xref: /freebsd/sys/netinet/tcp_ratelimit.c (revision da759cfa320d5076b075d15ff3f00ab3ba5634fd)
1 /*-
2  *
3  * SPDX-License-Identifier: BSD-3-Clause
4  *
5  * Copyright (c) 2018-2019
6  *	Netflix Inc.
7  *      All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  */
31 /**
32  * Author: Randall Stewart <rrs@netflix.com>
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
41 #include "opt_ratelimit.h"
42 #include <sys/param.h>
43 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/mbuf.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/eventhandler.h>
50 #include <sys/mutex.h>
51 #include <sys/ck.h>
52 #include <net/if.h>
53 #include <net/if_var.h>
54 #include <netinet/in.h>
55 #include <netinet/in_pcb.h>
56 #define TCPSTATES		/* for logging */
57 #include <netinet/tcp_var.h>
58 #ifdef INET6
59 #include <netinet6/tcp6_var.h>
60 #endif
61 #include <netinet/tcp_ratelimit.h>
62 #ifndef USECS_IN_SECOND
63 #define USECS_IN_SECOND 1000000
64 #endif
65 /*
66  * For the purposes of each send, what is the size
67  * of an ethernet frame.
68  */
69 MALLOC_DEFINE(M_TCPPACE, "tcp_hwpace", "TCP Hardware pacing memory");
70 #ifdef RATELIMIT
71 
72 /*
73  * The following preferred table will seem weird to
74  * the casual viewer. Why do we not have any rates below
75  * 1Mbps? Why do we have a rate at 1.44Mbps called common?
76  * Why do the rates cluster in the 1-100Mbps range more
77  * than others? Why does the table jump around at the beginnign
78  * and then be more consistently raising?
79  *
80  * Let me try to answer those questions. A lot of
81  * this is dependant on the hardware. We have three basic
82  * supporters of rate limiting
83  *
84  * Chelsio - Supporting 16 configurable rates.
85  * Mlx  - c4 supporting 13 fixed rates.
86  * Mlx  - c5 & c6 supporting 127 configurable rates.
87  *
88  * The c4 is why we have a common rate that is available
89  * in all rate tables. This is a selected rate from the
90  * c4 table and we assure its available in all ratelimit
91  * tables. This way the tcp_ratelimit code has an assured
92  * rate it should always be able to get. This answers a
93  * couple of the questions above.
94  *
95  * So what about the rest, well the table is built to
96  * try to get the most out of a joint hardware/software
97  * pacing system.  The software pacer will always pick
98  * a rate higher than the b/w that it is estimating
99  *
100  * on the path. This is done for two reasons.
101  * a) So we can discover more b/w
102  * and
103  * b) So we can send a block of MSS's down and then
104  *    have the software timer go off after the previous
105  *    send is completely out of the hardware.
106  *
107  * But when we do <b> we don't want to have the delay
108  * between the last packet sent by the hardware be
109  * excessively long (to reach our desired rate).
110  *
111  * So let me give an example for clarity.
112  *
113  * Lets assume that the tcp stack sees that 29,110,000 bps is
114  * what the bw of the path is. The stack would select the
115  * rate 31Mbps. 31Mbps means that each send that is done
116  * by the hardware will cause a 390 micro-second gap between
117  * the packets sent at that rate. For 29,110,000 bps we
118  * would need 416 micro-seconds gap between each send.
119  *
120  * Note that are calculating a complete time for pacing
121  * which includes the ethernet, IP and TCP overhead. So
122  * a full 1514 bytes is used for the above calculations.
123  * My testing has shown that both cards are also using this
124  * as their basis i.e. full payload size of the ethernet frame.
125  * The TCP stack caller needs to be aware of this and make the
126  * appropriate overhead calculations be included in its choices.
127  *
128  * Now, continuing our example, we pick a MSS size based on the
129  * delta between the two rates (416 - 390) divided into the rate
130  * we really wish to send at rounded up.  That results in a MSS
131  * send of 17 mss's at once. The hardware then will
132  * run out of data in a single 17MSS send in 6,630 micro-seconds.
133  *
134  * On the other hand the software pacer will send more data
135  * in 7,072 micro-seconds. This means that we will refill
136  * the hardware 52 microseconds after it would have sent
137  * next if it had not ran out of data. This is a win since we are
138  * only sending every 7ms or so and yet all the packets are spaced on
139  * the wire with 94% of what they should be and only
140  * the last packet is delayed extra to make up for the
141  * difference.
142  *
143  * Note that the above formula has two important caveat.
144  * If we are above (b/w wise) over 100Mbps we double the result
145  * of the MSS calculation. The second caveat is if we are 500Mbps
146  * or more we just send the maximum MSS at once i.e. 45MSS. At
147  * the higher b/w's even the cards have limits to what times (timer granularity)
148  * they can insert between packets and start to send more than one
149  * packet at a time on the wire.
150  *
151  */
152 #define COMMON_RATE 180500
153 const uint64_t desired_rates[] = {
154 	122500,			/* 1Mbps  - rate 1 */
155 	180500,			/* 1.44Mpbs - rate 2  common rate */
156 	375000,			/* 3Mbps    - rate 3 */
157 	625000,			/* 5Mbps    - rate 4 */
158 	875000,			/* 7Mbps    - rate 5 */
159 	1125000,		/* 9Mbps    - rate 6 */
160 	1375000,		/* 11Mbps   - rate 7 */
161 	1625000,	       	/* 13Mbps   - rate 8 */
162 	2625000,		/* 21Mbps   - rate 9 */
163 	3875000,		/* 31Mbps   - rate 10 */
164 	5125000,		/* 41Meg    - rate 11 */
165 	12500000,		/* 100Mbps  - rate 12 */
166 	25000000,		/* 200Mbps  - rate 13 */
167 	50000000,		/* 400Mbps  - rate 14 */
168 	63750000,		/* 51Mbps   - rate 15 */
169 	100000000,		/* 800Mbps  - rate 16 */
170 	1875000,		/* 15Mbps   - rate 17 */
171 	2125000,		/* 17Mbps   - rate 18 */
172 	2375000,		/* 19Mbps   - rate 19 */
173 	2875000,		/* 23Mbps   - rate 20 */
174 	3125000,		/* 25Mbps   - rate 21 */
175 	3375000,		/* 27Mbps   - rate 22 */
176 	3625000,		/* 29Mbps   - rate 23 */
177 	4125000,		/* 33Mbps   - rate 24 */
178 	4375000,		/* 35Mbps   - rate 25 */
179 	4625000,		/* 37Mbps   - rate 26 */
180 	4875000,		/* 39Mbps   - rate 27 */
181 	5375000,		/* 43Mbps   - rate 28 */
182 	5625000,		/* 45Mbps   - rate 29 */
183 	5875000,		/* 47Mbps   - rate 30 */
184 	6125000,		/* 49Mbps   - rate 31 */
185 	6625000,		/* 53Mbps   - rate 32 */
186 	6875000,		/* 55Mbps   - rate 33 */
187 	7125000,		/* 57Mbps   - rate 34 */
188 	7375000,		/* 59Mbps   - rate 35 */
189 	7625000,		/* 61Mbps   - rate 36 */
190 	7875000,		/* 63Mbps   - rate 37 */
191 	8125000,		/* 65Mbps   - rate 38 */
192 	8375000,		/* 67Mbps   - rate 39 */
193 	8625000,		/* 69Mbps   - rate 40 */
194 	8875000,		/* 71Mbps   - rate 41 */
195 	9125000,		/* 73Mbps   - rate 42 */
196 	9375000,		/* 75Mbps   - rate 43 */
197 	9625000,		/* 77Mbps   - rate 44 */
198 	9875000,		/* 79Mbps   - rate 45 */
199 	10125000,		/* 81Mbps   - rate 46 */
200 	10375000,		/* 83Mbps   - rate 47 */
201 	10625000,		/* 85Mbps   - rate 48 */
202 	10875000,		/* 87Mbps   - rate 49 */
203 	11125000,		/* 89Mbps   - rate 50 */
204 	11375000,		/* 91Mbps   - rate 51 */
205 	11625000,		/* 93Mbps   - rate 52 */
206 	11875000,		/* 95Mbps   - rate 53 */
207 	13125000,		/* 105Mbps  - rate 54 */
208 	13750000,		/* 110Mbps  - rate 55 */
209 	14375000,		/* 115Mbps  - rate 56 */
210 	15000000,		/* 120Mbps  - rate 57 */
211 	15625000,		/* 125Mbps  - rate 58 */
212 	16250000,		/* 130Mbps  - rate 59 */
213 	16875000,		/* 135Mbps  - rate 60 */
214 	17500000,		/* 140Mbps  - rate 61 */
215 	18125000,		/* 145Mbps  - rate 62 */
216 	18750000,		/* 150Mbps  - rate 64 */
217 	20000000,		/* 160Mbps  - rate 65 */
218 	21250000,		/* 170Mbps  - rate 66 */
219 	22500000,		/* 180Mbps  - rate 67 */
220 	23750000,		/* 190Mbps  - rate 68 */
221 	26250000,		/* 210Mbps  - rate 69 */
222 	27500000,		/* 220Mbps  - rate 70 */
223 	28750000,		/* 230Mbps  - rate 71 */
224 	30000000,	       	/* 240Mbps  - rate 72 */
225 	31250000,		/* 250Mbps  - rate 73 */
226 	34375000,		/* 275Mbps  - rate 74 */
227 	37500000,		/* 300Mbps  - rate 75 */
228 	40625000,		/* 325Mbps  - rate 76 */
229 	43750000,		/* 350Mbps  - rate 77 */
230 	46875000,		/* 375Mbps  - rate 78 */
231 	53125000,		/* 425Mbps  - rate 79 */
232 	56250000,		/* 450Mbps  - rate 80 */
233 	59375000,		/* 475Mbps  - rate 81 */
234 	62500000,		/* 500Mbps  - rate 82 */
235 	68750000,		/* 550Mbps  - rate 83 */
236 	75000000,		/* 600Mbps  - rate 84 */
237 	81250000,		/* 650Mbps  - rate 85 */
238 	87500000,		/* 700Mbps  - rate 86 */
239 	93750000,		/* 750Mbps  - rate 87 */
240 	106250000,		/* 850Mbps  - rate 88 */
241 	112500000,		/* 900Mbps  - rate 89 */
242 	125000000,		/* 1Gbps    - rate 90 */
243 	156250000,		/* 1.25Gps  - rate 91 */
244 	187500000,		/* 1.5Gps   - rate 92 */
245 	218750000,		/* 1.75Gps  - rate 93 */
246 	250000000,		/* 2Gbps    - rate 94 */
247 	281250000,		/* 2.25Gps  - rate 95 */
248 	312500000,		/* 2.5Gbps  - rate 96 */
249 	343750000,		/* 2.75Gbps - rate 97 */
250 	375000000,		/* 3Gbps    - rate 98 */
251 	500000000,		/* 4Gbps    - rate 99 */
252 	625000000,		/* 5Gbps    - rate 100 */
253 	750000000,		/* 6Gbps    - rate 101 */
254 	875000000,		/* 7Gbps    - rate 102 */
255 	1000000000,		/* 8Gbps    - rate 103 */
256 	1125000000,		/* 9Gbps    - rate 104 */
257 	1250000000,		/* 10Gbps   - rate 105 */
258 	1875000000,		/* 15Gbps   - rate 106 */
259 	2500000000		/* 20Gbps   - rate 107 */
260 };
261 
262 #define MAX_HDWR_RATES (sizeof(desired_rates)/sizeof(uint64_t))
263 #define RS_ORDERED_COUNT 16	/*
264 				 * Number that are in order
265 				 * at the beginning of the table,
266 				 * over this a sort is required.
267 				 */
268 #define RS_NEXT_ORDER_GROUP 16	/*
269 				 * The point in our table where
270 				 * we come fill in a second ordered
271 				 * group (index wise means -1).
272 				 */
273 #define ALL_HARDWARE_RATES 1004 /*
274 				 * 1Meg - 1Gig in 1 Meg steps
275 				 * plus 100, 200k  and 500k and
276 				 * 10Gig
277 				 */
278 
279 #define RS_ONE_MEGABIT_PERSEC 1000000
280 #define RS_ONE_GIGABIT_PERSEC 1000000000
281 #define RS_TEN_GIGABIT_PERSEC 10000000000
282 
283 static struct head_tcp_rate_set int_rs;
284 static struct mtx rs_mtx;
285 uint32_t rs_number_alive;
286 uint32_t rs_number_dead;
287 
288 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, rl, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
289     "TCP Ratelimit stats");
290 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, alive, CTLFLAG_RW,
291     &rs_number_alive, 0,
292     "Number of interfaces initialized for ratelimiting");
293 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, dead, CTLFLAG_RW,
294     &rs_number_dead, 0,
295     "Number of interfaces departing from ratelimiting");
296 
297 static void
298 rl_add_syctl_entries(struct sysctl_oid *rl_sysctl_root, struct tcp_rate_set *rs)
299 {
300 	/*
301 	 * Add sysctl entries for thus interface.
302 	 */
303 	if (rs->rs_flags & RS_INTF_NO_SUP) {
304 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
305 		   SYSCTL_CHILDREN(rl_sysctl_root),
306 		   OID_AUTO, "disable", CTLFLAG_RD,
307 		   &rs->rs_disable, 0,
308 		   "Disable this interface from new hdwr limiting?");
309 	} else {
310 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
311 		   SYSCTL_CHILDREN(rl_sysctl_root),
312 		   OID_AUTO, "disable", CTLFLAG_RW,
313 		   &rs->rs_disable, 0,
314 		   "Disable this interface from new hdwr limiting?");
315 	}
316 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
317 	    SYSCTL_CHILDREN(rl_sysctl_root),
318 	    OID_AUTO, "minseg", CTLFLAG_RW,
319 	    &rs->rs_min_seg, 0,
320 	    "What is the minimum we need to send on this interface?");
321 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
322 	    SYSCTL_CHILDREN(rl_sysctl_root),
323 	    OID_AUTO, "flow_limit", CTLFLAG_RW,
324 	    &rs->rs_flow_limit, 0,
325 	    "What is the limit for number of flows (0=unlimited)?");
326 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
327 	    SYSCTL_CHILDREN(rl_sysctl_root),
328 	    OID_AUTO, "highest", CTLFLAG_RD,
329 	    &rs->rs_highest_valid, 0,
330 	    "Highest valid rate");
331 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
332 	    SYSCTL_CHILDREN(rl_sysctl_root),
333 	    OID_AUTO, "lowest", CTLFLAG_RD,
334 	    &rs->rs_lowest_valid, 0,
335 	    "Lowest valid rate");
336 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
337 	    SYSCTL_CHILDREN(rl_sysctl_root),
338 	    OID_AUTO, "flags", CTLFLAG_RD,
339 	    &rs->rs_flags, 0,
340 	    "What lags are on the entry?");
341 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
342 	    SYSCTL_CHILDREN(rl_sysctl_root),
343 	    OID_AUTO, "numrates", CTLFLAG_RD,
344 	    &rs->rs_rate_cnt, 0,
345 	    "How many rates re there?");
346 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
347 	    SYSCTL_CHILDREN(rl_sysctl_root),
348 	    OID_AUTO, "flows_using", CTLFLAG_RD,
349 	    &rs->rs_flows_using, 0,
350 	    "How many flows are using this interface now?");
351 #ifdef DETAILED_RATELIMIT_SYSCTL
352 	if (rs->rs_rlt && rs->rs_rate_cnt > 0) {
353 		/*  Lets display the rates */
354 		int i;
355 		struct sysctl_oid *rl_rates;
356 		struct sysctl_oid *rl_rate_num;
357 		char rate_num[16];
358 		rl_rates = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
359 					    SYSCTL_CHILDREN(rl_sysctl_root),
360 					    OID_AUTO,
361 					    "rate",
362 					    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
363 					    "Ratelist");
364 		for( i = 0; i < rs->rs_rate_cnt; i++) {
365 			sprintf(rate_num, "%d", i);
366 			rl_rate_num = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
367 					    SYSCTL_CHILDREN(rl_rates),
368 					    OID_AUTO,
369 					    rate_num,
370 					    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
371 					    "Individual Rate");
372 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
373 				       SYSCTL_CHILDREN(rl_rate_num),
374 				       OID_AUTO, "flags", CTLFLAG_RD,
375 				       &rs->rs_rlt[i].flags, 0,
376 				       "Flags on this rate");
377 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
378 				       SYSCTL_CHILDREN(rl_rate_num),
379 				       OID_AUTO, "pacetime", CTLFLAG_RD,
380 				       &rs->rs_rlt[i].time_between, 0,
381 				       "Time hardware inserts between 1500 byte sends");
382 			SYSCTL_ADD_U64(&rs->sysctl_ctx,
383 				       SYSCTL_CHILDREN(rl_rate_num),
384 				       OID_AUTO, "rate", CTLFLAG_RD,
385 				       &rs->rs_rlt[i].rate, 0,
386 				       "Rate in bytes per second");
387 		}
388 	}
389 #endif
390 }
391 
392 static void
393 rs_destroy(epoch_context_t ctx)
394 {
395 	struct tcp_rate_set *rs;
396 	bool do_free_rs;
397 
398 	rs = __containerof(ctx, struct tcp_rate_set, rs_epoch_ctx);
399 
400 	mtx_lock(&rs_mtx);
401 	rs->rs_flags &= ~RS_FUNERAL_SCHD;
402 	/*
403 	 * In theory its possible (but unlikely)
404 	 * that while the delete was occuring
405 	 * and we were applying the DEAD flag
406 	 * someone slipped in and found the
407 	 * interface in a lookup. While we
408 	 * decided rs_flows_using were 0 and
409 	 * scheduling the epoch_call, the other
410 	 * thread incremented rs_flow_using. This
411 	 * is because users have a pointer and
412 	 * we only use the rs_flows_using in an
413 	 * atomic fashion, i.e. the other entities
414 	 * are not protected. To assure this did
415 	 * not occur, we check rs_flows_using here
416 	 * before deleting.
417 	 */
418 	do_free_rs = (rs->rs_flows_using == 0);
419 	rs_number_dead--;
420 	mtx_unlock(&rs_mtx);
421 
422 	if (do_free_rs) {
423 		sysctl_ctx_free(&rs->sysctl_ctx);
424 		free(rs->rs_rlt, M_TCPPACE);
425 		free(rs, M_TCPPACE);
426 	}
427 }
428 
429 static void
430 rs_defer_destroy(struct tcp_rate_set *rs)
431 {
432 
433 	mtx_assert(&rs_mtx, MA_OWNED);
434 
435 	/* Check if already pending. */
436 	if (rs->rs_flags & RS_FUNERAL_SCHD)
437 		return;
438 
439 	rs_number_dead++;
440 
441 	/* Set flag to only defer once. */
442 	rs->rs_flags |= RS_FUNERAL_SCHD;
443 	NET_EPOCH_CALL(rs_destroy, &rs->rs_epoch_ctx);
444 }
445 
446 #ifdef INET
447 extern counter_u64_t rate_limit_set_ok;
448 extern counter_u64_t rate_limit_active;
449 extern counter_u64_t rate_limit_alloc_fail;
450 #endif
451 
452 static int
453 rl_attach_txrtlmt(struct ifnet *ifp,
454     uint32_t flowtype,
455     int flowid,
456     uint64_t cfg_rate,
457     struct m_snd_tag **tag)
458 {
459 	int error;
460 	union if_snd_tag_alloc_params params = {
461 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
462 		.rate_limit.hdr.flowid = flowid,
463 		.rate_limit.hdr.flowtype = flowtype,
464 		.rate_limit.max_rate = cfg_rate,
465 		.rate_limit.flags = M_NOWAIT,
466 	};
467 
468 	if (ifp->if_snd_tag_alloc == NULL) {
469 		error = EOPNOTSUPP;
470 	} else {
471 		error = ifp->if_snd_tag_alloc(ifp, &params, tag);
472 #ifdef INET
473 		if (error == 0) {
474 			if_ref((*tag)->ifp);
475 			counter_u64_add(rate_limit_set_ok, 1);
476 			counter_u64_add(rate_limit_active, 1);
477 		} else
478 			counter_u64_add(rate_limit_alloc_fail, 1);
479 #endif
480 	}
481 	return (error);
482 }
483 
484 static void
485 populate_canned_table(struct tcp_rate_set *rs, const uint64_t *rate_table_act)
486 {
487 	/*
488 	 * The internal table is "special", it
489 	 * is two seperate ordered tables that
490 	 * must be merged. We get here when the
491 	 * adapter specifies a number of rates that
492 	 * covers both ranges in the table in some
493 	 * form.
494 	 */
495 	int i, at_low, at_high;
496 	uint8_t low_disabled = 0, high_disabled = 0;
497 
498 	for(i = 0, at_low = 0, at_high = RS_NEXT_ORDER_GROUP; i < rs->rs_rate_cnt; i++) {
499 		rs->rs_rlt[i].flags = 0;
500 		rs->rs_rlt[i].time_between = 0;
501 		if ((low_disabled == 0) &&
502 		    (high_disabled ||
503 		     (rate_table_act[at_low] < rate_table_act[at_high]))) {
504 			rs->rs_rlt[i].rate = rate_table_act[at_low];
505 			at_low++;
506 			if (at_low == RS_NEXT_ORDER_GROUP)
507 				low_disabled = 1;
508 		} else if (high_disabled == 0) {
509 			rs->rs_rlt[i].rate = rate_table_act[at_high];
510 			at_high++;
511 			if (at_high == MAX_HDWR_RATES)
512 				high_disabled = 1;
513 		}
514 	}
515 }
516 
517 static struct tcp_rate_set *
518 rt_setup_new_rs(struct ifnet *ifp, int *error)
519 {
520 	struct tcp_rate_set *rs;
521 	const uint64_t *rate_table_act;
522 	uint64_t lentim, res;
523 	size_t sz;
524 	uint32_t hash_type;
525 	int i;
526 	struct if_ratelimit_query_results rl;
527 	struct sysctl_oid *rl_sysctl_root;
528 	/*
529 	 * We expect to enter with the
530 	 * mutex locked.
531 	 */
532 
533 	if (ifp->if_ratelimit_query == NULL) {
534 		/*
535 		 * We can do nothing if we cannot
536 		 * get a query back from the driver.
537 		 */
538 		printf("Warning:No query functions for %s:%d-- failed\n",
539 		       ifp->if_dname, ifp->if_dunit);
540 		return (NULL);
541 	}
542 	rs = malloc(sizeof(struct tcp_rate_set), M_TCPPACE, M_NOWAIT | M_ZERO);
543 	if (rs == NULL) {
544 		if (error)
545 			*error = ENOMEM;
546 		printf("Warning:No memory for malloc of tcp_rate_set\n");
547 		return (NULL);
548 	}
549 	memset(&rl, 0, sizeof(rl));
550 	rl.flags = RT_NOSUPPORT;
551 	ifp->if_ratelimit_query(ifp, &rl);
552 	if (rl.flags & RT_IS_UNUSABLE) {
553 		/*
554 		 * The interface does not really support
555 		 * the rate-limiting.
556 		 */
557 		memset(rs, 0, sizeof(struct tcp_rate_set));
558 		rs->rs_ifp = ifp;
559 		rs->rs_if_dunit = ifp->if_dunit;
560 		rs->rs_flags = RS_INTF_NO_SUP;
561 		rs->rs_disable = 1;
562 		rs_number_alive++;
563 		sysctl_ctx_init(&rs->sysctl_ctx);
564 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
565 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
566 		    OID_AUTO,
567 		    rs->rs_ifp->if_xname,
568 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
569 		    "");
570 		rl_add_syctl_entries(rl_sysctl_root, rs);
571 		mtx_lock(&rs_mtx);
572 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
573 		mtx_unlock(&rs_mtx);
574 		return (rs);
575 	} else if ((rl.flags & RT_IS_INDIRECT) == RT_IS_INDIRECT) {
576 		memset(rs, 0, sizeof(struct tcp_rate_set));
577 		rs->rs_ifp = ifp;
578 		rs->rs_if_dunit = ifp->if_dunit;
579 		rs->rs_flags = RS_IS_DEFF;
580 		rs_number_alive++;
581 		sysctl_ctx_init(&rs->sysctl_ctx);
582 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
583 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
584 		    OID_AUTO,
585 		    rs->rs_ifp->if_xname,
586 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
587 		    "");
588 		rl_add_syctl_entries(rl_sysctl_root, rs);
589 		mtx_lock(&rs_mtx);
590 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
591 		mtx_unlock(&rs_mtx);
592 		return (rs);
593 	} else if ((rl.flags & RT_IS_FIXED_TABLE) == RT_IS_FIXED_TABLE) {
594 		/* Mellanox C4 likely */
595 		rs->rs_ifp = ifp;
596 		rs->rs_if_dunit = ifp->if_dunit;
597 		rs->rs_rate_cnt = rl.number_of_rates;
598 		rs->rs_min_seg = rl.min_segment_burst;
599 		rs->rs_highest_valid = 0;
600 		rs->rs_flow_limit = rl.max_flows;
601 		rs->rs_flags = RS_IS_INTF | RS_NO_PRE;
602 		rs->rs_disable = 0;
603 		rate_table_act = rl.rate_table;
604 	} else if ((rl.flags & RT_IS_SELECTABLE) == RT_IS_SELECTABLE) {
605 		/* Chelsio, C5 and C6 of Mellanox? */
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_disable = 0;
611 		rs->rs_flow_limit = rl.max_flows;
612 		rate_table_act = desired_rates;
613 		if ((rs->rs_rate_cnt > MAX_HDWR_RATES) &&
614 		    (rs->rs_rate_cnt < ALL_HARDWARE_RATES)) {
615 			/*
616 			 * Our desired table is not big
617 			 * enough, do what we can.
618 			 */
619 			rs->rs_rate_cnt = MAX_HDWR_RATES;
620 		 }
621 		if (rs->rs_rate_cnt <= RS_ORDERED_COUNT)
622 			rs->rs_flags = RS_IS_INTF;
623 		else
624 			rs->rs_flags = RS_IS_INTF | RS_INT_TBL;
625 		if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)
626 			rs->rs_rate_cnt = ALL_HARDWARE_RATES;
627 	} else {
628 		free(rs, M_TCPPACE);
629 		return (NULL);
630 	}
631 	sz = sizeof(struct tcp_hwrate_limit_table) * rs->rs_rate_cnt;
632 	rs->rs_rlt = malloc(sz, M_TCPPACE, M_NOWAIT);
633 	if (rs->rs_rlt == NULL) {
634 		if (error)
635 			*error = ENOMEM;
636 bail:
637 		free(rs, M_TCPPACE);
638 		return (NULL);
639 	}
640 	if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) {
641 		/*
642 		 * The interface supports all
643 		 * the rates we could possibly want.
644 		 */
645 		uint64_t rat;
646 
647 		rs->rs_rlt[0].rate = 12500;	/* 100k */
648 		rs->rs_rlt[1].rate = 25000;	/* 200k */
649 		rs->rs_rlt[2].rate = 62500;	/* 500k */
650 		/* Note 125000 == 1Megabit
651 		 * populate 1Meg - 1000meg.
652 		 */
653 		for(i = 3, rat = 125000; i< (ALL_HARDWARE_RATES-1); i++) {
654 			rs->rs_rlt[i].rate = rat;
655 			rat += 125000;
656 		}
657 		rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate = 1250000000;
658 	} else if (rs->rs_flags & RS_INT_TBL) {
659 		/* We populate this in a special way */
660 		populate_canned_table(rs, rate_table_act);
661 	} else {
662 		/*
663 		 * Just copy in the rates from
664 		 * the table, it is in order.
665 		 */
666 		for (i=0; i<rs->rs_rate_cnt; i++) {
667 			rs->rs_rlt[i].rate = rate_table_act[i];
668 			rs->rs_rlt[i].time_between = 0;
669 			rs->rs_rlt[i].flags = 0;
670 		}
671 	}
672 	for (i = (rs->rs_rate_cnt - 1); i >= 0; i--) {
673 		/*
674 		 * We go backwards through the list so that if we can't get
675 		 * a rate and fail to init one, we have at least a chance of
676 		 * getting the highest one.
677 		 */
678 		rs->rs_rlt[i].ptbl = rs;
679 		rs->rs_rlt[i].tag = NULL;
680 		/*
681 		 * Calculate the time between.
682 		 */
683 		lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
684 		res = lentim / rs->rs_rlt[i].rate;
685 		if (res > 0)
686 			rs->rs_rlt[i].time_between = res;
687 		else
688 			rs->rs_rlt[i].time_between = 1;
689 		if (rs->rs_flags & RS_NO_PRE) {
690 			rs->rs_rlt[i].flags = HDWRPACE_INITED;
691 			rs->rs_lowest_valid = i;
692 		} else {
693 			int err;
694 
695 			if ((rl.flags & RT_IS_SETUP_REQ)  &&
696 			    (ifp->if_ratelimit_query)) {
697 				err = ifp->if_ratelimit_setup(ifp,
698   				         rs->rs_rlt[i].rate, i);
699 				if (err)
700 					goto handle_err;
701 			}
702 #ifdef RSS
703 			hash_type = M_HASHTYPE_RSS_TCP_IPV4;
704 #else
705 			hash_type = M_HASHTYPE_OPAQUE_HASH;
706 #endif
707 			err = rl_attach_txrtlmt(ifp,
708 			    hash_type,
709 			    (i + 1),
710 			    rs->rs_rlt[i].rate,
711 			    &rs->rs_rlt[i].tag);
712 			if (err) {
713 handle_err:
714 				if (i == (rs->rs_rate_cnt - 1)) {
715 					/*
716 					 * Huh - first rate and we can't get
717 					 * it?
718 					 */
719 					free(rs->rs_rlt, M_TCPPACE);
720 					if (error)
721 						*error = err;
722 					goto bail;
723 				} else {
724 					if (error)
725 						*error = err;
726 				}
727 				break;
728 			} else {
729 				rs->rs_rlt[i].flags = HDWRPACE_INITED | HDWRPACE_TAGPRESENT;
730 				rs->rs_lowest_valid = i;
731 			}
732 		}
733 	}
734 	/* Did we get at least 1 rate? */
735 	if (rs->rs_rlt[(rs->rs_rate_cnt - 1)].flags & HDWRPACE_INITED)
736 		rs->rs_highest_valid = rs->rs_rate_cnt - 1;
737 	else {
738 		free(rs->rs_rlt, M_TCPPACE);
739 		goto bail;
740 	}
741 	rs_number_alive++;
742 	sysctl_ctx_init(&rs->sysctl_ctx);
743 	rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
744 	    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
745 	    OID_AUTO,
746 	    rs->rs_ifp->if_xname,
747 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
748 	    "");
749 	rl_add_syctl_entries(rl_sysctl_root, rs);
750 	mtx_lock(&rs_mtx);
751 	CK_LIST_INSERT_HEAD(&int_rs, rs, next);
752 	mtx_unlock(&rs_mtx);
753 	return (rs);
754 }
755 
756 static const struct tcp_hwrate_limit_table *
757 tcp_int_find_suitable_rate(const struct tcp_rate_set *rs,
758     uint64_t bytes_per_sec, uint32_t flags)
759 {
760 	struct tcp_hwrate_limit_table *arte = NULL, *rte = NULL;
761 	uint64_t mbits_per_sec, ind_calc;
762 	int i;
763 
764 	mbits_per_sec = (bytes_per_sec * 8);
765 	if (flags & RS_PACING_LT) {
766 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
767 		    (rs->rs_lowest_valid <= 2)){
768 			/*
769 			 * Smaller than 1Meg, only
770 			 * 3 entries can match it.
771 			 */
772 			for(i = rs->rs_lowest_valid; i < 3; i++) {
773 				if (bytes_per_sec <= rs->rs_rlt[i].rate) {
774 					rte = &rs->rs_rlt[i];
775 					break;
776 				} else if (rs->rs_rlt[i].flags & HDWRPACE_INITED) {
777 					arte = &rs->rs_rlt[i];
778 				}
779 			}
780 			goto done;
781 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
782 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
783 			/*
784 			 * Larger than 1G (the majority of
785 			 * our table.
786 			 */
787 			if (mbits_per_sec < RS_TEN_GIGABIT_PERSEC)
788 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
789 			else
790 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
791 			goto done;
792 		}
793 		/*
794 		 * If we reach here its in our table (between 1Meg - 1000Meg),
795 		 * just take the rounded down mbits per second, and add
796 		 * 1Megabit to it, from this we can calculate
797 		 * the index in the table.
798 		 */
799 		ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
800 		if ((ind_calc * RS_ONE_MEGABIT_PERSEC) != mbits_per_sec)
801 			ind_calc++;
802 		/* our table is offset by 3, we add 2 */
803 		ind_calc += 2;
804 		if (ind_calc > (ALL_HARDWARE_RATES-1)) {
805 			/* This should not happen */
806 			ind_calc = ALL_HARDWARE_RATES-1;
807 		}
808 		if ((ind_calc >= rs->rs_lowest_valid) &&
809 		    (ind_calc <= rs->rs_highest_valid))
810 		rte = &rs->rs_rlt[ind_calc];
811 	} else if (flags & RS_PACING_EXACT_MATCH) {
812 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
813 		    (rs->rs_lowest_valid <= 2)){
814 			for(i = rs->rs_lowest_valid; i < 3; i++) {
815 				if (bytes_per_sec == rs->rs_rlt[i].rate) {
816 					rte = &rs->rs_rlt[i];
817 					break;
818 				}
819 			}
820 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
821 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
822 			/* > 1Gbps only one rate */
823 			if (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) {
824 				/* Its 10G wow */
825 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
826 			}
827 		} else {
828 			/* Ok it must be a exact meg (its between 1G and 1Meg) */
829 			ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
830 			if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
831 				/* its an exact Mbps */
832 				ind_calc += 2;
833 				if (ind_calc > (ALL_HARDWARE_RATES-1)) {
834 					/* This should not happen */
835 					ind_calc = ALL_HARDWARE_RATES-1;
836 				}
837 				if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED)
838 					rte = &rs->rs_rlt[ind_calc];
839 			}
840 		}
841 	} else {
842 		/* we want greater than the requested rate */
843 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
844 		    (rs->rs_lowest_valid <= 2)){
845 			arte = &rs->rs_rlt[3]; /* set alternate to 1Meg */
846 			for (i=2; i>=rs->rs_lowest_valid; i--) {
847 				if (bytes_per_sec < rs->rs_rlt[i].rate) {
848 					rte = &rs->rs_rlt[i];
849 					break;
850 				} else if ((flags & RS_PACING_GEQ) &&
851 					   (bytes_per_sec == rs->rs_rlt[i].rate)) {
852 					rte = &rs->rs_rlt[i];
853 					break;
854 				} else {
855 					arte = &rs->rs_rlt[i]; /* new alternate */
856 				}
857 			}
858 		} else if (mbits_per_sec > RS_ONE_GIGABIT_PERSEC) {
859 			if ((bytes_per_sec < rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
860 			    (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
861 				/* Our top rate is larger than the request */
862 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
863 			} else if ((flags & RS_PACING_GEQ) &&
864 				   (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
865 				   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
866 				/* It matches our top rate */
867 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
868 			} else if (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED) {
869 				/* The top rate is an alternative */
870 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
871 			}
872 		} else {
873 			/* Its in our range 1Meg - 1Gig */
874 			if (flags & RS_PACING_GEQ) {
875 				ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
876 				if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
877 					if (ind_calc > (ALL_HARDWARE_RATES-1)) {
878 						/* This should not happen */
879 						ind_calc = (ALL_HARDWARE_RATES-1);
880 					}
881 					rte = &rs->rs_rlt[ind_calc];
882 				}
883 				goto done;
884 			}
885 			ind_calc = (mbits_per_sec + (RS_ONE_MEGABIT_PERSEC-1))/RS_ONE_MEGABIT_PERSEC;
886 			ind_calc += 2;
887 			if (ind_calc > (ALL_HARDWARE_RATES-1)) {
888 				/* This should not happen */
889 				ind_calc = ALL_HARDWARE_RATES-1;
890 			}
891 			if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED)
892 				rte = &rs->rs_rlt[ind_calc];
893 		}
894 	}
895 done:
896 	if ((rte == NULL) &&
897 	    (arte != NULL) &&
898 	    (flags & RS_PACING_SUB_OK)) {
899 		/* We can use the substitute */
900 		rte = arte;
901 	}
902 	return (rte);
903 }
904 
905 static const struct tcp_hwrate_limit_table *
906 tcp_find_suitable_rate(const struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags)
907 {
908 	/**
909 	 * Hunt the rate table with the restrictions in flags and find a
910 	 * suitable rate if possible.
911 	 * RS_PACING_EXACT_MATCH - look for an exact match to rate.
912 	 * RS_PACING_GT     - must be greater than.
913 	 * RS_PACING_GEQ    - must be greater than or equal.
914 	 * RS_PACING_LT     - must be less than.
915 	 * RS_PACING_SUB_OK - If we don't meet criteria a
916 	 *                    substitute is ok.
917 	 */
918 	int i, matched;
919 	struct tcp_hwrate_limit_table *rte = NULL;
920 
921 
922 	if ((rs->rs_flags & RS_INT_TBL) &&
923 	    (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)) {
924 		/*
925 		 * Here we don't want to paw thru
926 		 * a big table, we have everything
927 		 * from 1Meg - 1000Meg in 1Meg increments.
928 		 * Use an alternate method to "lookup".
929 		 */
930 		return (tcp_int_find_suitable_rate(rs, bytes_per_sec, flags));
931 	}
932 	if ((flags & RS_PACING_LT) ||
933 	    (flags & RS_PACING_EXACT_MATCH)) {
934 		/*
935 		 * For exact and less than we go forward through the table.
936 		 * This way when we find one larger we stop (exact was a
937 		 * toss up).
938 		 */
939 		for (i = rs->rs_lowest_valid, matched = 0; i <= rs->rs_highest_valid; i++) {
940 			if ((flags & RS_PACING_EXACT_MATCH) &&
941 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
942 				rte = &rs->rs_rlt[i];
943 				matched = 1;
944 				break;
945 			} else if ((flags & RS_PACING_LT) &&
946 			    (bytes_per_sec <= rs->rs_rlt[i].rate)) {
947 				rte = &rs->rs_rlt[i];
948 				matched = 1;
949 				break;
950 			}
951 			if (bytes_per_sec > rs->rs_rlt[i].rate)
952 				break;
953 		}
954 		if ((matched == 0) &&
955 		    (flags & RS_PACING_LT) &&
956 		    (flags & RS_PACING_SUB_OK)) {
957 			/* Kick in a substitute (the lowest) */
958 			rte = &rs->rs_rlt[rs->rs_lowest_valid];
959 		}
960 	} else {
961 		/*
962 		 * Here we go backward through the table so that we can find
963 		 * the one greater in theory faster (but its probably a
964 		 * wash).
965 		 */
966 		for (i = rs->rs_highest_valid, matched = 0; i >= rs->rs_lowest_valid; i--) {
967 			if (rs->rs_rlt[i].rate > bytes_per_sec) {
968 				/* A possible candidate */
969 				rte = &rs->rs_rlt[i];
970 			}
971 			if ((flags & RS_PACING_GEQ) &&
972 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
973 				/* An exact match and we want equal */
974 				matched = 1;
975 				rte = &rs->rs_rlt[i];
976 				break;
977 			} else if (rte) {
978 				/*
979 				 * Found one that is larger than but don't
980 				 * stop, there may be a more closer match.
981 				 */
982 				matched = 1;
983 			}
984 			if (rs->rs_rlt[i].rate < bytes_per_sec) {
985 				/*
986 				 * We found a table entry that is smaller,
987 				 * stop there will be none greater or equal.
988 				 */
989 				break;
990 			}
991 		}
992 		if ((matched == 0) &&
993 		    (flags & RS_PACING_SUB_OK)) {
994 			/* Kick in a substitute (the highest) */
995 			rte = &rs->rs_rlt[rs->rs_highest_valid];
996 		}
997 	}
998 	return (rte);
999 }
1000 
1001 static struct ifnet *
1002 rt_find_real_interface(struct ifnet *ifp, struct inpcb *inp, int *error)
1003 {
1004 	struct ifnet *tifp;
1005 	struct m_snd_tag *tag;
1006 	union if_snd_tag_alloc_params params = {
1007 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
1008 		.rate_limit.hdr.flowid = 1,
1009 		.rate_limit.max_rate = COMMON_RATE,
1010 		.rate_limit.flags = M_NOWAIT,
1011 	};
1012 	int err;
1013 #ifdef RSS
1014 	params.rate_limit.hdr.flowtype = ((inp->inp_vflag & INP_IPV6) ?
1015 	    M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_TCP_IPV4);
1016 #else
1017 	params.rate_limit.hdr.flowtype = M_HASHTYPE_OPAQUE_HASH;
1018 #endif
1019 	tag = NULL;
1020 	if (ifp->if_snd_tag_alloc) {
1021 		if (error)
1022 			*error = ENODEV;
1023 		return (NULL);
1024 	}
1025 	err = ifp->if_snd_tag_alloc(ifp, &params, &tag);
1026 	if (err) {
1027 		/* Failed to setup a tag? */
1028 		if (error)
1029 			*error = err;
1030 		return (NULL);
1031 	}
1032 	tifp = tag->ifp;
1033 	tifp->if_snd_tag_free(tag);
1034 	return (tifp);
1035 }
1036 
1037 static const struct tcp_hwrate_limit_table *
1038 rt_setup_rate(struct inpcb *inp, struct ifnet *ifp, uint64_t bytes_per_sec,
1039     uint32_t flags, int *error)
1040 {
1041 	/* First lets find the interface if it exists */
1042 	const struct tcp_hwrate_limit_table *rte;
1043 	struct tcp_rate_set *rs;
1044 	struct epoch_tracker et;
1045 	int err;
1046 
1047 	NET_EPOCH_ENTER(et);
1048 use_real_interface:
1049 	CK_LIST_FOREACH(rs, &int_rs, next) {
1050 		/*
1051 		 * Note we don't look with the lock since we either see a
1052 		 * new entry or will get one when we try to add it.
1053 		 */
1054 		if (rs->rs_flags & RS_IS_DEAD) {
1055 			/* The dead are not looked at */
1056 			continue;
1057 		}
1058 		if ((rs->rs_ifp == ifp) &&
1059 		    (rs->rs_if_dunit == ifp->if_dunit)) {
1060 			/* Ok we found it */
1061 			break;
1062 		}
1063 	}
1064 	if ((rs == NULL) ||
1065 	    (rs->rs_flags & RS_INTF_NO_SUP) ||
1066 	    (rs->rs_flags & RS_IS_DEAD)) {
1067 		/*
1068 		 * This means we got a packet *before*
1069 		 * the IF-UP was processed below, <or>
1070 		 * while or after we already received an interface
1071 		 * departed event. In either case we really don't
1072 		 * want to do anything with pacing, in
1073 		 * the departing case the packet is not
1074 		 * going to go very far. The new case
1075 		 * might be arguable, but its impossible
1076 		 * to tell from the departing case.
1077 		 */
1078 		if (rs->rs_disable && error)
1079 			*error = ENODEV;
1080 		NET_EPOCH_EXIT(et);
1081 		return (NULL);
1082 	}
1083 
1084 	if ((rs == NULL) || (rs->rs_disable != 0)) {
1085 		if (rs->rs_disable && error)
1086 			*error = ENOSPC;
1087 		NET_EPOCH_EXIT(et);
1088 		return (NULL);
1089 	}
1090 	if (rs->rs_flags & RS_IS_DEFF) {
1091 		/* We need to find the real interface */
1092 		struct ifnet *tifp;
1093 
1094 		tifp = rt_find_real_interface(ifp, inp, error);
1095 		if (tifp == NULL) {
1096 			if (rs->rs_disable && error)
1097 				*error = ENOTSUP;
1098 			NET_EPOCH_EXIT(et);
1099 			return (NULL);
1100 		}
1101 		goto use_real_interface;
1102 	}
1103 	if (rs->rs_flow_limit &&
1104 	    ((rs->rs_flows_using + 1) > rs->rs_flow_limit)) {
1105 		if (error)
1106 			*error = ENOSPC;
1107 		NET_EPOCH_EXIT(et);
1108 		return (NULL);
1109 	}
1110 	rte = tcp_find_suitable_rate(rs, bytes_per_sec, flags);
1111 	if (rte) {
1112 		err = in_pcbattach_txrtlmt(inp, rs->rs_ifp,
1113 		    inp->inp_flowtype,
1114 		    inp->inp_flowid,
1115 		    rte->rate,
1116 		    &inp->inp_snd_tag);
1117 		if (err) {
1118 			/* Failed to attach */
1119 			if (error)
1120 				*error = err;
1121 			rte = NULL;
1122 		}
1123 	}
1124 	if (rte) {
1125 		/*
1126 		 * We use an atomic here for accounting so we don't have to
1127 		 * use locks when freeing.
1128 		 */
1129 		atomic_add_64(&rs->rs_flows_using, 1);
1130 	}
1131 	NET_EPOCH_EXIT(et);
1132 	return (rte);
1133 }
1134 
1135 static void
1136 tcp_rl_ifnet_link(void *arg __unused, struct ifnet *ifp, int link_state)
1137 {
1138 	int error;
1139 	struct tcp_rate_set *rs;
1140 
1141 	if (((ifp->if_capabilities & IFCAP_TXRTLMT) == 0) ||
1142 	    (link_state != LINK_STATE_UP)) {
1143 		/*
1144 		 * We only care on an interface going up that is rate-limit
1145 		 * capable.
1146 		 */
1147 		return;
1148 	}
1149 	mtx_lock(&rs_mtx);
1150 	CK_LIST_FOREACH(rs, &int_rs, next) {
1151 		if ((rs->rs_ifp == ifp) &&
1152 		    (rs->rs_if_dunit == ifp->if_dunit)) {
1153 			/* We already have initialized this guy */
1154 			mtx_unlock(&rs_mtx);
1155 			return;
1156 		}
1157 	}
1158 	mtx_unlock(&rs_mtx);
1159 	rt_setup_new_rs(ifp, &error);
1160 }
1161 
1162 static void
1163 tcp_rl_ifnet_departure(void *arg __unused, struct ifnet *ifp)
1164 {
1165 	struct tcp_rate_set *rs, *nrs;
1166 	struct ifnet *tifp;
1167 	int i;
1168 
1169 	mtx_lock(&rs_mtx);
1170 	CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) {
1171 		if ((rs->rs_ifp == ifp) &&
1172 		    (rs->rs_if_dunit == ifp->if_dunit)) {
1173 			CK_LIST_REMOVE(rs, next);
1174 			rs_number_alive--;
1175 			rs->rs_flags |= RS_IS_DEAD;
1176 			for (i = 0; i < rs->rs_rate_cnt; i++) {
1177 				if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1178 					tifp = rs->rs_rlt[i].tag->ifp;
1179 					in_pcbdetach_tag(tifp, rs->rs_rlt[i].tag);
1180 					rs->rs_rlt[i].tag = NULL;
1181 				}
1182 				rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1183 			}
1184 			if (rs->rs_flows_using == 0)
1185 				rs_defer_destroy(rs);
1186 			break;
1187 		}
1188 	}
1189 	mtx_unlock(&rs_mtx);
1190 }
1191 
1192 static void
1193 tcp_rl_shutdown(void *arg __unused, int howto __unused)
1194 {
1195 	struct tcp_rate_set *rs, *nrs;
1196 	struct ifnet *tifp;
1197 	int i;
1198 
1199 	mtx_lock(&rs_mtx);
1200 	CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) {
1201 		CK_LIST_REMOVE(rs, next);
1202 		rs_number_alive--;
1203 		rs->rs_flags |= RS_IS_DEAD;
1204 		for (i = 0; i < rs->rs_rate_cnt; i++) {
1205 			if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1206 				tifp = rs->rs_rlt[i].tag->ifp;
1207 				in_pcbdetach_tag(tifp, rs->rs_rlt[i].tag);
1208 				rs->rs_rlt[i].tag = NULL;
1209 			}
1210 			rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1211 		}
1212 		if (rs->rs_flows_using == 0)
1213 			rs_defer_destroy(rs);
1214 	}
1215 	mtx_unlock(&rs_mtx);
1216 }
1217 
1218 const struct tcp_hwrate_limit_table *
1219 tcp_set_pacing_rate(struct tcpcb *tp, struct ifnet *ifp,
1220     uint64_t bytes_per_sec, int flags, int *error)
1221 {
1222 	const struct tcp_hwrate_limit_table *rte;
1223 
1224 	if (tp->t_inpcb->inp_snd_tag == NULL) {
1225 		/*
1226 		 * We are setting up a rate for the first time.
1227 		 */
1228 		if ((ifp->if_capabilities & IFCAP_TXRTLMT) == 0) {
1229 			/* Not supported by the egress */
1230 			if (error)
1231 				*error = ENODEV;
1232 			return (NULL);
1233 		}
1234 #ifdef KERN_TLS
1235 		if (tp->t_inpcb->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
1236 			/*
1237 			 * We currently can't do both TLS and hardware
1238 			 * pacing
1239 			 */
1240 			if (error)
1241 				*error = EINVAL;
1242 			return (NULL);
1243 		}
1244 #endif
1245 		rte = rt_setup_rate(tp->t_inpcb, ifp, bytes_per_sec, flags, error);
1246 	} else {
1247 		/*
1248 		 * We are modifying a rate, wrong interface?
1249 		 */
1250 		if (error)
1251 			*error = EINVAL;
1252 		rte = NULL;
1253 	}
1254 	*error = 0;
1255 	return (rte);
1256 }
1257 
1258 const struct tcp_hwrate_limit_table *
1259 tcp_chg_pacing_rate(const struct tcp_hwrate_limit_table *crte,
1260     struct tcpcb *tp, struct ifnet *ifp,
1261     uint64_t bytes_per_sec, int flags, int *error)
1262 {
1263 	const struct tcp_hwrate_limit_table *nrte;
1264 	const struct tcp_rate_set *rs;
1265 	int is_indirect = 0;
1266 	int err;
1267 
1268 
1269 	if ((tp->t_inpcb->inp_snd_tag == NULL) ||
1270 	    (crte == NULL)) {
1271 		/* Wrong interface */
1272 		if (error)
1273 			*error = EINVAL;
1274 		return (NULL);
1275 	}
1276 	rs = crte->ptbl;
1277 	if ((rs->rs_flags & RS_IS_DEAD) ||
1278 	    (crte->flags & HDWRPACE_IFPDEPARTED)) {
1279 		/* Release the rate, and try anew */
1280 re_rate:
1281 		tcp_rel_pacing_rate(crte, tp);
1282 		nrte = tcp_set_pacing_rate(tp, ifp,
1283 		    bytes_per_sec, flags, error);
1284 		return (nrte);
1285 	}
1286 	if ((rs->rs_flags & RT_IS_INDIRECT ) == RT_IS_INDIRECT)
1287 		is_indirect = 1;
1288 	else
1289 		is_indirect = 0;
1290 	if ((is_indirect == 0) &&
1291 	    ((ifp != rs->rs_ifp) ||
1292 	    (ifp->if_dunit != rs->rs_if_dunit))) {
1293 		/*
1294 		 * Something changed, the user is not pointing to the same
1295 		 * ifp? Maybe a route updated on this guy?
1296 		 */
1297 		goto re_rate;
1298 	} else if (is_indirect) {
1299 		/*
1300 		 * For indirect we have to dig in and find the real interface.
1301 		 */
1302 		struct ifnet *rifp;
1303 
1304 		rifp = rt_find_real_interface(ifp, tp->t_inpcb, error);
1305 		if (rifp == NULL) {
1306 			/* Can't find it? */
1307 			goto re_rate;
1308 		}
1309 		if ((rifp != rs->rs_ifp) ||
1310 		    (ifp->if_dunit != rs->rs_if_dunit)) {
1311 			goto re_rate;
1312 		}
1313 	}
1314 	nrte = tcp_find_suitable_rate(rs, bytes_per_sec, flags);
1315 	if (nrte == crte) {
1316 		/* No change */
1317 		if (error)
1318 			*error = 0;
1319 		return (crte);
1320 	}
1321 	if (nrte == NULL) {
1322 		/* Release the old rate */
1323 		tcp_rel_pacing_rate(crte, tp);
1324 		return (NULL);
1325 	}
1326 	/* Change rates to our new entry */
1327 	err = in_pcbmodify_txrtlmt(tp->t_inpcb, nrte->rate);
1328 	if (err) {
1329 		if (error)
1330 			*error = err;
1331 		return (NULL);
1332 	}
1333 	if (error)
1334 		*error = 0;
1335 	return (nrte);
1336 }
1337 
1338 void
1339 tcp_rel_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp)
1340 {
1341 	const struct tcp_rate_set *crs;
1342 	struct tcp_rate_set *rs;
1343 	uint64_t pre;
1344 
1345 	crs = crte->ptbl;
1346 	/*
1347 	 * Now we must break the const
1348 	 * in order to release our refcount.
1349 	 */
1350 	rs = __DECONST(struct tcp_rate_set *, crs);
1351 	pre = atomic_fetchadd_64(&rs->rs_flows_using, -1);
1352 	if (pre == 1) {
1353 		mtx_lock(&rs_mtx);
1354 		/*
1355 		 * Is it dead?
1356 		 */
1357 		if (rs->rs_flags & RS_IS_DEAD)
1358 			rs_defer_destroy(rs);
1359 		mtx_unlock(&rs_mtx);
1360 	}
1361 	in_pcbdetach_txrtlmt(tp->t_inpcb);
1362 }
1363 
1364 #define ONE_POINT_TWO_MEG 150000 /* 1.2 megabits in bytes */
1365 #define ONE_HUNDRED_MBPS 12500000	/* 100Mbps in bytes per second */
1366 #define FIVE_HUNDRED_MBPS 62500000	/* 500Mbps in bytes per second */
1367 #define MAX_MSS_SENT 43	/* 43 mss = 43 x 1500 = 64,500 bytes */
1368 
1369 
1370 uint32_t
1371 tcp_get_pacing_burst_size (uint64_t bw, uint32_t segsiz, int can_use_1mss,
1372    const struct tcp_hwrate_limit_table *te, int *err)
1373 {
1374 	/*
1375 	 * We use the google formula to calculate the
1376 	 * TSO size. I.E.
1377 	 * bw < 24Meg
1378 	 *   tso = 2mss
1379 	 * else
1380 	 *   tso = min(bw/1000, 64k)
1381 	 *
1382 	 * Note for these calculations we ignore the
1383 	 * packet overhead (enet hdr, ip hdr and tcp hdr).
1384 	 */
1385 	uint64_t lentim, res, bytes;
1386 	uint32_t new_tso, min_tso_segs;
1387 
1388 	bytes = bw / 1000;
1389 	if (bytes > (64 * 1000))
1390 		bytes = 64 * 1000;
1391 	/* Round up */
1392 	new_tso = (bytes + segsiz - 1) / segsiz;
1393 	if (can_use_1mss && (bw < ONE_POINT_TWO_MEG))
1394 		min_tso_segs = 1;
1395 	else
1396 		min_tso_segs = 2;
1397 	if (new_tso < min_tso_segs)
1398 		new_tso = min_tso_segs;
1399 	if (new_tso > MAX_MSS_SENT)
1400 		new_tso = MAX_MSS_SENT;
1401 	new_tso *= segsiz;
1402 	/*
1403 	 * If we are not doing hardware pacing
1404 	 * then we are done.
1405 	 */
1406 	if (te == NULL) {
1407 		if (err)
1408 			*err = 0;
1409 		return(new_tso);
1410 	}
1411 	/*
1412 	 * For hardware pacing we look at the
1413 	 * rate you are sending at and compare
1414 	 * that to the rate you have in hardware.
1415 	 *
1416 	 * If the hardware rate is slower than your
1417 	 * software rate then you are in error and
1418 	 * we will build a queue in our hardware whic
1419 	 * is probably not desired, in such a case
1420 	 * just return the non-hardware TSO size.
1421 	 *
1422 	 * If the rate in hardware is faster (which
1423 	 * it should be) then look at how long it
1424 	 * takes to send one ethernet segment size at
1425 	 * your b/w and compare that to the time it
1426 	 * takes to send at the rate you had selected.
1427 	 *
1428 	 * If your time is greater (which we hope it is)
1429 	 * we get the delta between the two, and then
1430 	 * divide that into your pacing time. This tells
1431 	 * us how many MSS you can send down at once (rounded up).
1432 	 *
1433 	 * Note we also double this value if the b/w is over
1434 	 * 100Mbps. If its over 500meg we just set you to the
1435 	 * max (43 segments).
1436 	 */
1437 	if (te->rate > FIVE_HUNDRED_MBPS)
1438 		return (segsiz * MAX_MSS_SENT);
1439 	if (te->rate == bw) {
1440 		/* We are pacing at exactly the hdwr rate */
1441 		return (segsiz * MAX_MSS_SENT);
1442 	}
1443 	lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
1444 	res = lentim / bw;
1445 	if (res > te->time_between) {
1446 		uint32_t delta, segs;
1447 
1448 		delta = res - te->time_between;
1449 		segs = (res + delta - 1)/delta;
1450 		if (te->rate > ONE_HUNDRED_MBPS)
1451 			segs *= 2;
1452 		if (segs < min_tso_segs)
1453 			segs = min_tso_segs;
1454 		if (segs > MAX_MSS_SENT)
1455 			segs = MAX_MSS_SENT;
1456 		segs *= segsiz;
1457 		if (err)
1458 			*err = 0;
1459 		if (segs < new_tso) {
1460 			/* unexpected ? */
1461 			return(new_tso);
1462 		} else {
1463 			return (segs);
1464 		}
1465 	} else {
1466 		/*
1467 		 * Your time is smaller which means
1468 		 * we will grow a queue on our
1469 		 * hardware. Send back the non-hardware
1470 		 * rate.
1471 		 */
1472 		if (err)
1473 			*err = -1;
1474 		return (new_tso);
1475 	}
1476 }
1477 
1478 static eventhandler_tag rl_ifnet_departs;
1479 static eventhandler_tag rl_ifnet_arrives;
1480 static eventhandler_tag rl_shutdown_start;
1481 
1482 static void
1483 tcp_rs_init(void *st __unused)
1484 {
1485 	CK_LIST_INIT(&int_rs);
1486 	rs_number_alive = 0;
1487 	rs_number_dead = 0;;
1488 	mtx_init(&rs_mtx, "tcp_rs_mtx", "rsmtx", MTX_DEF);
1489 	rl_ifnet_departs = EVENTHANDLER_REGISTER(ifnet_departure_event,
1490 	    tcp_rl_ifnet_departure,
1491 	    NULL, EVENTHANDLER_PRI_ANY);
1492 	rl_ifnet_arrives = EVENTHANDLER_REGISTER(ifnet_link_event,
1493 	    tcp_rl_ifnet_link,
1494 	    NULL, EVENTHANDLER_PRI_ANY);
1495 	rl_shutdown_start = EVENTHANDLER_REGISTER(shutdown_pre_sync,
1496 	    tcp_rl_shutdown, NULL,
1497 	    SHUTDOWN_PRI_FIRST);
1498 	printf("TCP_ratelimit: Is now initialized\n");
1499 }
1500 
1501 SYSINIT(tcp_rl_init, SI_SUB_SMP + 1, SI_ORDER_ANY, tcp_rs_init, NULL);
1502 #endif
1503