xref: /freebsd/sys/netinet/tcp_ratelimit.c (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
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 <net/if_private.h>
53 #include <netinet/in.h>
54 #include <netinet/in_pcb.h>
55 #define TCPSTATES		/* for logging */
56 #include <netinet/tcp_var.h>
57 #include <netinet/tcp_hpts.h>
58 #include <netinet/tcp_log_buf.h>
59 #include <netinet/tcp_ratelimit.h>
60 #ifndef USECS_IN_SECOND
61 #define USECS_IN_SECOND 1000000
62 #endif
63 /*
64  * For the purposes of each send, what is the size
65  * of an ethernet frame.
66  */
67 MALLOC_DEFINE(M_TCPPACE, "tcp_hwpace", "TCP Hardware pacing memory");
68 #ifdef RATELIMIT
69 
70 /*
71  * The following preferred table will seem weird to
72  * the casual viewer. Why do we not have any rates below
73  * 1Mbps? Why do we have a rate at 1.44Mbps called common?
74  * Why do the rates cluster in the 1-100Mbps range more
75  * than others? Why does the table jump around at the beginnign
76  * and then be more consistently raising?
77  *
78  * Let me try to answer those questions. A lot of
79  * this is dependant on the hardware. We have three basic
80  * supporters of rate limiting
81  *
82  * Chelsio - Supporting 16 configurable rates.
83  * Mlx  - c4 supporting 13 fixed rates.
84  * Mlx  - c5 & c6 supporting 127 configurable rates.
85  *
86  * The c4 is why we have a common rate that is available
87  * in all rate tables. This is a selected rate from the
88  * c4 table and we assure its available in all ratelimit
89  * tables. This way the tcp_ratelimit code has an assured
90  * rate it should always be able to get. This answers a
91  * couple of the questions above.
92  *
93  * So what about the rest, well the table is built to
94  * try to get the most out of a joint hardware/software
95  * pacing system.  The software pacer will always pick
96  * a rate higher than the b/w that it is estimating
97  *
98  * on the path. This is done for two reasons.
99  * a) So we can discover more b/w
100  * and
101  * b) So we can send a block of MSS's down and then
102  *    have the software timer go off after the previous
103  *    send is completely out of the hardware.
104  *
105  * But when we do <b> we don't want to have the delay
106  * between the last packet sent by the hardware be
107  * excessively long (to reach our desired rate).
108  *
109  * So let me give an example for clarity.
110  *
111  * Lets assume that the tcp stack sees that 29,110,000 bps is
112  * what the bw of the path is. The stack would select the
113  * rate 31Mbps. 31Mbps means that each send that is done
114  * by the hardware will cause a 390 micro-second gap between
115  * the packets sent at that rate. For 29,110,000 bps we
116  * would need 416 micro-seconds gap between each send.
117  *
118  * Note that are calculating a complete time for pacing
119  * which includes the ethernet, IP and TCP overhead. So
120  * a full 1514 bytes is used for the above calculations.
121  * My testing has shown that both cards are also using this
122  * as their basis i.e. full payload size of the ethernet frame.
123  * The TCP stack caller needs to be aware of this and make the
124  * appropriate overhead calculations be included in its choices.
125  *
126  * Now, continuing our example, we pick a MSS size based on the
127  * delta between the two rates (416 - 390) divided into the rate
128  * we really wish to send at rounded up.  That results in a MSS
129  * send of 17 mss's at once. The hardware then will
130  * run out of data in a single 17MSS send in 6,630 micro-seconds.
131  *
132  * On the other hand the software pacer will send more data
133  * in 7,072 micro-seconds. This means that we will refill
134  * the hardware 52 microseconds after it would have sent
135  * next if it had not ran out of data. This is a win since we are
136  * only sending every 7ms or so and yet all the packets are spaced on
137  * the wire with 94% of what they should be and only
138  * the last packet is delayed extra to make up for the
139  * difference.
140  *
141  * Note that the above formula has two important caveat.
142  * If we are above (b/w wise) over 100Mbps we double the result
143  * of the MSS calculation. The second caveat is if we are 500Mbps
144  * or more we just send the maximum MSS at once i.e. 45MSS. At
145  * the higher b/w's even the cards have limits to what times (timer granularity)
146  * they can insert between packets and start to send more than one
147  * packet at a time on the wire.
148  *
149  */
150 #define COMMON_RATE 180500
151 const uint64_t desired_rates[] = {
152 	122500,			/* 1Mbps  - rate 1 */
153 	180500,			/* 1.44Mpbs - rate 2  common rate */
154 	375000,			/* 3Mbps    - rate 3 */
155 	625000,			/* 5Mbps    - rate 4 */
156 	1250000,		/* 10Mbps   - rate 5 */
157 	1875000,		/* 15Mbps   - rate 6 */
158 	2500000,		/* 20Mbps   - rate 7 */
159 	3125000,	       	/* 25Mbps   - rate 8 */
160 	3750000,		/* 30Mbps   - rate 9 */
161 	4375000,		/* 35Mbps   - rate 10 */
162 	5000000,		/* 40Meg    - rate 11 */
163 	6250000,		/* 50Mbps   - rate 12 */
164 	12500000,		/* 100Mbps  - rate 13 */
165 	25000000,		/* 200Mbps  - rate 14 */
166 	50000000,		/* 400Mbps  - rate 15 */
167 	100000000,		/* 800Mbps  - rate 16 */
168 	5625000,		/* 45Mbps   - rate 17 */
169 	6875000,		/* 55Mbps   - rate 19 */
170 	7500000,		/* 60Mbps   - rate 20 */
171 	8125000,		/* 65Mbps   - rate 21 */
172 	8750000,		/* 70Mbps   - rate 22 */
173 	9375000,		/* 75Mbps   - rate 23 */
174 	10000000,		/* 80Mbps   - rate 24 */
175 	10625000,		/* 85Mbps   - rate 25 */
176 	11250000,		/* 90Mbps   - rate 26 */
177 	11875000,		/* 95Mbps   - rate 27 */
178 	12500000,		/* 100Mbps  - rate 28 */
179 	13750000,		/* 110Mbps  - rate 29 */
180 	15000000,		/* 120Mbps  - rate 30 */
181 	16250000,		/* 130Mbps  - rate 31 */
182 	17500000,		/* 140Mbps  - rate 32 */
183 	18750000,		/* 150Mbps  - rate 33 */
184 	20000000,		/* 160Mbps  - rate 34 */
185 	21250000,		/* 170Mbps  - rate 35 */
186 	22500000,		/* 180Mbps  - rate 36 */
187 	23750000,		/* 190Mbps  - rate 37 */
188 	26250000,		/* 210Mbps  - rate 38 */
189 	27500000,		/* 220Mbps  - rate 39 */
190 	28750000,		/* 230Mbps  - rate 40 */
191 	30000000,	       	/* 240Mbps  - rate 41 */
192 	31250000,		/* 250Mbps  - rate 42 */
193 	34375000,		/* 275Mbps  - rate 43 */
194 	37500000,		/* 300Mbps  - rate 44 */
195 	40625000,		/* 325Mbps  - rate 45 */
196 	43750000,		/* 350Mbps  - rate 46 */
197 	46875000,		/* 375Mbps  - rate 47 */
198 	53125000,		/* 425Mbps  - rate 48 */
199 	56250000,		/* 450Mbps  - rate 49 */
200 	59375000,		/* 475Mbps  - rate 50 */
201 	62500000,		/* 500Mbps  - rate 51 */
202 	68750000,		/* 550Mbps  - rate 52 */
203 	75000000,		/* 600Mbps  - rate 53 */
204 	81250000,		/* 650Mbps  - rate 54 */
205 	87500000,		/* 700Mbps  - rate 55 */
206 	93750000,		/* 750Mbps  - rate 56 */
207 	106250000,		/* 850Mbps  - rate 57 */
208 	112500000,		/* 900Mbps  - rate 58 */
209 	125000000,		/* 1Gbps    - rate 59 */
210 	156250000,		/* 1.25Gps  - rate 60 */
211 	187500000,		/* 1.5Gps   - rate 61 */
212 	218750000,		/* 1.75Gps  - rate 62 */
213 	250000000,		/* 2Gbps    - rate 63 */
214 	281250000,		/* 2.25Gps  - rate 64 */
215 	312500000,		/* 2.5Gbps  - rate 65 */
216 	343750000,		/* 2.75Gbps - rate 66 */
217 	375000000,		/* 3Gbps    - rate 67 */
218 	500000000,		/* 4Gbps    - rate 68 */
219 	625000000,		/* 5Gbps    - rate 69 */
220 	750000000,		/* 6Gbps    - rate 70 */
221 	875000000,		/* 7Gbps    - rate 71 */
222 	1000000000,		/* 8Gbps    - rate 72 */
223 	1125000000,		/* 9Gbps    - rate 73 */
224 	1250000000,		/* 10Gbps   - rate 74 */
225 	1875000000,		/* 15Gbps   - rate 75 */
226 	2500000000		/* 20Gbps   - rate 76 */
227 };
228 
229 #define MAX_HDWR_RATES (sizeof(desired_rates)/sizeof(uint64_t))
230 #define RS_ORDERED_COUNT 16	/*
231 				 * Number that are in order
232 				 * at the beginning of the table,
233 				 * over this a sort is required.
234 				 */
235 #define RS_NEXT_ORDER_GROUP 16	/*
236 				 * The point in our table where
237 				 * we come fill in a second ordered
238 				 * group (index wise means -1).
239 				 */
240 #define ALL_HARDWARE_RATES 1004 /*
241 				 * 1Meg - 1Gig in 1 Meg steps
242 				 * plus 100, 200k  and 500k and
243 				 * 10Gig
244 				 */
245 
246 #define RS_ONE_MEGABIT_PERSEC 1000000
247 #define RS_ONE_GIGABIT_PERSEC 1000000000
248 #define RS_TEN_GIGABIT_PERSEC 10000000000
249 
250 static struct head_tcp_rate_set int_rs;
251 static struct mtx rs_mtx;
252 uint32_t rs_number_alive;
253 uint32_t rs_number_dead;
254 static uint32_t rs_floor_mss = 0;
255 static uint32_t wait_time_floor = 8000;	/* 8 ms */
256 static uint32_t rs_hw_floor_mss = 16;
257 static uint32_t num_of_waits_allowed = 1; /* How many time blocks are we willing to wait */
258 
259 static uint32_t mss_divisor = RL_DEFAULT_DIVISOR;
260 static uint32_t even_num_segs = 1;
261 static uint32_t even_threshold = 4;
262 
263 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, rl, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
264     "TCP Ratelimit stats");
265 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, alive, CTLFLAG_RW,
266     &rs_number_alive, 0,
267     "Number of interfaces initialized for ratelimiting");
268 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, dead, CTLFLAG_RW,
269     &rs_number_dead, 0,
270     "Number of interfaces departing from ratelimiting");
271 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, floor_mss, CTLFLAG_RW,
272     &rs_floor_mss, 0,
273     "Number of MSS that will override the normal minimums (0 means don't enforce)");
274 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, wait_floor, CTLFLAG_RW,
275     &wait_time_floor, 2000,
276     "Has b/w increases what is the wait floor we are willing to wait at the end?");
277 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, time_blocks, CTLFLAG_RW,
278     &num_of_waits_allowed, 1,
279     "How many time blocks on the end should software pacing be willing to wait?");
280 
281 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, hw_floor_mss, CTLFLAG_RW,
282     &rs_hw_floor_mss, 16,
283     "Number of mss that are a minum for hardware pacing?");
284 
285 SYSCTL_INT(_net_inet_tcp_rl, OID_AUTO, divisor, CTLFLAG_RW,
286     &mss_divisor, RL_DEFAULT_DIVISOR,
287     "The value divided into bytes per second to help establish mss size");
288 SYSCTL_INT(_net_inet_tcp_rl, OID_AUTO, even, CTLFLAG_RW,
289     &even_num_segs, 1,
290     "Do we round mss size up to an even number of segments for delayed ack");
291 SYSCTL_INT(_net_inet_tcp_rl, OID_AUTO, eventhresh, CTLFLAG_RW,
292     &even_threshold, 4,
293     "At what number of mss do we start rounding up to an even number of mss?");
294 
295 static void
296 rl_add_syctl_entries(struct sysctl_oid *rl_sysctl_root, struct tcp_rate_set *rs)
297 {
298 	/*
299 	 * Add sysctl entries for thus interface.
300 	 */
301 	if (rs->rs_flags & RS_INTF_NO_SUP) {
302 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
303 		   SYSCTL_CHILDREN(rl_sysctl_root),
304 		   OID_AUTO, "disable", CTLFLAG_RD,
305 		   &rs->rs_disable, 0,
306 		   "Disable this interface from new hdwr limiting?");
307 	} else {
308 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
309 		   SYSCTL_CHILDREN(rl_sysctl_root),
310 		   OID_AUTO, "disable", CTLFLAG_RW,
311 		   &rs->rs_disable, 0,
312 		   "Disable this interface from new hdwr limiting?");
313 	}
314 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
315 	    SYSCTL_CHILDREN(rl_sysctl_root),
316 	    OID_AUTO, "minseg", CTLFLAG_RW,
317 	    &rs->rs_min_seg, 0,
318 	    "What is the minimum we need to send on this interface?");
319 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
320 	    SYSCTL_CHILDREN(rl_sysctl_root),
321 	    OID_AUTO, "flow_limit", CTLFLAG_RW,
322 	    &rs->rs_flow_limit, 0,
323 	    "What is the limit for number of flows (0=unlimited)?");
324 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
325 	    SYSCTL_CHILDREN(rl_sysctl_root),
326 	    OID_AUTO, "highest", CTLFLAG_RD,
327 	    &rs->rs_highest_valid, 0,
328 	    "Highest valid rate");
329 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
330 	    SYSCTL_CHILDREN(rl_sysctl_root),
331 	    OID_AUTO, "lowest", CTLFLAG_RD,
332 	    &rs->rs_lowest_valid, 0,
333 	    "Lowest valid rate");
334 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
335 	    SYSCTL_CHILDREN(rl_sysctl_root),
336 	    OID_AUTO, "flags", CTLFLAG_RD,
337 	    &rs->rs_flags, 0,
338 	    "What lags are on the entry?");
339 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
340 	    SYSCTL_CHILDREN(rl_sysctl_root),
341 	    OID_AUTO, "numrates", CTLFLAG_RD,
342 	    &rs->rs_rate_cnt, 0,
343 	    "How many rates re there?");
344 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
345 	    SYSCTL_CHILDREN(rl_sysctl_root),
346 	    OID_AUTO, "flows_using", CTLFLAG_RD,
347 	    &rs->rs_flows_using, 0,
348 	    "How many flows are using this interface now?");
349 #ifdef DETAILED_RATELIMIT_SYSCTL
350 	if (rs->rs_rlt && rs->rs_rate_cnt > 0) {
351 		/*  Lets display the rates */
352 		int i;
353 		struct sysctl_oid *rl_rates;
354 		struct sysctl_oid *rl_rate_num;
355 		char rate_num[16];
356 		rl_rates = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
357 					    SYSCTL_CHILDREN(rl_sysctl_root),
358 					    OID_AUTO,
359 					    "rate",
360 					    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
361 					    "Ratelist");
362 		for( i = 0; i < rs->rs_rate_cnt; i++) {
363 			sprintf(rate_num, "%d", i);
364 			rl_rate_num = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
365 					    SYSCTL_CHILDREN(rl_rates),
366 					    OID_AUTO,
367 					    rate_num,
368 					    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
369 					    "Individual Rate");
370 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
371 				       SYSCTL_CHILDREN(rl_rate_num),
372 				       OID_AUTO, "flags", CTLFLAG_RD,
373 				       &rs->rs_rlt[i].flags, 0,
374 				       "Flags on this rate");
375 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
376 				       SYSCTL_CHILDREN(rl_rate_num),
377 				       OID_AUTO, "pacetime", CTLFLAG_RD,
378 				       &rs->rs_rlt[i].time_between, 0,
379 				       "Time hardware inserts between 1500 byte sends");
380 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
381 				       SYSCTL_CHILDREN(rl_rate_num),
382 				       OID_AUTO, "rate", CTLFLAG_RD,
383 				       &rs->rs_rlt[i].rate,
384 				       "Rate in bytes per second");
385 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
386 				       SYSCTL_CHILDREN(rl_rate_num),
387 				       OID_AUTO, "using", CTLFLAG_RD,
388 				       &rs->rs_rlt[i].using,
389 				       "Number of flows using");
390 			SYSCTL_ADD_LONG(&rs->sysctl_ctx,
391 				       SYSCTL_CHILDREN(rl_rate_num),
392 				       OID_AUTO, "enobufs", CTLFLAG_RD,
393 				       &rs->rs_rlt[i].rs_num_enobufs,
394 				       "Number of enobufs logged on this rate");
395 
396 		}
397 	}
398 #endif
399 }
400 
401 static void
402 rs_destroy(epoch_context_t ctx)
403 {
404 	struct tcp_rate_set *rs;
405 	bool do_free_rs;
406 
407 	rs = __containerof(ctx, struct tcp_rate_set, rs_epoch_ctx);
408 
409 	mtx_lock(&rs_mtx);
410 	rs->rs_flags &= ~RS_FUNERAL_SCHD;
411 	/*
412 	 * In theory its possible (but unlikely)
413 	 * that while the delete was occuring
414 	 * and we were applying the DEAD flag
415 	 * someone slipped in and found the
416 	 * interface in a lookup. While we
417 	 * decided rs_flows_using were 0 and
418 	 * scheduling the epoch_call, the other
419 	 * thread incremented rs_flow_using. This
420 	 * is because users have a pointer and
421 	 * we only use the rs_flows_using in an
422 	 * atomic fashion, i.e. the other entities
423 	 * are not protected. To assure this did
424 	 * not occur, we check rs_flows_using here
425 	 * before deleting.
426 	 */
427 	do_free_rs = (rs->rs_flows_using == 0);
428 	rs_number_dead--;
429 	mtx_unlock(&rs_mtx);
430 
431 	if (do_free_rs) {
432 		sysctl_ctx_free(&rs->sysctl_ctx);
433 		free(rs->rs_rlt, M_TCPPACE);
434 		free(rs, M_TCPPACE);
435 	}
436 }
437 
438 static void
439 rs_defer_destroy(struct tcp_rate_set *rs)
440 {
441 
442 	mtx_assert(&rs_mtx, MA_OWNED);
443 
444 	/* Check if already pending. */
445 	if (rs->rs_flags & RS_FUNERAL_SCHD)
446 		return;
447 
448 	rs_number_dead++;
449 
450 	/* Set flag to only defer once. */
451 	rs->rs_flags |= RS_FUNERAL_SCHD;
452 	NET_EPOCH_CALL(rs_destroy, &rs->rs_epoch_ctx);
453 }
454 
455 #ifdef INET
456 extern counter_u64_t rate_limit_new;
457 extern counter_u64_t rate_limit_chg;
458 extern counter_u64_t rate_limit_set_ok;
459 extern counter_u64_t rate_limit_active;
460 extern counter_u64_t rate_limit_alloc_fail;
461 #endif
462 
463 static int
464 rl_attach_txrtlmt(struct ifnet *ifp,
465     uint32_t flowtype,
466     int flowid,
467     uint64_t cfg_rate,
468     struct m_snd_tag **tag)
469 {
470 	int error;
471 	union if_snd_tag_alloc_params params = {
472 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
473 		.rate_limit.hdr.flowid = flowid,
474 		.rate_limit.hdr.flowtype = flowtype,
475 		.rate_limit.max_rate = cfg_rate,
476 		.rate_limit.flags = M_NOWAIT,
477 	};
478 
479 	error = m_snd_tag_alloc(ifp, &params, tag);
480 #ifdef INET
481 	if (error == 0) {
482 		counter_u64_add(rate_limit_set_ok, 1);
483 		counter_u64_add(rate_limit_active, 1);
484 	} else if (error != EOPNOTSUPP)
485 		counter_u64_add(rate_limit_alloc_fail, 1);
486 #endif
487 	return (error);
488 }
489 
490 static void
491 populate_canned_table(struct tcp_rate_set *rs, const uint64_t *rate_table_act)
492 {
493 	/*
494 	 * The internal table is "special", it
495 	 * is two seperate ordered tables that
496 	 * must be merged. We get here when the
497 	 * adapter specifies a number of rates that
498 	 * covers both ranges in the table in some
499 	 * form.
500 	 */
501 	int i, at_low, at_high;
502 	uint8_t low_disabled = 0, high_disabled = 0;
503 
504 	for(i = 0, at_low = 0, at_high = RS_NEXT_ORDER_GROUP; i < rs->rs_rate_cnt; i++) {
505 		rs->rs_rlt[i].flags = 0;
506 		rs->rs_rlt[i].time_between = 0;
507 		if ((low_disabled == 0) &&
508 		    (high_disabled ||
509 		     (rate_table_act[at_low] < rate_table_act[at_high]))) {
510 			rs->rs_rlt[i].rate = rate_table_act[at_low];
511 			at_low++;
512 			if (at_low == RS_NEXT_ORDER_GROUP)
513 				low_disabled = 1;
514 		} else if (high_disabled == 0) {
515 			rs->rs_rlt[i].rate = rate_table_act[at_high];
516 			at_high++;
517 			if (at_high == MAX_HDWR_RATES)
518 				high_disabled = 1;
519 		}
520 	}
521 }
522 
523 static struct tcp_rate_set *
524 rt_setup_new_rs(struct ifnet *ifp, int *error)
525 {
526 	struct tcp_rate_set *rs;
527 	const uint64_t *rate_table_act;
528 	uint64_t lentim, res;
529 	size_t sz;
530 	uint32_t hash_type;
531 	int i;
532 	struct if_ratelimit_query_results rl;
533 	struct sysctl_oid *rl_sysctl_root;
534 	struct epoch_tracker et;
535 	/*
536 	 * We expect to enter with the
537 	 * mutex locked.
538 	 */
539 
540 	if (ifp->if_ratelimit_query == NULL) {
541 		/*
542 		 * We can do nothing if we cannot
543 		 * get a query back from the driver.
544 		 */
545 		printf("Warning:No query functions for %s:%d-- failed\n",
546 		       ifp->if_dname, ifp->if_dunit);
547 		return (NULL);
548 	}
549 	rs = malloc(sizeof(struct tcp_rate_set), M_TCPPACE, M_NOWAIT | M_ZERO);
550 	if (rs == NULL) {
551 		if (error)
552 			*error = ENOMEM;
553 		printf("Warning:No memory for malloc of tcp_rate_set\n");
554 		return (NULL);
555 	}
556 	memset(&rl, 0, sizeof(rl));
557 	rl.flags = RT_NOSUPPORT;
558 	ifp->if_ratelimit_query(ifp, &rl);
559 	if (rl.flags & RT_IS_UNUSABLE) {
560 		/*
561 		 * The interface does not really support
562 		 * the rate-limiting.
563 		 */
564 		memset(rs, 0, sizeof(struct tcp_rate_set));
565 		rs->rs_ifp = ifp;
566 		rs->rs_if_dunit = ifp->if_dunit;
567 		rs->rs_flags = RS_INTF_NO_SUP;
568 		rs->rs_disable = 1;
569 		rs_number_alive++;
570 		sysctl_ctx_init(&rs->sysctl_ctx);
571 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
572 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
573 		    OID_AUTO,
574 		    rs->rs_ifp->if_xname,
575 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
576 		    "");
577 		rl_add_syctl_entries(rl_sysctl_root, rs);
578 		NET_EPOCH_ENTER(et);
579 		mtx_lock(&rs_mtx);
580 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
581 		mtx_unlock(&rs_mtx);
582 		NET_EPOCH_EXIT(et);
583 		return (rs);
584 	} else if ((rl.flags & RT_IS_INDIRECT) == RT_IS_INDIRECT) {
585 		memset(rs, 0, sizeof(struct tcp_rate_set));
586 		rs->rs_ifp = ifp;
587 		rs->rs_if_dunit = ifp->if_dunit;
588 		rs->rs_flags = RS_IS_DEFF;
589 		rs_number_alive++;
590 		sysctl_ctx_init(&rs->sysctl_ctx);
591 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
592 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
593 		    OID_AUTO,
594 		    rs->rs_ifp->if_xname,
595 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
596 		    "");
597 		rl_add_syctl_entries(rl_sysctl_root, rs);
598 		NET_EPOCH_ENTER(et);
599 		mtx_lock(&rs_mtx);
600 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
601 		mtx_unlock(&rs_mtx);
602 		NET_EPOCH_EXIT(et);
603 		return (rs);
604 	} else if ((rl.flags & RT_IS_FIXED_TABLE) == RT_IS_FIXED_TABLE) {
605 		/* Mellanox C4 likely */
606 		rs->rs_ifp = ifp;
607 		rs->rs_if_dunit = ifp->if_dunit;
608 		rs->rs_rate_cnt = rl.number_of_rates;
609 		rs->rs_min_seg = rl.min_segment_burst;
610 		rs->rs_highest_valid = 0;
611 		rs->rs_flow_limit = rl.max_flows;
612 		rs->rs_flags = RS_IS_INTF | RS_NO_PRE;
613 		rs->rs_disable = 0;
614 		rate_table_act = rl.rate_table;
615 	} else if ((rl.flags & RT_IS_SELECTABLE) == RT_IS_SELECTABLE) {
616 		/* Chelsio, C5 and C6 of Mellanox? */
617 		rs->rs_ifp = ifp;
618 		rs->rs_if_dunit = ifp->if_dunit;
619 		rs->rs_rate_cnt = rl.number_of_rates;
620 		rs->rs_min_seg = rl.min_segment_burst;
621 		rs->rs_disable = 0;
622 		rs->rs_flow_limit = rl.max_flows;
623 		rate_table_act = desired_rates;
624 		if ((rs->rs_rate_cnt > MAX_HDWR_RATES) &&
625 		    (rs->rs_rate_cnt < ALL_HARDWARE_RATES)) {
626 			/*
627 			 * Our desired table is not big
628 			 * enough, do what we can.
629 			 */
630 			rs->rs_rate_cnt = MAX_HDWR_RATES;
631 		 }
632 		if (rs->rs_rate_cnt <= RS_ORDERED_COUNT)
633 			rs->rs_flags = RS_IS_INTF;
634 		else
635 			rs->rs_flags = RS_IS_INTF | RS_INT_TBL;
636 		if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)
637 			rs->rs_rate_cnt = ALL_HARDWARE_RATES;
638 	} else {
639 		free(rs, M_TCPPACE);
640 		return (NULL);
641 	}
642 	sz = sizeof(struct tcp_hwrate_limit_table) * rs->rs_rate_cnt;
643 	rs->rs_rlt = malloc(sz, M_TCPPACE, M_NOWAIT);
644 	if (rs->rs_rlt == NULL) {
645 		if (error)
646 			*error = ENOMEM;
647 bail:
648 		free(rs, M_TCPPACE);
649 		return (NULL);
650 	}
651 	if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) {
652 		/*
653 		 * The interface supports all
654 		 * the rates we could possibly want.
655 		 */
656 		uint64_t rat;
657 
658 		rs->rs_rlt[0].rate = 12500;	/* 100k */
659 		rs->rs_rlt[1].rate = 25000;	/* 200k */
660 		rs->rs_rlt[2].rate = 62500;	/* 500k */
661 		/* Note 125000 == 1Megabit
662 		 * populate 1Meg - 1000meg.
663 		 */
664 		for(i = 3, rat = 125000; i< (ALL_HARDWARE_RATES-1); i++) {
665 			rs->rs_rlt[i].rate = rat;
666 			rat += 125000;
667 		}
668 		rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate = 1250000000;
669 	} else if (rs->rs_flags & RS_INT_TBL) {
670 		/* We populate this in a special way */
671 		populate_canned_table(rs, rate_table_act);
672 	} else {
673 		/*
674 		 * Just copy in the rates from
675 		 * the table, it is in order.
676 		 */
677 		for (i=0; i<rs->rs_rate_cnt; i++) {
678 			rs->rs_rlt[i].rate = rate_table_act[i];
679 			rs->rs_rlt[i].time_between = 0;
680 			rs->rs_rlt[i].flags = 0;
681 		}
682 	}
683 	for (i = (rs->rs_rate_cnt - 1); i >= 0; i--) {
684 		/*
685 		 * We go backwards through the list so that if we can't get
686 		 * a rate and fail to init one, we have at least a chance of
687 		 * getting the highest one.
688 		 */
689 		rs->rs_rlt[i].ptbl = rs;
690 		rs->rs_rlt[i].tag = NULL;
691 		rs->rs_rlt[i].using = 0;
692 		rs->rs_rlt[i].rs_num_enobufs = 0;
693 		/*
694 		 * Calculate the time between.
695 		 */
696 		lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
697 		res = lentim / rs->rs_rlt[i].rate;
698 		if (res > 0)
699 			rs->rs_rlt[i].time_between = res;
700 		else
701 			rs->rs_rlt[i].time_between = 1;
702 		if (rs->rs_flags & RS_NO_PRE) {
703 			rs->rs_rlt[i].flags = HDWRPACE_INITED;
704 			rs->rs_lowest_valid = i;
705 		} else {
706 			int err;
707 
708 			if ((rl.flags & RT_IS_SETUP_REQ)  &&
709 			    (ifp->if_ratelimit_query)) {
710 				err = ifp->if_ratelimit_setup(ifp,
711   				         rs->rs_rlt[i].rate, i);
712 				if (err)
713 					goto handle_err;
714 			}
715 #ifdef RSS
716 			hash_type = M_HASHTYPE_RSS_TCP_IPV4;
717 #else
718 			hash_type = M_HASHTYPE_OPAQUE_HASH;
719 #endif
720 			err = rl_attach_txrtlmt(ifp,
721 			    hash_type,
722 			    (i + 1),
723 			    rs->rs_rlt[i].rate,
724 			    &rs->rs_rlt[i].tag);
725 			if (err) {
726 handle_err:
727 				if (i == (rs->rs_rate_cnt - 1)) {
728 					/*
729 					 * Huh - first rate and we can't get
730 					 * it?
731 					 */
732 					free(rs->rs_rlt, M_TCPPACE);
733 					if (error)
734 						*error = err;
735 					goto bail;
736 				} else {
737 					if (error)
738 						*error = err;
739 				}
740 				break;
741 			} else {
742 				rs->rs_rlt[i].flags = HDWRPACE_INITED | HDWRPACE_TAGPRESENT;
743 				rs->rs_lowest_valid = i;
744 			}
745 		}
746 	}
747 	/* Did we get at least 1 rate? */
748 	if (rs->rs_rlt[(rs->rs_rate_cnt - 1)].flags & HDWRPACE_INITED)
749 		rs->rs_highest_valid = rs->rs_rate_cnt - 1;
750 	else {
751 		free(rs->rs_rlt, M_TCPPACE);
752 		goto bail;
753 	}
754 	rs_number_alive++;
755 	sysctl_ctx_init(&rs->sysctl_ctx);
756 	rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
757 	    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
758 	    OID_AUTO,
759 	    rs->rs_ifp->if_xname,
760 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
761 	    "");
762 	rl_add_syctl_entries(rl_sysctl_root, rs);
763 	NET_EPOCH_ENTER(et);
764 	mtx_lock(&rs_mtx);
765 	CK_LIST_INSERT_HEAD(&int_rs, rs, next);
766 	mtx_unlock(&rs_mtx);
767 	NET_EPOCH_EXIT(et);
768 	return (rs);
769 }
770 
771 /*
772  * For an explanation of why the argument is volatile please
773  * look at the comments around rt_setup_rate().
774  */
775 static const struct tcp_hwrate_limit_table *
776 tcp_int_find_suitable_rate(const volatile struct tcp_rate_set *rs,
777     uint64_t bytes_per_sec, uint32_t flags, uint64_t *lower_rate)
778 {
779 	struct tcp_hwrate_limit_table *arte = NULL, *rte = NULL;
780 	uint64_t mbits_per_sec, ind_calc, previous_rate = 0;
781 	int i;
782 
783 	mbits_per_sec = (bytes_per_sec * 8);
784 	if (flags & RS_PACING_LT) {
785 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
786 		    (rs->rs_lowest_valid <= 2)){
787 			/*
788 			 * Smaller than 1Meg, only
789 			 * 3 entries can match it.
790 			 */
791 			previous_rate = 0;
792 			for(i = rs->rs_lowest_valid; i < 3; i++) {
793 				if (bytes_per_sec <= rs->rs_rlt[i].rate) {
794 					rte = &rs->rs_rlt[i];
795 					break;
796 				} else if (rs->rs_rlt[i].flags & HDWRPACE_INITED) {
797 					arte = &rs->rs_rlt[i];
798 				}
799 				previous_rate = rs->rs_rlt[i].rate;
800 			}
801 			goto done;
802 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
803 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
804 			/*
805 			 * Larger than 1G (the majority of
806 			 * our table.
807 			 */
808 			if (mbits_per_sec < RS_TEN_GIGABIT_PERSEC)
809 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
810 			else
811 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
812 			previous_rate = rs->rs_rlt[(ALL_HARDWARE_RATES-2)].rate;
813 			goto done;
814 		}
815 		/*
816 		 * If we reach here its in our table (between 1Meg - 1000Meg),
817 		 * just take the rounded down mbits per second, and add
818 		 * 1Megabit to it, from this we can calculate
819 		 * the index in the table.
820 		 */
821 		ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
822 		if ((ind_calc * RS_ONE_MEGABIT_PERSEC) != mbits_per_sec)
823 			ind_calc++;
824 		/* our table is offset by 3, we add 2 */
825 		ind_calc += 2;
826 		if (ind_calc > (ALL_HARDWARE_RATES-1)) {
827 			/* This should not happen */
828 			ind_calc = ALL_HARDWARE_RATES-1;
829 		}
830 		if ((ind_calc >= rs->rs_lowest_valid) &&
831 		    (ind_calc <= rs->rs_highest_valid)) {
832 			rte = &rs->rs_rlt[ind_calc];
833 			if (ind_calc >= 1)
834 				previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
835 		}
836 	} else if (flags & RS_PACING_EXACT_MATCH) {
837 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
838 		    (rs->rs_lowest_valid <= 2)){
839 			for(i = rs->rs_lowest_valid; i < 3; i++) {
840 				if (bytes_per_sec == rs->rs_rlt[i].rate) {
841 					rte = &rs->rs_rlt[i];
842 					break;
843 				}
844 			}
845 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
846 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
847 			/* > 1Gbps only one rate */
848 			if (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) {
849 				/* Its 10G wow */
850 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
851 			}
852 		} else {
853 			/* Ok it must be a exact meg (its between 1G and 1Meg) */
854 			ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
855 			if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
856 				/* its an exact Mbps */
857 				ind_calc += 2;
858 				if (ind_calc > (ALL_HARDWARE_RATES-1)) {
859 					/* This should not happen */
860 					ind_calc = ALL_HARDWARE_RATES-1;
861 				}
862 				if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED)
863 					rte = &rs->rs_rlt[ind_calc];
864 			}
865 		}
866 	} else {
867 		/* we want greater than the requested rate */
868 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
869 		    (rs->rs_lowest_valid <= 2)){
870 			arte = &rs->rs_rlt[3]; /* set alternate to 1Meg */
871 			for (i=2; i>=rs->rs_lowest_valid; i--) {
872 				if (bytes_per_sec < rs->rs_rlt[i].rate) {
873 					rte = &rs->rs_rlt[i];
874 					if (i >= 1) {
875 						previous_rate = rs->rs_rlt[(i-1)].rate;
876 					}
877 					break;
878 				} else if ((flags & RS_PACING_GEQ) &&
879 					   (bytes_per_sec == rs->rs_rlt[i].rate)) {
880 					rte = &rs->rs_rlt[i];
881 					if (i >= 1) {
882 						previous_rate = rs->rs_rlt[(i-1)].rate;
883 					}
884 					break;
885 				} else {
886 					arte = &rs->rs_rlt[i]; /* new alternate */
887 				}
888 			}
889 		} else if (mbits_per_sec > RS_ONE_GIGABIT_PERSEC) {
890 			if ((bytes_per_sec < rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
891 			    (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
892 				/* Our top rate is larger than the request */
893 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
894 			} else if ((flags & RS_PACING_GEQ) &&
895 				   (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
896 				   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
897 				/* It matches our top rate */
898 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
899 			} else if (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED) {
900 				/* The top rate is an alternative */
901 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
902 			}
903 			previous_rate = rs->rs_rlt[(ALL_HARDWARE_RATES-2)].rate;
904 		} else {
905 			/* Its in our range 1Meg - 1Gig */
906 			if (flags & RS_PACING_GEQ) {
907 				ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
908 				if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
909 					if (ind_calc > (ALL_HARDWARE_RATES-1)) {
910 						/* This should not happen */
911 						ind_calc = (ALL_HARDWARE_RATES-1);
912 					}
913 					rte = &rs->rs_rlt[ind_calc];
914 					if (ind_calc >= 1)
915 						previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
916 				}
917 				goto done;
918 			}
919 			ind_calc = (mbits_per_sec + (RS_ONE_MEGABIT_PERSEC-1))/RS_ONE_MEGABIT_PERSEC;
920 			ind_calc += 2;
921 			if (ind_calc > (ALL_HARDWARE_RATES-1)) {
922 				/* This should not happen */
923 				ind_calc = ALL_HARDWARE_RATES-1;
924 			}
925 			if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) {
926 				rte = &rs->rs_rlt[ind_calc];
927 				if (ind_calc >= 1)
928 					previous_rate = rs->rs_rlt[(ind_calc-1)].rate;
929 			}
930 		}
931 	}
932 done:
933 	if ((rte == NULL) &&
934 	    (arte != NULL) &&
935 	    (flags & RS_PACING_SUB_OK)) {
936 		/* We can use the substitute */
937 		rte = arte;
938 	}
939 	if (lower_rate)
940 		*lower_rate = previous_rate;
941 	return (rte);
942 }
943 
944 /*
945  * For an explanation of why the argument is volatile please
946  * look at the comments around rt_setup_rate().
947  */
948 static const struct tcp_hwrate_limit_table *
949 tcp_find_suitable_rate(const volatile struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags, uint64_t *lower_rate)
950 {
951 	/**
952 	 * Hunt the rate table with the restrictions in flags and find a
953 	 * suitable rate if possible.
954 	 * RS_PACING_EXACT_MATCH - look for an exact match to rate.
955 	 * RS_PACING_GT     - must be greater than.
956 	 * RS_PACING_GEQ    - must be greater than or equal.
957 	 * RS_PACING_LT     - must be less than.
958 	 * RS_PACING_SUB_OK - If we don't meet criteria a
959 	 *                    substitute is ok.
960 	 */
961 	int i, matched;
962 	struct tcp_hwrate_limit_table *rte = NULL;
963 	uint64_t previous_rate = 0;
964 
965 	if ((rs->rs_flags & RS_INT_TBL) &&
966 	    (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)) {
967 		/*
968 		 * Here we don't want to paw thru
969 		 * a big table, we have everything
970 		 * from 1Meg - 1000Meg in 1Meg increments.
971 		 * Use an alternate method to "lookup".
972 		 */
973 		return (tcp_int_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate));
974 	}
975 	if ((flags & RS_PACING_LT) ||
976 	    (flags & RS_PACING_EXACT_MATCH)) {
977 		/*
978 		 * For exact and less than we go forward through the table.
979 		 * This way when we find one larger we stop (exact was a
980 		 * toss up).
981 		 */
982 		for (i = rs->rs_lowest_valid, matched = 0; i <= rs->rs_highest_valid; i++) {
983 			if ((flags & RS_PACING_EXACT_MATCH) &&
984 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
985 				rte = &rs->rs_rlt[i];
986 				matched = 1;
987 				if (lower_rate != NULL)
988 					*lower_rate = previous_rate;
989 				break;
990 			} else if ((flags & RS_PACING_LT) &&
991 			    (bytes_per_sec <= rs->rs_rlt[i].rate)) {
992 				rte = &rs->rs_rlt[i];
993 				matched = 1;
994 				if (lower_rate != NULL)
995 					*lower_rate = previous_rate;
996 				break;
997 			}
998 			previous_rate = rs->rs_rlt[i].rate;
999 			if (bytes_per_sec > rs->rs_rlt[i].rate)
1000 				break;
1001 		}
1002 		if ((matched == 0) &&
1003 		    (flags & RS_PACING_LT) &&
1004 		    (flags & RS_PACING_SUB_OK)) {
1005 			/* Kick in a substitute (the lowest) */
1006 			rte = &rs->rs_rlt[rs->rs_lowest_valid];
1007 		}
1008 	} else {
1009 		/*
1010 		 * Here we go backward through the table so that we can find
1011 		 * the one greater in theory faster (but its probably a
1012 		 * wash).
1013 		 */
1014 		for (i = rs->rs_highest_valid, matched = 0; i >= rs->rs_lowest_valid; i--) {
1015 			if (rs->rs_rlt[i].rate > bytes_per_sec) {
1016 				/* A possible candidate */
1017 				rte = &rs->rs_rlt[i];
1018 			}
1019 			if ((flags & RS_PACING_GEQ) &&
1020 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
1021 				/* An exact match and we want equal */
1022 				matched = 1;
1023 				rte = &rs->rs_rlt[i];
1024 				break;
1025 			} else if (rte) {
1026 				/*
1027 				 * Found one that is larger than but don't
1028 				 * stop, there may be a more closer match.
1029 				 */
1030 				matched = 1;
1031 			}
1032 			if (rs->rs_rlt[i].rate < bytes_per_sec) {
1033 				/*
1034 				 * We found a table entry that is smaller,
1035 				 * stop there will be none greater or equal.
1036 				 */
1037 				if (lower_rate != NULL)
1038 					*lower_rate = rs->rs_rlt[i].rate;
1039 				break;
1040 			}
1041 		}
1042 		if ((matched == 0) &&
1043 		    (flags & RS_PACING_SUB_OK)) {
1044 			/* Kick in a substitute (the highest) */
1045 			rte = &rs->rs_rlt[rs->rs_highest_valid];
1046 		}
1047 	}
1048 	return (rte);
1049 }
1050 
1051 static struct ifnet *
1052 rt_find_real_interface(struct ifnet *ifp, struct inpcb *inp, int *error)
1053 {
1054 	struct ifnet *tifp;
1055 	struct m_snd_tag *tag, *ntag;
1056 	union if_snd_tag_alloc_params params = {
1057 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
1058 		.rate_limit.hdr.flowid = inp->inp_flowid,
1059 		.rate_limit.hdr.numa_domain = inp->inp_numa_domain,
1060 		.rate_limit.max_rate = COMMON_RATE,
1061 		.rate_limit.flags = M_NOWAIT,
1062 	};
1063 	int err;
1064 #ifdef RSS
1065 	params.rate_limit.hdr.flowtype = ((inp->inp_vflag & INP_IPV6) ?
1066 	    M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_TCP_IPV4);
1067 #else
1068 	params.rate_limit.hdr.flowtype = M_HASHTYPE_OPAQUE_HASH;
1069 #endif
1070 	err = m_snd_tag_alloc(ifp, &params, &tag);
1071 	if (err) {
1072 		/* Failed to setup a tag? */
1073 		if (error)
1074 			*error = err;
1075 		return (NULL);
1076 	}
1077 	ntag = tag;
1078 	while (ntag->sw->next_snd_tag != NULL) {
1079 		ntag = ntag->sw->next_snd_tag(ntag);
1080 	}
1081 	tifp = ntag->ifp;
1082 	m_snd_tag_rele(tag);
1083 	return (tifp);
1084 }
1085 
1086 static void
1087 rl_increment_using(const struct tcp_hwrate_limit_table *rte)
1088 {
1089 	struct tcp_hwrate_limit_table *decon_rte;
1090 
1091 	decon_rte = __DECONST(struct tcp_hwrate_limit_table *, rte);
1092 	atomic_add_long(&decon_rte->using, 1);
1093 }
1094 
1095 static void
1096 rl_decrement_using(const struct tcp_hwrate_limit_table *rte)
1097 {
1098 	struct tcp_hwrate_limit_table *decon_rte;
1099 
1100 	decon_rte = __DECONST(struct tcp_hwrate_limit_table *, rte);
1101 	atomic_subtract_long(&decon_rte->using, 1);
1102 }
1103 
1104 void
1105 tcp_rl_log_enobuf(const struct tcp_hwrate_limit_table *rte)
1106 {
1107 	struct tcp_hwrate_limit_table *decon_rte;
1108 
1109 	decon_rte = __DECONST(struct tcp_hwrate_limit_table *, rte);
1110 	atomic_add_long(&decon_rte->rs_num_enobufs, 1);
1111 }
1112 
1113 /*
1114  * Do NOT take the __noinline out of the
1115  * find_rs_for_ifp() function. If you do the inline
1116  * of it for the rt_setup_rate() will show you a
1117  * compiler bug. For some reason the compiler thinks
1118  * the list can never be empty. The consequence of
1119  * this will be a crash when we dereference NULL
1120  * if an ifp is removed just has a hw rate limit
1121  * is attempted. If you are working on the compiler
1122  * and want to "test" this go ahead and take the noinline
1123  * out otherwise let sleeping dogs ly until such time
1124  * as we get a compiler fix 10/2/20 -- RRS
1125  */
1126 static __noinline struct tcp_rate_set *
1127 find_rs_for_ifp(struct ifnet *ifp)
1128 {
1129 	struct tcp_rate_set *rs;
1130 
1131 	CK_LIST_FOREACH(rs, &int_rs, next) {
1132 		if ((rs->rs_ifp == ifp) &&
1133 		    (rs->rs_if_dunit == ifp->if_dunit)) {
1134 			/* Ok we found it */
1135 			return (rs);
1136 		}
1137 	}
1138 	return (NULL);
1139 }
1140 
1141 
1142 static const struct tcp_hwrate_limit_table *
1143 rt_setup_rate(struct inpcb *inp, struct ifnet *ifp, uint64_t bytes_per_sec,
1144     uint32_t flags, int *error, uint64_t *lower_rate)
1145 {
1146 	/* First lets find the interface if it exists */
1147 	const struct tcp_hwrate_limit_table *rte;
1148 	/*
1149 	 * So why is rs volatile? This is to defeat a
1150 	 * compiler bug where in the compiler is convinced
1151 	 * that rs can never be NULL (which is not true). Because
1152 	 * of its conviction it nicely optimizes out the if ((rs == NULL
1153 	 * below which means if you get a NULL back you dereference it.
1154 	 */
1155 	volatile struct tcp_rate_set *rs;
1156 	struct epoch_tracker et;
1157 	struct ifnet *oifp = ifp;
1158 	int err;
1159 
1160 	NET_EPOCH_ENTER(et);
1161 use_real_interface:
1162 	rs = find_rs_for_ifp(ifp);
1163 	if ((rs == NULL) ||
1164 	    (rs->rs_flags & RS_INTF_NO_SUP) ||
1165 	    (rs->rs_flags & RS_IS_DEAD)) {
1166 		/*
1167 		 * This means we got a packet *before*
1168 		 * the IF-UP was processed below, <or>
1169 		 * while or after we already received an interface
1170 		 * departed event. In either case we really don't
1171 		 * want to do anything with pacing, in
1172 		 * the departing case the packet is not
1173 		 * going to go very far. The new case
1174 		 * might be arguable, but its impossible
1175 		 * to tell from the departing case.
1176 		 */
1177 		if (error)
1178 			*error = ENODEV;
1179 		NET_EPOCH_EXIT(et);
1180 		return (NULL);
1181 	}
1182 
1183 	if ((rs == NULL) || (rs->rs_disable != 0)) {
1184 		if (error)
1185 			*error = ENOSPC;
1186 		NET_EPOCH_EXIT(et);
1187 		return (NULL);
1188 	}
1189 	if (rs->rs_flags & RS_IS_DEFF) {
1190 		/* We need to find the real interface */
1191 		struct ifnet *tifp;
1192 
1193 		tifp = rt_find_real_interface(ifp, inp, error);
1194 		if (tifp == NULL) {
1195 			if (rs->rs_disable && error)
1196 				*error = ENOTSUP;
1197 			NET_EPOCH_EXIT(et);
1198 			return (NULL);
1199 		}
1200 		KASSERT((tifp != ifp),
1201 			("Lookup failure ifp:%p inp:%p rt_find_real_interface() returns the same interface tifp:%p?\n",
1202 			 ifp, inp, tifp));
1203 		ifp = tifp;
1204 		goto use_real_interface;
1205 	}
1206 	if (rs->rs_flow_limit &&
1207 	    ((rs->rs_flows_using + 1) > rs->rs_flow_limit)) {
1208 		if (error)
1209 			*error = ENOSPC;
1210 		NET_EPOCH_EXIT(et);
1211 		return (NULL);
1212 	}
1213 	rte = tcp_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate);
1214 	if (rte) {
1215 		err = in_pcbattach_txrtlmt(inp, oifp,
1216 		    inp->inp_flowtype,
1217 		    inp->inp_flowid,
1218 		    rte->rate,
1219 		    &inp->inp_snd_tag);
1220 		if (err) {
1221 			/* Failed to attach */
1222 			if (error)
1223 				*error = err;
1224 			rte = NULL;
1225 		} else {
1226 			KASSERT((inp->inp_snd_tag != NULL) ,
1227 				("Setup rate has no snd_tag inp:%p rte:%p rate:%llu rs:%p",
1228 				 inp, rte, (unsigned long long)rte->rate, rs));
1229 #ifdef INET
1230 			counter_u64_add(rate_limit_new, 1);
1231 #endif
1232 		}
1233 	}
1234 	if (rte) {
1235 		/*
1236 		 * We use an atomic here for accounting so we don't have to
1237 		 * use locks when freeing.
1238 		 */
1239 		atomic_add_64(&rs->rs_flows_using, 1);
1240 	}
1241 	NET_EPOCH_EXIT(et);
1242 	return (rte);
1243 }
1244 
1245 static void
1246 tcp_rl_ifnet_link(void *arg __unused, struct ifnet *ifp, int link_state)
1247 {
1248 	int error;
1249 	struct tcp_rate_set *rs;
1250 	struct epoch_tracker et;
1251 
1252 	if (((ifp->if_capenable & IFCAP_TXRTLMT) == 0) ||
1253 	    (link_state != LINK_STATE_UP)) {
1254 		/*
1255 		 * We only care on an interface going up that is rate-limit
1256 		 * capable.
1257 		 */
1258 		return;
1259 	}
1260 	NET_EPOCH_ENTER(et);
1261 	mtx_lock(&rs_mtx);
1262 	rs = find_rs_for_ifp(ifp);
1263 	if (rs) {
1264 		/* We already have initialized this guy */
1265 		mtx_unlock(&rs_mtx);
1266 		NET_EPOCH_EXIT(et);
1267 		return;
1268 	}
1269 	mtx_unlock(&rs_mtx);
1270 	NET_EPOCH_EXIT(et);
1271 	rt_setup_new_rs(ifp, &error);
1272 }
1273 
1274 static void
1275 tcp_rl_ifnet_departure(void *arg __unused, struct ifnet *ifp)
1276 {
1277 	struct tcp_rate_set *rs;
1278 	struct epoch_tracker et;
1279 	int i;
1280 
1281 	NET_EPOCH_ENTER(et);
1282 	mtx_lock(&rs_mtx);
1283 	rs = find_rs_for_ifp(ifp);
1284 	if (rs) {
1285 		CK_LIST_REMOVE(rs, next);
1286 		rs_number_alive--;
1287 		rs->rs_flags |= RS_IS_DEAD;
1288 		for (i = 0; i < rs->rs_rate_cnt; i++) {
1289 			if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1290 				in_pcbdetach_tag(rs->rs_rlt[i].tag);
1291 				rs->rs_rlt[i].tag = NULL;
1292 			}
1293 			rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1294 		}
1295 		if (rs->rs_flows_using == 0)
1296 			rs_defer_destroy(rs);
1297 	}
1298 	mtx_unlock(&rs_mtx);
1299 	NET_EPOCH_EXIT(et);
1300 }
1301 
1302 static void
1303 tcp_rl_shutdown(void *arg __unused, int howto __unused)
1304 {
1305 	struct tcp_rate_set *rs, *nrs;
1306 	struct epoch_tracker et;
1307 	int i;
1308 
1309 	NET_EPOCH_ENTER(et);
1310 	mtx_lock(&rs_mtx);
1311 	CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) {
1312 		CK_LIST_REMOVE(rs, next);
1313 		rs_number_alive--;
1314 		rs->rs_flags |= RS_IS_DEAD;
1315 		for (i = 0; i < rs->rs_rate_cnt; i++) {
1316 			if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1317 				in_pcbdetach_tag(rs->rs_rlt[i].tag);
1318 				rs->rs_rlt[i].tag = NULL;
1319 			}
1320 			rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1321 		}
1322 		if (rs->rs_flows_using == 0)
1323 			rs_defer_destroy(rs);
1324 	}
1325 	mtx_unlock(&rs_mtx);
1326 	NET_EPOCH_EXIT(et);
1327 }
1328 
1329 const struct tcp_hwrate_limit_table *
1330 tcp_set_pacing_rate(struct tcpcb *tp, struct ifnet *ifp,
1331     uint64_t bytes_per_sec, int flags, int *error, uint64_t *lower_rate)
1332 {
1333 	struct inpcb *inp = tptoinpcb(tp);
1334 	const struct tcp_hwrate_limit_table *rte;
1335 #ifdef KERN_TLS
1336 	struct ktls_session *tls;
1337 #endif
1338 
1339 	INP_WLOCK_ASSERT(inp);
1340 
1341 	if (inp->inp_snd_tag == NULL) {
1342 		/*
1343 		 * We are setting up a rate for the first time.
1344 		 */
1345 		if ((ifp->if_capenable & IFCAP_TXRTLMT) == 0) {
1346 			/* Not supported by the egress */
1347 			if (error)
1348 				*error = ENODEV;
1349 			return (NULL);
1350 		}
1351 #ifdef KERN_TLS
1352 		tls = NULL;
1353 		if (tp->t_nic_ktls_xmit != 0) {
1354 			tls = tptosocket(tp)->so_snd.sb_tls_info;
1355 
1356 			if ((ifp->if_capenable & IFCAP_TXTLS_RTLMT) == 0 ||
1357 			    tls->mode != TCP_TLS_MODE_IFNET) {
1358 				if (error)
1359 					*error = ENODEV;
1360 				return (NULL);
1361 			}
1362 		}
1363 #endif
1364 		rte = rt_setup_rate(inp, ifp, bytes_per_sec, flags, error, lower_rate);
1365 		if (rte)
1366 			rl_increment_using(rte);
1367 #ifdef KERN_TLS
1368 		if (rte != NULL && tls != NULL && tls->snd_tag != NULL) {
1369 			/*
1370 			 * Fake a route change error to reset the TLS
1371 			 * send tag.  This will convert the existing
1372 			 * tag to a TLS ratelimit tag.
1373 			 */
1374 			MPASS(tls->snd_tag->sw->type == IF_SND_TAG_TYPE_TLS);
1375 			ktls_output_eagain(inp, tls);
1376 		}
1377 #endif
1378 	} else {
1379 		/*
1380 		 * We are modifying a rate, wrong interface?
1381 		 */
1382 		if (error)
1383 			*error = EINVAL;
1384 		rte = NULL;
1385 	}
1386 	if (rte != NULL) {
1387 		tp->t_pacing_rate = rte->rate;
1388 		*error = 0;
1389 	}
1390 	return (rte);
1391 }
1392 
1393 const struct tcp_hwrate_limit_table *
1394 tcp_chg_pacing_rate(const struct tcp_hwrate_limit_table *crte,
1395     struct tcpcb *tp, struct ifnet *ifp,
1396     uint64_t bytes_per_sec, int flags, int *error, uint64_t *lower_rate)
1397 {
1398 	struct inpcb *inp = tptoinpcb(tp);
1399 	const struct tcp_hwrate_limit_table *nrte;
1400 	const struct tcp_rate_set *rs;
1401 #ifdef KERN_TLS
1402 	struct ktls_session *tls = NULL;
1403 #endif
1404 	int err;
1405 
1406 	INP_WLOCK_ASSERT(inp);
1407 
1408 	if (crte == NULL) {
1409 		/* Wrong interface */
1410 		if (error)
1411 			*error = EINVAL;
1412 		return (NULL);
1413 	}
1414 
1415 #ifdef KERN_TLS
1416 	if (tp->t_nic_ktls_xmit) {
1417 		tls = tptosocket(tp)->so_snd.sb_tls_info;
1418 		if (tls->mode != TCP_TLS_MODE_IFNET)
1419 			tls = NULL;
1420 		else if (tls->snd_tag != NULL &&
1421 		    tls->snd_tag->sw->type != IF_SND_TAG_TYPE_TLS_RATE_LIMIT) {
1422 			if (!tls->reset_pending) {
1423 				/*
1424 				 * NIC probably doesn't support
1425 				 * ratelimit TLS tags if it didn't
1426 				 * allocate one when an existing rate
1427 				 * was present, so ignore.
1428 				 */
1429 				tcp_rel_pacing_rate(crte, tp);
1430 				if (error)
1431 					*error = EOPNOTSUPP;
1432 				return (NULL);
1433 			}
1434 
1435 			/*
1436 			 * The send tag is being converted, so set the
1437 			 * rate limit on the inpcb tag.  There is a
1438 			 * race that the new NIC send tag might use
1439 			 * the current rate instead of this one.
1440 			 */
1441 			tls = NULL;
1442 		}
1443 	}
1444 #endif
1445 	if (inp->inp_snd_tag == NULL) {
1446 		/* Wrong interface */
1447 		tcp_rel_pacing_rate(crte, tp);
1448 		if (error)
1449 			*error = EINVAL;
1450 		return (NULL);
1451 	}
1452 	rs = crte->ptbl;
1453 	if ((rs->rs_flags & RS_IS_DEAD) ||
1454 	    (crte->flags & HDWRPACE_IFPDEPARTED)) {
1455 		/* Release the rate, and try anew */
1456 
1457 		tcp_rel_pacing_rate(crte, tp);
1458 		nrte = tcp_set_pacing_rate(tp, ifp,
1459 		    bytes_per_sec, flags, error, lower_rate);
1460 		return (nrte);
1461 	}
1462 	nrte = tcp_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate);
1463 	if (nrte == crte) {
1464 		/* No change */
1465 		if (error)
1466 			*error = 0;
1467 		return (crte);
1468 	}
1469 	if (nrte == NULL) {
1470 		/* Release the old rate */
1471 		if (error)
1472 			*error = ENOENT;
1473 		tcp_rel_pacing_rate(crte, tp);
1474 		return (NULL);
1475 	}
1476 	rl_decrement_using(crte);
1477 	rl_increment_using(nrte);
1478 	/* Change rates to our new entry */
1479 #ifdef KERN_TLS
1480 	if (tls != NULL)
1481 		err = ktls_modify_txrtlmt(tls, nrte->rate);
1482 	else
1483 #endif
1484 		err = in_pcbmodify_txrtlmt(inp, nrte->rate);
1485 	if (err) {
1486 		struct tcp_rate_set *lrs;
1487 		uint64_t pre;
1488 
1489 		rl_decrement_using(nrte);
1490 		lrs = __DECONST(struct tcp_rate_set *, rs);
1491 		pre = atomic_fetchadd_64(&lrs->rs_flows_using, -1);
1492 		/* Do we still have a snd-tag attached? */
1493 		if (inp->inp_snd_tag)
1494 			in_pcbdetach_txrtlmt(inp);
1495 
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 (lrs->rs_flags & RS_IS_DEAD)
1505 				rs_defer_destroy(lrs);
1506 			mtx_unlock(&rs_mtx);
1507 			NET_EPOCH_EXIT(et);
1508 		}
1509 		if (error)
1510 			*error = err;
1511 		return (NULL);
1512 	} else {
1513 #ifdef INET
1514 		counter_u64_add(rate_limit_chg, 1);
1515 #endif
1516 	}
1517 	if (error)
1518 		*error = 0;
1519 	tp->t_pacing_rate = nrte->rate;
1520 	return (nrte);
1521 }
1522 
1523 void
1524 tcp_rel_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp)
1525 {
1526 	struct inpcb *inp = tptoinpcb(tp);
1527 	const struct tcp_rate_set *crs;
1528 	struct tcp_rate_set *rs;
1529 	uint64_t pre;
1530 
1531 	INP_WLOCK_ASSERT(inp);
1532 
1533 	tp->t_pacing_rate = -1;
1534 	crs = crte->ptbl;
1535 	/*
1536 	 * Now we must break the const
1537 	 * in order to release our refcount.
1538 	 */
1539 	rs = __DECONST(struct tcp_rate_set *, crs);
1540 	rl_decrement_using(crte);
1541 	pre = atomic_fetchadd_64(&rs->rs_flows_using, -1);
1542 	if (pre == 1) {
1543 		struct epoch_tracker et;
1544 
1545 		NET_EPOCH_ENTER(et);
1546 		mtx_lock(&rs_mtx);
1547 		/*
1548 		 * Is it dead?
1549 		 */
1550 		if (rs->rs_flags & RS_IS_DEAD)
1551 			rs_defer_destroy(rs);
1552 		mtx_unlock(&rs_mtx);
1553 		NET_EPOCH_EXIT(et);
1554 	}
1555 
1556 	/*
1557 	 * XXX: If this connection is using ifnet TLS, should we
1558 	 * switch it to using an unlimited rate, or perhaps use
1559 	 * ktls_output_eagain() to reset the send tag to a plain
1560 	 * TLS tag?
1561 	 */
1562 	in_pcbdetach_txrtlmt(inp);
1563 }
1564 
1565 #define ONE_POINT_TWO_MEG 150000 /* 1.2 megabits in bytes */
1566 #define ONE_HUNDRED_MBPS 12500000	/* 100Mbps in bytes per second */
1567 #define FIVE_HUNDRED_MBPS 62500000	/* 500Mbps in bytes per second */
1568 #define MAX_MSS_SENT 43	/* 43 mss = 43 x 1500 = 64,500 bytes */
1569 
1570 static void
1571 tcp_log_pacing_size(struct tcpcb *tp, uint64_t bw, uint32_t segsiz, uint32_t new_tso,
1572 		    uint64_t hw_rate, uint32_t time_between, uint32_t calc_time_between,
1573 		    uint32_t segs, uint32_t res_div, uint16_t mult, uint8_t mod)
1574 {
1575 	if (tcp_bblogging_on(tp)) {
1576 		union tcp_log_stackspecific log;
1577 		struct timeval tv;
1578 
1579 		memset(&log, 0, sizeof(log));
1580 		log.u_bbr.flex1 = segsiz;
1581 		log.u_bbr.flex2 = new_tso;
1582 		log.u_bbr.flex3 = time_between;
1583 		log.u_bbr.flex4 = calc_time_between;
1584 		log.u_bbr.flex5 = segs;
1585 		log.u_bbr.flex6 = res_div;
1586 		log.u_bbr.flex7 = mult;
1587 		log.u_bbr.flex8 = mod;
1588 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1589 		log.u_bbr.cur_del_rate = bw;
1590 		log.u_bbr.delRate = hw_rate;
1591 		TCP_LOG_EVENTP(tp, NULL,
1592 		    &tptosocket(tp)->so_rcv,
1593 		    &tptosocket(tp)->so_snd,
1594 		    TCP_HDWR_PACE_SIZE, 0,
1595 		    0, &log, false, &tv);
1596 	}
1597 }
1598 
1599 uint32_t
1600 tcp_get_pacing_burst_size_w_divisor(struct tcpcb *tp, uint64_t bw, uint32_t segsiz, int can_use_1mss,
1601    const struct tcp_hwrate_limit_table *te, int *err, int divisor)
1602 {
1603 	/*
1604 	 * We use the google formula to calculate the
1605 	 * TSO size. I.E.
1606 	 * bw < 24Meg
1607 	 *   tso = 2mss
1608 	 * else
1609 	 *   tso = min(bw/(div=1000), 64k)
1610 	 *
1611 	 * Note for these calculations we ignore the
1612 	 * packet overhead (enet hdr, ip hdr and tcp hdr).
1613 	 * We only get the google formula when we have
1614 	 * divisor = 1000, which is the default for now.
1615 	 */
1616 	uint64_t lentim, res, bytes;
1617 	uint32_t new_tso, min_tso_segs;
1618 
1619 	/* It can't be zero */
1620 	if ((divisor == 0) ||
1621 	    (divisor < RL_MIN_DIVISOR)) {
1622 		if (mss_divisor)
1623 			bytes = bw / mss_divisor;
1624 		else
1625 			bytes = bw / 1000;
1626 	} else
1627 		bytes = bw / divisor;
1628 	/* We can't ever send more than 65k in a TSO */
1629 	if (bytes > 0xffff) {
1630 		bytes = 0xffff;
1631 	}
1632 	/* Round up */
1633 	new_tso = (bytes + segsiz - 1) / segsiz;
1634 	/* Are we enforcing even boundaries? */
1635 	if (even_num_segs && (new_tso & 1) && (new_tso > even_threshold))
1636 		new_tso++;
1637 	if (can_use_1mss)
1638 		min_tso_segs = 1;
1639 	else
1640 		min_tso_segs = 2;
1641 	if (rs_floor_mss && (new_tso < rs_floor_mss))
1642 		new_tso = rs_floor_mss;
1643 	else if (new_tso < min_tso_segs)
1644 		new_tso = min_tso_segs;
1645 	if (new_tso > MAX_MSS_SENT)
1646 		new_tso = MAX_MSS_SENT;
1647 	new_tso *= segsiz;
1648  	tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1649 			    0, 0, 0, 0, 0, 0, 1);
1650 	/*
1651 	 * If we are not doing hardware pacing
1652 	 * then we are done.
1653 	 */
1654 	if (te == NULL) {
1655 		if (err)
1656 			*err = 0;
1657 		return(new_tso);
1658 	}
1659 	/*
1660 	 * For hardware pacing we look at the
1661 	 * rate you are sending at and compare
1662 	 * that to the rate you have in hardware.
1663 	 *
1664 	 * If the hardware rate is slower than your
1665 	 * software rate then you are in error and
1666 	 * we will build a queue in our hardware whic
1667 	 * is probably not desired, in such a case
1668 	 * just return the non-hardware TSO size.
1669 	 *
1670 	 * If the rate in hardware is faster (which
1671 	 * it should be) then look at how long it
1672 	 * takes to send one ethernet segment size at
1673 	 * your b/w and compare that to the time it
1674 	 * takes to send at the rate you had selected.
1675 	 *
1676 	 * If your time is greater (which we hope it is)
1677 	 * we get the delta between the two, and then
1678 	 * divide that into your pacing time. This tells
1679 	 * us how many MSS you can send down at once (rounded up).
1680 	 *
1681 	 * Note we also double this value if the b/w is over
1682 	 * 100Mbps. If its over 500meg we just set you to the
1683 	 * max (43 segments).
1684 	 */
1685 	if (te->rate > FIVE_HUNDRED_MBPS)
1686 		goto max;
1687 	if (te->rate == bw) {
1688 		/* We are pacing at exactly the hdwr rate */
1689 max:
1690 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1691 				    te->rate, te->time_between, (uint32_t)0,
1692 				    (segsiz * MAX_MSS_SENT), 0, 0, 3);
1693 		return (segsiz * MAX_MSS_SENT);
1694 	}
1695 	lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
1696 	res = lentim / bw;
1697 	if (res > te->time_between) {
1698 		uint32_t delta, segs, res_div;
1699 
1700 		res_div = ((res * num_of_waits_allowed) + wait_time_floor);
1701 		delta = res - te->time_between;
1702 		segs = (res_div + delta - 1)/delta;
1703 		if (segs < min_tso_segs)
1704 			segs = min_tso_segs;
1705 		if (segs < rs_hw_floor_mss)
1706 			segs = rs_hw_floor_mss;
1707 		if (segs > MAX_MSS_SENT)
1708 			segs = MAX_MSS_SENT;
1709 		segs *= segsiz;
1710 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1711 				    te->rate, te->time_between, (uint32_t)res,
1712 				    segs, res_div, 1, 3);
1713 		if (err)
1714 			*err = 0;
1715 		if (segs < new_tso) {
1716 			/* unexpected ? */
1717 			return(new_tso);
1718 		} else {
1719 			return (segs);
1720 		}
1721 	} else {
1722 		/*
1723 		 * Your time is smaller which means
1724 		 * we will grow a queue on our
1725 		 * hardware. Send back the non-hardware
1726 		 * rate.
1727 		 */
1728 		tcp_log_pacing_size(tp, bw, segsiz, new_tso,
1729 				    te->rate, te->time_between, (uint32_t)res,
1730 				    0, 0, 0, 4);
1731 		if (err)
1732 			*err = -1;
1733 		return (new_tso);
1734 	}
1735 }
1736 
1737 uint64_t
1738 tcp_hw_highest_rate_ifp(struct ifnet *ifp, struct inpcb *inp)
1739 {
1740 	struct epoch_tracker et;
1741 	struct tcp_rate_set *rs;
1742 	uint64_t rate_ret;
1743 
1744 	NET_EPOCH_ENTER(et);
1745 use_next_interface:
1746 	rs = find_rs_for_ifp(ifp);
1747 	if (rs == NULL) {
1748 		/* This interface does not do ratelimiting */
1749 		rate_ret = 0;
1750 	} else if (rs->rs_flags & RS_IS_DEFF) {
1751 		/* We need to find the real interface */
1752 		struct ifnet *tifp;
1753 
1754 		tifp = rt_find_real_interface(ifp, inp, NULL);
1755 		if (tifp == NULL) {
1756 			NET_EPOCH_EXIT(et);
1757 			return (0);
1758 		}
1759 		ifp = tifp;
1760 		goto use_next_interface;
1761 	} else {
1762 		/* Lets return the highest rate this guy has */
1763 		rate_ret = rs->rs_rlt[rs->rs_highest_valid].rate;
1764 	}
1765 	NET_EPOCH_EXIT(et);
1766 	return(rate_ret);
1767 }
1768 
1769 static eventhandler_tag rl_ifnet_departs;
1770 static eventhandler_tag rl_ifnet_arrives;
1771 static eventhandler_tag rl_shutdown_start;
1772 
1773 static void
1774 tcp_rs_init(void *st __unused)
1775 {
1776 	CK_LIST_INIT(&int_rs);
1777 	rs_number_alive = 0;
1778 	rs_number_dead = 0;
1779 	mtx_init(&rs_mtx, "tcp_rs_mtx", "rsmtx", MTX_DEF);
1780 	rl_ifnet_departs = EVENTHANDLER_REGISTER(ifnet_departure_event,
1781 	    tcp_rl_ifnet_departure,
1782 	    NULL, EVENTHANDLER_PRI_ANY);
1783 	rl_ifnet_arrives = EVENTHANDLER_REGISTER(ifnet_link_event,
1784 	    tcp_rl_ifnet_link,
1785 	    NULL, EVENTHANDLER_PRI_ANY);
1786 	rl_shutdown_start = EVENTHANDLER_REGISTER(shutdown_pre_sync,
1787 	    tcp_rl_shutdown, NULL,
1788 	    SHUTDOWN_PRI_FIRST);
1789 	printf("TCP_ratelimit: Is now initialized\n");
1790 }
1791 
1792 SYSINIT(tcp_rl_init, SI_SUB_SMP + 1, SI_ORDER_ANY, tcp_rs_init, NULL);
1793 #endif
1794