xref: /freebsd/sys/netinet/tcp_ratelimit.c (revision 4b50c451720d8b427757a6da1dd2bb4c52cd9e35)
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 #define TCPSTATES		/* for logging */
53 #include <netinet/in.h>
54 #include <netinet/in_pcb.h>
55 #include <netinet/tcp_var.h>
56 #ifdef INET6
57 #include <netinet6/tcp6_var.h>
58 #endif
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 #ifndef ETHERNET_SEGMENT_SIZE
68 #define ETHERNET_SEGMENT_SIZE 1500
69 #endif
70 MALLOC_DEFINE(M_TCPPACE, "tcp_hwpace", "TCP Hardware pacing memory");
71 #ifdef RATELIMIT
72 
73 #define COMMON_RATE 180500
74 uint64_t desired_rates[] = {
75 	62500,			/* 500Kbps */
76 	180500,			/* 1.44Mpbs */
77 	375000,			/* 3Mbps */
78 	500000,			/* 4Mbps */
79 	625000,			/* 5Mbps */
80 	750000,			/* 6Mbps */
81 	1000000,		/* 8Mbps */
82 	1250000,		/* 10Mbps */
83 	2500000,		/* 20Mbps */
84 	3750000,		/* 30Mbps */
85 	5000000,		/* 40Meg */
86 	6250000,		/* 50Mbps */
87 	12500000,		/* 100Mbps */
88 	25000000,		/* 200Mbps */
89 	50000000,		/* 400Mbps */
90 	100000000,		/* 800Mbps */
91 	12500,			/* 100kbps */
92 	25000,			/* 200kbps */
93 	875000,			/* 7Mbps */
94 	1125000,		/* 9Mbps */
95 	1875000,		/* 15Mbps */
96 	3125000,		/* 25Mbps */
97 	8125000,		/* 65Mbps */
98 	10000000,		/* 80Mbps */
99 	18750000,		/* 150Mbps */
100 	20000000,		/* 250Mbps */
101 	37500000,		/* 350Mbps */
102 	62500000,		/* 500Mbps */
103 	78125000,		/* 625Mbps */
104 	125000000,		/* 1Gbps */
105 };
106 #define MAX_HDWR_RATES (sizeof(desired_rates)/sizeof(uint64_t))
107 #define RS_ORDERED_COUNT 16	/*
108 				 * Number that are in order
109 				 * at the beginning of the table,
110 				 * over this a sort is required.
111 				 */
112 #define RS_NEXT_ORDER_GROUP 16	/*
113 				 * The point in our table where
114 				 * we come fill in a second ordered
115 				 * group (index wise means -1).
116 				 */
117 #define ALL_HARDWARE_RATES 1004 /*
118 				 * 1Meg - 1Gig in 1 Meg steps
119 				 * plus 100, 200k  and 500k and
120 				 * 10Gig
121 				 */
122 
123 #define RS_ONE_MEGABIT_PERSEC 1000000
124 #define RS_ONE_GIGABIT_PERSEC 1000000000
125 #define RS_TEN_GIGABIT_PERSEC 10000000000
126 
127 static struct head_tcp_rate_set int_rs;
128 static struct mtx rs_mtx;
129 uint32_t rs_number_alive;
130 uint32_t rs_number_dead;
131 
132 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, rl, CTLFLAG_RW, 0,
133     "TCP Ratelimit stats");
134 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, alive, CTLFLAG_RW,
135     &rs_number_alive, 0,
136     "Number of interfaces initialized for ratelimiting");
137 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, dead, CTLFLAG_RW,
138     &rs_number_dead, 0,
139     "Number of interfaces departing from ratelimiting");
140 
141 static void
142 rl_add_syctl_entries(struct sysctl_oid *rl_sysctl_root, struct tcp_rate_set *rs)
143 {
144 	/*
145 	 * Add sysctl entries for thus interface.
146 	 */
147 	if (rs->rs_flags & RS_INTF_NO_SUP) {
148 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
149 		   SYSCTL_CHILDREN(rl_sysctl_root),
150 		   OID_AUTO, "disable", CTLFLAG_RD,
151 		   &rs->rs_disable, 0,
152 		   "Disable this interface from new hdwr limiting?");
153 	} else {
154 		SYSCTL_ADD_S32(&rs->sysctl_ctx,
155 		   SYSCTL_CHILDREN(rl_sysctl_root),
156 		   OID_AUTO, "disable", CTLFLAG_RW,
157 		   &rs->rs_disable, 0,
158 		   "Disable this interface from new hdwr limiting?");
159 	}
160 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
161 	    SYSCTL_CHILDREN(rl_sysctl_root),
162 	    OID_AUTO, "minseg", CTLFLAG_RW,
163 	    &rs->rs_min_seg, 0,
164 	    "What is the minimum we need to send on this interface?");
165 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
166 	    SYSCTL_CHILDREN(rl_sysctl_root),
167 	    OID_AUTO, "flow_limit", CTLFLAG_RW,
168 	    &rs->rs_flow_limit, 0,
169 	    "What is the limit for number of flows (0=unlimited)?");
170 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
171 	    SYSCTL_CHILDREN(rl_sysctl_root),
172 	    OID_AUTO, "highest", CTLFLAG_RD,
173 	    &rs->rs_highest_valid, 0,
174 	    "Highest valid rate");
175 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
176 	    SYSCTL_CHILDREN(rl_sysctl_root),
177 	    OID_AUTO, "lowest", CTLFLAG_RD,
178 	    &rs->rs_lowest_valid, 0,
179 	    "Lowest valid rate");
180 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
181 	    SYSCTL_CHILDREN(rl_sysctl_root),
182 	    OID_AUTO, "flags", CTLFLAG_RD,
183 	    &rs->rs_flags, 0,
184 	    "What lags are on the entry?");
185 	SYSCTL_ADD_S32(&rs->sysctl_ctx,
186 	    SYSCTL_CHILDREN(rl_sysctl_root),
187 	    OID_AUTO, "numrates", CTLFLAG_RD,
188 	    &rs->rs_rate_cnt, 0,
189 	    "How many rates re there?");
190 	SYSCTL_ADD_U64(&rs->sysctl_ctx,
191 	    SYSCTL_CHILDREN(rl_sysctl_root),
192 	    OID_AUTO, "flows_using", CTLFLAG_RD,
193 	    &rs->rs_flows_using, 0,
194 	    "How many flows are using this interface now?");
195 #ifdef DETAILED_RATELIMIT_SYSCTL
196 	if (rs->rs_rlt && rs->rs_rate_cnt > 0) {
197 		/*  Lets display the rates */
198 		int i;
199 		struct sysctl_oid *rl_rates;
200 		struct sysctl_oid *rl_rate_num;
201 		char rate_num[16];
202 		rl_rates = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
203 					    SYSCTL_CHILDREN(rl_sysctl_root),
204 					    OID_AUTO,
205 					    "rate",
206 					    CTLFLAG_RW, 0,
207 					    "Ratelist");
208 		for( i = 0; i < rs->rs_rate_cnt; i++) {
209 			sprintf(rate_num, "%d", i);
210 			rl_rate_num = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
211 					    SYSCTL_CHILDREN(rl_rates),
212 					    OID_AUTO,
213 					    rate_num,
214 					    CTLFLAG_RW, 0,
215 					    "Individual Rate");
216 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
217 				       SYSCTL_CHILDREN(rl_rate_num),
218 				       OID_AUTO, "flags", CTLFLAG_RD,
219 				       &rs->rs_rlt[i].flags, 0,
220 				       "Flags on this rate");
221 			SYSCTL_ADD_U32(&rs->sysctl_ctx,
222 				       SYSCTL_CHILDREN(rl_rate_num),
223 				       OID_AUTO, "pacetime", CTLFLAG_RD,
224 				       &rs->rs_rlt[i].time_between, 0,
225 				       "Time hardware inserts between 1500 byte sends");
226 			SYSCTL_ADD_U64(&rs->sysctl_ctx,
227 				       SYSCTL_CHILDREN(rl_rate_num),
228 				       OID_AUTO, "rate", CTLFLAG_RD,
229 				       &rs->rs_rlt[i].rate, 0,
230 				       "Rate in bytes per second");
231 		}
232 	}
233 #endif
234 }
235 
236 static void
237 rs_destroy(epoch_context_t ctx)
238 {
239 	struct tcp_rate_set *rs;
240 	bool do_free_rs;
241 
242 	rs = __containerof(ctx, struct tcp_rate_set, rs_epoch_ctx);
243 
244 	mtx_lock(&rs_mtx);
245 	rs->rs_flags &= ~RS_FUNERAL_SCHD;
246 	/*
247 	 * In theory its possible (but unlikely)
248 	 * that while the delete was occuring
249 	 * and we were applying the DEAD flag
250 	 * someone slipped in and found the
251 	 * interface in a lookup. While we
252 	 * decided rs_flows_using were 0 and
253 	 * scheduling the epoch_call, the other
254 	 * thread incremented rs_flow_using. This
255 	 * is because users have a pointer and
256 	 * we only use the rs_flows_using in an
257 	 * atomic fashion, i.e. the other entities
258 	 * are not protected. To assure this did
259 	 * not occur, we check rs_flows_using here
260 	 * before deleting.
261 	 */
262 	do_free_rs = (rs->rs_flows_using == 0);
263 	rs_number_dead--;
264 	mtx_unlock(&rs_mtx);
265 
266 	if (do_free_rs) {
267 		sysctl_ctx_free(&rs->sysctl_ctx);
268 		free(rs->rs_rlt, M_TCPPACE);
269 		free(rs, M_TCPPACE);
270 	}
271 }
272 
273 static void
274 rs_defer_destroy(struct tcp_rate_set *rs)
275 {
276 
277 	mtx_assert(&rs_mtx, MA_OWNED);
278 
279 	/* Check if already pending. */
280 	if (rs->rs_flags & RS_FUNERAL_SCHD)
281 		return;
282 
283 	rs_number_dead++;
284 
285 	/* Set flag to only defer once. */
286 	rs->rs_flags |= RS_FUNERAL_SCHD;
287 	epoch_call(net_epoch, &rs->rs_epoch_ctx, rs_destroy);
288 }
289 
290 #ifdef INET
291 extern counter_u64_t rate_limit_set_ok;
292 extern counter_u64_t rate_limit_active;
293 extern counter_u64_t rate_limit_alloc_fail;
294 #endif
295 
296 static int
297 rl_attach_txrtlmt(struct ifnet *ifp,
298     uint32_t flowtype,
299     int flowid,
300     uint64_t cfg_rate,
301     struct m_snd_tag **tag)
302 {
303 	int error;
304 	union if_snd_tag_alloc_params params = {
305 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
306 		.rate_limit.hdr.flowid = flowid,
307 		.rate_limit.hdr.flowtype = flowtype,
308 		.rate_limit.max_rate = cfg_rate,
309 		.rate_limit.flags = M_NOWAIT,
310 	};
311 
312 	if (ifp->if_snd_tag_alloc == NULL) {
313 		error = EOPNOTSUPP;
314 	} else {
315 		error = ifp->if_snd_tag_alloc(ifp, &params, tag);
316 #ifdef INET
317 		if (error == 0) {
318 			if_ref((*tag)->ifp);
319 			counter_u64_add(rate_limit_set_ok, 1);
320 			counter_u64_add(rate_limit_active, 1);
321 		} else
322 			counter_u64_add(rate_limit_alloc_fail, 1);
323 #endif
324 	}
325 	return (error);
326 }
327 
328 static void
329 populate_canned_table(struct tcp_rate_set *rs, const uint64_t *rate_table_act)
330 {
331 	/*
332 	 * The internal table is "special", it
333 	 * is two seperate ordered tables that
334 	 * must be merged. We get here when the
335 	 * adapter specifies a number of rates that
336 	 * covers both ranges in the table in some
337 	 * form.
338 	 */
339 	int i, at_low, at_high;
340 	uint8_t low_disabled = 0, high_disabled = 0;
341 
342 	for(i = 0, at_low = 0, at_high = RS_NEXT_ORDER_GROUP; i < rs->rs_rate_cnt; i++) {
343 		rs->rs_rlt[i].flags = 0;
344 		rs->rs_rlt[i].time_between = 0;
345 		if ((low_disabled == 0) &&
346 		    (high_disabled ||
347 		     (rate_table_act[at_low] < rate_table_act[at_high]))) {
348 			rs->rs_rlt[i].rate = rate_table_act[at_low];
349 			at_low++;
350 			if (at_low == RS_NEXT_ORDER_GROUP)
351 				low_disabled = 1;
352 		} else if (high_disabled == 0) {
353 			rs->rs_rlt[i].rate = rate_table_act[at_high];
354 			at_high++;
355 			if (at_high == MAX_HDWR_RATES)
356 				high_disabled = 1;
357 		}
358 	}
359 }
360 
361 static struct tcp_rate_set *
362 rt_setup_new_rs(struct ifnet *ifp, int *error)
363 {
364 	struct tcp_rate_set *rs;
365 	const uint64_t *rate_table_act;
366 	uint64_t lentim, res;
367 	size_t sz;
368 	uint32_t hash_type;
369 	int i;
370 	struct if_ratelimit_query_results rl;
371 	struct sysctl_oid *rl_sysctl_root;
372 	/*
373 	 * We expect to enter with the
374 	 * mutex locked.
375 	 */
376 
377 	if (ifp->if_ratelimit_query == NULL) {
378 		/*
379 		 * We can do nothing if we cannot
380 		 * get a query back from the driver.
381 		 */
382 		return (NULL);
383 	}
384 	rs = malloc(sizeof(struct tcp_rate_set), M_TCPPACE, M_NOWAIT | M_ZERO);
385 	if (rs == NULL) {
386 		if (error)
387 			*error = ENOMEM;
388 		return (NULL);
389 	}
390 	rl.flags = RT_NOSUPPORT;
391 	ifp->if_ratelimit_query(ifp, &rl);
392 	if (rl.flags & RT_IS_UNUSABLE) {
393 		/*
394 		 * The interface does not really support
395 		 * the rate-limiting.
396 		 */
397 		memset(rs, 0, sizeof(struct tcp_rate_set));
398 		rs->rs_ifp = ifp;
399 		rs->rs_if_dunit = ifp->if_dunit;
400 		rs->rs_flags = RS_INTF_NO_SUP;
401 		rs->rs_disable = 1;
402 		rs_number_alive++;
403 		sysctl_ctx_init(&rs->sysctl_ctx);
404 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
405 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
406 		    OID_AUTO,
407 		    rs->rs_ifp->if_xname,
408 		    CTLFLAG_RW, 0,
409 		    "");
410 		rl_add_syctl_entries(rl_sysctl_root, rs);
411 		mtx_lock(&rs_mtx);
412 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
413 		mtx_unlock(&rs_mtx);
414 		return (rs);
415 	} else if ((rl.flags & RT_IS_INDIRECT) == RT_IS_INDIRECT) {
416 		memset(rs, 0, sizeof(struct tcp_rate_set));
417 		rs->rs_ifp = ifp;
418 		rs->rs_if_dunit = ifp->if_dunit;
419 		rs->rs_flags = RS_IS_DEFF;
420 		rs_number_alive++;
421 		sysctl_ctx_init(&rs->sysctl_ctx);
422 		rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
423 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
424 		    OID_AUTO,
425 		    rs->rs_ifp->if_xname,
426 		    CTLFLAG_RW, 0,
427 		    "");
428 		rl_add_syctl_entries(rl_sysctl_root, rs);
429 		mtx_lock(&rs_mtx);
430 		CK_LIST_INSERT_HEAD(&int_rs, rs, next);
431 		mtx_unlock(&rs_mtx);
432 		return (rs);
433 	} else if ((rl.flags & RT_IS_FIXED_TABLE) == RT_IS_FIXED_TABLE) {
434 		/* Mellanox most likely */
435 		rs->rs_ifp = ifp;
436 		rs->rs_if_dunit = ifp->if_dunit;
437 		rs->rs_rate_cnt = rl.number_of_rates;
438 		rs->rs_min_seg = rl.min_segment_burst;
439 		rs->rs_highest_valid = 0;
440 		rs->rs_flow_limit = rl.max_flows;
441 		rs->rs_flags = RS_IS_INTF | RS_NO_PRE;
442 		rs->rs_disable = 0;
443 		rate_table_act = rl.rate_table;
444 	} else if ((rl.flags & RT_IS_SELECTABLE) == RT_IS_SELECTABLE) {
445 		/* Chelsio */
446 		rs->rs_ifp = ifp;
447 		rs->rs_if_dunit = ifp->if_dunit;
448 		rs->rs_rate_cnt = rl.number_of_rates;
449 		rs->rs_min_seg = rl.min_segment_burst;
450 		rs->rs_disable = 0;
451 		rs->rs_flow_limit = rl.max_flows;
452 		rate_table_act = desired_rates;
453 		if ((rs->rs_rate_cnt > MAX_HDWR_RATES) &&
454 		    (rs->rs_rate_cnt < ALL_HARDWARE_RATES)) {
455 			/*
456 			 * Our desired table is not big
457 			 * enough, do what we can.
458 			 */
459 			rs->rs_rate_cnt = MAX_HDWR_RATES;
460 		 }
461 		if (rs->rs_rate_cnt <= RS_ORDERED_COUNT)
462 			rs->rs_flags = RS_IS_INTF;
463 		else
464 			rs->rs_flags = RS_IS_INTF | RS_INT_TBL;
465 		if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)
466 			rs->rs_rate_cnt = ALL_HARDWARE_RATES;
467 	} else {
468 		printf("Interface:%s unit:%d not one known to have rate-limits\n",
469 		    ifp->if_dname,
470 		    ifp->if_dunit);
471 		free(rs, M_TCPPACE);
472 		return (NULL);
473 	}
474 	sz = sizeof(struct tcp_hwrate_limit_table) * rs->rs_rate_cnt;
475 	rs->rs_rlt = malloc(sz, M_TCPPACE, M_NOWAIT);
476 	if (rs->rs_rlt == NULL) {
477 		if (error)
478 			*error = ENOMEM;
479 bail:
480 		free(rs, M_TCPPACE);
481 		return (NULL);
482 	}
483 	if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) {
484 		/*
485 		 * The interface supports all
486 		 * the rates we could possibly want.
487 		 */
488 		uint64_t rat;
489 
490 		rs->rs_rlt[0].rate = 12500;	/* 100k */
491 		rs->rs_rlt[1].rate = 25000;	/* 200k */
492 		rs->rs_rlt[2].rate = 62500;	/* 500k */
493 		/* Note 125000 == 1Megabit
494 		 * populate 1Meg - 1000meg.
495 		 */
496 		for(i = 3, rat = 125000; i< (ALL_HARDWARE_RATES-1); i++) {
497 			rs->rs_rlt[i].rate = rat;
498 			rat += 125000;
499 		}
500 		rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate = 1250000000;
501 	} else if (rs->rs_flags & RS_INT_TBL) {
502 		/* We populate this in a special way */
503 		populate_canned_table(rs, rate_table_act);
504 	} else {
505 		/*
506 		 * Just copy in the rates from
507 		 * the table, it is in order.
508 		 */
509 		for (i=0; i<rs->rs_rate_cnt; i++) {
510 			rs->rs_rlt[i].rate = rate_table_act[i];
511 			rs->rs_rlt[i].time_between = 0;
512 			rs->rs_rlt[i].flags = 0;
513 		}
514 	}
515 	for (i = (rs->rs_rate_cnt - 1); i >= 0; i--) {
516 		/*
517 		 * We go backwards through the list so that if we can't get
518 		 * a rate and fail to init one, we have at least a chance of
519 		 * getting the highest one.
520 		 */
521 		rs->rs_rlt[i].ptbl = rs;
522 		rs->rs_rlt[i].tag = NULL;
523 		/*
524 		 * Calculate the time between.
525 		 */
526 		lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND;
527 		res = lentim / rs->rs_rlt[i].rate;
528 		if (res > 0)
529 			rs->rs_rlt[i].time_between = res;
530 		else
531 			rs->rs_rlt[i].time_between = 1;
532 		if (rs->rs_flags & RS_NO_PRE) {
533 			rs->rs_rlt[i].flags = HDWRPACE_INITED;
534 			rs->rs_lowest_valid = i;
535 		} else {
536 			int err;
537 #ifdef RSS
538 			hash_type = M_HASHTYPE_RSS_TCP_IPV4;
539 #else
540 			hash_type = M_HASHTYPE_OPAQUE_HASH;
541 #endif
542 			err = rl_attach_txrtlmt(ifp,
543 			    hash_type,
544 			    (i + 1),
545 			    rs->rs_rlt[i].rate,
546 			    &rs->rs_rlt[i].tag);
547 			if (err) {
548 				if (i == (rs->rs_rate_cnt - 1)) {
549 					/*
550 					 * Huh - first rate and we can't get
551 					 * it?
552 					 */
553 					free(rs->rs_rlt, M_TCPPACE);
554 					if (error)
555 						*error = err;
556 					goto bail;
557 				} else {
558 					if (error)
559 						*error = err;
560 				}
561 				break;
562 			} else {
563 				rs->rs_rlt[i].flags = HDWRPACE_INITED | HDWRPACE_TAGPRESENT;
564 				rs->rs_lowest_valid = i;
565 			}
566 		}
567 	}
568 	/* Did we get at least 1 rate? */
569 	if (rs->rs_rlt[(rs->rs_rate_cnt - 1)].flags & HDWRPACE_INITED)
570 		rs->rs_highest_valid = rs->rs_rate_cnt - 1;
571 	else {
572 		free(rs->rs_rlt, M_TCPPACE);
573 		goto bail;
574 	}
575 	rs_number_alive++;
576 	sysctl_ctx_init(&rs->sysctl_ctx);
577 	rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx,
578 	    SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl),
579 	    OID_AUTO,
580 	    rs->rs_ifp->if_xname,
581 	    CTLFLAG_RW, 0,
582 	    "");
583 	rl_add_syctl_entries(rl_sysctl_root, rs);
584 	mtx_lock(&rs_mtx);
585 	CK_LIST_INSERT_HEAD(&int_rs, rs, next);
586 	mtx_unlock(&rs_mtx);
587 	return (rs);
588 }
589 
590 static const struct tcp_hwrate_limit_table *
591 tcp_int_find_suitable_rate(const struct tcp_rate_set *rs,
592     uint64_t bytes_per_sec, uint32_t flags)
593 {
594 	struct tcp_hwrate_limit_table *arte = NULL, *rte = NULL;
595 	uint64_t mbits_per_sec, ind_calc;
596 	int i;
597 
598 	mbits_per_sec = (bytes_per_sec * 8);
599 	if (flags & RS_PACING_LT) {
600 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
601 		    (rs->rs_lowest_valid <= 2)){
602 			/*
603 			 * Smaller than 1Meg, only
604 			 * 3 entries can match it.
605 			 */
606 			for(i = rs->rs_lowest_valid; i < 3; i++) {
607 				if (bytes_per_sec <= rs->rs_rlt[i].rate) {
608 					rte = &rs->rs_rlt[i];
609 					break;
610 				} else if (rs->rs_rlt[i].flags & HDWRPACE_INITED) {
611 					arte = &rs->rs_rlt[i];
612 				}
613 			}
614 			goto done;
615 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
616 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
617 			/*
618 			 * Larger than 1G (the majority of
619 			 * our table.
620 			 */
621 			if (mbits_per_sec < RS_TEN_GIGABIT_PERSEC)
622 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
623 			else
624 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
625 			goto done;
626 		}
627 		/*
628 		 * If we reach here its in our table (between 1Meg - 1000Meg),
629 		 * just take the rounded down mbits per second, and add
630 		 * 1Megabit to it, from this we can calculate
631 		 * the index in the table.
632 		 */
633 		ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
634 		if ((ind_calc * RS_ONE_MEGABIT_PERSEC) != mbits_per_sec)
635 			ind_calc++;
636 		/* our table is offset by 3, we add 2 */
637 		ind_calc += 2;
638 		if (ind_calc > (ALL_HARDWARE_RATES-1)) {
639 			/* This should not happen */
640 			ind_calc = ALL_HARDWARE_RATES-1;
641 		}
642 		if ((ind_calc >= rs->rs_lowest_valid) &&
643 		    (ind_calc <= rs->rs_highest_valid))
644 		rte = &rs->rs_rlt[ind_calc];
645 	} else if (flags & RS_PACING_EXACT_MATCH) {
646 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
647 		    (rs->rs_lowest_valid <= 2)){
648 			for(i = rs->rs_lowest_valid; i < 3; i++) {
649 				if (bytes_per_sec == rs->rs_rlt[i].rate) {
650 					rte = &rs->rs_rlt[i];
651 					break;
652 				}
653 			}
654 		} else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) &&
655 			   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
656 			/* > 1Gbps only one rate */
657 			if (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) {
658 				/* Its 10G wow */
659 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
660 			}
661 		} else {
662 			/* Ok it must be a exact meg (its between 1G and 1Meg) */
663 			ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
664 			if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
665 				/* its an exact Mbps */
666 				ind_calc += 2;
667 				if (ind_calc > (ALL_HARDWARE_RATES-1)) {
668 					/* This should not happen */
669 					ind_calc = ALL_HARDWARE_RATES-1;
670 				}
671 				if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED)
672 					rte = &rs->rs_rlt[ind_calc];
673 			}
674 		}
675 	} else {
676 		/* we want greater than the requested rate */
677 		if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) &&
678 		    (rs->rs_lowest_valid <= 2)){
679 			arte = &rs->rs_rlt[3]; /* set alternate to 1Meg */
680 			for (i=2; i>=rs->rs_lowest_valid; i--) {
681 				if (bytes_per_sec < rs->rs_rlt[i].rate) {
682 					rte = &rs->rs_rlt[i];
683 					break;
684 				} else if ((flags & RS_PACING_GEQ) &&
685 					   (bytes_per_sec == rs->rs_rlt[i].rate)) {
686 					rte = &rs->rs_rlt[i];
687 					break;
688 				} else {
689 					arte = &rs->rs_rlt[i]; /* new alternate */
690 				}
691 			}
692 		} else if (mbits_per_sec > RS_ONE_GIGABIT_PERSEC) {
693 			if ((bytes_per_sec < rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
694 			    (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){
695 				/* Our top rate is larger than the request */
696 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
697 			} else if ((flags & RS_PACING_GEQ) &&
698 				   (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) &&
699 				   (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) {
700 				/* It matches our top rate */
701 				rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
702 			} else if (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED) {
703 				/* The top rate is an alternative */
704 				arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)];
705 			}
706 		} else {
707 			/* Its in our range 1Meg - 1Gig */
708 			if (flags & RS_PACING_GEQ) {
709 				ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC;
710 				if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) {
711 					if (ind_calc > (ALL_HARDWARE_RATES-1)) {
712 						/* This should not happen */
713 						ind_calc = (ALL_HARDWARE_RATES-1);
714 					}
715 					rte = &rs->rs_rlt[ind_calc];
716 				}
717 				goto done;
718 			}
719 			ind_calc = (mbits_per_sec + (RS_ONE_MEGABIT_PERSEC-1))/RS_ONE_MEGABIT_PERSEC;
720 			ind_calc += 2;
721 			if (ind_calc > (ALL_HARDWARE_RATES-1)) {
722 				/* This should not happen */
723 				ind_calc = ALL_HARDWARE_RATES-1;
724 			}
725 			if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED)
726 				rte = &rs->rs_rlt[ind_calc];
727 		}
728 	}
729 done:
730 	if ((rte == NULL) &&
731 	    (arte != NULL) &&
732 	    (flags & RS_PACING_SUB_OK)) {
733 		/* We can use the substitute */
734 		rte = arte;
735 	}
736 	return (rte);
737 }
738 
739 static const struct tcp_hwrate_limit_table *
740 tcp_find_suitable_rate(const struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags)
741 {
742 	/**
743 	 * Hunt the rate table with the restrictions in flags and find a
744 	 * suitable rate if possible.
745 	 * RS_PACING_EXACT_MATCH - look for an exact match to rate.
746 	 * RS_PACING_GT     - must be greater than.
747 	 * RS_PACING_GEQ    - must be greater than or equal.
748 	 * RS_PACING_LT     - must be less than.
749 	 * RS_PACING_SUB_OK - If we don't meet criteria a
750 	 *                    substitute is ok.
751 	 */
752 	int i, matched;
753 	struct tcp_hwrate_limit_table *rte = NULL;
754 
755 
756 	if ((rs->rs_flags & RS_INT_TBL) &&
757 	    (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)) {
758 		/*
759 		 * Here we don't want to paw thru
760 		 * a big table, we have everything
761 		 * from 1Meg - 1000Meg in 1Meg increments.
762 		 * Use an alternate method to "lookup".
763 		 */
764 		return (tcp_int_find_suitable_rate(rs, bytes_per_sec, flags));
765 	}
766 	if ((flags & RS_PACING_LT) ||
767 	    (flags & RS_PACING_EXACT_MATCH)) {
768 		/*
769 		 * For exact and less than we go forward through the table.
770 		 * This way when we find one larger we stop (exact was a
771 		 * toss up).
772 		 */
773 		for (i = rs->rs_lowest_valid, matched = 0; i <= rs->rs_highest_valid; i++) {
774 			if ((flags & RS_PACING_EXACT_MATCH) &&
775 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
776 				rte = &rs->rs_rlt[i];
777 				matched = 1;
778 				break;
779 			} else if ((flags & RS_PACING_LT) &&
780 			    (bytes_per_sec <= rs->rs_rlt[i].rate)) {
781 				rte = &rs->rs_rlt[i];
782 				matched = 1;
783 				break;
784 			}
785 			if (bytes_per_sec > rs->rs_rlt[i].rate)
786 				break;
787 		}
788 		if ((matched == 0) &&
789 		    (flags & RS_PACING_LT) &&
790 		    (flags & RS_PACING_SUB_OK)) {
791 			/* Kick in a substitute (the lowest) */
792 			rte = &rs->rs_rlt[rs->rs_lowest_valid];
793 		}
794 	} else {
795 		/*
796 		 * Here we go backward through the table so that we can find
797 		 * the one greater in theory faster (but its probably a
798 		 * wash).
799 		 */
800 		for (i = rs->rs_highest_valid, matched = 0; i >= rs->rs_lowest_valid; i--) {
801 			if (rs->rs_rlt[i].rate > bytes_per_sec) {
802 				/* A possible candidate */
803 				rte = &rs->rs_rlt[i];
804 			}
805 			if ((flags & RS_PACING_GEQ) &&
806 			    (bytes_per_sec == rs->rs_rlt[i].rate)) {
807 				/* An exact match and we want equal */
808 				matched = 1;
809 				rte = &rs->rs_rlt[i];
810 				break;
811 			} else if (rte) {
812 				/*
813 				 * Found one that is larger than but don't
814 				 * stop, there may be a more closer match.
815 				 */
816 				matched = 1;
817 			}
818 			if (rs->rs_rlt[i].rate < bytes_per_sec) {
819 				/*
820 				 * We found a table entry that is smaller,
821 				 * stop there will be none greater or equal.
822 				 */
823 				break;
824 			}
825 		}
826 		if ((matched == 0) &&
827 		    (flags & RS_PACING_SUB_OK)) {
828 			/* Kick in a substitute (the highest) */
829 			rte = &rs->rs_rlt[rs->rs_highest_valid];
830 		}
831 	}
832 	return (rte);
833 }
834 
835 static struct ifnet *
836 rt_find_real_interface(struct ifnet *ifp, struct inpcb *inp, int *error)
837 {
838 	struct ifnet *tifp;
839 	struct m_snd_tag *tag;
840 	union if_snd_tag_alloc_params params = {
841 		.rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT,
842 		.rate_limit.hdr.flowid = 1,
843 		.rate_limit.max_rate = COMMON_RATE,
844 		.rate_limit.flags = M_NOWAIT,
845 	};
846 	int err;
847 #ifdef RSS
848 	params.rate_limit.hdr.flowtype = ((inp->inp_vflag & INP_IPV6) ?
849 	    M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_TCP_IPV4);
850 #else
851 	params.rate_limit.hdr.flowtype = M_HASHTYPE_OPAQUE_HASH;
852 #endif
853 	tag = NULL;
854 	if (ifp->if_snd_tag_alloc) {
855 		if (error)
856 			*error = ENODEV;
857 		return (NULL);
858 	}
859 	err = ifp->if_snd_tag_alloc(ifp, &params, &tag);
860 	if (err) {
861 		/* Failed to setup a tag? */
862 		if (error)
863 			*error = err;
864 		return (NULL);
865 	}
866 	tifp = tag->ifp;
867 	tifp->if_snd_tag_free(tag);
868 	return (tifp);
869 }
870 
871 static const struct tcp_hwrate_limit_table *
872 rt_setup_rate(struct inpcb *inp, struct ifnet *ifp, uint64_t bytes_per_sec,
873     uint32_t flags, int *error)
874 {
875 	/* First lets find the interface if it exists */
876 	const struct tcp_hwrate_limit_table *rte;
877 	struct tcp_rate_set *rs;
878 	struct epoch_tracker et;
879 	int err;
880 
881 	epoch_enter_preempt(net_epoch_preempt, &et);
882 use_real_interface:
883 	CK_LIST_FOREACH(rs, &int_rs, next) {
884 		/*
885 		 * Note we don't look with the lock since we either see a
886 		 * new entry or will get one when we try to add it.
887 		 */
888 		if (rs->rs_flags & RS_IS_DEAD) {
889 			/* The dead are not looked at */
890 			continue;
891 		}
892 		if ((rs->rs_ifp == ifp) &&
893 		    (rs->rs_if_dunit == ifp->if_dunit)) {
894 			/* Ok we found it */
895 			break;
896 		}
897 	}
898 	if ((rs == NULL) ||
899 	    (rs->rs_flags & RS_INTF_NO_SUP) ||
900 	    (rs->rs_flags & RS_IS_DEAD)) {
901 		/*
902 		 * This means we got a packet *before*
903 		 * the IF-UP was processed below, <or>
904 		 * while or after we already received an interface
905 		 * departed event. In either case we really don't
906 		 * want to do anything with pacing, in
907 		 * the departing case the packet is not
908 		 * going to go very far. The new case
909 		 * might be arguable, but its impossible
910 		 * to tell from the departing case.
911 		 */
912 		if (rs->rs_disable && error)
913 			*error = ENODEV;
914 		epoch_exit_preempt(net_epoch_preempt, &et);
915 		return (NULL);
916 	}
917 
918 	if ((rs == NULL) || (rs->rs_disable != 0)) {
919 		if (rs->rs_disable && error)
920 			*error = ENOSPC;
921 		epoch_exit_preempt(net_epoch_preempt, &et);
922 		return (NULL);
923 	}
924 	if (rs->rs_flags & RS_IS_DEFF) {
925 		/* We need to find the real interface */
926 		struct ifnet *tifp;
927 
928 		tifp = rt_find_real_interface(ifp, inp, error);
929 		if (tifp == NULL) {
930 			if (rs->rs_disable && error)
931 				*error = ENOTSUP;
932 			epoch_exit_preempt(net_epoch_preempt, &et);
933 			return (NULL);
934 		}
935 		goto use_real_interface;
936 	}
937 	if (rs->rs_flow_limit &&
938 	    ((rs->rs_flows_using + 1) > rs->rs_flow_limit)) {
939 		if (error)
940 			*error = ENOSPC;
941 		epoch_exit_preempt(net_epoch_preempt, &et);
942 		return (NULL);
943 	}
944 	rte = tcp_find_suitable_rate(rs, bytes_per_sec, flags);
945 	if (rte) {
946 		err = in_pcbattach_txrtlmt(inp, rs->rs_ifp,
947 		    inp->inp_flowtype,
948 		    inp->inp_flowid,
949 		    rte->rate,
950 		    &inp->inp_snd_tag);
951 		if (err) {
952 			/* Failed to attach */
953 			if (error)
954 				*error = err;
955 			rte = NULL;
956 		}
957 	}
958 	if (rte) {
959 		/*
960 		 * We use an atomic here for accounting so we don't have to
961 		 * use locks when freeing.
962 		 */
963 		atomic_add_64(&rs->rs_flows_using, 1);
964 	}
965 	epoch_exit_preempt(net_epoch_preempt, &et);
966 	return (rte);
967 }
968 
969 static void
970 tcp_rl_ifnet_link(void *arg __unused, struct ifnet *ifp, int link_state)
971 {
972 	int error;
973 	struct tcp_rate_set *rs;
974 
975 	if (((ifp->if_capabilities & IFCAP_TXRTLMT) == 0) ||
976 	    (link_state != LINK_STATE_UP)) {
977 		/*
978 		 * We only care on an interface going up that is rate-limit
979 		 * capable.
980 		 */
981 		return;
982 	}
983 	mtx_lock(&rs_mtx);
984 	CK_LIST_FOREACH(rs, &int_rs, next) {
985 		if ((rs->rs_ifp == ifp) &&
986 		    (rs->rs_if_dunit == ifp->if_dunit)) {
987 			/* We already have initialized this guy */
988 			mtx_unlock(&rs_mtx);
989 			return;
990 		}
991 	}
992 	mtx_unlock(&rs_mtx);
993 	rt_setup_new_rs(ifp, &error);
994 }
995 
996 static void
997 tcp_rl_ifnet_departure(void *arg __unused, struct ifnet *ifp)
998 {
999 	struct tcp_rate_set *rs, *nrs;
1000 	struct ifnet *tifp;
1001 	int i;
1002 
1003 	mtx_lock(&rs_mtx);
1004 	CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) {
1005 		if ((rs->rs_ifp == ifp) &&
1006 		    (rs->rs_if_dunit == ifp->if_dunit)) {
1007 			CK_LIST_REMOVE(rs, next);
1008 			rs_number_alive--;
1009 			rs->rs_flags |= RS_IS_DEAD;
1010 			for (i = 0; i < rs->rs_rate_cnt; i++) {
1011 				if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1012 					tifp = rs->rs_rlt[i].tag->ifp;
1013 					in_pcbdetach_tag(tifp, rs->rs_rlt[i].tag);
1014 					rs->rs_rlt[i].tag = NULL;
1015 				}
1016 				rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1017 			}
1018 			if (rs->rs_flows_using == 0)
1019 				rs_defer_destroy(rs);
1020 			break;
1021 		}
1022 	}
1023 	mtx_unlock(&rs_mtx);
1024 }
1025 
1026 static void
1027 tcp_rl_shutdown(void *arg __unused, int howto __unused)
1028 {
1029 	struct tcp_rate_set *rs, *nrs;
1030 	struct ifnet *tifp;
1031 	int i;
1032 
1033 	mtx_lock(&rs_mtx);
1034 	CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) {
1035 		CK_LIST_REMOVE(rs, next);
1036 		rs_number_alive--;
1037 		rs->rs_flags |= RS_IS_DEAD;
1038 		for (i = 0; i < rs->rs_rate_cnt; i++) {
1039 			if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) {
1040 				tifp = rs->rs_rlt[i].tag->ifp;
1041 				in_pcbdetach_tag(tifp, rs->rs_rlt[i].tag);
1042 				rs->rs_rlt[i].tag = NULL;
1043 			}
1044 			rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED;
1045 		}
1046 		if (rs->rs_flows_using == 0)
1047 			rs_defer_destroy(rs);
1048 	}
1049 	mtx_unlock(&rs_mtx);
1050 }
1051 
1052 const struct tcp_hwrate_limit_table *
1053 tcp_set_pacing_rate(struct tcpcb *tp, struct ifnet *ifp,
1054     uint64_t bytes_per_sec, int flags, int *error)
1055 {
1056 	const struct tcp_hwrate_limit_table *rte;
1057 
1058 	if (tp->t_inpcb->inp_snd_tag == NULL) {
1059 		/*
1060 		 * We are setting up a rate for the first time.
1061 		 */
1062 		if ((ifp->if_capabilities & IFCAP_TXRTLMT) == 0) {
1063 			/* Not supported by the egress */
1064 			if (error)
1065 				*error = ENODEV;
1066 			return (NULL);
1067 		}
1068 #ifdef KERN_TLS
1069 		if (tp->t_inpcb->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
1070 			/*
1071 			 * We currently can't do both TLS and hardware
1072 			 * pacing
1073 			 */
1074 			if (error)
1075 				*error = EINVAL;
1076 			return (NULL);
1077 		}
1078 #endif
1079 		rte = rt_setup_rate(tp->t_inpcb, ifp, bytes_per_sec, flags, error);
1080 	} else {
1081 		/*
1082 		 * We are modifying a rate, wrong interface?
1083 		 */
1084 		if (error)
1085 			*error = EINVAL;
1086 		rte = NULL;
1087 	}
1088 	return (rte);
1089 }
1090 
1091 const struct tcp_hwrate_limit_table *
1092 tcp_chg_pacing_rate(const struct tcp_hwrate_limit_table *crte,
1093     struct tcpcb *tp, struct ifnet *ifp,
1094     uint64_t bytes_per_sec, int flags, int *error)
1095 {
1096 	const struct tcp_hwrate_limit_table *nrte;
1097 	const struct tcp_rate_set *rs;
1098 	int is_indirect = 0;
1099 	int err;
1100 
1101 
1102 	if ((tp->t_inpcb->inp_snd_tag == NULL) ||
1103 	    (crte == NULL)) {
1104 		/* Wrong interface */
1105 		if (error)
1106 			*error = EINVAL;
1107 		return (NULL);
1108 	}
1109 	rs = crte->ptbl;
1110 	if ((rs->rs_flags & RS_IS_DEAD) ||
1111 	    (crte->flags & HDWRPACE_IFPDEPARTED)) {
1112 		/* Release the rate, and try anew */
1113 re_rate:
1114 		tcp_rel_pacing_rate(crte, tp);
1115 		nrte = tcp_set_pacing_rate(tp, ifp,
1116 		    bytes_per_sec, flags, error);
1117 		return (nrte);
1118 	}
1119 	if ((rs->rs_flags & RT_IS_INDIRECT ) == RT_IS_INDIRECT)
1120 		is_indirect = 1;
1121 	else
1122 		is_indirect = 0;
1123 	if ((is_indirect == 0) &&
1124 	    ((ifp != rs->rs_ifp) ||
1125 	    (ifp->if_dunit != rs->rs_if_dunit))) {
1126 		/*
1127 		 * Something changed, the user is not pointing to the same
1128 		 * ifp? Maybe a route updated on this guy?
1129 		 */
1130 		goto re_rate;
1131 	} else if (is_indirect) {
1132 		/*
1133 		 * For indirect we have to dig in and find the real interface.
1134 		 */
1135 		struct ifnet *rifp;
1136 
1137 		rifp = rt_find_real_interface(ifp, tp->t_inpcb, error);
1138 		if (rifp == NULL) {
1139 			/* Can't find it? */
1140 			goto re_rate;
1141 		}
1142 		if ((rifp != rs->rs_ifp) ||
1143 		    (ifp->if_dunit != rs->rs_if_dunit)) {
1144 			goto re_rate;
1145 		}
1146 	}
1147 	nrte = tcp_find_suitable_rate(rs, bytes_per_sec, flags);
1148 	if (nrte == crte) {
1149 		/* No change */
1150 		if (error)
1151 			*error = 0;
1152 		return (crte);
1153 	}
1154 	if (nrte == NULL) {
1155 		/* Release the old rate */
1156 		tcp_rel_pacing_rate(crte, tp);
1157 		return (NULL);
1158 	}
1159 	/* Change rates to our new entry */
1160 	err = in_pcbmodify_txrtlmt(tp->t_inpcb, nrte->rate);
1161 	if (err) {
1162 		if (error)
1163 			*error = err;
1164 		return (NULL);
1165 	}
1166 	if (error)
1167 		*error = 0;
1168 	return (nrte);
1169 }
1170 
1171 void
1172 tcp_rel_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp)
1173 {
1174 	const struct tcp_rate_set *crs;
1175 	struct tcp_rate_set *rs;
1176 	uint64_t pre;
1177 
1178 	crs = crte->ptbl;
1179 	/*
1180 	 * Now we must break the const
1181 	 * in order to release our refcount.
1182 	 */
1183 	rs = __DECONST(struct tcp_rate_set *, crs);
1184 	pre = atomic_fetchadd_64(&rs->rs_flows_using, -1);
1185 	if (pre == 1) {
1186 		mtx_lock(&rs_mtx);
1187 		/*
1188 		 * Is it dead?
1189 		 */
1190 		if (rs->rs_flags & RS_IS_DEAD)
1191 			rs_defer_destroy(rs);
1192 		mtx_unlock(&rs_mtx);
1193 	}
1194 	in_pcbdetach_txrtlmt(tp->t_inpcb);
1195 }
1196 
1197 static eventhandler_tag rl_ifnet_departs;
1198 static eventhandler_tag rl_ifnet_arrives;
1199 static eventhandler_tag rl_shutdown_start;
1200 
1201 static void
1202 tcp_rs_init(void *st __unused)
1203 {
1204 	CK_LIST_INIT(&int_rs);
1205 	rs_number_alive = 0;
1206 	rs_number_dead = 0;;
1207 	mtx_init(&rs_mtx, "tcp_rs_mtx", "rsmtx", MTX_DEF);
1208 	rl_ifnet_departs = EVENTHANDLER_REGISTER(ifnet_departure_event,
1209 	    tcp_rl_ifnet_departure,
1210 	    NULL, EVENTHANDLER_PRI_ANY);
1211 	rl_ifnet_arrives = EVENTHANDLER_REGISTER(ifnet_link_event,
1212 	    tcp_rl_ifnet_link,
1213 	    NULL, EVENTHANDLER_PRI_ANY);
1214 	rl_shutdown_start = EVENTHANDLER_REGISTER(shutdown_pre_sync,
1215 	    tcp_rl_shutdown, NULL,
1216 	    SHUTDOWN_PRI_FIRST);
1217 	printf("TCP_ratelimit: Is now initialized\n");
1218 }
1219 
1220 SYSINIT(tcp_rl_init, SI_SUB_SMP + 1, SI_ORDER_ANY, tcp_rs_init, NULL);
1221 #endif
1222