */
157 	375000,			/* 3Mbps    - rate 3 */
158 	625000,			/* 5Mbps    - rate 4 */
159 	1250000,		/* 10Mbps   - rate 5 */
160 	1875000,		/* 15Mbps   - rate 6 */
161 	2500000,		/* 20Mbps   - rate 7 */
162 	3125000,	       	/* 25Mbps   - rate 8 */
163 	3750000,		/* 30Mbps   - rate 9 */
164 	4375000,		/* 35Mbps   - rate 10 */
165 	5000000,		/* 40Meg    - rate 11 */
166 	6250000,		/* 50Mbps   - rate 12 */
167 	12500000,		/* 100Mbps  - rate 13 */
168 	25000000,		/* 200Mbps  - rate 14 */
169 	50000000,		/* 400Mbps  - rate 15 */
170 	100000000,		/* 800Mbps  - rate 16 */
171 	5625000,		/* 45Mbps   - rate 17 */
172 	6875000,		/* 55Mbps   - rate 19 */
173 	7500000,		/* 60Mbps   - rate 20 */
174 	8125000,		/* 65Mbps   - rate 21 */
175 	8750000,		/* 70Mbps   - rate 22 */
176 	9375000,		/* 75Mbps   - rate 23 */
177 	10000000,		/* 80Mbps   - rate 24 */
178 	10625000,		/* 85Mbps   - rate 25 */
179 	11250000,		/* 90Mbps   - rate 26 */
180 	11875000,		/* 95Mbps   - rate 27 */
181 	12500000,		/* 100Mbps  - rate 28 */
182 	13750000,		/* 110Mbps  - rate 29 */
183 	15000000,		/* 120Mbps  - rate 30 */
184 	16250000,		/* 130Mbps  - rate 31 */
185 	17500000,		/* 140Mbps  - rate 32 */
186 	18750000,		/* 150Mbps  - rate 33 */
187 	20000000,		/* 160Mbps  - rate 34 */
188 	21250000,		/* 170Mbps  - rate 35 */
189 	22500000,		/* 180Mbps  - rate 36 */
190 	23750000,		/* 190Mbps  - rate 37 */
191 	26250000,		/* 210Mbps  - rate 38 */
192 	27500000,		/* 220Mbps  - rate 39 */
193 	28750000,		/* 230Mbps  - rate 40 */
194 	30000000,	       	/* 240Mbps  - rate 41 */
195 	31250000,		/* 250Mbps  - rate 42 */
196 	34375000,		/* 275Mbps  - rate 43 */
197 	37500000,		/* 300Mbps  - rate 44 */
198 	40625000,		/* 325Mbps  - rate 45 */
199 	43750000,		/* 350Mbps  - rate 46 */
200 	46875000,		/* 375Mbps  - rate 47 */
201 	53125000,		/* 425Mbps  - rate 48 */
202 	56250000,		/* 450Mbps  - rate 49 */
203 	59375000,		/* 475Mbps  - rate 50 */
204 	62500000,		/* 500Mbps  - rate 51 */
205 	68750000,		/* 550Mbps  - rate 52 */
206 	75000000,		/* 600Mbps  - rate 53 */
207 	81250000,		/* 650Mbps  - rate 54 */
208 	87500000,		/* 700Mbps  - rate 55 */
209 	93750000,		/* 750Mbps  - rate 56 */
210 	106250000,		/* 850Mbps  - rate 57 */
211 	112500000,		/* 900Mbps  - rate 58 */
212 	125000000,		/* 1Gbps    - rate 59 */
213 	156250000,		/* 1.25Gps  - rate 60 */
214 	187500000,		/* 1.5Gps   - rate 61 */
215 	218750000,		/* 1.75Gps  - rate 62 */
216 	250000000,		/* 2Gbps    - rate 63 */
217 	281250000,		/* 2.25Gps  - rate 64 */
218 	312500000,		/* 2.5Gbps  - rate 65 */
219 	343750000,		/* 2.75Gbps - rate 66 */
220 	375000000,		/* 3Gbps    - rate 67 */
221 	500000000,		/* 4Gbps    - rate 68 */
222 	625000000,		/* 5Gbps    - rate 69 */
223 	750000000,		/* 6Gbps    - rate 70 */
224 	875000000,		/* 7Gbps    - rate 71 */
225 	1000000000,		/* 8Gbps    - rate 72 */
226 	1125000000,		/* 9Gbps    - rate 73 */
227 	1250000000,		/* 10Gbps   - rate 74 */
228 	1875000000,		/* 15Gbps   - rate 75 */
229 	2500000000		/* 20Gbps   - rate 76 */
230 };
231 
232 #define MAX_HDWR_RATES (sizeof(desired_rates)/sizeof(uint64_t))
233 #define RS_ORDERED_COUNT 16	/*
234 				 * Number that are in order
235 				 * at the beginning of the table,
236 				 * over this a sort is required.
237 				 */
238 #define RS_NEXT_ORDER_GROUP 16	/*
239 				 * The point in our table where
240 				 * we come fill in a second ordered
241 				 * group (index wise means -1).
242 				 */
243 #define ALL_HARDWARE_RATES 1004 /*
244 				 * 1Meg - 1Gig in 1 Meg steps
245 				 * plus 100, 200k  and 500k and
246 				 * 10Gig
247 				 */
248 
249 #define RS_ONE_MEGABIT_PERSEC 1000000
250 #define RS_ONE_GIGABIT_PERSEC 1000000000
251 #define RS_TEN_GIGABIT_PERSEC 10000000000
252 
253 static struct head_tcp_rate_set int_rs;
254 static struct mtx rs_mtx;
255 uint32_t rs_number_alive;
256 uint32_t rs_number_dead;
257 static uint32_t rs_floor_mss = 0;
258 static uint32_t wait_time_floor = 8000;	/* 8 ms */
259 static uint32_t rs_hw_floor_mss = 16;
260 static uint32_t num_of_waits_allowed = 1; /* How many time blocks are we willing to wait */
261 
262 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, rl, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
263     "TCP Ratelimit stats");
264 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, alive, CTLFLAG_RW,
265     &rs_number_alive, 0,
266     "Number of interfaces initialized for ratelimiting");
267 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, dead, CTLFLAG_RW,
268     &rs_number_dead, 0,
269     "Number of interfaces departing from ratelimiting");
270 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, floor_mss, CTLFLAG_RW,
271     &rs_floor_mss, 0,
272     "Number of MSS that will override the normal minimums (0 means don't enforce)");
273 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, wait_floor, CTLFLAG_RW,
274     &wait_time_floor, 2000,
275     "Has b/w increases what is the wait floor we are willing to wait at the end?");
276 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, time_blocks, CTLFLAG_RW,
277     &num_of_waits_allowed, 1,
278     "How many time blocks on the end should software pacing be willing to wait?");
279 
280 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, hw_floor_mss, CTLFLAG_RW,
281     &rs_hw_floor_mss, 16,
282     "Number of mss that are a minum for hardware pacing?");
283 
284 
285 static void
286 rl_add_syctl_entries(struct sysctl_oid *rl_sysctl_root, struct tcp_rate_set *rs)
287 {
288 	/*
289 	 * Add sysctl entries for thus interface.
290 	 */
291 	if (rs->rs_flags & RS_INTF_NO_SUP) {
292 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
293 		   SYSCTL_CHILDREN(rl_sysctl_root),
294 		   OID_AUTO, "disable", CTLFLAG_RD,
295 		   &rs->rs_disable, 0,
296 		   "Disable this interface from new hdwr limiting?");
297 	} else {
298 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
299 		   SYSCTL_CHILDREN(rl_sysctl_root),
300 		   OID_AUTO, "disable", CTLFLAG_RW,
301 		   &rs->rs_disable, 0,
302 		   "Disable this interface from new hdwr limiting?");
303 	}
304 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
305 	    SYSCTL_CHILDREN(rl_sysctl_root),
306 	    OID_AUTO, "minseg", CTLFLAG_RW,
307 	    &rs->rs_min_seg, 0,
308 	    "What is the minimum we need to send on this interface?");
309 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
310 	    SYSCTL_CHILDREN(rl_sysctl_root),
311 	    OID_AUTO, "flow_limit", CTLFLAG_RW,
312 	    &rs->rs_flow_limit, 0,
313 	    "What is the limit for number of flows (0=unlimited)?");
314 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
315 	    SYSCTL_CHILDREN(rl_sysctl_root),
316 	    OID_AUTO, "highest", CTLFLAG_RD,
317 	    &rs->rs_highest_valid, 0,
318 	    "Highest valid rate");
319 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
320 	    SYSCTL_CHILDREN(rl_sysctl_root),
321 	    OID_AUTO, "lowest", CTLFLAG_RD,
322 	    &rs->rs_lowest_valid, 0,
323 	    "Lowest valid rate");
324 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
325 	    SYSCTL_CHILDREN(rl_sysctl_root),
326 	    OID_AUTO, "flags", CTLFLAG_RD,
327 	    &rs->rs_flags, 0,
328 	    "What lags are on the entry?");
329 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
330 	    SYSCTL_CHILDREN(rl_sysctl_root),
331 	    OID_AUTO, "numrates", CTLFLAG_RD,
332 	    &rs->rs_rate_cnt, 0,
333 	    "How many rates re there?");
334 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
335 	    SYSCTL_CHILDREN(rl_sysctl_root),
336 	    OID_AUTO, "flows_using", CTLFLAG_RD,
337 	    &rs->rs_flows_using, 0,
338 	    "How many flows are using this interface now?");
339 #ifdef DETAILED_RATELIMIT_SYSCTL
340 	if (rs->rs_rlt && rs->rs_rate_cnt > 0) {
341 		/*  Lets display the rates */
342 		int i;
343 		struct sysctl_oid *rl_rates;
344 		struct sysctl_oid *rl_rate_num;
345 		char rate_num[16];
346 		rl_rates = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
347 					    SYSCTL_CHILDREN(rl_sysctl_root),
348 					    OID_AUTO,
349 					    "rate",
350 					    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
351 					    "Ratelist");
352 		for( i = 0; i < rs->rs_rate_cnt; i++) {
353 			sprintf(rate_num, "%d", i);
354 			rl_rate_num = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
355 					    SYSCTL_CHILDREN(rl_rates),
356 					    OID_AUTO,
357 					    rate_num,
358 					    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
359 					    "Individual Rate");
360 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
361 				       SYSCTL_CHILDREN(rl_rate_num),
362 				       OID_AUTO, "flags", CTLFLAG_RD,
363 				       &rs->rs_rlt[i].flags, 0,
364 				       "Flags on this rate");
365 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
366 				       SYSCTL_CHILDREN(rl_rate_num),
367 				       OID_AUTO, "pacetime", CTLFLAG_RD,
368 				       &rs->rs_rlt[i].time_between, 0,
369 				       "Time hardware inserts between 1500 byte sends");
370 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
371 				       SYSCTL_CHILDREN(rl_rate_num),
372 				       OID_AUTO, "rate", CTLFLAG_RD,
373 				       &rs->rs_rlt[i].rate,
374 				       "Rate in bytes per second");
375 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
376 				       SYSCTL_CHILDREN(rl_rate_num),
377 				       OID_AUTO, "using", CTLFLAG_RD,
378 				       &rs->rs_rlt[i].using,
379 				       "Number of flows using");
380 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
381 				       SYSCTL_CHILDREN(rl_rate_num),
382 				       OID_AUTO, "enobufs", CTLFLAG_RD,
383 				       &rs->rs_rlt[i].rs_num_enobufs,
384 				       "Number of enobufs logged on this rate");
385 
386 		}
387 	}
388 #endif
389 }
390 
391 static void
392 rs_destroy(epoch_context_t ctx)
393 {
394 	struct tcp_rate_set *rs;
395 	bool do_free_rs;
396 
397 	rs = __containerof(ctx, struct tcp_rate_set, rs_epoch_ctx);
398 
399 	mtx_lock(&rs_mtx);
400 	rs->rs_flags &= ~RS_FUNERAL_SCHD;
401 	/*
402 	 * In theory its possible (but unlikely)
403 	 * that while the delete was occuring
404 	 * and we were applying the DEAD flag
405 	 * someone slipped in and found the
406 	 * interface in a lookup. While we
407 	 * decided rs_flows_using were 0 and
408 	 * scheduling the epoch_call, the other
409 	 * thread incremented rs_flow_using. This
410 	 * is because users have a pointer and
411 	 * we only use the rs_flows_using in an
412 	 * atomic fashion, i.e. the other entities
413 	 * are not protected. To assure this did
414 	 * not occur, we check rs_flows_using here
415 	 * before deleting.
416 	 */
417 	do_free_rs = (rs->rs_flows_using == 0);
418 	rs_number_dead--;
419 	mtx_unlock(&rs_mtx);
420 
421 	if (do_free_rs) {
422 		sysctl_ctx_free(&rs->sysctl_ctx);
423 		free(rs->rs_rlt, M_TCPPACE);
424 		free(rs, M_TCPPACE);
425 	}
426 }
427 
428 static void
429 rs_defer_destroy(struct tcp_rate_set *rs)
430 {
431 
432 	mtx_assert(&rs_mtx, MA_OWNED);
433 
434 	/* Check if already pending. */
435 	if (rs->rs_flags & RS_FUNERAL_SCHD)
436 		return;
437 
438 	rs_number_dead++;
439 
440 	/* Set flag to only defer once. */
441 	rs->rs_flags |= RS_FUNERAL_SCHD;
442 	NET_EPOCH_CALL(rs_destroy, &rs->rs_epoch_ctx);
443 }
444 
445 #ifdef INET
446 extern counter_u64_t rate_limit_new;
447 extern counter_u64_t rate_limit_chg;
448 extern counter_u64_t rate_limit_set_ok;
449 extern counter_u64_t rate_limit_active;
450 extern counter_u64_t rate_limit_alloc_fail;
451 #endif
452 
453 static int
454 rl_attach_txrtlmt(struct ifnet *ifp,
455     uint32_t flowtype,
456     int flowid,
457     uint64_t cfg_rate,
458     struct m_snd_tag **tag)
459 {
460 	int error;
461 	union if_snd_tag_alloc_params params = {
462 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
463 		.rate_limit.hdr.flowid = flowid,
464 		.rate_limit.hdr.flowtype = flowtype,
465 		.rate_limit.max_rate = cfg_rate,
466 		.rate_limit.flags = M_NOWAIT,
467 	};
468 
469 	error = m_snd_tag_alloc(ifp, &params, tag);
470 #ifdef INET
471 	if (error == 0) {
472 		counter_u64_add(rate_limit_set_ok, 1);
473 		counter_u64_add(rate_limit_active, 1);
474 	} else if (error != EOPNOTSUPP)
475 		counter_u64_add(rate_limit_alloc_fail, 1);
476 #endif
477 	return (error);
478 }
479 
480 static void
481 populate_canned_table(struct tcp_rate_set *rs, const uint64_t *rate_table_act)
482 {
483 	/*
484 	 * The internal table is "special", it
485 	 * is two seperate ordered tables that
486 	 * must be merged. We get here when the
487 	 * adapter specifies a number of rates that
488 	 * covers both ranges in the table in some
489 	 * form.
490 	 */
491 	int i, at_low, at_high;
492 	uint8_t low_disabled = 0, high_disabled = 0;
493 
494 	for(i = 0, at_low = 0, at_high = RS_NEXT_ORDER_GROUP; i < rs->rs_rate_cnt; i++) {
495 		rs->rs_rlt[i].flags = 0;
496 		rs->rs_rlt[i].time_between = 0;
497 		if ((low_disabled == 0) &&
498 		    (high_disabled ||
499 		     (rate_table_act[at_low] < rate_table_act[at_high]))) {
500 			rs->rs_rlt[i].rate = rate_table_act[at_low];
501 			at_low++;
502 			if (at_low == RS_NEXT_ORDER_GROUP)
503 				low_disabled = 1;
504 		} else if (high_disabled == 0) {
505 			rs->rs_rlt[i].rate = rate_table_act[at_high];
506 			at_high++;
507 			if (at_high == MAX_HDWR_RATES)
508 				high_disabled = 1;
509 		}
510 	}
511 }
512 
513 static struct tcp_rate_set *
514 rt_setup_new_rs(struct ifnet *ifp, int *error)
515 {
516 	struct tcp_rate_set *rs;
517 	const uint64_t *rate_table_act;
518 	uint64_t lentim, res;
519 	size_t sz;
520 	uint32_t hash_type;
521 	int i;
522 	struct if_ratelimit_query_results rl;
523 	struct sysctl_oid *rl_sysctl_root;
524 	struct epoch_tracker et;
525 	/*
526 	 * We expect to enter with the
527 	 * mutex locked.
528 	 */
529 
530 	if (ifp->if_ratelimit_query == NULL) {
531 		/*
532 		 * We can do nothing if we cannot
533 		 * get a query back from the driver.
534 		 */
535 		printf("Warning:No query functions for %s:%d-- failed\n",
536 		       ifp->if_dname, ifp->if_dunit);
537 		return (NULL);
538 	}
539 	rs = malloc(sizeof(struct tcp_rate_set), M_TCPPACE, M_NOWAIT | M_ZERO);
540 	if (rs == NULL) {
541 		if (error)
542 			*error = ENOMEM;
543 		printf("Warning:No memory for malloc of tcp_rate_set\n");
544 		return (NULL);
545 	}
546 	memset(&rl, 0, sizeof(rl));
547 	rl.flags = RT_NOSUPPORT;
548 	ifp->if_ratelimit_query(ifp, &rl);
549 	if (rl.flags & RT_IS_UNUSABLE) {
550 		/*
551 		 * The interface does not really support
552 		 * the rate-limiting.
553 		 */
554 		memset(rs, 0, sizeof(struct tcp_rate_set));
555 		rs->rs_ifp = ifp;
556 		rs->rs_if_dunit = ifp->if_dunit;
557 		rs->rs_flags = RS_INTF_NO_SUP;
558 		rs->rs_disable = 1;
559 		rs_number_alive++;
560 		sysctl_ctx_init(&rs->sysctl_ctx);
561 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
562 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
563 		    OID_AUTO,
564 		    rs->rs_ifp->if_xname,
565 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
566 		    "");
567 		rl_add_syctl_entries(rl_sysctl_root, rs);
568 		NET_EPOCH_ENTER(et);
569 		mtx_lock(&rs_mtx);
570 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
571 		mtx_unlock(&rs_mtx);
572 		NET_EPOCH_EXIT(et);
573 		return (rs);
574 	} else if ((rl.flags & RT_IS_INDIRECT) == RT_IS_INDIRECT) {
575 		memset(rs, 0, sizeof(struct tcp_rate_set));
576 		rs->rs_ifp = ifp;
577 		rs->rs_if_dunit = ifp->if_dunit;
578 		rs->rs_flags = RS_IS_DEFF;
579 		rs_number_alive++;
580 		sysctl_ctx_init(&rs->sysctl_ctx);
581 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
582 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
583 		    OID_AUTO,
584 		    rs->rs_ifp->if_xname,
585 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
586 		    "");
587 		rl_add_syctl_entries(rl_sysctl_root, rs);
588 		NET_EPOCH_ENTER(et);
589 		mtx_lock(&rs_mtx);
590 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
591 		mtx_unlock(&rs_mtx);
592 		NET_EPOCH_EXIT(et);
593 		return (rs);
594 	} else if ((rl.flags & RT_IS_FIXED_TABLE) == RT_IS_FIXED_TABLE) {
595 		/* Mellanox C4 likely */
596 		rs->rs_ifp = ifp;
597 		rs->rs_if_dunit = ifp->if_dunit;
598 		rs->rs_rate_cnt = rl.number_of_rates;
599 		rs->rs_min_seg = rl.min_segment_burst;
600 		rs->rs_highest_valid = 0;
601 		rs->rs_flow_limit = rl.max_flows;
602 		rs->rs_flags = RS_IS_INTF | RS_NO_PRE;
603 		rs->rs_disable = 0;
604 		rate_table_act = rl.rate_table;
605 	} else if ((rl.flags & RT_IS_SELECTABLE) == RT_IS_SELECTABLE) {
606 		/* Chelsio, C5 and C6 of Mellanox? */
607 		rs->rs_ifp = ifp;
608 		rs->rs_if_dunit = ifp->if_dunit;
609 		rs->rs_rate_cnt = rl.number_of_rates;
610 		rs->rs_min_seg = rl.min_segment_burst;
611 		rs->rs_disable = 0;
612 		rs->rs_flow_limit = rl.max_flows;
613 		rate_table_act = desired_rates;
614 		if ((rs->rs_rate_cnt > MAX_HDWR_RATES) &&
615 		    (rs->rs_rate_cnt < ALL_HARDWARE_RATES)) {
616 			/*
617 			 * Our desired table is not big
618 			 * enough, do what we can.
619 			 */
620 			rs->rs_rate_cnt = MAX_HDWR_RATES;
621 		 }
622 		if (rs->rs_rate_cnt <= RS_ORDERED_COUNT)
623 			rs->rs_flags = RS_IS_INTF;
624 		else
625 			rs->rs_flags = RS_IS_INTF | RS_INT_TBL;
626 		if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)
627 			rs->rs_rate_cnt = ALL_HARDWARE_RATES;
628 	} else {
629 		free(rs, M_TCPPACE);
630 		return (NULL);
631 	}
632 	sz = sizeof(struct tcp_hwrate_limit_table) * rs->rs_rate_cnt;
633 	rs->rs_rlt = malloc(sz, M_TCPPACE, M_NOWAIT);
634 	if (rs->rs_rlt == NULL) {
635 		if (error)
636 			*error = ENOMEM;
637 bail:
638 		free(rs, M_TCPPACE);
639 		return (NULL);
640 	}
641 	if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) {
642 		/*
643 		 * The interface supports all
644 		 * the rates we could possibly want.
645 		 */
646 		uint64_t rat;
647 
648 		rs->rs_rlt[0].rate = 12500;	/* 100k */
649 		rs->rs_rlt[1].rate = 25000;	/* 200k */
650 		rs->rs_rlt[2].rate = 62500;	/* 500k */
651 		/* Note 125000 == 1Megabit
652 		 * populate 1Meg - 1000meg.
653 		 */
654 		for(i = 3, rat = 125000; i< (ALL_HARDWARE_RATES-1); i++) {
655 			rs->rs_rlt[i].rate = rat;
656 			rat += 125000;
657 		}
658 		rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate = 1250000000;
659 	} else if (rs->rs_flags & RS_INT_TBL) {
660 		/* We populate this in a special way */
661 		populate_canned_table(rs, rate_table_act);
662 	} else {
663 		/*
664 		 * Just copy in the rates from
665 		 * the table, it is in order.
666 		 */
667 		for (i=0; i<rs->rs_rate_cnt; i++) {
668 			rs->rs_rlt[i].rate = rate_table_act[i];
669 			rs->rs_rlt[i].time_between = 0;
670 			rs->rs_rlt[i].flags = 0;
671 		}
672 	}
673 	for (i = (rs->rs_rate_cnt - 1); i >= 0; i--) {
674 		/*
675 		 * We go backwards through the list so that if we can't get
676 		 * a rate and fail to init one, we have at least a chance of
677 		 * getting the highest one.
678 		 */
679 		rs->rs_rlt[i].ptbl = rs;
680 		rs->rs_rlt[i].tag = NULL;
681 		rs->rs_rlt[i].using = 0;
682 		rs->rs_rlt[i].rs_num_enobufs = 0;
683 		/*
684 		 * Calculate the time between.
685 		 */
686 		lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
687 		res = lentim / rs->rs_rlt[i].rate;
688 		if (res > 0)
689 			rs->rs_rlt[i].time_between = res;
690 		else
691 			rs->rs_rlt[i].time_between = 1;
692 		if (rs->rs_flags & RS_NO_PRE) {
693 			rs->rs_rlt[i].flags = HDWRPACE_INITED;
694 			rs->rs_lowest_valid = i;
695 		} else {
696 			int err;
697 
698 			if ((rl.flags & RT_IS_SETUP_REQ)  &&
699 			    (ifp->if_ratelimit_query)) {
700 				err = ifp->if_ratelimit_setup(ifp,
701   				         rs->rs_rlt[i].rate, i);
702 				if (err)
703 					goto handle_err;
704 			}
705 #ifdef RSS
706 			hash_type = M_HASHTYPE_RSS_TCP_IPV4;
707 #else
708 			hash_type = M_HASHTYPE_OPAQUE_HASH;
709 #endif
710 			err = rl_attach_txrtlmt(ifp,
711 			    hash_type,
712 			    (i + 1),
713 			    rs->rs_rlt[i].rate,
714 			    &rs->rs_rlt[i].tag);
715 			if (err) {
716 handle_err:
717 				if (i == (rs->rs_rate_cnt - 1)) {
718 					/*
719 					 * Huh - first rate and we can't get
720 					 * it?
721 					 */
722 					free(rs->rs_rlt, M_TCPPACE);
723 					if (error)
724 						*error = err;
725 					goto bail;
726 				} else {
727 					if (error)
728 						*error = err;
729 				}
730 				break;
731 			} else {
732 				rs->rs_rlt[i].flags = HDWRPACE_INITED | HDWRPACE_TAGPRESENT;
733 				rs->rs_lowest_valid = i;
734 			}
735 		}
736 	}
737 	/* Did we get at least 1 rate? */
738 	if (rs->rs_rlt[(rs->rs_rate_cnt - 1)].flags & HDWRPACE_INITED)
739 		rs->rs_highest_valid = rs->rs_rate_cnt - 1;
740 	else {
741 		free(rs->rs_rlt, M_TCPPACE);
742 		goto bail;
743 	}
744 	rs_number_alive++;
745 	sysctl_ctx_init(&rs->sysctl_ctx);
746 	rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
747 	    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
748 	    OID_AUTO,
749 	    rs->rs_ifp->if_xname,
750 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
751 	    "");
752 	rl_add_syctl_entries(rl_sysctl_root, rs);
753 	NET_EPOCH_ENTER(et);
754 	mtx_lock(&rs_mtx);
755 	CK_LIST_INSERT_HEAD(&int_rs, rs, next);
756 	mtx_unlock(&rs_mtx);
757 	NET_EPOCH_EXIT(et);
758 	return (rs);
759 }
760 
761 /*
762  * For an explanation of why the argument is volatile please
763  * look at the comments around rt_setup_rate().
764  */
765 static const struct tcp_hwrate_limit_table *
766 tcp_int_find_suitable_rate(const volatile struct tcp_rate_set *rs,
767     uint64_t bytes_per_sec, uint32_t flags, uint64_t *lower_rate)
768 {
769 	struct tcp_hwrate_limit_table *arte = NULL, *rte = NULL;
770 	uint64_t mbits_per_sec, ind_calc, previous_rate = 0;
771 	int i;
772 
773 	mbits_per_sec = (bytes_per_sec * 8);
774 	if (flags & RS_PACING_LT) {
775 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
776 		    (rs->rs_lowest_valid <= 2)){
777 			/*
778 			 * Smaller than 1Meg, only
779 			 * 3 entries can match it.
780 			 */
781 			previous_rate = 0;
782 			for(i = rs->rs_lowest_valid; i < 3; i++) {
783 				if (bytes_per_sec <= rs->rs_rlt[i].rate) {
784 					rte = &rs->rs_rlt[i];
785 					break;
786 				} else if (rs->rs_rlt[i].flags & HDWRPACE_INITED) {
787 					arte = &rs->rs_rlt[i];
788 				}
789 				previous_rate = rs->rs_rlt[i].rate;
790 			}
791 			goto done;
792 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
793 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
794 			/*
795 			 * Larger than 1G (the majority of
796 			 * our table.
797 			 */
798 			if (mbits_per_sec < RS_TEN_GIGABIT_PERSEC)
799 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
800 			else
801 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
802 			previous_rate = rs->rs_rlt[(ALL_HARDWARE_RATES-2)].rate;
803 			goto done;
804 		}
805 		/*
806 		 * If we reach here its in our table (between 1Meg - 1000Meg),
807 		 * just take the rounded down mbits per second, and add
808 		 * 1Megabit to it, from this we can calculate
809 		 * the index in the table.
810 		 */
811 		ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
812 		if ((ind_calc * RS_ONE_MEGABIT_PERSEC) != mbits_per_sec)
813 			ind_calc++;
814 		/* our table is offset by 3, we add 2 */
815 		ind_calc += 2;
816 		if (ind_calc > (ALL_HARDWARE_RATES-1)) {
817 			/* This should not happen */
818 			ind_calc = ALL_HARDWARE_RATES-1;
819 		}
820 		if ((ind_calc >= rs->rs_lowest_valid) &&
821 		    (ind_calc <= rs->rs_highest_valid)) {
822 			rte = &rs->rs_rlt[ind_calc];
823 			if (ind_calc >= 1)
824 				previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
825 		}
826 	} else if (flags & RS_PACING_EXACT_MATCH) {
827 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
828 		    (rs->rs_lowest_valid <= 2)){
829 			for(i = rs->rs_lowest_valid; i < 3; i++) {
830 				if (bytes_per_sec == rs->rs_rlt[i].rate) {
831 					rte = &rs->rs_rlt[i];
832 					break;
833 				}
834 			}
835 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
836 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
837 			/* > 1Gbps only one rate */
838 			if (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) {
839 				/* Its 10G wow */
840 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
841 			}
842 		} else {
843 			/* Ok it must be a exact meg (its between 1G and 1Meg) */
844 			ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
845 			if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
846 				/* its an exact Mbps */
847 				ind_calc += 2;
848 				if (ind_calc > (ALL_HARDWARE_RATES-1)) {
849 					/* This should not happen */
850 					ind_calc = ALL_HARDWARE_RATES-1;
851 				}
852 				if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED)
853 					rte = &rs->rs_rlt[ind_calc];
854 			}
855 		}
856 	} else {
857 		/* we want greater than the requested rate */
858 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
859 		    (rs->rs_lowest_valid <= 2)){
860 			arte = &rs->rs_rlt[3]; /* set alternate to 1Meg */
861 			for (i=2; i>=rs->rs_lowest_valid; i--) {
862 				if (bytes_per_sec < rs->rs_rlt[i].rate) {
863 					rte = &rs->rs_rlt[i];
864 					if (i >= 1) {
865 						previous_rate = rs->rs_rlt[(i-1)].rate;
866 					}
867 					break;
868 				} else if ((flags & RS_PACING_GEQ) &&
869 					   (bytes_per_sec == rs->rs_rlt[i].rate)) {
870 					rte = &rs->rs_rlt[i];
871 					if (i >= 1) {
872 						previous_rate = rs->rs_rlt[(i-1)].rate;
873 					}
874 					break;
875 				} else {
876 					arte = &rs->rs_rlt[i]; /* new alternate */
877 				}
878 			}
879 		} else if (mbits_per_sec > RS_ONE_GIGABIT_PERSEC) {
880 			if ((bytes_per_sec < rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
881 			    (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
882 				/* Our top rate is larger than the request */
883 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
884 			} else if ((flags & RS_PACING_GEQ) &&
885 				   (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
886 				   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
887 				/* It matches our top rate */
888 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
889 			} else if (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED) {
890 				/* The top rate is an alternative */
891 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
892 			}
893 			previous_rate = rs->rs_rlt[(ALL_HARDWARE_RATES-2)].rate;
894 		} else {
895 			/* Its in our range 1Meg - 1Gig */
896 			if (flags & RS_PACING_GEQ) {
897 				ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
898 				if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
899 					if (ind_calc > (ALL_HARDWARE_RATES-1)) {
900 						/* This should not happen */
901 						ind_calc = (ALL_HARDWARE_RATES-1);
902 					}
903 					rte = &rs->rs_rlt[ind_calc];
904 					if (ind_calc >= 1)
905 						previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
906 				}
907 				goto done;
908 			}
909 			ind_calc = (mbits_per_sec + (RS_ONE_MEGABIT_PERSEC-1))/RS_ONE_MEGABIT_PERSEC;
910 			ind_calc += 2;
911 			if (ind_calc > (ALL_HARDWARE_RATES-1)) {
912 				/* This should not happen */
913 				ind_calc = ALL_HARDWARE_RATES-1;
914 			}
915 			if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) {
916 				rte = &rs->rs_rlt[ind_calc];
917 				if (ind_calc >= 1)
918 					previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
919 			}
920 		}
921 	}
922 done:
923 	if ((rte == NULL) &&
924 	    (arte != NULL) &&
925 	    (flags & RS_PACING_SUB_OK)) {
926 		/* We can use the substitute */
927 		rte = arte;
928 	}
929 	if (lower_rate)
930 		*lower_rate = previous_rate;
931 	return (rte);
932 }
933 
934 /*
935  * For an explanation of why the argument is volatile please
936  * look at the comments around rt_setup_rate().
937  */
938 static const struct tcp_hwrate_limit_table *
939 tcp_find_suitable_rate(const volatile struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags, uint64_t *lower_rate)
940 {
941 	/**
942 	 * Hunt the rate table with the restrictions in flags and find a
943 	 * suitable rate if possible.
944 	 * RS_PACING_EXACT_MATCH - look for an exact match to rate.
945 	 * RS_PACING_GT     - must be greater than.
946 	 * RS_PACING_GEQ    - must be greater than or equal.
947 	 * RS_PACING_LT     - must be less than.
948 	 * RS_PACING_SUB_OK - If we don't meet criteria a
949 	 *                    substitute is ok.
950 	 */
951 	int i, matched;
952 	struct tcp_hwrate_limit_table *rte = NULL;
953 	uint64_t previous_rate = 0;
954 
955 	if ((rs->rs_flags & RS_INT_TBL) &&
956 	    (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)) {
957 		/*
958 		 * Here we don't want to paw thru
959 		 * a big table, we have everything
960 		 * from 1Meg - 1000Meg in 1Meg increments.
961 		 * Use an alternate method to "lookup".
962 		 */
963 		return (tcp_int_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate));
964 	}
965 	if ((flags & RS_PACING_LT) ||
966 	    (flags & RS_PACING_EXACT_MATCH)) {
967 		/*
968 		 * For exact and less than we go forward through the table.
969 		 * This way when we find one larger we stop (exact was a
970 		 * toss up).
971 		 */
972 		for (i = rs->rs_lowest_valid, matched = 0; i <= rs->rs_highest_valid; i++) {
973 			if ((flags & RS_PACING_EXACT_MATCH) &&
974 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
975 				rte = &rs->rs_rlt[i];
976 				matched = 1;
977 				if (lower_rate != NULL)
978 					*lower_rate = previous_rate;
979 				break;
980 			} else if ((flags & RS_PACING_LT) &&
981 			    (bytes_per_sec <= rs->rs_rlt[i].rate)) {
982 				rte = &rs->rs_rlt[i];
983 				matched = 1;
984 				if (lower_rate != NULL)
985 					*lower_rate = previous_rate;
986 				break;
987 			}
988 			previous_rate = rs->rs_rlt[i].rate;
989 			if (bytes_per_sec > rs->rs_rlt[i].rate)
990 				break;
991 		}
992 		if ((matched == 0) &&
993 		    (flags & RS_PACING_LT) &&
994 		    (flags & RS_PACING_SUB_OK)) {
995 			/* Kick in a substitute (the lowest) */
996 			rte = &rs->rs_rlt[rs->rs_lowest_valid];
997 		}
998 	} else {
999 		/*
1000 		 * Here we go backward through the table so that we can find
1001 		 * the one greater in theory faster (but its probably a
1002 		 * wash).
1003 		 */
1004 		for (i = rs->rs_highest_valid, matched = 0; i >= rs->rs_lowest_valid; i--) {
1005 			if (rs->rs_rlt[i].rate > bytes_per_sec) {
1006 				/* A possible candidate */
1007 				rte = &rs->rs_rlt[i];
1008 			}
1009 			if ((flags & RS_PACING_GEQ) &&
1010 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
1011 				/* An exact match and we want equal */
1012 				matched = 1;
1013 				rte = &rs->rs_rlt[i];
1014 				break;
1015 			} else if (rte) {
1016 				/*
1017 				 * Found one that is larger than but don't
1018 				 * stop, there may be a more closer match.
1019 				 */
1020 				matched = 1;
1021 			}
1022 			if (rs->rs_rlt[i].rate < bytes_per_sec) {
1023 				/*
1024 				 * We found a table entry that is smaller,
1025 				 * stop there will be none greater or equal.
1026 				 */
1027 				if (lower_rate != NULL)
1028 					*lower_rate = rs->rs_rlt[i].rate;
1029 				break;
1030 			}
1031 		}
1032 		if ((matched == 0) &&
1033 		    (flags & RS_PACING_SUB_OK)) {
1034 			/* Kick in a substitute (the highest) */
1035 			rte = &rs->rs_rlt[rs->rs_highest_valid];
1036 		}
1037 	}
1038 	return (rte);
1039 }
1040 
1041 static struct ifnet *
1042 rt_find_real_interface(struct ifnet *ifp, struct inpcb *inp, int *error)
1043 {
1044 	struct ifnet *tifp;
1045 	struct m_snd_tag *tag, *ntag;
1046 	union if_snd_tag_alloc_params params = {
1047 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
1048 		.rate_limit.hdr.flowid = inp->inp_flowid,
1049 		.rate_limit.hdr.numa_domain = inp->inp_numa_domain,
1050 		.rate_limit.max_rate = COMMON_RATE,
1051 		.rate_limit.flags = M_NOWAIT,
1052 	};
1053 	int err;
1054 #ifdef RSS
1055 	params.rate_limit.hdr.flowtype = ((inp->inp_vflag & INP_IPV6) ?
1056 	    M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_TCP_IPV4);
1057 #else
1058 	params.rate_limit.hdr.flowtype = M_HASHTYPE_OPAQUE_HASH;
1059 #endif
1060 	err = m_snd_tag_alloc(ifp, &params, &tag);
1061 	if (err) {
1062 		/* Failed to setup a tag? */
1063 		if (error)
1064 			*error = err;
1065 		return (NULL);
1066 	}
1067 	ntag = tag;
1068 	while (ntag->sw->next_snd_tag != NULL) {
1069 		ntag = ntag->sw->next_snd_tag(ntag);
1070 	}
1071 	tifp = ntag->ifp;
1072 	m_snd_tag_rele(tag);
1073 	return (tifp);
1074 }
1075 
1076 static void
1077 rl_increment_using(const struct tcp_hwrate_limit_table *rte)
1078 {
1079 	struct tcp_hwrate_limit_table *decon_rte;
1080 
1081 	decon_rte = __DECONST(struct tcp_hwrate_limit_table *, rte);
1082 	atomic_add_long(&decon_rte->using, 1);
1083 }
1084 
1085 static void
1086 rl_decrement_using(const struct tcp_hwrate_limit_table *rte)
1087 {
1088 	struct tcp_hwrate_limit_table *decon_rte;
1089 
1090 	decon_rte = __DECONST(struct tcp_hwrate_limit_table *, rte);
1091 	atomic_subtract_long(&decon_rte->using, 1);
1092 }
1093 
1094 void
1095 tcp_rl_log_enobuf(const struct tcp_hwrate_limit_table *rte)
1096 {
1097 	struct tcp_hwrate_limit_table *decon_rte;
1098 
1099 	decon_rte = __DECONST(struct tcp_hwrate_limit_table *, rte);
1100 	atomic_add_long(&decon_rte->rs_num_enobufs, 1);
1101 }
1102 
1103 /*
1104  * Do NOT take the __noinline out of the
1105  * find_rs_for_ifp() function. If you do the inline
1106  * of it for the rt_setup_rate() will show you a
1107  * compiler bug. For some reason the compiler thinks
1108  * the list can never be empty. The consequence of
1109  * this will be a crash when we dereference NULL
1110  * if an ifp is removed just has a hw rate limit
1111  * is attempted. If you are working on the compiler
1112  * and want to "test" this go ahead and take the noinline
1113  * out otherwise let sleeping dogs ly until such time
1114  * as we get a compiler fix 10/2/20 -- RRS
1115  */
1116 static __noinline struct tcp_rate_set *
1117 find_rs_for_ifp(struct ifnet *ifp)
1118 {
1119 	struct tcp_rate_set *rs;
1120 
1121 	CK_LIST_FOREACH(rs, &int_rs, next) {
1122 		if ((rs->rs_ifp == ifp) &&
1123 		    (rs->rs_if_dunit == ifp->if_dunit)) {
1124 			/* Ok we found it */
1125 			return (rs);
1126 		}
1127 	}
1128 	return (NULL);
1129 }
1130 
1131 
1132 static const struct tcp_hwrate_limit_table *
1133 rt_setup_rate(struct inpcb *inp, struct ifnet *ifp, uint64_t bytes_per_sec,
1134     uint32_t flags, int *error, uint64_t *lower_rate)
1135 {
1136 	/* First lets find the interface if it exists */
1137 	const struct tcp_hwrate_limit_table *rte;
1138 	/*
1139 	 * So why is rs volatile? This is to defeat a
1140 	 * compiler bug where in the compiler is convinced
1141 	 * that rs can never be NULL (which is not true). Because
1142 	 * of its conviction it nicely optimizes out the if ((rs == NULL
1143 	 * below which means if you get a NULL back you dereference it.
1144 	 */
1145 	volatile struct tcp_rate_set *rs;
1146 	struct epoch_tracker et;
1147 	struct ifnet *oifp = ifp;
1148 	int err;
1149 
1150 	NET_EPOCH_ENTER(et);
1151 use_real_interface:
1152 	rs = find_rs_for_ifp(ifp);
1153 	if ((rs == NULL) ||
1154 	    (rs->rs_flags & RS_INTF_NO_SUP) ||
1155 	    (rs->rs_flags & RS_IS_DEAD)) {
1156 		/*
1157 		 * This means we got a packet *before*
1158 		 * the IF-UP was processed below, <or>
1159 		 * while or after we already received an interface
1160 		 * departed event. In either case we really don't
1161 		 * want to do anything with pacing, in
1162 		 * the departing case the packet is not
1163 		 * going to go very far. The new case
1164 		 * might be arguable, but its impossible
1165 		 * to tell from the departing case.
1166 		 */
1167 		if (error)
1168 			*error = ENODEV;
1169 		NET_EPOCH_EXIT(et);
1170 		return (NULL);
1171 	}
1172 
1173 	if ((rs == NULL) || (rs->rs_disable != 0)) {
1174 		if (error)
1175 			*error = ENOSPC;
1176 		NET_EPOCH_EXIT(et);
1177 		return (NULL);
1178 	}
1179 	if (rs->rs_flags & RS_IS_DEFF) {
1180 		/* We need to find the real interface */
1181 		struct ifnet *tifp;
1182 
1183 		tifp = rt_find_real_interface(ifp, inp, error);
1184 		if (tifp == NULL) {
1185 			if (rs->rs_disable && error)
1186 				*error = ENOTSUP;
1187 			NET_EPOCH_EXIT(et);
1188 			return (NULL);
1189 		}
1190 		KASSERT((tifp != ifp),
1191 			("Lookup failure ifp:%p inp:%p rt_find_real_interface() returns the same interface tifp:%p?\n",
1192 			 ifp, inp, tifp));
1193 		ifp = tifp;
1194 		goto use_real_interface;
1195 	}
1196 	if (rs->rs_flow_limit &&
1197 	    ((rs->rs_flows_using + 1) > rs->rs_flow_limit)) {
1198 		if (error)
1199 			*error = ENOSPC;
1200 		NET_EPOCH_EXIT(et);
1201 		return (NULL);
1202 	}
1203 	rte = tcp_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate);
1204 	if (rte) {
1205 		err = in_pcbattach_txrtlmt(inp, oifp,
1206 		    inp->inp_flowtype,
1207 		    inp->inp_flowid,
1208 		    rte->rate,
1209 		    &inp->inp_snd_tag);
1210 		if (err) {
1211 			/* Failed to attach */
1212 			if (error)
1213 				*error = err;
1214 			rte = NULL;
1215 		} else {
1216 			KASSERT((inp->inp_snd_tag != NULL) ,
1217 				("Setup rate has no snd_tag inp:%p rte:%p rate:%llu rs:%p",
1218 				 inp, rte, (unsigned long long)rte->rate, rs));
1219 #ifdef INET
1220 			counter_u64_add(rate_limit_new, 1);
1221 #endif
1222 		}
1223 	}
1224 	if (rte) {
1225 		/*
1226 		 * We use an atomic here for accounting so we don't have to
1227 		 * use locks when freeing.
1228 		 */
1229 		atomic_add_64(&rs->rs_flows_using, 1);
1230 	}
1231 	NET_EPOCH_EXIT(et);
1232 	return (rte);
1233 }
1234 
1235 static void
1236 tcp_rl_ifnet_link(void *arg __unused, struct ifnet *ifp, int link_state)
1237 {
1238 	int error;
1239 	struct tcp_rate_set *rs;
1240 	struct epoch_tracker et;
1241 
1242 	if (((ifp->if_capenable & IFCAP_TXRTLMT) == 0) ||
1243 	    (link_state != LINK_STATE_UP)) {
1244 		/*
1245 		 * We only care on an interface going up that is rate-limit
1246 		 * capable.
1247 		 */
1248 		return;
1249 	}
1250 	NET_EPOCH_ENTER(et);
1251 	mtx_lock(&rs_mtx);
1252 	rs = find_rs_for_ifp(ifp);
1253 	if (rs) {
1254 		/* We already have initialized this guy */
1255 		mtx_unlock(&rs_mtx);
1256 		NET_EPOCH_EXIT(et);
1257 		return;
1258 	}
1259 	mtx_unlock(&rs_mtx);
1260 	NET_EPOCH_EXIT(et);
1261 	rt_setup_new_rs(ifp, &error);
1262 }
1263 
1264 static void
1265 tcp_rl_ifnet_departure(void *arg __unused, struct ifnet *ifp)
1266 {
1267 	struct tcp_rate_set *rs;
1268 	struct epoch_tracker et;
1269 	int i;
1270 
1271 	NET_EPOCH_ENTER(et);
1272 	mtx_lock(&rs_mtx);
1273 	rs = find_rs_for_ifp(ifp);
1274 	if (rs) {
1275 		CK_LIST_REMOVE(rs, next);
1276 		rs_number_alive--;
1277 		rs->rs_flags |= RS_IS_DEAD;
1278 		for (i = 0; i < rs->rs_rate_cnt; i++) {
1279 			if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1280 				in_pcbdetach_tag(rs->rs_rlt[i].tag);
1281 				rs->rs_rlt[i].tag = NULL;
1282 			}
1283 			rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1284 		}
1285 		if (rs->rs_flows_using == 0)
1286 			rs_defer_destroy(rs);
1287 	}
1288 	mtx_unlock(&rs_mtx);
1289 	NET_EPOCH_EXIT(et);
1290 }
1291 
1292 static void
1293 tcp_rl_shutdown(void *arg __unused, int howto __unused)
1294 {
1295 	struct tcp_rate_set *rs, *nrs;
1296 	struct epoch_tracker et;
1297 	int i;
1298 
1299 	NET_EPOCH_ENTER(et);
1300 	mtx_lock(&rs_mtx);
1301 	CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) {
1302 		CK_LIST_REMOVE(rs, next);
1303 		rs_number_alive--;
1304 		rs->rs_flags |= RS_IS_DEAD;
1305 		for (i = 0; i < rs->rs_rate_cnt; i++) {
1306 			if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1307 				in_pcbdetach_tag(rs->rs_rlt[i].tag);
1308 				rs->rs_rlt[i].tag = NULL;
1309 			}
1310 			rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1311 		}
1312 		if (rs->rs_flows_using == 0)
1313 			rs_defer_destroy(rs);
1314 	}
1315 	mtx_unlock(&rs_mtx);
1316 	NET_EPOCH_EXIT(et);
1317 }
1318 
1319 const struct tcp_hwrate_limit_table *
1320 tcp_set_pacing_rate(struct tcpcb *tp, struct ifnet *ifp,
1321     uint64_t bytes_per_sec, int flags, int *error, uint64_t *lower_rate)
1322 {
1323 	const struct tcp_hwrate_limit_table *rte;
1324 #ifdef KERN_TLS
1325 	struct ktls_session *tls;
1326 #endif
1327 
1328 	INP_WLOCK_ASSERT(tp->t_inpcb);
1329 
1330 	if (tp->t_inpcb->inp_snd_tag == NULL) {
1331 		/*
1332 		 * We are setting up a rate for the first time.
1333 		 */
1334 		if ((ifp->if_capenable & IFCAP_TXRTLMT) == 0) {
1335 			/* Not supported by the egress */
1336 			if (error)
1337 				*error = ENODEV;
1338 			return (NULL);
1339 		}
1340 #ifdef KERN_TLS
1341 		tls = NULL;
1342 		if (tp->t_inpcb->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
1343 			tls = tp->t_inpcb->inp_socket->so_snd.sb_tls_info;
1344 
1345 			if ((ifp->if_capenable & IFCAP_TXTLS_RTLMT) == 0 ||
1346 			    tls->mode != TCP_TLS_MODE_IFNET) {
1347 				if (error)
1348 					*error = ENODEV;
1349 				return (NULL);
1350 			}
1351 		}
1352 #endif
1353 		rte = rt_setup_rate(tp->t_inpcb, ifp, bytes_per_sec, flags, error, lower_rate);
1354 		if (rte)
1355 			rl_increment_using(rte);
1356 #ifdef KERN_TLS
1357 		if (rte != NULL && tls != NULL && tls->snd_tag != NULL) {
1358 			/*
1359 			 * Fake a route change error to reset the TLS
1360 			 * send tag.  This will convert the existing
1361 			 * tag to a TLS ratelimit tag.
1362 			 */
1363 			MPASS(tls->snd_tag->sw->type == IF_SND_TAG_TYPE_TLS);
1364 			ktls_output_eagain(tp->t_inpcb, tls);
1365 		}
1366 #endif
1367 	} else {
1368 		/*
1369 		 * We are modifying a rate, wrong interface?
1370 		 */
1371 		if (error)
1372 			*error = EINVAL;
1373 		rte = NULL;
1374 	}
1375 	if (rte != NULL) {
1376 		tp->t_pacing_rate = rte->rate;
1377 		*error = 0;
1378 	}
1379 	return (rte);
1380 }
1381 
1382 const struct tcp_hwrate_limit_table *
1383 tcp_chg_pacing_rate(const struct tcp_hwrate_limit_table *crte,
1384     struct tcpcb *tp, struct ifnet *ifp,
1385     uint64_t bytes_per_sec, int flags, int *error, uint64_t *lower_rate)
1386 {
1387 	const struct tcp_hwrate_limit_table *nrte;
1388 	const struct tcp_rate_set *rs;
1389 #ifdef KERN_TLS
1390 	struct ktls_session *tls = NULL;
1391 #endif
1392 	int err;
1393 
1394 	INP_WLOCK_ASSERT(tp->t_inpcb);
1395 
1396 	if (crte == NULL) {
1397 		/* Wrong interface */
1398 		if (error)
1399 			*error = EINVAL;
1400 		return (NULL);
1401 	}
1402 
1403 #ifdef KERN_TLS
1404 	if (tp->t_inpcb->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
1405 		tls = tp->t_inpcb->inp_socket->so_snd.sb_tls_info;
1406 		MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1407 		if (tls->snd_tag != NULL &&
1408 		    tls->snd_tag->sw->type != IF_SND_TAG_TYPE_TLS_RATE_LIMIT) {
1409 			/*
1410 			 * NIC probably doesn't support ratelimit TLS
1411 			 * tags if it didn't allocate one when an
1412 			 * existing rate was present, so ignore.
1413 			 */
1414 			if (error)
1415 				*error = EOPNOTSUPP;
1416 			return (NULL);
1417 		}
1418 	}
1419 #endif
1420 	if (tp->t_inpcb->inp_snd_tag == NULL) {
1421 		/* Wrong interface */
1422 		if (error)
1423 			*error = EINVAL;
1424 		return (NULL);
1425 	}
1426 	rs = crte->ptbl;
1427 	if ((rs->rs_flags & RS_IS_DEAD) ||
1428 	    (crte->flags & HDWRPACE_IFPDEPARTED)) {
1429 		/* Release the rate, and try anew */
1430 
1431 		tcp_rel_pacing_rate(crte, tp);
1432 		nrte = tcp_set_pacing_rate(tp, ifp,
1433 		    bytes_per_sec, flags, error, lower_rate);
1434 		return (nrte);
1435 	}
1436 	nrte = tcp_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate);
1437 	if (nrte == crte) {
1438 		/* No change */
1439 		if (error)
1440 			*error = 0;
1441 		return (crte);
1442 	}
1443 	if (nrte == NULL) {
1444 		/* Release the old rate */
1445 		if (error)
1446 			*error = ENOENT;
1447 		tcp_rel_pacing_rate(crte, tp);
1448 		return (NULL);
1449 	}
1450 	rl_decrement_using(crte);
1451 	rl_increment_using(nrte);
1452 	/* Change rates to our new entry */
1453 #ifdef KERN_TLS
1454 	if (tls != NULL)
1455 		err = ktls_modify_txrtlmt(tls, nrte->rate);
1456 	else
1457 #endif
1458 		err = in_pcbmodify_txrtlmt(tp->t_inpcb, nrte->rate);
1459 	if (err) {
1460 		rl_decrement_using(nrte);
1461 		/* Do we still have a snd-tag attached? */
1462 		if (tp->t_inpcb->inp_snd_tag)
1463 			in_pcbdetach_txrtlmt(tp->t_inpcb);
1464 		if (error)
1465 			*error = err;
1466 		return (NULL);
1467 	} else {
1468 #ifdef INET
1469 		counter_u64_add(rate_limit_chg, 1);
1470 #endif
1471 	}
1472 	if (error)
1473 		*error = 0;
1474 	tp->t_pacing_rate = nrte->rate;
1475 	return (nrte);
1476 }
1477 
1478 void
1479 tcp_rel_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp)
1480 {
1481 	const struct tcp_rate_set *crs;
1482 	struct tcp_rate_set *rs;
1483 	uint64_t pre;
1484 
1485 	INP_WLOCK_ASSERT(tp->t_inpcb);
1486 
1487 	tp->t_pacing_rate = -1;
1488 	crs = crte->ptbl;
1489 	/*
1490 	 * Now we must break the const
1491 	 * in order to release our refcount.
1492 	 */
1493 	rs = __DECONST(struct tcp_rate_set *, crs);
1494 	rl_decrement_using(crte);
1495 	pre = atomic_fetchadd_64(&rs->rs_flows_using, -1);
1496 	if (pre == 1) {
1497 		struct epoch_tracker et;
1498 
1499 		NET_EPOCH_ENTER(et);
1500 		mtx_lock(&rs_mtx);
1501 		/*
1502 		 * Is it dead?
1503 		 */
1504 		if (rs->rs_flags & RS_IS_DEAD)
1505 			rs_defer_destroy(rs);
1506 		mtx_unlock(&rs_mtx);
1507 		NET_EPOCH_EXIT(et);
1508 	}
1509 
1510 	/*
1511 	 * XXX: If this connection is using ifnet TLS, should we
1512 	 * switch it to using an unlimited rate, or perhaps use
1513 	 * ktls_output_eagain() to reset the send tag to a plain
1514 	 * TLS tag?
1515 	 */
1516 	in_pcbdetach_txrtlmt(tp->t_inpcb);
1517 }
1518 
1519 #define ONE_POINT_TWO_MEG 150000 /* 1.2 megabits in bytes */
1520 #define ONE_HUNDRED_MBPS 12500000	/* 100Mbps in bytes per second */
1521 #define FIVE_HUNDRED_MBPS 62500000	/* 500Mbps in bytes per second */
1522 #define MAX_MSS_SENT 43	/* 43 mss = 43 x 1500 = 64,500 bytes */
1523 
1524 static void
1525 tcp_log_pacing_size(struct tcpcb *tp, uint64_t bw, uint32_t segsiz, uint32_t new_tso,
1526 		    uint64_t hw_rate, uint32_t time_between, uint32_t calc_time_between,
1527 		    uint32_t segs, uint32_t res_div, uint16_t mult, uint8_t mod)
1528 {
1529 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1530 		union tcp_log_stackspecific log;
1531 		struct timeval tv;
1532 		uint32_t cts;
1533 
1534 		memset(&log, 0, sizeof(log));
1535 		cts = tcp_get_usecs(&tv);
1536 		log.u_bbr.flex1 = segsiz;
1537 		log.u_bbr.flex2 = new_tso;
1538 		log.u_bbr.flex3 = time_between;
1539 		log.u_bbr.flex4 = calc_time_between;
1540 		log.u_bbr.flex5 = segs;
1541 		log.u_bbr.flex6 = res_div;
1542 		log.u_bbr.flex7 = mult;
1543 		log.u_bbr.flex8 = mod;
1544 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1545 		log.u_bbr.cur_del_rate = bw;
1546 		log.u_bbr.delRate = hw_rate;
1547 		TCP_LOG_EVENTP(tp, NULL,
1548 		    &tp->t_inpcb->inp_socket->so_rcv,
1549 		    &tp->t_inpcb->inp_socket->so_snd,
1550 		    TCP_HDWR_PACE_SIZE, 0,
1551 		    0, &log, false, &tv);
1552 	}
1553 }
1554 
1555 uint32_t
1556 tcp_get_pacing_burst_size (struct tcpcb *tp, uint64_t bw, uint32_t segsiz, int can_use_1mss,
1557    const struct tcp_hwrate_limit_table *te, int *err)
1558 {
1559 	/*
1560 	 * We use the google formula to calculate the
1561 	 * TSO size. I.E.
1562 	 * bw < 24Meg
1563 	 *   tso = 2mss
1564 	 * else
1565 	 *   tso = min(bw/1000, 64k)
1566 	 *
1567 	 * Note for these calculations we ignore the
1568 	 * packet overhead (enet hdr, ip hdr and tcp hdr).
1569 	 */
1570 	uint64_t lentim, res, bytes;
1571 	uint32_t new_tso, min_tso_segs;
1572 
1573 	bytes = bw / 1000;
1574 	if (bytes > (64 * 1000))
1575 		bytes = 64 * 1000;
1576 	/* Round up */
1577 	new_tso = (bytes + segsiz - 1) / segsiz;
1578 	if (can_use_1mss && (bw < ONE_POINT_TWO_MEG))
1579 		min_tso_segs = 1;
1580 	else
1581 		min_tso_segs = 2;
1582 	if (rs_floor_mss && (new_tso < rs_floor_mss))
1583 		new_tso = rs_floor_mss;
1584 	else if (new_tso < min_tso_segs)
1585 		new_tso = min_tso_segs;
1586 	if (new_tso > MAX_MSS_SENT)
1587 		new_tso = MAX_MSS_SENT;
1588 	new_tso *= segsiz;
1589  	tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1590 			    0, 0, 0, 0, 0, 0, 1);
1591 	/*
1592 	 * If we are not doing hardware pacing
1593 	 * then we are done.
1594 	 */
1595 	if (te == NULL) {
1596 		if (err)
1597 			*err = 0;
1598 		return(new_tso);
1599 	}
1600 	/*
1601 	 * For hardware pacing we look at the
1602 	 * rate you are sending at and compare
1603 	 * that to the rate you have in hardware.
1604 	 *
1605 	 * If the hardware rate is slower than your
1606 	 * software rate then you are in error and
1607 	 * we will build a queue in our hardware whic
1608 	 * is probably not desired, in such a case
1609 	 * just return the non-hardware TSO size.
1610 	 *
1611 	 * If the rate in hardware is faster (which
1612 	 * it should be) then look at how long it
1613 	 * takes to send one ethernet segment size at
1614 	 * your b/w and compare that to the time it
1615 	 * takes to send at the rate you had selected.
1616 	 *
1617 	 * If your time is greater (which we hope it is)
1618 	 * we get the delta between the two, and then
1619 	 * divide that into your pacing time. This tells
1620 	 * us how many MSS you can send down at once (rounded up).
1621 	 *
1622 	 * Note we also double this value if the b/w is over
1623 	 * 100Mbps. If its over 500meg we just set you to the
1624 	 * max (43 segments).
1625 	 */
1626 	if (te->rate > FIVE_HUNDRED_MBPS)
1627 		goto max;
1628 	if (te->rate == bw) {
1629 		/* We are pacing at exactly the hdwr rate */
1630 max:
1631 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1632 				    te->rate, te->time_between, (uint32_t)0,
1633 				    (segsiz * MAX_MSS_SENT), 0, 0, 3);
1634 		return (segsiz * MAX_MSS_SENT);
1635 	}
1636 	lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
1637 	res = lentim / bw;
1638 	if (res > te->time_between) {
1639 		uint32_t delta, segs, res_div;
1640 
1641 		res_div = ((res * num_of_waits_allowed) + wait_time_floor);
1642 		delta = res - te->time_between;
1643 		segs = (res_div + delta - 1)/delta;
1644 		if (segs < min_tso_segs)
1645 			segs = min_tso_segs;
1646 		if (segs < rs_hw_floor_mss)
1647 			segs = rs_hw_floor_mss;
1648 		if (segs > MAX_MSS_SENT)
1649 			segs = MAX_MSS_SENT;
1650 		segs *= segsiz;
1651 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1652 				    te->rate, te->time_between, (uint32_t)res,
1653 				    segs, res_div, 1, 3);
1654 		if (err)
1655 			*err = 0;
1656 		if (segs < new_tso) {
1657 			/* unexpected ? */
1658 			return(new_tso);
1659 		} else {
1660 			return (segs);
1661 		}
1662 	} else {
1663 		/*
1664 		 * Your time is smaller which means
1665 		 * we will grow a queue on our
1666 		 * hardware. Send back the non-hardware
1667 		 * rate.
1668 		 */
1669 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1670 				    te->rate, te->time_between, (uint32_t)res,
1671 				    0, 0, 0, 4);
1672 		if (err)
1673 			*err = -1;
1674 		return (new_tso);
1675 	}
1676 }
1677 
1678 uint64_t
1679 tcp_hw_highest_rate_ifp(struct ifnet *ifp, struct inpcb *inp)
1680 {
1681 	struct epoch_tracker et;
1682 	struct tcp_rate_set *rs;
1683 	uint64_t rate_ret;
1684 
1685 	NET_EPOCH_ENTER(et);
1686 use_next_interface:
1687 	rs = find_rs_for_ifp(ifp);
1688 	if (rs == NULL) {
1689 		/* This interface does not do ratelimiting */
1690 		rate_ret = 0;
1691 	} else if (rs->rs_flags & RS_IS_DEFF) {
1692 		/* We need to find the real interface */
1693 		struct ifnet *tifp;
1694 
1695 		tifp = rt_find_real_interface(ifp, inp, NULL);
1696 		if (tifp == NULL) {
1697 			NET_EPOCH_EXIT(et);
1698 			return (0);
1699 		}
1700 		ifp = tifp;
1701 		goto use_next_interface;
1702 	} else {
1703 		/* Lets return the highest rate this guy has */
1704 		rate_ret = rs->rs_rlt[rs->rs_highest_valid].rate;
1705 	}
1706 	NET_EPOCH_EXIT(et);
1707 	return(rate_ret);
1708 }
1709 
1710 static eventhandler_tag rl_ifnet_departs;
1711 static eventhandler_tag rl_ifnet_arrives;
1712 static eventhandler_tag rl_shutdown_start;
1713 
1714 static void
1715 tcp_rs_init(void *st __unused)
1716 {
1717 	CK_LIST_INIT(&int_rs);
1718 	rs_number_alive = 0;
1719 	rs_number_dead = 0;
1720 	mtx_init(&rs_mtx, "tcp_rs_mtx", "rsmtx", MTX_DEF);
1721 	rl_ifnet_departs = EVENTHANDLER_REGISTER(ifnet_departure_event,
1722 	    tcp_rl_ifnet_departure,
1723 	    NULL, EVENTHANDLER_PRI_ANY);
1724 	rl_ifnet_arrives = EVENTHANDLER_REGISTER(ifnet_link_event,
1725 	    tcp_rl_ifnet_link,
1726 	    NULL, EVENTHANDLER_PRI_ANY);
1727 	rl_shutdown_start = EVENTHANDLER_REGISTER(shutdown_pre_sync,
1728 	    tcp_rl_shutdown, NULL,
1729 	    SHUTDOWN_PRI_FIRST);
1730 	printf("TCP_ratelimit: Is now initialized\n");
1731 }
1732 
1733 SYSINIT(tcp_rl_init, SI_SUB_SMP + 1, SI_ORDER_ANY, tcp_rs_init, NULL);
1734 #endif
1735