xref: /freebsd/sys/netinet/tcp_ratelimit.c (revision ec965063070e5753c166cf592c9336444b74720a)
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_ratelimit.h"
40 #include <sys/param.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/mbuf.h>
44 #include <sys/socket.h>
45 #include <sys/socketvar.h>
46 #include <sys/sysctl.h>
47 #include <sys/eventhandler.h>
48 #include <sys/mutex.h>
49 #include <sys/ck.h>
50 #include <net/if.h>
51 #include <net/if_var.h>
52 #include <netinet/in.h>
53 #include <netinet/in_pcb.h>
54 #define TCPSTATES		/* for logging */
55 #include <netinet/tcp_var.h>
56 #include <netinet/tcp_hpts.h>
57 #include <netinet/tcp_log_buf.h>
58 #include <netinet/tcp_ratelimit.h>
59 #ifndef USECS_IN_SECOND
60 #define USECS_IN_SECOND 1000000
61 #endif
62 /*
63  * For the purposes of each send, what is the size
64  * of an ethernet frame.
65  */
66 MALLOC_DEFINE(M_TCPPACE, "tcp_hwpace", "TCP Hardware pacing memory");
67 #ifdef RATELIMIT
68 
69 /*
70  * The following preferred table will seem weird to
71  * the casual viewer. Why do we not have any rates below
72  * 1Mbps? Why do we have a rate at 1.44Mbps called common?
73  * Why do the rates cluster in the 1-100Mbps range more
74  * than others? Why does the table jump around at the beginnign
75  * and then be more consistently raising?
76  *
77  * Let me try to answer those questions. A lot of
78  * this is dependant on the hardware. We have three basic
79  * supporters of rate limiting
80  *
81  * Chelsio - Supporting 16 configurable rates.
82  * Mlx  - c4 supporting 13 fixed rates.
83  * Mlx  - c5 & c6 supporting 127 configurable rates.
84  *
85  * The c4 is why we have a common rate that is available
86  * in all rate tables. This is a selected rate from the
87  * c4 table and we assure its available in all ratelimit
88  * tables. This way the tcp_ratelimit code has an assured
89  * rate it should always be able to get. This answers a
90  * couple of the questions above.
91  *
92  * So what about the rest, well the table is built to
93  * try to get the most out of a joint hardware/software
94  * pacing system.  The software pacer will always pick
95  * a rate higher than the b/w that it is estimating
96  *
97  * on the path. This is done for two reasons.
98  * a) So we can discover more b/w
99  * and
100  * b) So we can send a block of MSS's down and then
101  *    have the software timer go off after the previous
102  *    send is completely out of the hardware.
103  *
104  * But when we do <b> we don't want to have the delay
105  * between the last packet sent by the hardware be
106  * excessively long (to reach our desired rate).
107  *
108  * So let me give an example for clarity.
109  *
110  * Lets assume that the tcp stack sees that 29,110,000 bps is
111  * what the bw of the path is. The stack would select the
112  * rate 31Mbps. 31Mbps means that each send that is done
113  * by the hardware will cause a 390 micro-second gap between
114  * the packets sent at that rate. For 29,110,000 bps we
115  * would need 416 micro-seconds gap between each send.
116  *
117  * Note that are calculating a complete time for pacing
118  * which includes the ethernet, IP and TCP overhead. So
119  * a full 1514 bytes is used for the above calculations.
120  * My testing has shown that both cards are also using this
121  * as their basis i.e. full payload size of the ethernet frame.
122  * The TCP stack caller needs to be aware of this and make the
123  * appropriate overhead calculations be included in its choices.
124  *
125  * Now, continuing our example, we pick a MSS size based on the
126  * delta between the two rates (416 - 390) divided into the rate
127  * we really wish to send at rounded up.  That results in a MSS
128  * send of 17 mss's at once. The hardware then will
129  * run out of data in a single 17MSS send in 6,630 micro-seconds.
130  *
131  * On the other hand the software pacer will send more data
132  * in 7,072 micro-seconds. This means that we will refill
133  * the hardware 52 microseconds after it would have sent
134  * next if it had not ran out of data. This is a win since we are
135  * only sending every 7ms or so and yet all the packets are spaced on
136  * the wire with 94% of what they should be and only
137  * the last packet is delayed extra to make up for the
138  * difference.
139  *
140  * Note that the above formula has two important caveat.
141  * If we are above (b/w wise) over 100Mbps we double the result
142  * of the MSS calculation. The second caveat is if we are 500Mbps
143  * or more we just send the maximum MSS at once i.e. 45MSS. At
144  * the higher b/w's even the cards have limits to what times (timer granularity)
145  * they can insert between packets and start to send more than one
146  * packet at a time on the wire.
147  *
148  */
149 #define COMMON_RATE 180500
150 const uint64_t desired_rates[] = {
151 	122500,			/* 1Mbps  - rate 1 */
152 	180500,			/* 1.44Mpbs - rate 2  common rate */
153 	375000,			/* 3Mbps    - rate 3 */
154 	625000,			/* 5Mbps    - rate 4 */
155 	1250000,		/* 10Mbps   - rate 5 */
156 	1875000,		/* 15Mbps   - rate 6 */
157 	2500000,		/* 20Mbps   - rate 7 */
158 	3125000,	       	/* 25Mbps   - rate 8 */
159 	3750000,		/* 30Mbps   - rate 9 */
160 	4375000,		/* 35Mbps   - rate 10 */
161 	5000000,		/* 40Meg    - rate 11 */
162 	6250000,		/* 50Mbps   - rate 12 */
163 	12500000,		/* 100Mbps  - rate 13 */
164 	25000000,		/* 200Mbps  - rate 14 */
165 	50000000,		/* 400Mbps  - rate 15 */
166 	100000000,		/* 800Mbps  - rate 16 */
167 	5625000,		/* 45Mbps   - rate 17 */
168 	6875000,		/* 55Mbps   - rate 19 */
169 	7500000,		/* 60Mbps   - rate 20 */
170 	8125000,		/* 65Mbps   - rate 21 */
171 	8750000,		/* 70Mbps   - rate 22 */
172 	9375000,		/* 75Mbps   - rate 23 */
173 	10000000,		/* 80Mbps   - rate 24 */
174 	10625000,		/* 85Mbps   - rate 25 */
175 	11250000,		/* 90Mbps   - rate 26 */
176 	11875000,		/* 95Mbps   - rate 27 */
177 	12500000,		/* 100Mbps  - rate 28 */
178 	13750000,		/* 110Mbps  - rate 29 */
179 	15000000,		/* 120Mbps  - rate 30 */
180 	16250000,		/* 130Mbps  - rate 31 */
181 	17500000,		/* 140Mbps  - rate 32 */
182 	18750000,		/* 150Mbps  - rate 33 */
183 	20000000,		/* 160Mbps  - rate 34 */
184 	21250000,		/* 170Mbps  - rate 35 */
185 	22500000,		/* 180Mbps  - rate 36 */
186 	23750000,		/* 190Mbps  - rate 37 */
187 	26250000,		/* 210Mbps  - rate 38 */
188 	27500000,		/* 220Mbps  - rate 39 */
189 	28750000,		/* 230Mbps  - rate 40 */
190 	30000000,	       	/* 240Mbps  - rate 41 */
191 	31250000,		/* 250Mbps  - rate 42 */
192 	34375000,		/* 275Mbps  - rate 43 */
193 	37500000,		/* 300Mbps  - rate 44 */
194 	40625000,		/* 325Mbps  - rate 45 */
195 	43750000,		/* 350Mbps  - rate 46 */
196 	46875000,		/* 375Mbps  - rate 47 */
197 	53125000,		/* 425Mbps  - rate 48 */
198 	56250000,		/* 450Mbps  - rate 49 */
199 	59375000,		/* 475Mbps  - rate 50 */
200 	62500000,		/* 500Mbps  - rate 51 */
201 	68750000,		/* 550Mbps  - rate 52 */
202 	75000000,		/* 600Mbps  - rate 53 */
203 	81250000,		/* 650Mbps  - rate 54 */
204 	87500000,		/* 700Mbps  - rate 55 */
205 	93750000,		/* 750Mbps  - rate 56 */
206 	106250000,		/* 850Mbps  - rate 57 */
207 	112500000,		/* 900Mbps  - rate 58 */
208 	125000000,		/* 1Gbps    - rate 59 */
209 	156250000,		/* 1.25Gps  - rate 60 */
210 	187500000,		/* 1.5Gps   - rate 61 */
211 	218750000,		/* 1.75Gps  - rate 62 */
212 	250000000,		/* 2Gbps    - rate 63 */
213 	281250000,		/* 2.25Gps  - rate 64 */
214 	312500000,		/* 2.5Gbps  - rate 65 */
215 	343750000,		/* 2.75Gbps - rate 66 */
216 	375000000,		/* 3Gbps    - rate 67 */
217 	500000000,		/* 4Gbps    - rate 68 */
218 	625000000,		/* 5Gbps    - rate 69 */
219 	750000000,		/* 6Gbps    - rate 70 */
220 	875000000,		/* 7Gbps    - rate 71 */
221 	1000000000,		/* 8Gbps    - rate 72 */
222 	1125000000,		/* 9Gbps    - rate 73 */
223 	1250000000,		/* 10Gbps   - rate 74 */
224 	1875000000,		/* 15Gbps   - rate 75 */
225 	2500000000		/* 20Gbps   - rate 76 */
226 };
227 
228 #define MAX_HDWR_RATES (sizeof(desired_rates)/sizeof(uint64_t))
229 #define RS_ORDERED_COUNT 16	/*
230 				 * Number that are in order
231 				 * at the beginning of the table,
232 				 * over this a sort is required.
233 				 */
234 #define RS_NEXT_ORDER_GROUP 16	/*
235 				 * The point in our table where
236 				 * we come fill in a second ordered
237 				 * group (index wise means -1).
238 				 */
239 #define ALL_HARDWARE_RATES 1004 /*
240 				 * 1Meg - 1Gig in 1 Meg steps
241 				 * plus 100, 200k  and 500k and
242 				 * 10Gig
243 				 */
244 
245 #define RS_ONE_MEGABIT_PERSEC 1000000
246 #define RS_ONE_GIGABIT_PERSEC 1000000000
247 #define RS_TEN_GIGABIT_PERSEC 10000000000
248 
249 static struct head_tcp_rate_set int_rs;
250 static struct mtx rs_mtx;
251 uint32_t rs_number_alive;
252 uint32_t rs_number_dead;
253 static uint32_t rs_floor_mss = 0;
254 static uint32_t wait_time_floor = 8000;	/* 8 ms */
255 static uint32_t rs_hw_floor_mss = 16;
256 static uint32_t num_of_waits_allowed = 1; /* How many time blocks are we willing to wait */
257 
258 static uint32_t mss_divisor = RL_DEFAULT_DIVISOR;
259 static uint32_t even_num_segs = 1;
260 static uint32_t even_threshold = 4;
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 SYSCTL_INT(_net_inet_tcp_rl, OID_AUTO, divisor, CTLFLAG_RW,
285     &mss_divisor, RL_DEFAULT_DIVISOR,
286     "The value divided into bytes per second to help establish mss size");
287 SYSCTL_INT(_net_inet_tcp_rl, OID_AUTO, even, CTLFLAG_RW,
288     &even_num_segs, 1,
289     "Do we round mss size up to an even number of segments for delayed ack");
290 SYSCTL_INT(_net_inet_tcp_rl, OID_AUTO, eventhresh, CTLFLAG_RW,
291     &even_threshold, 4,
292     "At what number of mss do we start rounding up to an even number of mss?");
293 
294 static void
295 rl_add_syctl_entries(struct sysctl_oid *rl_sysctl_root, struct tcp_rate_set *rs)
296 {
297 	/*
298 	 * Add sysctl entries for thus interface.
299 	 */
300 	if (rs->rs_flags & RS_INTF_NO_SUP) {
301 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
302 		   SYSCTL_CHILDREN(rl_sysctl_root),
303 		   OID_AUTO, "disable", CTLFLAG_RD,
304 		   &rs->rs_disable, 0,
305 		   "Disable this interface from new hdwr limiting?");
306 	} else {
307 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
308 		   SYSCTL_CHILDREN(rl_sysctl_root),
309 		   OID_AUTO, "disable", CTLFLAG_RW,
310 		   &rs->rs_disable, 0,
311 		   "Disable this interface from new hdwr limiting?");
312 	}
313 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
314 	    SYSCTL_CHILDREN(rl_sysctl_root),
315 	    OID_AUTO, "minseg", CTLFLAG_RW,
316 	    &rs->rs_min_seg, 0,
317 	    "What is the minimum we need to send on this interface?");
318 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
319 	    SYSCTL_CHILDREN(rl_sysctl_root),
320 	    OID_AUTO, "flow_limit", CTLFLAG_RW,
321 	    &rs->rs_flow_limit, 0,
322 	    "What is the limit for number of flows (0=unlimited)?");
323 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
324 	    SYSCTL_CHILDREN(rl_sysctl_root),
325 	    OID_AUTO, "highest", CTLFLAG_RD,
326 	    &rs->rs_highest_valid, 0,
327 	    "Highest valid rate");
328 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
329 	    SYSCTL_CHILDREN(rl_sysctl_root),
330 	    OID_AUTO, "lowest", CTLFLAG_RD,
331 	    &rs->rs_lowest_valid, 0,
332 	    "Lowest valid rate");
333 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
334 	    SYSCTL_CHILDREN(rl_sysctl_root),
335 	    OID_AUTO, "flags", CTLFLAG_RD,
336 	    &rs->rs_flags, 0,
337 	    "What lags are on the entry?");
338 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
339 	    SYSCTL_CHILDREN(rl_sysctl_root),
340 	    OID_AUTO, "numrates", CTLFLAG_RD,
341 	    &rs->rs_rate_cnt, 0,
342 	    "How many rates re there?");
343 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
344 	    SYSCTL_CHILDREN(rl_sysctl_root),
345 	    OID_AUTO, "flows_using", CTLFLAG_RD,
346 	    &rs->rs_flows_using, 0,
347 	    "How many flows are using this interface now?");
348 #ifdef DETAILED_RATELIMIT_SYSCTL
349 	if (rs->rs_rlt && rs->rs_rate_cnt > 0) {
350 		/*  Lets display the rates */
351 		int i;
352 		struct sysctl_oid *rl_rates;
353 		struct sysctl_oid *rl_rate_num;
354 		char rate_num[16];
355 		rl_rates = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
356 					    SYSCTL_CHILDREN(rl_sysctl_root),
357 					    OID_AUTO,
358 					    "rate",
359 					    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
360 					    "Ratelist");
361 		for( i = 0; i < rs->rs_rate_cnt; i++) {
362 			sprintf(rate_num, "%d", i);
363 			rl_rate_num = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
364 					    SYSCTL_CHILDREN(rl_rates),
365 					    OID_AUTO,
366 					    rate_num,
367 					    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
368 					    "Individual Rate");
369 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
370 				       SYSCTL_CHILDREN(rl_rate_num),
371 				       OID_AUTO, "flags", CTLFLAG_RD,
372 				       &rs->rs_rlt[i].flags, 0,
373 				       "Flags on this rate");
374 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
375 				       SYSCTL_CHILDREN(rl_rate_num),
376 				       OID_AUTO, "pacetime", CTLFLAG_RD,
377 				       &rs->rs_rlt[i].time_between, 0,
378 				       "Time hardware inserts between 1500 byte sends");
379 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
380 				       SYSCTL_CHILDREN(rl_rate_num),
381 				       OID_AUTO, "rate", CTLFLAG_RD,
382 				       &rs->rs_rlt[i].rate,
383 				       "Rate in bytes per second");
384 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
385 				       SYSCTL_CHILDREN(rl_rate_num),
386 				       OID_AUTO, "using", CTLFLAG_RD,
387 				       &rs->rs_rlt[i].using,
388 				       "Number of flows using");
389 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
390 				       SYSCTL_CHILDREN(rl_rate_num),
391 				       OID_AUTO, "enobufs", CTLFLAG_RD,
392 				       &rs->rs_rlt[i].rs_num_enobufs,
393 				       "Number of enobufs logged on this rate");
394 
395 		}
396 	}
397 #endif
398 }
399 
400 static void
401 rs_destroy(epoch_context_t ctx)
402 {
403 	struct tcp_rate_set *rs;
404 	bool do_free_rs;
405 
406 	rs = __containerof(ctx, struct tcp_rate_set, rs_epoch_ctx);
407 
408 	mtx_lock(&rs_mtx);
409 	rs->rs_flags &= ~RS_FUNERAL_SCHD;
410 	/*
411 	 * In theory its possible (but unlikely)
412 	 * that while the delete was occuring
413 	 * and we were applying the DEAD flag
414 	 * someone slipped in and found the
415 	 * interface in a lookup. While we
416 	 * decided rs_flows_using were 0 and
417 	 * scheduling the epoch_call, the other
418 	 * thread incremented rs_flow_using. This
419 	 * is because users have a pointer and
420 	 * we only use the rs_flows_using in an
421 	 * atomic fashion, i.e. the other entities
422 	 * are not protected. To assure this did
423 	 * not occur, we check rs_flows_using here
424 	 * before deleting.
425 	 */
426 	do_free_rs = (rs->rs_flows_using == 0);
427 	rs_number_dead--;
428 	mtx_unlock(&rs_mtx);
429 
430 	if (do_free_rs) {
431 		sysctl_ctx_free(&rs->sysctl_ctx);
432 		free(rs->rs_rlt, M_TCPPACE);
433 		free(rs, M_TCPPACE);
434 	}
435 }
436 
437 static void
438 rs_defer_destroy(struct tcp_rate_set *rs)
439 {
440 
441 	mtx_assert(&rs_mtx, MA_OWNED);
442 
443 	/* Check if already pending. */
444 	if (rs->rs_flags & RS_FUNERAL_SCHD)
445 		return;
446 
447 	rs_number_dead++;
448 
449 	/* Set flag to only defer once. */
450 	rs->rs_flags |= RS_FUNERAL_SCHD;
451 	NET_EPOCH_CALL(rs_destroy, &rs->rs_epoch_ctx);
452 }
453 
454 #ifdef INET
455 extern counter_u64_t rate_limit_new;
456 extern counter_u64_t rate_limit_chg;
457 extern counter_u64_t rate_limit_set_ok;
458 extern counter_u64_t rate_limit_active;
459 extern counter_u64_t rate_limit_alloc_fail;
460 #endif
461 
462 static int
463 rl_attach_txrtlmt(struct ifnet *ifp,
464     uint32_t flowtype,
465     int flowid,
466     uint64_t cfg_rate,
467     struct m_snd_tag **tag)
468 {
469 	int error;
470 	union if_snd_tag_alloc_params params = {
471 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
472 		.rate_limit.hdr.flowid = flowid,
473 		.rate_limit.hdr.flowtype = flowtype,
474 		.rate_limit.max_rate = cfg_rate,
475 		.rate_limit.flags = M_NOWAIT,
476 	};
477 
478 	error = m_snd_tag_alloc(ifp, &params, tag);
479 #ifdef INET
480 	if (error == 0) {
481 		counter_u64_add(rate_limit_set_ok, 1);
482 		counter_u64_add(rate_limit_active, 1);
483 	} else if (error != EOPNOTSUPP)
484 		counter_u64_add(rate_limit_alloc_fail, 1);
485 #endif
486 	return (error);
487 }
488 
489 static void
490 populate_canned_table(struct tcp_rate_set *rs, const uint64_t *rate_table_act)
491 {
492 	/*
493 	 * The internal table is "special", it
494 	 * is two seperate ordered tables that
495 	 * must be merged. We get here when the
496 	 * adapter specifies a number of rates that
497 	 * covers both ranges in the table in some
498 	 * form.
499 	 */
500 	int i, at_low, at_high;
501 	uint8_t low_disabled = 0, high_disabled = 0;
502 
503 	for(i = 0, at_low = 0, at_high = RS_NEXT_ORDER_GROUP; i < rs->rs_rate_cnt; i++) {
504 		rs->rs_rlt[i].flags = 0;
505 		rs->rs_rlt[i].time_between = 0;
506 		if ((low_disabled == 0) &&
507 		    (high_disabled ||
508 		     (rate_table_act[at_low] < rate_table_act[at_high]))) {
509 			rs->rs_rlt[i].rate = rate_table_act[at_low];
510 			at_low++;
511 			if (at_low == RS_NEXT_ORDER_GROUP)
512 				low_disabled = 1;
513 		} else if (high_disabled == 0) {
514 			rs->rs_rlt[i].rate = rate_table_act[at_high];
515 			at_high++;
516 			if (at_high == MAX_HDWR_RATES)
517 				high_disabled = 1;
518 		}
519 	}
520 }
521 
522 static struct tcp_rate_set *
523 rt_setup_new_rs(struct ifnet *ifp, int *error)
524 {
525 	struct tcp_rate_set *rs;
526 	const uint64_t *rate_table_act;
527 	uint64_t lentim, res;
528 	size_t sz;
529 	uint32_t hash_type;
530 	int i;
531 	struct if_ratelimit_query_results rl;
532 	struct sysctl_oid *rl_sysctl_root;
533 	struct epoch_tracker et;
534 	/*
535 	 * We expect to enter with the
536 	 * mutex locked.
537 	 */
538 
539 	if (ifp->if_ratelimit_query == NULL) {
540 		/*
541 		 * We can do nothing if we cannot
542 		 * get a query back from the driver.
543 		 */
544 		printf("Warning:No query functions for %s:%d-- failed\n",
545 		       ifp->if_dname, ifp->if_dunit);
546 		return (NULL);
547 	}
548 	rs = malloc(sizeof(struct tcp_rate_set), M_TCPPACE, M_NOWAIT | M_ZERO);
549 	if (rs == NULL) {
550 		if (error)
551 			*error = ENOMEM;
552 		printf("Warning:No memory for malloc of tcp_rate_set\n");
553 		return (NULL);
554 	}
555 	memset(&rl, 0, sizeof(rl));
556 	rl.flags = RT_NOSUPPORT;
557 	ifp->if_ratelimit_query(ifp, &rl);
558 	if (rl.flags & RT_IS_UNUSABLE) {
559 		/*
560 		 * The interface does not really support
561 		 * the rate-limiting.
562 		 */
563 		memset(rs, 0, sizeof(struct tcp_rate_set));
564 		rs->rs_ifp = ifp;
565 		rs->rs_if_dunit = ifp->if_dunit;
566 		rs->rs_flags = RS_INTF_NO_SUP;
567 		rs->rs_disable = 1;
568 		rs_number_alive++;
569 		sysctl_ctx_init(&rs->sysctl_ctx);
570 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
571 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
572 		    OID_AUTO,
573 		    rs->rs_ifp->if_xname,
574 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
575 		    "");
576 		rl_add_syctl_entries(rl_sysctl_root, rs);
577 		NET_EPOCH_ENTER(et);
578 		mtx_lock(&rs_mtx);
579 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
580 		mtx_unlock(&rs_mtx);
581 		NET_EPOCH_EXIT(et);
582 		return (rs);
583 	} else if ((rl.flags & RT_IS_INDIRECT) == RT_IS_INDIRECT) {
584 		memset(rs, 0, sizeof(struct tcp_rate_set));
585 		rs->rs_ifp = ifp;
586 		rs->rs_if_dunit = ifp->if_dunit;
587 		rs->rs_flags = RS_IS_DEFF;
588 		rs_number_alive++;
589 		sysctl_ctx_init(&rs->sysctl_ctx);
590 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
591 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
592 		    OID_AUTO,
593 		    rs->rs_ifp->if_xname,
594 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
595 		    "");
596 		rl_add_syctl_entries(rl_sysctl_root, rs);
597 		NET_EPOCH_ENTER(et);
598 		mtx_lock(&rs_mtx);
599 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
600 		mtx_unlock(&rs_mtx);
601 		NET_EPOCH_EXIT(et);
602 		return (rs);
603 	} else if ((rl.flags & RT_IS_FIXED_TABLE) == RT_IS_FIXED_TABLE) {
604 		/* Mellanox C4 likely */
605 		rs->rs_ifp = ifp;
606 		rs->rs_if_dunit = ifp->if_dunit;
607 		rs->rs_rate_cnt = rl.number_of_rates;
608 		rs->rs_min_seg = rl.min_segment_burst;
609 		rs->rs_highest_valid = 0;
610 		rs->rs_flow_limit = rl.max_flows;
611 		rs->rs_flags = RS_IS_INTF | RS_NO_PRE;
612 		rs->rs_disable = 0;
613 		rate_table_act = rl.rate_table;
614 	} else if ((rl.flags & RT_IS_SELECTABLE) == RT_IS_SELECTABLE) {
615 		/* Chelsio, C5 and C6 of Mellanox? */
616 		rs->rs_ifp = ifp;
617 		rs->rs_if_dunit = ifp->if_dunit;
618 		rs->rs_rate_cnt = rl.number_of_rates;
619 		rs->rs_min_seg = rl.min_segment_burst;
620 		rs->rs_disable = 0;
621 		rs->rs_flow_limit = rl.max_flows;
622 		rate_table_act = desired_rates;
623 		if ((rs->rs_rate_cnt > MAX_HDWR_RATES) &&
624 		    (rs->rs_rate_cnt < ALL_HARDWARE_RATES)) {
625 			/*
626 			 * Our desired table is not big
627 			 * enough, do what we can.
628 			 */
629 			rs->rs_rate_cnt = MAX_HDWR_RATES;
630 		 }
631 		if (rs->rs_rate_cnt <= RS_ORDERED_COUNT)
632 			rs->rs_flags = RS_IS_INTF;
633 		else
634 			rs->rs_flags = RS_IS_INTF | RS_INT_TBL;
635 		if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)
636 			rs->rs_rate_cnt = ALL_HARDWARE_RATES;
637 	} else {
638 		free(rs, M_TCPPACE);
639 		return (NULL);
640 	}
641 	sz = sizeof(struct tcp_hwrate_limit_table) * rs->rs_rate_cnt;
642 	rs->rs_rlt = malloc(sz, M_TCPPACE, M_NOWAIT);
643 	if (rs->rs_rlt == NULL) {
644 		if (error)
645 			*error = ENOMEM;
646 bail:
647 		free(rs, M_TCPPACE);
648 		return (NULL);
649 	}
650 	if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) {
651 		/*
652 		 * The interface supports all
653 		 * the rates we could possibly want.
654 		 */
655 		uint64_t rat;
656 
657 		rs->rs_rlt[0].rate = 12500;	/* 100k */
658 		rs->rs_rlt[1].rate = 25000;	/* 200k */
659 		rs->rs_rlt[2].rate = 62500;	/* 500k */
660 		/* Note 125000 == 1Megabit
661 		 * populate 1Meg - 1000meg.
662 		 */
663 		for(i = 3, rat = 125000; i< (ALL_HARDWARE_RATES-1); i++) {
664 			rs->rs_rlt[i].rate = rat;
665 			rat += 125000;
666 		}
667 		rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate = 1250000000;
668 	} else if (rs->rs_flags & RS_INT_TBL) {
669 		/* We populate this in a special way */
670 		populate_canned_table(rs, rate_table_act);
671 	} else {
672 		/*
673 		 * Just copy in the rates from
674 		 * the table, it is in order.
675 		 */
676 		for (i=0; i<rs->rs_rate_cnt; i++) {
677 			rs->rs_rlt[i].rate = rate_table_act[i];
678 			rs->rs_rlt[i].time_between = 0;
679 			rs->rs_rlt[i].flags = 0;
680 		}
681 	}
682 	for (i = (rs->rs_rate_cnt - 1); i >= 0; i--) {
683 		/*
684 		 * We go backwards through the list so that if we can't get
685 		 * a rate and fail to init one, we have at least a chance of
686 		 * getting the highest one.
687 		 */
688 		rs->rs_rlt[i].ptbl = rs;
689 		rs->rs_rlt[i].tag = NULL;
690 		rs->rs_rlt[i].using = 0;
691 		rs->rs_rlt[i].rs_num_enobufs = 0;
692 		/*
693 		 * Calculate the time between.
694 		 */
695 		lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
696 		res = lentim / rs->rs_rlt[i].rate;
697 		if (res > 0)
698 			rs->rs_rlt[i].time_between = res;
699 		else
700 			rs->rs_rlt[i].time_between = 1;
701 		if (rs->rs_flags & RS_NO_PRE) {
702 			rs->rs_rlt[i].flags = HDWRPACE_INITED;
703 			rs->rs_lowest_valid = i;
704 		} else {
705 			int err;
706 
707 			if ((rl.flags & RT_IS_SETUP_REQ)  &&
708 			    (ifp->if_ratelimit_query)) {
709 				err = ifp->if_ratelimit_setup(ifp,
710   				         rs->rs_rlt[i].rate, i);
711 				if (err)
712 					goto handle_err;
713 			}
714 #ifdef RSS
715 			hash_type = M_HASHTYPE_RSS_TCP_IPV4;
716 #else
717 			hash_type = M_HASHTYPE_OPAQUE_HASH;
718 #endif
719 			err = rl_attach_txrtlmt(ifp,
720 			    hash_type,
721 			    (i + 1),
722 			    rs->rs_rlt[i].rate,
723 			    &rs->rs_rlt[i].tag);
724 			if (err) {
725 handle_err:
726 				if (i == (rs->rs_rate_cnt - 1)) {
727 					/*
728 					 * Huh - first rate and we can't get
729 					 * it?
730 					 */
731 					free(rs->rs_rlt, M_TCPPACE);
732 					if (error)
733 						*error = err;
734 					goto bail;
735 				} else {
736 					if (error)
737 						*error = err;
738 				}
739 				break;
740 			} else {
741 				rs->rs_rlt[i].flags = HDWRPACE_INITED | HDWRPACE_TAGPRESENT;
742 				rs->rs_lowest_valid = i;
743 			}
744 		}
745 	}
746 	/* Did we get at least 1 rate? */
747 	if (rs->rs_rlt[(rs->rs_rate_cnt - 1)].flags & HDWRPACE_INITED)
748 		rs->rs_highest_valid = rs->rs_rate_cnt - 1;
749 	else {
750 		free(rs->rs_rlt, M_TCPPACE);
751 		goto bail;
752 	}
753 	rs_number_alive++;
754 	sysctl_ctx_init(&rs->sysctl_ctx);
755 	rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
756 	    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
757 	    OID_AUTO,
758 	    rs->rs_ifp->if_xname,
759 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
760 	    "");
761 	rl_add_syctl_entries(rl_sysctl_root, rs);
762 	NET_EPOCH_ENTER(et);
763 	mtx_lock(&rs_mtx);
764 	CK_LIST_INSERT_HEAD(&int_rs, rs, next);
765 	mtx_unlock(&rs_mtx);
766 	NET_EPOCH_EXIT(et);
767 	return (rs);
768 }
769 
770 /*
771  * For an explanation of why the argument is volatile please
772  * look at the comments around rt_setup_rate().
773  */
774 static const struct tcp_hwrate_limit_table *
775 tcp_int_find_suitable_rate(const volatile struct tcp_rate_set *rs,
776     uint64_t bytes_per_sec, uint32_t flags, uint64_t *lower_rate)
777 {
778 	struct tcp_hwrate_limit_table *arte = NULL, *rte = NULL;
779 	uint64_t mbits_per_sec, ind_calc, previous_rate = 0;
780 	int i;
781 
782 	mbits_per_sec = (bytes_per_sec * 8);
783 	if (flags & RS_PACING_LT) {
784 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
785 		    (rs->rs_lowest_valid <= 2)){
786 			/*
787 			 * Smaller than 1Meg, only
788 			 * 3 entries can match it.
789 			 */
790 			previous_rate = 0;
791 			for(i = rs->rs_lowest_valid; i < 3; i++) {
792 				if (bytes_per_sec <= rs->rs_rlt[i].rate) {
793 					rte = &rs->rs_rlt[i];
794 					break;
795 				} else if (rs->rs_rlt[i].flags & HDWRPACE_INITED) {
796 					arte = &rs->rs_rlt[i];
797 				}
798 				previous_rate = rs->rs_rlt[i].rate;
799 			}
800 			goto done;
801 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
802 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
803 			/*
804 			 * Larger than 1G (the majority of
805 			 * our table.
806 			 */
807 			if (mbits_per_sec < RS_TEN_GIGABIT_PERSEC)
808 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
809 			else
810 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
811 			previous_rate = rs->rs_rlt[(ALL_HARDWARE_RATES-2)].rate;
812 			goto done;
813 		}
814 		/*
815 		 * If we reach here its in our table (between 1Meg - 1000Meg),
816 		 * just take the rounded down mbits per second, and add
817 		 * 1Megabit to it, from this we can calculate
818 		 * the index in the table.
819 		 */
820 		ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
821 		if ((ind_calc * RS_ONE_MEGABIT_PERSEC) != mbits_per_sec)
822 			ind_calc++;
823 		/* our table is offset by 3, we add 2 */
824 		ind_calc += 2;
825 		if (ind_calc > (ALL_HARDWARE_RATES-1)) {
826 			/* This should not happen */
827 			ind_calc = ALL_HARDWARE_RATES-1;
828 		}
829 		if ((ind_calc >= rs->rs_lowest_valid) &&
830 		    (ind_calc <= rs->rs_highest_valid)) {
831 			rte = &rs->rs_rlt[ind_calc];
832 			if (ind_calc >= 1)
833 				previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
834 		}
835 	} else if (flags & RS_PACING_EXACT_MATCH) {
836 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
837 		    (rs->rs_lowest_valid <= 2)){
838 			for(i = rs->rs_lowest_valid; i < 3; i++) {
839 				if (bytes_per_sec == rs->rs_rlt[i].rate) {
840 					rte = &rs->rs_rlt[i];
841 					break;
842 				}
843 			}
844 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
845 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
846 			/* > 1Gbps only one rate */
847 			if (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) {
848 				/* Its 10G wow */
849 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
850 			}
851 		} else {
852 			/* Ok it must be a exact meg (its between 1G and 1Meg) */
853 			ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
854 			if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
855 				/* its an exact Mbps */
856 				ind_calc += 2;
857 				if (ind_calc > (ALL_HARDWARE_RATES-1)) {
858 					/* This should not happen */
859 					ind_calc = ALL_HARDWARE_RATES-1;
860 				}
861 				if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED)
862 					rte = &rs->rs_rlt[ind_calc];
863 			}
864 		}
865 	} else {
866 		/* we want greater than the requested rate */
867 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
868 		    (rs->rs_lowest_valid <= 2)){
869 			arte = &rs->rs_rlt[3]; /* set alternate to 1Meg */
870 			for (i=2; i>=rs->rs_lowest_valid; i--) {
871 				if (bytes_per_sec < rs->rs_rlt[i].rate) {
872 					rte = &rs->rs_rlt[i];
873 					if (i >= 1) {
874 						previous_rate = rs->rs_rlt[(i-1)].rate;
875 					}
876 					break;
877 				} else if ((flags & RS_PACING_GEQ) &&
878 					   (bytes_per_sec == rs->rs_rlt[i].rate)) {
879 					rte = &rs->rs_rlt[i];
880 					if (i >= 1) {
881 						previous_rate = rs->rs_rlt[(i-1)].rate;
882 					}
883 					break;
884 				} else {
885 					arte = &rs->rs_rlt[i]; /* new alternate */
886 				}
887 			}
888 		} else if (mbits_per_sec > RS_ONE_GIGABIT_PERSEC) {
889 			if ((bytes_per_sec < rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
890 			    (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
891 				/* Our top rate is larger than the request */
892 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
893 			} else if ((flags & RS_PACING_GEQ) &&
894 				   (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
895 				   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
896 				/* It matches our top rate */
897 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
898 			} else if (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED) {
899 				/* The top rate is an alternative */
900 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
901 			}
902 			previous_rate = rs->rs_rlt[(ALL_HARDWARE_RATES-2)].rate;
903 		} else {
904 			/* Its in our range 1Meg - 1Gig */
905 			if (flags & RS_PACING_GEQ) {
906 				ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
907 				if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
908 					if (ind_calc > (ALL_HARDWARE_RATES-1)) {
909 						/* This should not happen */
910 						ind_calc = (ALL_HARDWARE_RATES-1);
911 					}
912 					rte = &rs->rs_rlt[ind_calc];
913 					if (ind_calc >= 1)
914 						previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
915 				}
916 				goto done;
917 			}
918 			ind_calc = (mbits_per_sec + (RS_ONE_MEGABIT_PERSEC-1))/RS_ONE_MEGABIT_PERSEC;
919 			ind_calc += 2;
920 			if (ind_calc > (ALL_HARDWARE_RATES-1)) {
921 				/* This should not happen */
922 				ind_calc = ALL_HARDWARE_RATES-1;
923 			}
924 			if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) {
925 				rte = &rs->rs_rlt[ind_calc];
926 				if (ind_calc >= 1)
927 					previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
928 			}
929 		}
930 	}
931 done:
932 	if ((rte == NULL) &&
933 	    (arte != NULL) &&
934 	    (flags & RS_PACING_SUB_OK)) {
935 		/* We can use the substitute */
936 		rte = arte;
937 	}
938 	if (lower_rate)
939 		*lower_rate = previous_rate;
940 	return (rte);
941 }
942 
943 /*
944  * For an explanation of why the argument is volatile please
945  * look at the comments around rt_setup_rate().
946  */
947 static const struct tcp_hwrate_limit_table *
948 tcp_find_suitable_rate(const volatile struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags, uint64_t *lower_rate)
949 {
950 	/**
951 	 * Hunt the rate table with the restrictions in flags and find a
952 	 * suitable rate if possible.
953 	 * RS_PACING_EXACT_MATCH - look for an exact match to rate.
954 	 * RS_PACING_GT     - must be greater than.
955 	 * RS_PACING_GEQ    - must be greater than or equal.
956 	 * RS_PACING_LT     - must be less than.
957 	 * RS_PACING_SUB_OK - If we don't meet criteria a
958 	 *                    substitute is ok.
959 	 */
960 	int i, matched;
961 	struct tcp_hwrate_limit_table *rte = NULL;
962 	uint64_t previous_rate = 0;
963 
964 	if ((rs->rs_flags & RS_INT_TBL) &&
965 	    (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)) {
966 		/*
967 		 * Here we don't want to paw thru
968 		 * a big table, we have everything
969 		 * from 1Meg - 1000Meg in 1Meg increments.
970 		 * Use an alternate method to "lookup".
971 		 */
972 		return (tcp_int_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate));
973 	}
974 	if ((flags & RS_PACING_LT) ||
975 	    (flags & RS_PACING_EXACT_MATCH)) {
976 		/*
977 		 * For exact and less than we go forward through the table.
978 		 * This way when we find one larger we stop (exact was a
979 		 * toss up).
980 		 */
981 		for (i = rs->rs_lowest_valid, matched = 0; i <= rs->rs_highest_valid; i++) {
982 			if ((flags & RS_PACING_EXACT_MATCH) &&
983 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
984 				rte = &rs->rs_rlt[i];
985 				matched = 1;
986 				if (lower_rate != NULL)
987 					*lower_rate = previous_rate;
988 				break;
989 			} else if ((flags & RS_PACING_LT) &&
990 			    (bytes_per_sec <= rs->rs_rlt[i].rate)) {
991 				rte = &rs->rs_rlt[i];
992 				matched = 1;
993 				if (lower_rate != NULL)
994 					*lower_rate = previous_rate;
995 				break;
996 			}
997 			previous_rate = rs->rs_rlt[i].rate;
998 			if (bytes_per_sec > rs->rs_rlt[i].rate)
999 				break;
1000 		}
1001 		if ((matched == 0) &&
1002 		    (flags & RS_PACING_LT) &&
1003 		    (flags & RS_PACING_SUB_OK)) {
1004 			/* Kick in a substitute (the lowest) */
1005 			rte = &rs->rs_rlt[rs->rs_lowest_valid];
1006 		}
1007 	} else {
1008 		/*
1009 		 * Here we go backward through the table so that we can find
1010 		 * the one greater in theory faster (but its probably a
1011 		 * wash).
1012 		 */
1013 		for (i = rs->rs_highest_valid, matched = 0; i >= rs->rs_lowest_valid; i--) {
1014 			if (rs->rs_rlt[i].rate > bytes_per_sec) {
1015 				/* A possible candidate */
1016 				rte = &rs->rs_rlt[i];
1017 			}
1018 			if ((flags & RS_PACING_GEQ) &&
1019 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
1020 				/* An exact match and we want equal */
1021 				matched = 1;
1022 				rte = &rs->rs_rlt[i];
1023 				break;
1024 			} else if (rte) {
1025 				/*
1026 				 * Found one that is larger than but don't
1027 				 * stop, there may be a more closer match.
1028 				 */
1029 				matched = 1;
1030 			}
1031 			if (rs->rs_rlt[i].rate < bytes_per_sec) {
1032 				/*
1033 				 * We found a table entry that is smaller,
1034 				 * stop there will be none greater or equal.
1035 				 */
1036 				if (lower_rate != NULL)
1037 					*lower_rate = rs->rs_rlt[i].rate;
1038 				break;
1039 			}
1040 		}
1041 		if ((matched == 0) &&
1042 		    (flags & RS_PACING_SUB_OK)) {
1043 			/* Kick in a substitute (the highest) */
1044 			rte = &rs->rs_rlt[rs->rs_highest_valid];
1045 		}
1046 	}
1047 	return (rte);
1048 }
1049 
1050 static struct ifnet *
1051 rt_find_real_interface(struct ifnet *ifp, struct inpcb *inp, int *error)
1052 {
1053 	struct ifnet *tifp;
1054 	struct m_snd_tag *tag, *ntag;
1055 	union if_snd_tag_alloc_params params = {
1056 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
1057 		.rate_limit.hdr.flowid = inp->inp_flowid,
1058 		.rate_limit.hdr.numa_domain = inp->inp_numa_domain,
1059 		.rate_limit.max_rate = COMMON_RATE,
1060 		.rate_limit.flags = M_NOWAIT,
1061 	};
1062 	int err;
1063 #ifdef RSS
1064 	params.rate_limit.hdr.flowtype = ((inp->inp_vflag & INP_IPV6) ?
1065 	    M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_TCP_IPV4);
1066 #else
1067 	params.rate_limit.hdr.flowtype = M_HASHTYPE_OPAQUE_HASH;
1068 #endif
1069 	err = m_snd_tag_alloc(ifp, &params, &tag);
1070 	if (err) {
1071 		/* Failed to setup a tag? */
1072 		if (error)
1073 			*error = err;
1074 		return (NULL);
1075 	}
1076 	ntag = tag;
1077 	while (ntag->sw->next_snd_tag != NULL) {
1078 		ntag = ntag->sw->next_snd_tag(ntag);
1079 	}
1080 	tifp = ntag->ifp;
1081 	m_snd_tag_rele(tag);
1082 	return (tifp);
1083 }
1084 
1085 static void
1086 rl_increment_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_add_long(&decon_rte->using, 1);
1092 }
1093 
1094 static void
1095 rl_decrement_using(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_subtract_long(&decon_rte->using, 1);
1101 }
1102 
1103 void
1104 tcp_rl_log_enobuf(const struct tcp_hwrate_limit_table *rte)
1105 {
1106 	struct tcp_hwrate_limit_table *decon_rte;
1107 
1108 	decon_rte = __DECONST(struct tcp_hwrate_limit_table *, rte);
1109 	atomic_add_long(&decon_rte->rs_num_enobufs, 1);
1110 }
1111 
1112 /*
1113  * Do NOT take the __noinline out of the
1114  * find_rs_for_ifp() function. If you do the inline
1115  * of it for the rt_setup_rate() will show you a
1116  * compiler bug. For some reason the compiler thinks
1117  * the list can never be empty. The consequence of
1118  * this will be a crash when we dereference NULL
1119  * if an ifp is removed just has a hw rate limit
1120  * is attempted. If you are working on the compiler
1121  * and want to "test" this go ahead and take the noinline
1122  * out otherwise let sleeping dogs ly until such time
1123  * as we get a compiler fix 10/2/20 -- RRS
1124  */
1125 static __noinline struct tcp_rate_set *
1126 find_rs_for_ifp(struct ifnet *ifp)
1127 {
1128 	struct tcp_rate_set *rs;
1129 
1130 	CK_LIST_FOREACH(rs, &int_rs, next) {
1131 		if ((rs->rs_ifp == ifp) &&
1132 		    (rs->rs_if_dunit == ifp->if_dunit)) {
1133 			/* Ok we found it */
1134 			return (rs);
1135 		}
1136 	}
1137 	return (NULL);
1138 }
1139 
1140 
1141 static const struct tcp_hwrate_limit_table *
1142 rt_setup_rate(struct inpcb *inp, struct ifnet *ifp, uint64_t bytes_per_sec,
1143     uint32_t flags, int *error, uint64_t *lower_rate)
1144 {
1145 	/* First lets find the interface if it exists */
1146 	const struct tcp_hwrate_limit_table *rte;
1147 	/*
1148 	 * So why is rs volatile? This is to defeat a
1149 	 * compiler bug where in the compiler is convinced
1150 	 * that rs can never be NULL (which is not true). Because
1151 	 * of its conviction it nicely optimizes out the if ((rs == NULL
1152 	 * below which means if you get a NULL back you dereference it.
1153 	 */
1154 	volatile struct tcp_rate_set *rs;
1155 	struct epoch_tracker et;
1156 	struct ifnet *oifp = ifp;
1157 	int err;
1158 
1159 	NET_EPOCH_ENTER(et);
1160 use_real_interface:
1161 	rs = find_rs_for_ifp(ifp);
1162 	if ((rs == NULL) ||
1163 	    (rs->rs_flags & RS_INTF_NO_SUP) ||
1164 	    (rs->rs_flags & RS_IS_DEAD)) {
1165 		/*
1166 		 * This means we got a packet *before*
1167 		 * the IF-UP was processed below, <or>
1168 		 * while or after we already received an interface
1169 		 * departed event. In either case we really don't
1170 		 * want to do anything with pacing, in
1171 		 * the departing case the packet is not
1172 		 * going to go very far. The new case
1173 		 * might be arguable, but its impossible
1174 		 * to tell from the departing case.
1175 		 */
1176 		if (error)
1177 			*error = ENODEV;
1178 		NET_EPOCH_EXIT(et);
1179 		return (NULL);
1180 	}
1181 
1182 	if ((rs == NULL) || (rs->rs_disable != 0)) {
1183 		if (error)
1184 			*error = ENOSPC;
1185 		NET_EPOCH_EXIT(et);
1186 		return (NULL);
1187 	}
1188 	if (rs->rs_flags & RS_IS_DEFF) {
1189 		/* We need to find the real interface */
1190 		struct ifnet *tifp;
1191 
1192 		tifp = rt_find_real_interface(ifp, inp, error);
1193 		if (tifp == NULL) {
1194 			if (rs->rs_disable && error)
1195 				*error = ENOTSUP;
1196 			NET_EPOCH_EXIT(et);
1197 			return (NULL);
1198 		}
1199 		KASSERT((tifp != ifp),
1200 			("Lookup failure ifp:%p inp:%p rt_find_real_interface() returns the same interface tifp:%p?\n",
1201 			 ifp, inp, tifp));
1202 		ifp = tifp;
1203 		goto use_real_interface;
1204 	}
1205 	if (rs->rs_flow_limit &&
1206 	    ((rs->rs_flows_using + 1) > rs->rs_flow_limit)) {
1207 		if (error)
1208 			*error = ENOSPC;
1209 		NET_EPOCH_EXIT(et);
1210 		return (NULL);
1211 	}
1212 	rte = tcp_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate);
1213 	if (rte) {
1214 		err = in_pcbattach_txrtlmt(inp, oifp,
1215 		    inp->inp_flowtype,
1216 		    inp->inp_flowid,
1217 		    rte->rate,
1218 		    &inp->inp_snd_tag);
1219 		if (err) {
1220 			/* Failed to attach */
1221 			if (error)
1222 				*error = err;
1223 			rte = NULL;
1224 		} else {
1225 			KASSERT((inp->inp_snd_tag != NULL) ,
1226 				("Setup rate has no snd_tag inp:%p rte:%p rate:%llu rs:%p",
1227 				 inp, rte, (unsigned long long)rte->rate, rs));
1228 #ifdef INET
1229 			counter_u64_add(rate_limit_new, 1);
1230 #endif
1231 		}
1232 	}
1233 	if (rte) {
1234 		/*
1235 		 * We use an atomic here for accounting so we don't have to
1236 		 * use locks when freeing.
1237 		 */
1238 		atomic_add_64(&rs->rs_flows_using, 1);
1239 	}
1240 	NET_EPOCH_EXIT(et);
1241 	return (rte);
1242 }
1243 
1244 static void
1245 tcp_rl_ifnet_link(void *arg __unused, struct ifnet *ifp, int link_state)
1246 {
1247 	int error;
1248 	struct tcp_rate_set *rs;
1249 	struct epoch_tracker et;
1250 
1251 	if (((ifp->if_capenable & IFCAP_TXRTLMT) == 0) ||
1252 	    (link_state != LINK_STATE_UP)) {
1253 		/*
1254 		 * We only care on an interface going up that is rate-limit
1255 		 * capable.
1256 		 */
1257 		return;
1258 	}
1259 	NET_EPOCH_ENTER(et);
1260 	mtx_lock(&rs_mtx);
1261 	rs = find_rs_for_ifp(ifp);
1262 	if (rs) {
1263 		/* We already have initialized this guy */
1264 		mtx_unlock(&rs_mtx);
1265 		NET_EPOCH_EXIT(et);
1266 		return;
1267 	}
1268 	mtx_unlock(&rs_mtx);
1269 	NET_EPOCH_EXIT(et);
1270 	rt_setup_new_rs(ifp, &error);
1271 }
1272 
1273 static void
1274 tcp_rl_ifnet_departure(void *arg __unused, struct ifnet *ifp)
1275 {
1276 	struct tcp_rate_set *rs;
1277 	struct epoch_tracker et;
1278 	int i;
1279 
1280 	NET_EPOCH_ENTER(et);
1281 	mtx_lock(&rs_mtx);
1282 	rs = find_rs_for_ifp(ifp);
1283 	if (rs) {
1284 		CK_LIST_REMOVE(rs, next);
1285 		rs_number_alive--;
1286 		rs->rs_flags |= RS_IS_DEAD;
1287 		for (i = 0; i < rs->rs_rate_cnt; i++) {
1288 			if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1289 				in_pcbdetach_tag(rs->rs_rlt[i].tag);
1290 				rs->rs_rlt[i].tag = NULL;
1291 			}
1292 			rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1293 		}
1294 		if (rs->rs_flows_using == 0)
1295 			rs_defer_destroy(rs);
1296 	}
1297 	mtx_unlock(&rs_mtx);
1298 	NET_EPOCH_EXIT(et);
1299 }
1300 
1301 static void
1302 tcp_rl_shutdown(void *arg __unused, int howto __unused)
1303 {
1304 	struct tcp_rate_set *rs, *nrs;
1305 	struct epoch_tracker et;
1306 	int i;
1307 
1308 	NET_EPOCH_ENTER(et);
1309 	mtx_lock(&rs_mtx);
1310 	CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) {
1311 		CK_LIST_REMOVE(rs, next);
1312 		rs_number_alive--;
1313 		rs->rs_flags |= RS_IS_DEAD;
1314 		for (i = 0; i < rs->rs_rate_cnt; i++) {
1315 			if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1316 				in_pcbdetach_tag(rs->rs_rlt[i].tag);
1317 				rs->rs_rlt[i].tag = NULL;
1318 			}
1319 			rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1320 		}
1321 		if (rs->rs_flows_using == 0)
1322 			rs_defer_destroy(rs);
1323 	}
1324 	mtx_unlock(&rs_mtx);
1325 	NET_EPOCH_EXIT(et);
1326 }
1327 
1328 const struct tcp_hwrate_limit_table *
1329 tcp_set_pacing_rate(struct tcpcb *tp, struct ifnet *ifp,
1330     uint64_t bytes_per_sec, int flags, int *error, uint64_t *lower_rate)
1331 {
1332 	struct inpcb *inp = tptoinpcb(tp);
1333 	const struct tcp_hwrate_limit_table *rte;
1334 #ifdef KERN_TLS
1335 	struct ktls_session *tls;
1336 #endif
1337 
1338 	INP_WLOCK_ASSERT(inp);
1339 
1340 	if (inp->inp_snd_tag == NULL) {
1341 		/*
1342 		 * We are setting up a rate for the first time.
1343 		 */
1344 		if ((ifp->if_capenable & IFCAP_TXRTLMT) == 0) {
1345 			/* Not supported by the egress */
1346 			if (error)
1347 				*error = ENODEV;
1348 			return (NULL);
1349 		}
1350 #ifdef KERN_TLS
1351 		tls = NULL;
1352 		if (tptosocket(tp)->so_snd.sb_flags & SB_TLS_IFNET) {
1353 			tls = tptosocket(tp)->so_snd.sb_tls_info;
1354 
1355 			if ((ifp->if_capenable & IFCAP_TXTLS_RTLMT) == 0 ||
1356 			    tls->mode != TCP_TLS_MODE_IFNET) {
1357 				if (error)
1358 					*error = ENODEV;
1359 				return (NULL);
1360 			}
1361 		}
1362 #endif
1363 		rte = rt_setup_rate(inp, ifp, bytes_per_sec, flags, error, lower_rate);
1364 		if (rte)
1365 			rl_increment_using(rte);
1366 #ifdef KERN_TLS
1367 		if (rte != NULL && tls != NULL && tls->snd_tag != NULL) {
1368 			/*
1369 			 * Fake a route change error to reset the TLS
1370 			 * send tag.  This will convert the existing
1371 			 * tag to a TLS ratelimit tag.
1372 			 */
1373 			MPASS(tls->snd_tag->sw->type == IF_SND_TAG_TYPE_TLS);
1374 			ktls_output_eagain(inp, tls);
1375 		}
1376 #endif
1377 	} else {
1378 		/*
1379 		 * We are modifying a rate, wrong interface?
1380 		 */
1381 		if (error)
1382 			*error = EINVAL;
1383 		rte = NULL;
1384 	}
1385 	if (rte != NULL) {
1386 		tp->t_pacing_rate = rte->rate;
1387 		*error = 0;
1388 	}
1389 	return (rte);
1390 }
1391 
1392 const struct tcp_hwrate_limit_table *
1393 tcp_chg_pacing_rate(const struct tcp_hwrate_limit_table *crte,
1394     struct tcpcb *tp, struct ifnet *ifp,
1395     uint64_t bytes_per_sec, int flags, int *error, uint64_t *lower_rate)
1396 {
1397 	struct inpcb *inp = tptoinpcb(tp);
1398 	const struct tcp_hwrate_limit_table *nrte;
1399 	const struct tcp_rate_set *rs;
1400 #ifdef KERN_TLS
1401 	struct ktls_session *tls = NULL;
1402 #endif
1403 	int err;
1404 
1405 	INP_WLOCK_ASSERT(inp);
1406 
1407 	if (crte == NULL) {
1408 		/* Wrong interface */
1409 		if (error)
1410 			*error = EINVAL;
1411 		return (NULL);
1412 	}
1413 
1414 #ifdef KERN_TLS
1415 	if (tptosocket(tp)->so_snd.sb_flags & SB_TLS_IFNET) {
1416 		tls = tptosocket(tp)->so_snd.sb_tls_info;
1417 		if (tls->mode != TCP_TLS_MODE_IFNET)
1418 			tls = NULL;
1419 		else if (tls->snd_tag != NULL &&
1420 		    tls->snd_tag->sw->type != IF_SND_TAG_TYPE_TLS_RATE_LIMIT) {
1421 			if (!tls->reset_pending) {
1422 				/*
1423 				 * NIC probably doesn't support
1424 				 * ratelimit TLS tags if it didn't
1425 				 * allocate one when an existing rate
1426 				 * was present, so ignore.
1427 				 */
1428 				tcp_rel_pacing_rate(crte, tp);
1429 				if (error)
1430 					*error = EOPNOTSUPP;
1431 				return (NULL);
1432 			}
1433 
1434 			/*
1435 			 * The send tag is being converted, so set the
1436 			 * rate limit on the inpcb tag.  There is a
1437 			 * race that the new NIC send tag might use
1438 			 * the current rate instead of this one.
1439 			 */
1440 			tls = NULL;
1441 		}
1442 	}
1443 #endif
1444 	if (inp->inp_snd_tag == NULL) {
1445 		/* Wrong interface */
1446 		tcp_rel_pacing_rate(crte, tp);
1447 		if (error)
1448 			*error = EINVAL;
1449 		return (NULL);
1450 	}
1451 	rs = crte->ptbl;
1452 	if ((rs->rs_flags & RS_IS_DEAD) ||
1453 	    (crte->flags & HDWRPACE_IFPDEPARTED)) {
1454 		/* Release the rate, and try anew */
1455 
1456 		tcp_rel_pacing_rate(crte, tp);
1457 		nrte = tcp_set_pacing_rate(tp, ifp,
1458 		    bytes_per_sec, flags, error, lower_rate);
1459 		return (nrte);
1460 	}
1461 	nrte = tcp_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate);
1462 	if (nrte == crte) {
1463 		/* No change */
1464 		if (error)
1465 			*error = 0;
1466 		return (crte);
1467 	}
1468 	if (nrte == NULL) {
1469 		/* Release the old rate */
1470 		if (error)
1471 			*error = ENOENT;
1472 		tcp_rel_pacing_rate(crte, tp);
1473 		return (NULL);
1474 	}
1475 	rl_decrement_using(crte);
1476 	rl_increment_using(nrte);
1477 	/* Change rates to our new entry */
1478 #ifdef KERN_TLS
1479 	if (tls != NULL)
1480 		err = ktls_modify_txrtlmt(tls, nrte->rate);
1481 	else
1482 #endif
1483 		err = in_pcbmodify_txrtlmt(inp, nrte->rate);
1484 	if (err) {
1485 		struct tcp_rate_set *lrs;
1486 		uint64_t pre;
1487 
1488 		rl_decrement_using(nrte);
1489 		lrs = __DECONST(struct tcp_rate_set *, rs);
1490 		pre = atomic_fetchadd_64(&lrs->rs_flows_using, -1);
1491 		/* Do we still have a snd-tag attached? */
1492 		if (inp->inp_snd_tag)
1493 			in_pcbdetach_txrtlmt(inp);
1494 
1495 		if (pre == 1) {
1496 			struct epoch_tracker et;
1497 
1498 			NET_EPOCH_ENTER(et);
1499 			mtx_lock(&rs_mtx);
1500 			/*
1501 			 * Is it dead?
1502 			 */
1503 			if (lrs->rs_flags & RS_IS_DEAD)
1504 				rs_defer_destroy(lrs);
1505 			mtx_unlock(&rs_mtx);
1506 			NET_EPOCH_EXIT(et);
1507 		}
1508 		if (error)
1509 			*error = err;
1510 		return (NULL);
1511 	} else {
1512 #ifdef INET
1513 		counter_u64_add(rate_limit_chg, 1);
1514 #endif
1515 	}
1516 	if (error)
1517 		*error = 0;
1518 	tp->t_pacing_rate = nrte->rate;
1519 	return (nrte);
1520 }
1521 
1522 void
1523 tcp_rel_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp)
1524 {
1525 	struct inpcb *inp = tptoinpcb(tp);
1526 	const struct tcp_rate_set *crs;
1527 	struct tcp_rate_set *rs;
1528 	uint64_t pre;
1529 
1530 	INP_WLOCK_ASSERT(inp);
1531 
1532 	tp->t_pacing_rate = -1;
1533 	crs = crte->ptbl;
1534 	/*
1535 	 * Now we must break the const
1536 	 * in order to release our refcount.
1537 	 */
1538 	rs = __DECONST(struct tcp_rate_set *, crs);
1539 	rl_decrement_using(crte);
1540 	pre = atomic_fetchadd_64(&rs->rs_flows_using, -1);
1541 	if (pre == 1) {
1542 		struct epoch_tracker et;
1543 
1544 		NET_EPOCH_ENTER(et);
1545 		mtx_lock(&rs_mtx);
1546 		/*
1547 		 * Is it dead?
1548 		 */
1549 		if (rs->rs_flags & RS_IS_DEAD)
1550 			rs_defer_destroy(rs);
1551 		mtx_unlock(&rs_mtx);
1552 		NET_EPOCH_EXIT(et);
1553 	}
1554 
1555 	/*
1556 	 * XXX: If this connection is using ifnet TLS, should we
1557 	 * switch it to using an unlimited rate, or perhaps use
1558 	 * ktls_output_eagain() to reset the send tag to a plain
1559 	 * TLS tag?
1560 	 */
1561 	in_pcbdetach_txrtlmt(inp);
1562 }
1563 
1564 #define ONE_POINT_TWO_MEG 150000 /* 1.2 megabits in bytes */
1565 #define ONE_HUNDRED_MBPS 12500000	/* 100Mbps in bytes per second */
1566 #define FIVE_HUNDRED_MBPS 62500000	/* 500Mbps in bytes per second */
1567 #define MAX_MSS_SENT 43	/* 43 mss = 43 x 1500 = 64,500 bytes */
1568 
1569 static void
1570 tcp_log_pacing_size(struct tcpcb *tp, uint64_t bw, uint32_t segsiz, uint32_t new_tso,
1571 		    uint64_t hw_rate, uint32_t time_between, uint32_t calc_time_between,
1572 		    uint32_t segs, uint32_t res_div, uint16_t mult, uint8_t mod)
1573 {
1574 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1575 		union tcp_log_stackspecific log;
1576 		struct timeval tv;
1577 
1578 		memset(&log, 0, sizeof(log));
1579 		log.u_bbr.flex1 = segsiz;
1580 		log.u_bbr.flex2 = new_tso;
1581 		log.u_bbr.flex3 = time_between;
1582 		log.u_bbr.flex4 = calc_time_between;
1583 		log.u_bbr.flex5 = segs;
1584 		log.u_bbr.flex6 = res_div;
1585 		log.u_bbr.flex7 = mult;
1586 		log.u_bbr.flex8 = mod;
1587 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1588 		log.u_bbr.cur_del_rate = bw;
1589 		log.u_bbr.delRate = hw_rate;
1590 		TCP_LOG_EVENTP(tp, NULL,
1591 		    &tptosocket(tp)->so_rcv,
1592 		    &tptosocket(tp)->so_snd,
1593 		    TCP_HDWR_PACE_SIZE, 0,
1594 		    0, &log, false, &tv);
1595 	}
1596 }
1597 
1598 uint32_t
1599 tcp_get_pacing_burst_size_w_divisor(struct tcpcb *tp, uint64_t bw, uint32_t segsiz, int can_use_1mss,
1600    const struct tcp_hwrate_limit_table *te, int *err, int divisor)
1601 {
1602 	/*
1603 	 * We use the google formula to calculate the
1604 	 * TSO size. I.E.
1605 	 * bw < 24Meg
1606 	 *   tso = 2mss
1607 	 * else
1608 	 *   tso = min(bw/(div=1000), 64k)
1609 	 *
1610 	 * Note for these calculations we ignore the
1611 	 * packet overhead (enet hdr, ip hdr and tcp hdr).
1612 	 * We only get the google formula when we have
1613 	 * divisor = 1000, which is the default for now.
1614 	 */
1615 	uint64_t lentim, res, bytes;
1616 	uint32_t new_tso, min_tso_segs;
1617 
1618 	/* It can't be zero */
1619 	if ((divisor == 0) ||
1620 	    (divisor < RL_MIN_DIVISOR)) {
1621 		if (mss_divisor)
1622 			bytes = bw / mss_divisor;
1623 		else
1624 			bytes = bw / 1000;
1625 	} else
1626 		bytes = bw / divisor;
1627 	/* We can't ever send more than 65k in a TSO */
1628 	if (bytes > 0xffff) {
1629 		bytes = 0xffff;
1630 	}
1631 	/* Round up */
1632 	new_tso = (bytes + segsiz - 1) / segsiz;
1633 	/* Are we enforcing even boundaries? */
1634 	if (even_num_segs && (new_tso & 1) && (new_tso > even_threshold))
1635 		new_tso++;
1636 	if (can_use_1mss)
1637 		min_tso_segs = 1;
1638 	else
1639 		min_tso_segs = 2;
1640 	if (rs_floor_mss && (new_tso < rs_floor_mss))
1641 		new_tso = rs_floor_mss;
1642 	else if (new_tso < min_tso_segs)
1643 		new_tso = min_tso_segs;
1644 	if (new_tso > MAX_MSS_SENT)
1645 		new_tso = MAX_MSS_SENT;
1646 	new_tso *= segsiz;
1647  	tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1648 			    0, 0, 0, 0, 0, 0, 1);
1649 	/*
1650 	 * If we are not doing hardware pacing
1651 	 * then we are done.
1652 	 */
1653 	if (te == NULL) {
1654 		if (err)
1655 			*err = 0;
1656 		return(new_tso);
1657 	}
1658 	/*
1659 	 * For hardware pacing we look at the
1660 	 * rate you are sending at and compare
1661 	 * that to the rate you have in hardware.
1662 	 *
1663 	 * If the hardware rate is slower than your
1664 	 * software rate then you are in error and
1665 	 * we will build a queue in our hardware whic
1666 	 * is probably not desired, in such a case
1667 	 * just return the non-hardware TSO size.
1668 	 *
1669 	 * If the rate in hardware is faster (which
1670 	 * it should be) then look at how long it
1671 	 * takes to send one ethernet segment size at
1672 	 * your b/w and compare that to the time it
1673 	 * takes to send at the rate you had selected.
1674 	 *
1675 	 * If your time is greater (which we hope it is)
1676 	 * we get the delta between the two, and then
1677 	 * divide that into your pacing time. This tells
1678 	 * us how many MSS you can send down at once (rounded up).
1679 	 *
1680 	 * Note we also double this value if the b/w is over
1681 	 * 100Mbps. If its over 500meg we just set you to the
1682 	 * max (43 segments).
1683 	 */
1684 	if (te->rate > FIVE_HUNDRED_MBPS)
1685 		goto max;
1686 	if (te->rate == bw) {
1687 		/* We are pacing at exactly the hdwr rate */
1688 max:
1689 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1690 				    te->rate, te->time_between, (uint32_t)0,
1691 				    (segsiz * MAX_MSS_SENT), 0, 0, 3);
1692 		return (segsiz * MAX_MSS_SENT);
1693 	}
1694 	lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
1695 	res = lentim / bw;
1696 	if (res > te->time_between) {
1697 		uint32_t delta, segs, res_div;
1698 
1699 		res_div = ((res * num_of_waits_allowed) + wait_time_floor);
1700 		delta = res - te->time_between;
1701 		segs = (res_div + delta - 1)/delta;
1702 		if (segs < min_tso_segs)
1703 			segs = min_tso_segs;
1704 		if (segs < rs_hw_floor_mss)
1705 			segs = rs_hw_floor_mss;
1706 		if (segs > MAX_MSS_SENT)
1707 			segs = MAX_MSS_SENT;
1708 		segs *= segsiz;
1709 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1710 				    te->rate, te->time_between, (uint32_t)res,
1711 				    segs, res_div, 1, 3);
1712 		if (err)
1713 			*err = 0;
1714 		if (segs < new_tso) {
1715 			/* unexpected ? */
1716 			return(new_tso);
1717 		} else {
1718 			return (segs);
1719 		}
1720 	} else {
1721 		/*
1722 		 * Your time is smaller which means
1723 		 * we will grow a queue on our
1724 		 * hardware. Send back the non-hardware
1725 		 * rate.
1726 		 */
1727 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1728 				    te->rate, te->time_between, (uint32_t)res,
1729 				    0, 0, 0, 4);
1730 		if (err)
1731 			*err = -1;
1732 		return (new_tso);
1733 	}
1734 }
1735 
1736 uint64_t
1737 tcp_hw_highest_rate_ifp(struct ifnet *ifp, struct inpcb *inp)
1738 {
1739 	struct epoch_tracker et;
1740 	struct tcp_rate_set *rs;
1741 	uint64_t rate_ret;
1742 
1743 	NET_EPOCH_ENTER(et);
1744 use_next_interface:
1745 	rs = find_rs_for_ifp(ifp);
1746 	if (rs == NULL) {
1747 		/* This interface does not do ratelimiting */
1748 		rate_ret = 0;
1749 	} else if (rs->rs_flags & RS_IS_DEFF) {
1750 		/* We need to find the real interface */
1751 		struct ifnet *tifp;
1752 
1753 		tifp = rt_find_real_interface(ifp, inp, NULL);
1754 		if (tifp == NULL) {
1755 			NET_EPOCH_EXIT(et);
1756 			return (0);
1757 		}
1758 		ifp = tifp;
1759 		goto use_next_interface;
1760 	} else {
1761 		/* Lets return the highest rate this guy has */
1762 		rate_ret = rs->rs_rlt[rs->rs_highest_valid].rate;
1763 	}
1764 	NET_EPOCH_EXIT(et);
1765 	return(rate_ret);
1766 }
1767 
1768 static eventhandler_tag rl_ifnet_departs;
1769 static eventhandler_tag rl_ifnet_arrives;
1770 static eventhandler_tag rl_shutdown_start;
1771 
1772 static void
1773 tcp_rs_init(void *st __unused)
1774 {
1775 	CK_LIST_INIT(&int_rs);
1776 	rs_number_alive = 0;
1777 	rs_number_dead = 0;
1778 	mtx_init(&rs_mtx, "tcp_rs_mtx", "rsmtx", MTX_DEF);
1779 	rl_ifnet_departs = EVENTHANDLER_REGISTER(ifnet_departure_event,
1780 	    tcp_rl_ifnet_departure,
1781 	    NULL, EVENTHANDLER_PRI_ANY);
1782 	rl_ifnet_arrives = EVENTHANDLER_REGISTER(ifnet_link_event,
1783 	    tcp_rl_ifnet_link,
1784 	    NULL, EVENTHANDLER_PRI_ANY);
1785 	rl_shutdown_start = EVENTHANDLER_REGISTER(shutdown_pre_sync,
1786 	    tcp_rl_shutdown, NULL,
1787 	    SHUTDOWN_PRI_FIRST);
1788 	printf("TCP_ratelimit: Is now initialized\n");
1789 }
1790 
1791 SYSINIT(tcp_rl_init, SI_SUB_SMP + 1, SI_ORDER_ANY, tcp_rs_init, NULL);
1792 #endif
1793