1 /*- 2 * 3 * SPDX-License-Identifier: BSD-3-Clause 4 * 5 * Copyright (c) 2018-2020 6 * Netflix Inc. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 * 29 */ 30 /** 31 * Author: Randall Stewart <rrs@netflix.com> 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_tcpdebug.h" 40 #include "opt_ratelimit.h" 41 #include <sys/param.h> 42 #include <sys/kernel.h> 43 #include <sys/malloc.h> 44 #include <sys/mbuf.h> 45 #include <sys/socket.h> 46 #include <sys/socketvar.h> 47 #include <sys/sysctl.h> 48 #include <sys/eventhandler.h> 49 #include <sys/mutex.h> 50 #include <sys/ck.h> 51 #include <net/if.h> 52 #include <net/if_var.h> 53 #include <netinet/in.h> 54 #include <netinet/in_pcb.h> 55 #define TCPSTATES /* for logging */ 56 #include <netinet/tcp_var.h> 57 #ifdef INET6 58 #include <netinet6/tcp6_var.h> 59 #endif 60 #include <netinet/tcp_ratelimit.h> 61 #ifndef USECS_IN_SECOND 62 #define USECS_IN_SECOND 1000000 63 #endif 64 /* 65 * For the purposes of each send, what is the size 66 * of an ethernet frame. 67 */ 68 MALLOC_DEFINE(M_TCPPACE, "tcp_hwpace", "TCP Hardware pacing memory"); 69 #ifdef RATELIMIT 70 71 /* 72 * The following preferred table will seem weird to 73 * the casual viewer. Why do we not have any rates below 74 * 1Mbps? Why do we have a rate at 1.44Mbps called common? 75 * Why do the rates cluster in the 1-100Mbps range more 76 * than others? Why does the table jump around at the beginnign 77 * and then be more consistently raising? 78 * 79 * Let me try to answer those questions. A lot of 80 * this is dependant on the hardware. We have three basic 81 * supporters of rate limiting 82 * 83 * Chelsio - Supporting 16 configurable rates. 84 * Mlx - c4 supporting 13 fixed rates. 85 * Mlx - c5 & c6 supporting 127 configurable rates. 86 * 87 * The c4 is why we have a common rate that is available 88 * in all rate tables. This is a selected rate from the 89 * c4 table and we assure its available in all ratelimit 90 * tables. This way the tcp_ratelimit code has an assured 91 * rate it should always be able to get. This answers a 92 * couple of the questions above. 93 * 94 * So what about the rest, well the table is built to 95 * try to get the most out of a joint hardware/software 96 * pacing system. The software pacer will always pick 97 * a rate higher than the b/w that it is estimating 98 * 99 * on the path. This is done for two reasons. 100 * a) So we can discover more b/w 101 * and 102 * b) So we can send a block of MSS's down and then 103 * have the software timer go off after the previous 104 * send is completely out of the hardware. 105 * 106 * But when we do <b> we don't want to have the delay 107 * between the last packet sent by the hardware be 108 * excessively long (to reach our desired rate). 109 * 110 * So let me give an example for clarity. 111 * 112 * Lets assume that the tcp stack sees that 29,110,000 bps is 113 * what the bw of the path is. The stack would select the 114 * rate 31Mbps. 31Mbps means that each send that is done 115 * by the hardware will cause a 390 micro-second gap between 116 * the packets sent at that rate. For 29,110,000 bps we 117 * would need 416 micro-seconds gap between each send. 118 * 119 * Note that are calculating a complete time for pacing 120 * which includes the ethernet, IP and TCP overhead. So 121 * a full 1514 bytes is used for the above calculations. 122 * My testing has shown that both cards are also using this 123 * as their basis i.e. full payload size of the ethernet frame. 124 * The TCP stack caller needs to be aware of this and make the 125 * appropriate overhead calculations be included in its choices. 126 * 127 * Now, continuing our example, we pick a MSS size based on the 128 * delta between the two rates (416 - 390) divided into the rate 129 * we really wish to send at rounded up. That results in a MSS 130 * send of 17 mss's at once. The hardware then will 131 * run out of data in a single 17MSS send in 6,630 micro-seconds. 132 * 133 * On the other hand the software pacer will send more data 134 * in 7,072 micro-seconds. This means that we will refill 135 * the hardware 52 microseconds after it would have sent 136 * next if it had not ran out of data. This is a win since we are 137 * only sending every 7ms or so and yet all the packets are spaced on 138 * the wire with 94% of what they should be and only 139 * the last packet is delayed extra to make up for the 140 * difference. 141 * 142 * Note that the above formula has two important caveat. 143 * If we are above (b/w wise) over 100Mbps we double the result 144 * of the MSS calculation. The second caveat is if we are 500Mbps 145 * or more we just send the maximum MSS at once i.e. 45MSS. At 146 * the higher b/w's even the cards have limits to what times (timer granularity) 147 * they can insert between packets and start to send more than one 148 * packet at a time on the wire. 149 * 150 */ 151 #define COMMON_RATE 180500 152 const uint64_t desired_rates[] = { 153 122500, /* 1Mbps - rate 1 */ 154 180500, /* 1.44Mpbs - rate 2 common rate */ 155 375000, /* 3Mbps - rate 3 */ 156 625000, /* 5Mbps - rate 4 */ 157 875000, /* 7Mbps - rate 5 */ 158 1125000, /* 9Mbps - rate 6 */ 159 1375000, /* 11Mbps - rate 7 */ 160 1625000, /* 13Mbps - rate 8 */ 161 2625000, /* 21Mbps - rate 9 */ 162 3875000, /* 31Mbps - rate 10 */ 163 5125000, /* 41Meg - rate 11 */ 164 12500000, /* 100Mbps - rate 12 */ 165 25000000, /* 200Mbps - rate 13 */ 166 50000000, /* 400Mbps - rate 14 */ 167 63750000, /* 51Mbps - rate 15 */ 168 100000000, /* 800Mbps - rate 16 */ 169 1875000, /* 15Mbps - rate 17 */ 170 2125000, /* 17Mbps - rate 18 */ 171 2375000, /* 19Mbps - rate 19 */ 172 2875000, /* 23Mbps - rate 20 */ 173 3125000, /* 25Mbps - rate 21 */ 174 3375000, /* 27Mbps - rate 22 */ 175 3625000, /* 29Mbps - rate 23 */ 176 4125000, /* 33Mbps - rate 24 */ 177 4375000, /* 35Mbps - rate 25 */ 178 4625000, /* 37Mbps - rate 26 */ 179 4875000, /* 39Mbps - rate 27 */ 180 5375000, /* 43Mbps - rate 28 */ 181 5625000, /* 45Mbps - rate 29 */ 182 5875000, /* 47Mbps - rate 30 */ 183 6125000, /* 49Mbps - rate 31 */ 184 6625000, /* 53Mbps - rate 32 */ 185 6875000, /* 55Mbps - rate 33 */ 186 7125000, /* 57Mbps - rate 34 */ 187 7375000, /* 59Mbps - rate 35 */ 188 7625000, /* 61Mbps - rate 36 */ 189 7875000, /* 63Mbps - rate 37 */ 190 8125000, /* 65Mbps - rate 38 */ 191 8375000, /* 67Mbps - rate 39 */ 192 8625000, /* 69Mbps - rate 40 */ 193 8875000, /* 71Mbps - rate 41 */ 194 9125000, /* 73Mbps - rate 42 */ 195 9375000, /* 75Mbps - rate 43 */ 196 9625000, /* 77Mbps - rate 44 */ 197 9875000, /* 79Mbps - rate 45 */ 198 10125000, /* 81Mbps - rate 46 */ 199 10375000, /* 83Mbps - rate 47 */ 200 10625000, /* 85Mbps - rate 48 */ 201 10875000, /* 87Mbps - rate 49 */ 202 11125000, /* 89Mbps - rate 50 */ 203 11375000, /* 91Mbps - rate 51 */ 204 11625000, /* 93Mbps - rate 52 */ 205 11875000, /* 95Mbps - rate 53 */ 206 13125000, /* 105Mbps - rate 54 */ 207 13750000, /* 110Mbps - rate 55 */ 208 14375000, /* 115Mbps - rate 56 */ 209 15000000, /* 120Mbps - rate 57 */ 210 15625000, /* 125Mbps - rate 58 */ 211 16250000, /* 130Mbps - rate 59 */ 212 16875000, /* 135Mbps - rate 60 */ 213 17500000, /* 140Mbps - rate 61 */ 214 18125000, /* 145Mbps - rate 62 */ 215 18750000, /* 150Mbps - rate 64 */ 216 20000000, /* 160Mbps - rate 65 */ 217 21250000, /* 170Mbps - rate 66 */ 218 22500000, /* 180Mbps - rate 67 */ 219 23750000, /* 190Mbps - rate 68 */ 220 26250000, /* 210Mbps - rate 69 */ 221 27500000, /* 220Mbps - rate 70 */ 222 28750000, /* 230Mbps - rate 71 */ 223 30000000, /* 240Mbps - rate 72 */ 224 31250000, /* 250Mbps - rate 73 */ 225 34375000, /* 275Mbps - rate 74 */ 226 37500000, /* 300Mbps - rate 75 */ 227 40625000, /* 325Mbps - rate 76 */ 228 43750000, /* 350Mbps - rate 77 */ 229 46875000, /* 375Mbps - rate 78 */ 230 53125000, /* 425Mbps - rate 79 */ 231 56250000, /* 450Mbps - rate 80 */ 232 59375000, /* 475Mbps - rate 81 */ 233 62500000, /* 500Mbps - rate 82 */ 234 68750000, /* 550Mbps - rate 83 */ 235 75000000, /* 600Mbps - rate 84 */ 236 81250000, /* 650Mbps - rate 85 */ 237 87500000, /* 700Mbps - rate 86 */ 238 93750000, /* 750Mbps - rate 87 */ 239 106250000, /* 850Mbps - rate 88 */ 240 112500000, /* 900Mbps - rate 89 */ 241 125000000, /* 1Gbps - rate 90 */ 242 156250000, /* 1.25Gps - rate 91 */ 243 187500000, /* 1.5Gps - rate 92 */ 244 218750000, /* 1.75Gps - rate 93 */ 245 250000000, /* 2Gbps - rate 94 */ 246 281250000, /* 2.25Gps - rate 95 */ 247 312500000, /* 2.5Gbps - rate 96 */ 248 343750000, /* 2.75Gbps - rate 97 */ 249 375000000, /* 3Gbps - rate 98 */ 250 500000000, /* 4Gbps - rate 99 */ 251 625000000, /* 5Gbps - rate 100 */ 252 750000000, /* 6Gbps - rate 101 */ 253 875000000, /* 7Gbps - rate 102 */ 254 1000000000, /* 8Gbps - rate 103 */ 255 1125000000, /* 9Gbps - rate 104 */ 256 1250000000, /* 10Gbps - rate 105 */ 257 1875000000, /* 15Gbps - rate 106 */ 258 2500000000 /* 20Gbps - rate 107 */ 259 }; 260 261 #define MAX_HDWR_RATES (sizeof(desired_rates)/sizeof(uint64_t)) 262 #define RS_ORDERED_COUNT 16 /* 263 * Number that are in order 264 * at the beginning of the table, 265 * over this a sort is required. 266 */ 267 #define RS_NEXT_ORDER_GROUP 16 /* 268 * The point in our table where 269 * we come fill in a second ordered 270 * group (index wise means -1). 271 */ 272 #define ALL_HARDWARE_RATES 1004 /* 273 * 1Meg - 1Gig in 1 Meg steps 274 * plus 100, 200k and 500k and 275 * 10Gig 276 */ 277 278 #define RS_ONE_MEGABIT_PERSEC 1000000 279 #define RS_ONE_GIGABIT_PERSEC 1000000000 280 #define RS_TEN_GIGABIT_PERSEC 10000000000 281 282 static struct head_tcp_rate_set int_rs; 283 static struct mtx rs_mtx; 284 uint32_t rs_number_alive; 285 uint32_t rs_number_dead; 286 287 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, rl, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 288 "TCP Ratelimit stats"); 289 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, alive, CTLFLAG_RW, 290 &rs_number_alive, 0, 291 "Number of interfaces initialized for ratelimiting"); 292 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, dead, CTLFLAG_RW, 293 &rs_number_dead, 0, 294 "Number of interfaces departing from ratelimiting"); 295 296 static void 297 rl_add_syctl_entries(struct sysctl_oid *rl_sysctl_root, struct tcp_rate_set *rs) 298 { 299 /* 300 * Add sysctl entries for thus interface. 301 */ 302 if (rs->rs_flags & RS_INTF_NO_SUP) { 303 SYSCTL_ADD_S32(&rs->sysctl_ctx, 304 SYSCTL_CHILDREN(rl_sysctl_root), 305 OID_AUTO, "disable", CTLFLAG_RD, 306 &rs->rs_disable, 0, 307 "Disable this interface from new hdwr limiting?"); 308 } else { 309 SYSCTL_ADD_S32(&rs->sysctl_ctx, 310 SYSCTL_CHILDREN(rl_sysctl_root), 311 OID_AUTO, "disable", CTLFLAG_RW, 312 &rs->rs_disable, 0, 313 "Disable this interface from new hdwr limiting?"); 314 } 315 SYSCTL_ADD_S32(&rs->sysctl_ctx, 316 SYSCTL_CHILDREN(rl_sysctl_root), 317 OID_AUTO, "minseg", CTLFLAG_RW, 318 &rs->rs_min_seg, 0, 319 "What is the minimum we need to send on this interface?"); 320 SYSCTL_ADD_U64(&rs->sysctl_ctx, 321 SYSCTL_CHILDREN(rl_sysctl_root), 322 OID_AUTO, "flow_limit", CTLFLAG_RW, 323 &rs->rs_flow_limit, 0, 324 "What is the limit for number of flows (0=unlimited)?"); 325 SYSCTL_ADD_S32(&rs->sysctl_ctx, 326 SYSCTL_CHILDREN(rl_sysctl_root), 327 OID_AUTO, "highest", CTLFLAG_RD, 328 &rs->rs_highest_valid, 0, 329 "Highest valid rate"); 330 SYSCTL_ADD_S32(&rs->sysctl_ctx, 331 SYSCTL_CHILDREN(rl_sysctl_root), 332 OID_AUTO, "lowest", CTLFLAG_RD, 333 &rs->rs_lowest_valid, 0, 334 "Lowest valid rate"); 335 SYSCTL_ADD_S32(&rs->sysctl_ctx, 336 SYSCTL_CHILDREN(rl_sysctl_root), 337 OID_AUTO, "flags", CTLFLAG_RD, 338 &rs->rs_flags, 0, 339 "What lags are on the entry?"); 340 SYSCTL_ADD_S32(&rs->sysctl_ctx, 341 SYSCTL_CHILDREN(rl_sysctl_root), 342 OID_AUTO, "numrates", CTLFLAG_RD, 343 &rs->rs_rate_cnt, 0, 344 "How many rates re there?"); 345 SYSCTL_ADD_U64(&rs->sysctl_ctx, 346 SYSCTL_CHILDREN(rl_sysctl_root), 347 OID_AUTO, "flows_using", CTLFLAG_RD, 348 &rs->rs_flows_using, 0, 349 "How many flows are using this interface now?"); 350 #ifdef DETAILED_RATELIMIT_SYSCTL 351 if (rs->rs_rlt && rs->rs_rate_cnt > 0) { 352 /* Lets display the rates */ 353 int i; 354 struct sysctl_oid *rl_rates; 355 struct sysctl_oid *rl_rate_num; 356 char rate_num[16]; 357 rl_rates = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 358 SYSCTL_CHILDREN(rl_sysctl_root), 359 OID_AUTO, 360 "rate", 361 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 362 "Ratelist"); 363 for( i = 0; i < rs->rs_rate_cnt; i++) { 364 sprintf(rate_num, "%d", i); 365 rl_rate_num = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 366 SYSCTL_CHILDREN(rl_rates), 367 OID_AUTO, 368 rate_num, 369 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 370 "Individual Rate"); 371 SYSCTL_ADD_U32(&rs->sysctl_ctx, 372 SYSCTL_CHILDREN(rl_rate_num), 373 OID_AUTO, "flags", CTLFLAG_RD, 374 &rs->rs_rlt[i].flags, 0, 375 "Flags on this rate"); 376 SYSCTL_ADD_U32(&rs->sysctl_ctx, 377 SYSCTL_CHILDREN(rl_rate_num), 378 OID_AUTO, "pacetime", CTLFLAG_RD, 379 &rs->rs_rlt[i].time_between, 0, 380 "Time hardware inserts between 1500 byte sends"); 381 SYSCTL_ADD_U64(&rs->sysctl_ctx, 382 SYSCTL_CHILDREN(rl_rate_num), 383 OID_AUTO, "rate", CTLFLAG_RD, 384 &rs->rs_rlt[i].rate, 0, 385 "Rate in bytes per second"); 386 } 387 } 388 #endif 389 } 390 391 static void 392 rs_destroy(epoch_context_t ctx) 393 { 394 struct tcp_rate_set *rs; 395 bool do_free_rs; 396 397 rs = __containerof(ctx, struct tcp_rate_set, rs_epoch_ctx); 398 399 mtx_lock(&rs_mtx); 400 rs->rs_flags &= ~RS_FUNERAL_SCHD; 401 /* 402 * In theory its possible (but unlikely) 403 * that while the delete was occuring 404 * and we were applying the DEAD flag 405 * someone slipped in and found the 406 * interface in a lookup. While we 407 * decided rs_flows_using were 0 and 408 * scheduling the epoch_call, the other 409 * thread incremented rs_flow_using. This 410 * is because users have a pointer and 411 * we only use the rs_flows_using in an 412 * atomic fashion, i.e. the other entities 413 * are not protected. To assure this did 414 * not occur, we check rs_flows_using here 415 * before deleting. 416 */ 417 do_free_rs = (rs->rs_flows_using == 0); 418 rs_number_dead--; 419 mtx_unlock(&rs_mtx); 420 421 if (do_free_rs) { 422 sysctl_ctx_free(&rs->sysctl_ctx); 423 free(rs->rs_rlt, M_TCPPACE); 424 free(rs, M_TCPPACE); 425 } 426 } 427 428 static void 429 rs_defer_destroy(struct tcp_rate_set *rs) 430 { 431 432 mtx_assert(&rs_mtx, MA_OWNED); 433 434 /* Check if already pending. */ 435 if (rs->rs_flags & RS_FUNERAL_SCHD) 436 return; 437 438 rs_number_dead++; 439 440 /* Set flag to only defer once. */ 441 rs->rs_flags |= RS_FUNERAL_SCHD; 442 NET_EPOCH_CALL(rs_destroy, &rs->rs_epoch_ctx); 443 } 444 445 #ifdef INET 446 extern counter_u64_t rate_limit_set_ok; 447 extern counter_u64_t rate_limit_active; 448 extern counter_u64_t rate_limit_alloc_fail; 449 #endif 450 451 static int 452 rl_attach_txrtlmt(struct ifnet *ifp, 453 uint32_t flowtype, 454 int flowid, 455 uint64_t cfg_rate, 456 struct m_snd_tag **tag) 457 { 458 int error; 459 union if_snd_tag_alloc_params params = { 460 .rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT, 461 .rate_limit.hdr.flowid = flowid, 462 .rate_limit.hdr.flowtype = flowtype, 463 .rate_limit.max_rate = cfg_rate, 464 .rate_limit.flags = M_NOWAIT, 465 }; 466 467 if (ifp->if_snd_tag_alloc == NULL) { 468 error = EOPNOTSUPP; 469 } else { 470 error = ifp->if_snd_tag_alloc(ifp, ¶ms, tag); 471 #ifdef INET 472 if (error == 0) { 473 if_ref((*tag)->ifp); 474 counter_u64_add(rate_limit_set_ok, 1); 475 counter_u64_add(rate_limit_active, 1); 476 } else 477 counter_u64_add(rate_limit_alloc_fail, 1); 478 #endif 479 } 480 return (error); 481 } 482 483 static void 484 populate_canned_table(struct tcp_rate_set *rs, const uint64_t *rate_table_act) 485 { 486 /* 487 * The internal table is "special", it 488 * is two seperate ordered tables that 489 * must be merged. We get here when the 490 * adapter specifies a number of rates that 491 * covers both ranges in the table in some 492 * form. 493 */ 494 int i, at_low, at_high; 495 uint8_t low_disabled = 0, high_disabled = 0; 496 497 for(i = 0, at_low = 0, at_high = RS_NEXT_ORDER_GROUP; i < rs->rs_rate_cnt; i++) { 498 rs->rs_rlt[i].flags = 0; 499 rs->rs_rlt[i].time_between = 0; 500 if ((low_disabled == 0) && 501 (high_disabled || 502 (rate_table_act[at_low] < rate_table_act[at_high]))) { 503 rs->rs_rlt[i].rate = rate_table_act[at_low]; 504 at_low++; 505 if (at_low == RS_NEXT_ORDER_GROUP) 506 low_disabled = 1; 507 } else if (high_disabled == 0) { 508 rs->rs_rlt[i].rate = rate_table_act[at_high]; 509 at_high++; 510 if (at_high == MAX_HDWR_RATES) 511 high_disabled = 1; 512 } 513 } 514 } 515 516 static struct tcp_rate_set * 517 rt_setup_new_rs(struct ifnet *ifp, int *error) 518 { 519 struct tcp_rate_set *rs; 520 const uint64_t *rate_table_act; 521 uint64_t lentim, res; 522 size_t sz; 523 uint32_t hash_type; 524 int i; 525 struct if_ratelimit_query_results rl; 526 struct sysctl_oid *rl_sysctl_root; 527 /* 528 * We expect to enter with the 529 * mutex locked. 530 */ 531 532 if (ifp->if_ratelimit_query == NULL) { 533 /* 534 * We can do nothing if we cannot 535 * get a query back from the driver. 536 */ 537 printf("Warning:No query functions for %s:%d-- failed\n", 538 ifp->if_dname, ifp->if_dunit); 539 return (NULL); 540 } 541 rs = malloc(sizeof(struct tcp_rate_set), M_TCPPACE, M_NOWAIT | M_ZERO); 542 if (rs == NULL) { 543 if (error) 544 *error = ENOMEM; 545 printf("Warning:No memory for malloc of tcp_rate_set\n"); 546 return (NULL); 547 } 548 memset(&rl, 0, sizeof(rl)); 549 rl.flags = RT_NOSUPPORT; 550 ifp->if_ratelimit_query(ifp, &rl); 551 if (rl.flags & RT_IS_UNUSABLE) { 552 /* 553 * The interface does not really support 554 * the rate-limiting. 555 */ 556 memset(rs, 0, sizeof(struct tcp_rate_set)); 557 rs->rs_ifp = ifp; 558 rs->rs_if_dunit = ifp->if_dunit; 559 rs->rs_flags = RS_INTF_NO_SUP; 560 rs->rs_disable = 1; 561 rs_number_alive++; 562 sysctl_ctx_init(&rs->sysctl_ctx); 563 rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 564 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), 565 OID_AUTO, 566 rs->rs_ifp->if_xname, 567 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 568 ""); 569 rl_add_syctl_entries(rl_sysctl_root, rs); 570 mtx_lock(&rs_mtx); 571 CK_LIST_INSERT_HEAD(&int_rs, rs, next); 572 mtx_unlock(&rs_mtx); 573 return (rs); 574 } else if ((rl.flags & RT_IS_INDIRECT) == RT_IS_INDIRECT) { 575 memset(rs, 0, sizeof(struct tcp_rate_set)); 576 rs->rs_ifp = ifp; 577 rs->rs_if_dunit = ifp->if_dunit; 578 rs->rs_flags = RS_IS_DEFF; 579 rs_number_alive++; 580 sysctl_ctx_init(&rs->sysctl_ctx); 581 rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 582 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), 583 OID_AUTO, 584 rs->rs_ifp->if_xname, 585 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 586 ""); 587 rl_add_syctl_entries(rl_sysctl_root, rs); 588 mtx_lock(&rs_mtx); 589 CK_LIST_INSERT_HEAD(&int_rs, rs, next); 590 mtx_unlock(&rs_mtx); 591 return (rs); 592 } else if ((rl.flags & RT_IS_FIXED_TABLE) == RT_IS_FIXED_TABLE) { 593 /* Mellanox C4 likely */ 594 rs->rs_ifp = ifp; 595 rs->rs_if_dunit = ifp->if_dunit; 596 rs->rs_rate_cnt = rl.number_of_rates; 597 rs->rs_min_seg = rl.min_segment_burst; 598 rs->rs_highest_valid = 0; 599 rs->rs_flow_limit = rl.max_flows; 600 rs->rs_flags = RS_IS_INTF | RS_NO_PRE; 601 rs->rs_disable = 0; 602 rate_table_act = rl.rate_table; 603 } else if ((rl.flags & RT_IS_SELECTABLE) == RT_IS_SELECTABLE) { 604 /* Chelsio, C5 and C6 of Mellanox? */ 605 rs->rs_ifp = ifp; 606 rs->rs_if_dunit = ifp->if_dunit; 607 rs->rs_rate_cnt = rl.number_of_rates; 608 rs->rs_min_seg = rl.min_segment_burst; 609 rs->rs_disable = 0; 610 rs->rs_flow_limit = rl.max_flows; 611 rate_table_act = desired_rates; 612 if ((rs->rs_rate_cnt > MAX_HDWR_RATES) && 613 (rs->rs_rate_cnt < ALL_HARDWARE_RATES)) { 614 /* 615 * Our desired table is not big 616 * enough, do what we can. 617 */ 618 rs->rs_rate_cnt = MAX_HDWR_RATES; 619 } 620 if (rs->rs_rate_cnt <= RS_ORDERED_COUNT) 621 rs->rs_flags = RS_IS_INTF; 622 else 623 rs->rs_flags = RS_IS_INTF | RS_INT_TBL; 624 if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) 625 rs->rs_rate_cnt = ALL_HARDWARE_RATES; 626 } else { 627 free(rs, M_TCPPACE); 628 return (NULL); 629 } 630 sz = sizeof(struct tcp_hwrate_limit_table) * rs->rs_rate_cnt; 631 rs->rs_rlt = malloc(sz, M_TCPPACE, M_NOWAIT); 632 if (rs->rs_rlt == NULL) { 633 if (error) 634 *error = ENOMEM; 635 bail: 636 free(rs, M_TCPPACE); 637 return (NULL); 638 } 639 if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) { 640 /* 641 * The interface supports all 642 * the rates we could possibly want. 643 */ 644 uint64_t rat; 645 646 rs->rs_rlt[0].rate = 12500; /* 100k */ 647 rs->rs_rlt[1].rate = 25000; /* 200k */ 648 rs->rs_rlt[2].rate = 62500; /* 500k */ 649 /* Note 125000 == 1Megabit 650 * populate 1Meg - 1000meg. 651 */ 652 for(i = 3, rat = 125000; i< (ALL_HARDWARE_RATES-1); i++) { 653 rs->rs_rlt[i].rate = rat; 654 rat += 125000; 655 } 656 rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate = 1250000000; 657 } else if (rs->rs_flags & RS_INT_TBL) { 658 /* We populate this in a special way */ 659 populate_canned_table(rs, rate_table_act); 660 } else { 661 /* 662 * Just copy in the rates from 663 * the table, it is in order. 664 */ 665 for (i=0; i<rs->rs_rate_cnt; i++) { 666 rs->rs_rlt[i].rate = rate_table_act[i]; 667 rs->rs_rlt[i].time_between = 0; 668 rs->rs_rlt[i].flags = 0; 669 } 670 } 671 for (i = (rs->rs_rate_cnt - 1); i >= 0; i--) { 672 /* 673 * We go backwards through the list so that if we can't get 674 * a rate and fail to init one, we have at least a chance of 675 * getting the highest one. 676 */ 677 rs->rs_rlt[i].ptbl = rs; 678 rs->rs_rlt[i].tag = NULL; 679 /* 680 * Calculate the time between. 681 */ 682 lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND; 683 res = lentim / rs->rs_rlt[i].rate; 684 if (res > 0) 685 rs->rs_rlt[i].time_between = res; 686 else 687 rs->rs_rlt[i].time_between = 1; 688 if (rs->rs_flags & RS_NO_PRE) { 689 rs->rs_rlt[i].flags = HDWRPACE_INITED; 690 rs->rs_lowest_valid = i; 691 } else { 692 int err; 693 694 if ((rl.flags & RT_IS_SETUP_REQ) && 695 (ifp->if_ratelimit_query)) { 696 err = ifp->if_ratelimit_setup(ifp, 697 rs->rs_rlt[i].rate, i); 698 if (err) 699 goto handle_err; 700 } 701 #ifdef RSS 702 hash_type = M_HASHTYPE_RSS_TCP_IPV4; 703 #else 704 hash_type = M_HASHTYPE_OPAQUE_HASH; 705 #endif 706 err = rl_attach_txrtlmt(ifp, 707 hash_type, 708 (i + 1), 709 rs->rs_rlt[i].rate, 710 &rs->rs_rlt[i].tag); 711 if (err) { 712 handle_err: 713 if (i == (rs->rs_rate_cnt - 1)) { 714 /* 715 * Huh - first rate and we can't get 716 * it? 717 */ 718 free(rs->rs_rlt, M_TCPPACE); 719 if (error) 720 *error = err; 721 goto bail; 722 } else { 723 if (error) 724 *error = err; 725 } 726 break; 727 } else { 728 rs->rs_rlt[i].flags = HDWRPACE_INITED | HDWRPACE_TAGPRESENT; 729 rs->rs_lowest_valid = i; 730 } 731 } 732 } 733 /* Did we get at least 1 rate? */ 734 if (rs->rs_rlt[(rs->rs_rate_cnt - 1)].flags & HDWRPACE_INITED) 735 rs->rs_highest_valid = rs->rs_rate_cnt - 1; 736 else { 737 free(rs->rs_rlt, M_TCPPACE); 738 goto bail; 739 } 740 rs_number_alive++; 741 sysctl_ctx_init(&rs->sysctl_ctx); 742 rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 743 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), 744 OID_AUTO, 745 rs->rs_ifp->if_xname, 746 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 747 ""); 748 rl_add_syctl_entries(rl_sysctl_root, rs); 749 mtx_lock(&rs_mtx); 750 CK_LIST_INSERT_HEAD(&int_rs, rs, next); 751 mtx_unlock(&rs_mtx); 752 return (rs); 753 } 754 755 static const struct tcp_hwrate_limit_table * 756 tcp_int_find_suitable_rate(const struct tcp_rate_set *rs, 757 uint64_t bytes_per_sec, uint32_t flags) 758 { 759 struct tcp_hwrate_limit_table *arte = NULL, *rte = NULL; 760 uint64_t mbits_per_sec, ind_calc; 761 int i; 762 763 mbits_per_sec = (bytes_per_sec * 8); 764 if (flags & RS_PACING_LT) { 765 if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && 766 (rs->rs_lowest_valid <= 2)){ 767 /* 768 * Smaller than 1Meg, only 769 * 3 entries can match it. 770 */ 771 for(i = rs->rs_lowest_valid; i < 3; i++) { 772 if (bytes_per_sec <= rs->rs_rlt[i].rate) { 773 rte = &rs->rs_rlt[i]; 774 break; 775 } else if (rs->rs_rlt[i].flags & HDWRPACE_INITED) { 776 arte = &rs->rs_rlt[i]; 777 } 778 } 779 goto done; 780 } else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) && 781 (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){ 782 /* 783 * Larger than 1G (the majority of 784 * our table. 785 */ 786 if (mbits_per_sec < RS_TEN_GIGABIT_PERSEC) 787 rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 788 else 789 arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 790 goto done; 791 } 792 /* 793 * If we reach here its in our table (between 1Meg - 1000Meg), 794 * just take the rounded down mbits per second, and add 795 * 1Megabit to it, from this we can calculate 796 * the index in the table. 797 */ 798 ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC; 799 if ((ind_calc * RS_ONE_MEGABIT_PERSEC) != mbits_per_sec) 800 ind_calc++; 801 /* our table is offset by 3, we add 2 */ 802 ind_calc += 2; 803 if (ind_calc > (ALL_HARDWARE_RATES-1)) { 804 /* This should not happen */ 805 ind_calc = ALL_HARDWARE_RATES-1; 806 } 807 if ((ind_calc >= rs->rs_lowest_valid) && 808 (ind_calc <= rs->rs_highest_valid)) 809 rte = &rs->rs_rlt[ind_calc]; 810 } else if (flags & RS_PACING_EXACT_MATCH) { 811 if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && 812 (rs->rs_lowest_valid <= 2)){ 813 for(i = rs->rs_lowest_valid; i < 3; i++) { 814 if (bytes_per_sec == rs->rs_rlt[i].rate) { 815 rte = &rs->rs_rlt[i]; 816 break; 817 } 818 } 819 } else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) && 820 (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) { 821 /* > 1Gbps only one rate */ 822 if (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) { 823 /* Its 10G wow */ 824 rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 825 } 826 } else { 827 /* Ok it must be a exact meg (its between 1G and 1Meg) */ 828 ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC; 829 if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) { 830 /* its an exact Mbps */ 831 ind_calc += 2; 832 if (ind_calc > (ALL_HARDWARE_RATES-1)) { 833 /* This should not happen */ 834 ind_calc = ALL_HARDWARE_RATES-1; 835 } 836 if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) 837 rte = &rs->rs_rlt[ind_calc]; 838 } 839 } 840 } else { 841 /* we want greater than the requested rate */ 842 if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && 843 (rs->rs_lowest_valid <= 2)){ 844 arte = &rs->rs_rlt[3]; /* set alternate to 1Meg */ 845 for (i=2; i>=rs->rs_lowest_valid; i--) { 846 if (bytes_per_sec < rs->rs_rlt[i].rate) { 847 rte = &rs->rs_rlt[i]; 848 break; 849 } else if ((flags & RS_PACING_GEQ) && 850 (bytes_per_sec == rs->rs_rlt[i].rate)) { 851 rte = &rs->rs_rlt[i]; 852 break; 853 } else { 854 arte = &rs->rs_rlt[i]; /* new alternate */ 855 } 856 } 857 } else if (mbits_per_sec > RS_ONE_GIGABIT_PERSEC) { 858 if ((bytes_per_sec < rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) && 859 (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){ 860 /* Our top rate is larger than the request */ 861 rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 862 } else if ((flags & RS_PACING_GEQ) && 863 (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) && 864 (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) { 865 /* It matches our top rate */ 866 rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 867 } else if (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED) { 868 /* The top rate is an alternative */ 869 arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 870 } 871 } else { 872 /* Its in our range 1Meg - 1Gig */ 873 if (flags & RS_PACING_GEQ) { 874 ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC; 875 if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) { 876 if (ind_calc > (ALL_HARDWARE_RATES-1)) { 877 /* This should not happen */ 878 ind_calc = (ALL_HARDWARE_RATES-1); 879 } 880 rte = &rs->rs_rlt[ind_calc]; 881 } 882 goto done; 883 } 884 ind_calc = (mbits_per_sec + (RS_ONE_MEGABIT_PERSEC-1))/RS_ONE_MEGABIT_PERSEC; 885 ind_calc += 2; 886 if (ind_calc > (ALL_HARDWARE_RATES-1)) { 887 /* This should not happen */ 888 ind_calc = ALL_HARDWARE_RATES-1; 889 } 890 if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) 891 rte = &rs->rs_rlt[ind_calc]; 892 } 893 } 894 done: 895 if ((rte == NULL) && 896 (arte != NULL) && 897 (flags & RS_PACING_SUB_OK)) { 898 /* We can use the substitute */ 899 rte = arte; 900 } 901 return (rte); 902 } 903 904 static const struct tcp_hwrate_limit_table * 905 tcp_find_suitable_rate(const struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags) 906 { 907 /** 908 * Hunt the rate table with the restrictions in flags and find a 909 * suitable rate if possible. 910 * RS_PACING_EXACT_MATCH - look for an exact match to rate. 911 * RS_PACING_GT - must be greater than. 912 * RS_PACING_GEQ - must be greater than or equal. 913 * RS_PACING_LT - must be less than. 914 * RS_PACING_SUB_OK - If we don't meet criteria a 915 * substitute is ok. 916 */ 917 int i, matched; 918 struct tcp_hwrate_limit_table *rte = NULL; 919 920 if ((rs->rs_flags & RS_INT_TBL) && 921 (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)) { 922 /* 923 * Here we don't want to paw thru 924 * a big table, we have everything 925 * from 1Meg - 1000Meg in 1Meg increments. 926 * Use an alternate method to "lookup". 927 */ 928 return (tcp_int_find_suitable_rate(rs, bytes_per_sec, flags)); 929 } 930 if ((flags & RS_PACING_LT) || 931 (flags & RS_PACING_EXACT_MATCH)) { 932 /* 933 * For exact and less than we go forward through the table. 934 * This way when we find one larger we stop (exact was a 935 * toss up). 936 */ 937 for (i = rs->rs_lowest_valid, matched = 0; i <= rs->rs_highest_valid; i++) { 938 if ((flags & RS_PACING_EXACT_MATCH) && 939 (bytes_per_sec == rs->rs_rlt[i].rate)) { 940 rte = &rs->rs_rlt[i]; 941 matched = 1; 942 break; 943 } else if ((flags & RS_PACING_LT) && 944 (bytes_per_sec <= rs->rs_rlt[i].rate)) { 945 rte = &rs->rs_rlt[i]; 946 matched = 1; 947 break; 948 } 949 if (bytes_per_sec > rs->rs_rlt[i].rate) 950 break; 951 } 952 if ((matched == 0) && 953 (flags & RS_PACING_LT) && 954 (flags & RS_PACING_SUB_OK)) { 955 /* Kick in a substitute (the lowest) */ 956 rte = &rs->rs_rlt[rs->rs_lowest_valid]; 957 } 958 } else { 959 /* 960 * Here we go backward through the table so that we can find 961 * the one greater in theory faster (but its probably a 962 * wash). 963 */ 964 for (i = rs->rs_highest_valid, matched = 0; i >= rs->rs_lowest_valid; i--) { 965 if (rs->rs_rlt[i].rate > bytes_per_sec) { 966 /* A possible candidate */ 967 rte = &rs->rs_rlt[i]; 968 } 969 if ((flags & RS_PACING_GEQ) && 970 (bytes_per_sec == rs->rs_rlt[i].rate)) { 971 /* An exact match and we want equal */ 972 matched = 1; 973 rte = &rs->rs_rlt[i]; 974 break; 975 } else if (rte) { 976 /* 977 * Found one that is larger than but don't 978 * stop, there may be a more closer match. 979 */ 980 matched = 1; 981 } 982 if (rs->rs_rlt[i].rate < bytes_per_sec) { 983 /* 984 * We found a table entry that is smaller, 985 * stop there will be none greater or equal. 986 */ 987 break; 988 } 989 } 990 if ((matched == 0) && 991 (flags & RS_PACING_SUB_OK)) { 992 /* Kick in a substitute (the highest) */ 993 rte = &rs->rs_rlt[rs->rs_highest_valid]; 994 } 995 } 996 return (rte); 997 } 998 999 static struct ifnet * 1000 rt_find_real_interface(struct ifnet *ifp, struct inpcb *inp, int *error) 1001 { 1002 struct ifnet *tifp; 1003 struct m_snd_tag *tag; 1004 union if_snd_tag_alloc_params params = { 1005 .rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT, 1006 .rate_limit.hdr.flowid = 1, 1007 .rate_limit.hdr.numa_domain = inp->inp_numa_domain, 1008 .rate_limit.max_rate = COMMON_RATE, 1009 .rate_limit.flags = M_NOWAIT, 1010 }; 1011 int err; 1012 #ifdef RSS 1013 params.rate_limit.hdr.flowtype = ((inp->inp_vflag & INP_IPV6) ? 1014 M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_TCP_IPV4); 1015 #else 1016 params.rate_limit.hdr.flowtype = M_HASHTYPE_OPAQUE_HASH; 1017 #endif 1018 tag = NULL; 1019 if (ifp->if_snd_tag_alloc) { 1020 if (error) 1021 *error = ENODEV; 1022 return (NULL); 1023 } 1024 err = ifp->if_snd_tag_alloc(ifp, ¶ms, &tag); 1025 if (err) { 1026 /* Failed to setup a tag? */ 1027 if (error) 1028 *error = err; 1029 return (NULL); 1030 } 1031 tifp = tag->ifp; 1032 tifp->if_snd_tag_free(tag); 1033 return (tifp); 1034 } 1035 1036 static const struct tcp_hwrate_limit_table * 1037 rt_setup_rate(struct inpcb *inp, struct ifnet *ifp, uint64_t bytes_per_sec, 1038 uint32_t flags, int *error) 1039 { 1040 /* First lets find the interface if it exists */ 1041 const struct tcp_hwrate_limit_table *rte; 1042 struct tcp_rate_set *rs; 1043 struct epoch_tracker et; 1044 int err; 1045 1046 NET_EPOCH_ENTER(et); 1047 use_real_interface: 1048 CK_LIST_FOREACH(rs, &int_rs, next) { 1049 /* 1050 * Note we don't look with the lock since we either see a 1051 * new entry or will get one when we try to add it. 1052 */ 1053 if (rs->rs_flags & RS_IS_DEAD) { 1054 /* The dead are not looked at */ 1055 continue; 1056 } 1057 if ((rs->rs_ifp == ifp) && 1058 (rs->rs_if_dunit == ifp->if_dunit)) { 1059 /* Ok we found it */ 1060 break; 1061 } 1062 } 1063 if ((rs == NULL) || 1064 (rs->rs_flags & RS_INTF_NO_SUP) || 1065 (rs->rs_flags & RS_IS_DEAD)) { 1066 /* 1067 * This means we got a packet *before* 1068 * the IF-UP was processed below, <or> 1069 * while or after we already received an interface 1070 * departed event. In either case we really don't 1071 * want to do anything with pacing, in 1072 * the departing case the packet is not 1073 * going to go very far. The new case 1074 * might be arguable, but its impossible 1075 * to tell from the departing case. 1076 */ 1077 if (rs->rs_disable && error) 1078 *error = ENODEV; 1079 NET_EPOCH_EXIT(et); 1080 return (NULL); 1081 } 1082 1083 if ((rs == NULL) || (rs->rs_disable != 0)) { 1084 if (rs->rs_disable && error) 1085 *error = ENOSPC; 1086 NET_EPOCH_EXIT(et); 1087 return (NULL); 1088 } 1089 if (rs->rs_flags & RS_IS_DEFF) { 1090 /* We need to find the real interface */ 1091 struct ifnet *tifp; 1092 1093 tifp = rt_find_real_interface(ifp, inp, error); 1094 if (tifp == NULL) { 1095 if (rs->rs_disable && error) 1096 *error = ENOTSUP; 1097 NET_EPOCH_EXIT(et); 1098 return (NULL); 1099 } 1100 goto use_real_interface; 1101 } 1102 if (rs->rs_flow_limit && 1103 ((rs->rs_flows_using + 1) > rs->rs_flow_limit)) { 1104 if (error) 1105 *error = ENOSPC; 1106 NET_EPOCH_EXIT(et); 1107 return (NULL); 1108 } 1109 rte = tcp_find_suitable_rate(rs, bytes_per_sec, flags); 1110 if (rte) { 1111 err = in_pcbattach_txrtlmt(inp, rs->rs_ifp, 1112 inp->inp_flowtype, 1113 inp->inp_flowid, 1114 rte->rate, 1115 &inp->inp_snd_tag); 1116 if (err) { 1117 /* Failed to attach */ 1118 if (error) 1119 *error = err; 1120 rte = NULL; 1121 } 1122 } 1123 if (rte) { 1124 /* 1125 * We use an atomic here for accounting so we don't have to 1126 * use locks when freeing. 1127 */ 1128 atomic_add_64(&rs->rs_flows_using, 1); 1129 } 1130 NET_EPOCH_EXIT(et); 1131 return (rte); 1132 } 1133 1134 static void 1135 tcp_rl_ifnet_link(void *arg __unused, struct ifnet *ifp, int link_state) 1136 { 1137 int error; 1138 struct tcp_rate_set *rs; 1139 1140 if (((ifp->if_capabilities & IFCAP_TXRTLMT) == 0) || 1141 (link_state != LINK_STATE_UP)) { 1142 /* 1143 * We only care on an interface going up that is rate-limit 1144 * capable. 1145 */ 1146 return; 1147 } 1148 mtx_lock(&rs_mtx); 1149 CK_LIST_FOREACH(rs, &int_rs, next) { 1150 if ((rs->rs_ifp == ifp) && 1151 (rs->rs_if_dunit == ifp->if_dunit)) { 1152 /* We already have initialized this guy */ 1153 mtx_unlock(&rs_mtx); 1154 return; 1155 } 1156 } 1157 mtx_unlock(&rs_mtx); 1158 rt_setup_new_rs(ifp, &error); 1159 } 1160 1161 static void 1162 tcp_rl_ifnet_departure(void *arg __unused, struct ifnet *ifp) 1163 { 1164 struct tcp_rate_set *rs, *nrs; 1165 struct ifnet *tifp; 1166 int i; 1167 1168 mtx_lock(&rs_mtx); 1169 CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) { 1170 if ((rs->rs_ifp == ifp) && 1171 (rs->rs_if_dunit == ifp->if_dunit)) { 1172 CK_LIST_REMOVE(rs, next); 1173 rs_number_alive--; 1174 rs->rs_flags |= RS_IS_DEAD; 1175 for (i = 0; i < rs->rs_rate_cnt; i++) { 1176 if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) { 1177 tifp = rs->rs_rlt[i].tag->ifp; 1178 in_pcbdetach_tag(tifp, rs->rs_rlt[i].tag); 1179 rs->rs_rlt[i].tag = NULL; 1180 } 1181 rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED; 1182 } 1183 if (rs->rs_flows_using == 0) 1184 rs_defer_destroy(rs); 1185 break; 1186 } 1187 } 1188 mtx_unlock(&rs_mtx); 1189 } 1190 1191 static void 1192 tcp_rl_shutdown(void *arg __unused, int howto __unused) 1193 { 1194 struct tcp_rate_set *rs, *nrs; 1195 struct ifnet *tifp; 1196 int i; 1197 1198 mtx_lock(&rs_mtx); 1199 CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) { 1200 CK_LIST_REMOVE(rs, next); 1201 rs_number_alive--; 1202 rs->rs_flags |= RS_IS_DEAD; 1203 for (i = 0; i < rs->rs_rate_cnt; i++) { 1204 if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) { 1205 tifp = rs->rs_rlt[i].tag->ifp; 1206 in_pcbdetach_tag(tifp, rs->rs_rlt[i].tag); 1207 rs->rs_rlt[i].tag = NULL; 1208 } 1209 rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED; 1210 } 1211 if (rs->rs_flows_using == 0) 1212 rs_defer_destroy(rs); 1213 } 1214 mtx_unlock(&rs_mtx); 1215 } 1216 1217 const struct tcp_hwrate_limit_table * 1218 tcp_set_pacing_rate(struct tcpcb *tp, struct ifnet *ifp, 1219 uint64_t bytes_per_sec, int flags, int *error) 1220 { 1221 const struct tcp_hwrate_limit_table *rte; 1222 1223 if (tp->t_inpcb->inp_snd_tag == NULL) { 1224 /* 1225 * We are setting up a rate for the first time. 1226 */ 1227 if ((ifp->if_capabilities & IFCAP_TXRTLMT) == 0) { 1228 /* Not supported by the egress */ 1229 if (error) 1230 *error = ENODEV; 1231 return (NULL); 1232 } 1233 #ifdef KERN_TLS 1234 if (tp->t_inpcb->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) { 1235 /* 1236 * We currently can't do both TLS and hardware 1237 * pacing 1238 */ 1239 if (error) 1240 *error = EINVAL; 1241 return (NULL); 1242 } 1243 #endif 1244 rte = rt_setup_rate(tp->t_inpcb, ifp, bytes_per_sec, flags, error); 1245 } else { 1246 /* 1247 * We are modifying a rate, wrong interface? 1248 */ 1249 if (error) 1250 *error = EINVAL; 1251 rte = NULL; 1252 } 1253 *error = 0; 1254 return (rte); 1255 } 1256 1257 const struct tcp_hwrate_limit_table * 1258 tcp_chg_pacing_rate(const struct tcp_hwrate_limit_table *crte, 1259 struct tcpcb *tp, struct ifnet *ifp, 1260 uint64_t bytes_per_sec, int flags, int *error) 1261 { 1262 const struct tcp_hwrate_limit_table *nrte; 1263 const struct tcp_rate_set *rs; 1264 int is_indirect = 0; 1265 int err; 1266 1267 if ((tp->t_inpcb->inp_snd_tag == NULL) || 1268 (crte == NULL)) { 1269 /* Wrong interface */ 1270 if (error) 1271 *error = EINVAL; 1272 return (NULL); 1273 } 1274 rs = crte->ptbl; 1275 if ((rs->rs_flags & RS_IS_DEAD) || 1276 (crte->flags & HDWRPACE_IFPDEPARTED)) { 1277 /* Release the rate, and try anew */ 1278 re_rate: 1279 tcp_rel_pacing_rate(crte, tp); 1280 nrte = tcp_set_pacing_rate(tp, ifp, 1281 bytes_per_sec, flags, error); 1282 return (nrte); 1283 } 1284 if ((rs->rs_flags & RT_IS_INDIRECT ) == RT_IS_INDIRECT) 1285 is_indirect = 1; 1286 else 1287 is_indirect = 0; 1288 if ((is_indirect == 0) && 1289 ((ifp != rs->rs_ifp) || 1290 (ifp->if_dunit != rs->rs_if_dunit))) { 1291 /* 1292 * Something changed, the user is not pointing to the same 1293 * ifp? Maybe a route updated on this guy? 1294 */ 1295 goto re_rate; 1296 } else if (is_indirect) { 1297 /* 1298 * For indirect we have to dig in and find the real interface. 1299 */ 1300 struct ifnet *rifp; 1301 1302 rifp = rt_find_real_interface(ifp, tp->t_inpcb, error); 1303 if (rifp == NULL) { 1304 /* Can't find it? */ 1305 goto re_rate; 1306 } 1307 if ((rifp != rs->rs_ifp) || 1308 (ifp->if_dunit != rs->rs_if_dunit)) { 1309 goto re_rate; 1310 } 1311 } 1312 nrte = tcp_find_suitable_rate(rs, bytes_per_sec, flags); 1313 if (nrte == crte) { 1314 /* No change */ 1315 if (error) 1316 *error = 0; 1317 return (crte); 1318 } 1319 if (nrte == NULL) { 1320 /* Release the old rate */ 1321 tcp_rel_pacing_rate(crte, tp); 1322 return (NULL); 1323 } 1324 /* Change rates to our new entry */ 1325 err = in_pcbmodify_txrtlmt(tp->t_inpcb, nrte->rate); 1326 if (err) { 1327 if (error) 1328 *error = err; 1329 return (NULL); 1330 } 1331 if (error) 1332 *error = 0; 1333 return (nrte); 1334 } 1335 1336 void 1337 tcp_rel_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp) 1338 { 1339 const struct tcp_rate_set *crs; 1340 struct tcp_rate_set *rs; 1341 uint64_t pre; 1342 1343 crs = crte->ptbl; 1344 /* 1345 * Now we must break the const 1346 * in order to release our refcount. 1347 */ 1348 rs = __DECONST(struct tcp_rate_set *, crs); 1349 pre = atomic_fetchadd_64(&rs->rs_flows_using, -1); 1350 if (pre == 1) { 1351 mtx_lock(&rs_mtx); 1352 /* 1353 * Is it dead? 1354 */ 1355 if (rs->rs_flags & RS_IS_DEAD) 1356 rs_defer_destroy(rs); 1357 mtx_unlock(&rs_mtx); 1358 } 1359 in_pcbdetach_txrtlmt(tp->t_inpcb); 1360 } 1361 1362 #define ONE_POINT_TWO_MEG 150000 /* 1.2 megabits in bytes */ 1363 #define ONE_HUNDRED_MBPS 12500000 /* 100Mbps in bytes per second */ 1364 #define FIVE_HUNDRED_MBPS 62500000 /* 500Mbps in bytes per second */ 1365 #define MAX_MSS_SENT 43 /* 43 mss = 43 x 1500 = 64,500 bytes */ 1366 1367 uint32_t 1368 tcp_get_pacing_burst_size (uint64_t bw, uint32_t segsiz, int can_use_1mss, 1369 const struct tcp_hwrate_limit_table *te, int *err) 1370 { 1371 /* 1372 * We use the google formula to calculate the 1373 * TSO size. I.E. 1374 * bw < 24Meg 1375 * tso = 2mss 1376 * else 1377 * tso = min(bw/1000, 64k) 1378 * 1379 * Note for these calculations we ignore the 1380 * packet overhead (enet hdr, ip hdr and tcp hdr). 1381 */ 1382 uint64_t lentim, res, bytes; 1383 uint32_t new_tso, min_tso_segs; 1384 1385 bytes = bw / 1000; 1386 if (bytes > (64 * 1000)) 1387 bytes = 64 * 1000; 1388 /* Round up */ 1389 new_tso = (bytes + segsiz - 1) / segsiz; 1390 if (can_use_1mss && (bw < ONE_POINT_TWO_MEG)) 1391 min_tso_segs = 1; 1392 else 1393 min_tso_segs = 2; 1394 if (new_tso < min_tso_segs) 1395 new_tso = min_tso_segs; 1396 if (new_tso > MAX_MSS_SENT) 1397 new_tso = MAX_MSS_SENT; 1398 new_tso *= segsiz; 1399 /* 1400 * If we are not doing hardware pacing 1401 * then we are done. 1402 */ 1403 if (te == NULL) { 1404 if (err) 1405 *err = 0; 1406 return(new_tso); 1407 } 1408 /* 1409 * For hardware pacing we look at the 1410 * rate you are sending at and compare 1411 * that to the rate you have in hardware. 1412 * 1413 * If the hardware rate is slower than your 1414 * software rate then you are in error and 1415 * we will build a queue in our hardware whic 1416 * is probably not desired, in such a case 1417 * just return the non-hardware TSO size. 1418 * 1419 * If the rate in hardware is faster (which 1420 * it should be) then look at how long it 1421 * takes to send one ethernet segment size at 1422 * your b/w and compare that to the time it 1423 * takes to send at the rate you had selected. 1424 * 1425 * If your time is greater (which we hope it is) 1426 * we get the delta between the two, and then 1427 * divide that into your pacing time. This tells 1428 * us how many MSS you can send down at once (rounded up). 1429 * 1430 * Note we also double this value if the b/w is over 1431 * 100Mbps. If its over 500meg we just set you to the 1432 * max (43 segments). 1433 */ 1434 if (te->rate > FIVE_HUNDRED_MBPS) 1435 return (segsiz * MAX_MSS_SENT); 1436 if (te->rate == bw) { 1437 /* We are pacing at exactly the hdwr rate */ 1438 return (segsiz * MAX_MSS_SENT); 1439 } 1440 lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND; 1441 res = lentim / bw; 1442 if (res > te->time_between) { 1443 uint32_t delta, segs; 1444 1445 delta = res - te->time_between; 1446 segs = (res + delta - 1)/delta; 1447 if (te->rate > ONE_HUNDRED_MBPS) 1448 segs *= 2; 1449 if (segs < min_tso_segs) 1450 segs = min_tso_segs; 1451 if (segs > MAX_MSS_SENT) 1452 segs = MAX_MSS_SENT; 1453 segs *= segsiz; 1454 if (err) 1455 *err = 0; 1456 if (segs < new_tso) { 1457 /* unexpected ? */ 1458 return(new_tso); 1459 } else { 1460 return (segs); 1461 } 1462 } else { 1463 /* 1464 * Your time is smaller which means 1465 * we will grow a queue on our 1466 * hardware. Send back the non-hardware 1467 * rate. 1468 */ 1469 if (err) 1470 *err = -1; 1471 return (new_tso); 1472 } 1473 } 1474 1475 static eventhandler_tag rl_ifnet_departs; 1476 static eventhandler_tag rl_ifnet_arrives; 1477 static eventhandler_tag rl_shutdown_start; 1478 1479 static void 1480 tcp_rs_init(void *st __unused) 1481 { 1482 CK_LIST_INIT(&int_rs); 1483 rs_number_alive = 0; 1484 rs_number_dead = 0; 1485 mtx_init(&rs_mtx, "tcp_rs_mtx", "rsmtx", MTX_DEF); 1486 rl_ifnet_departs = EVENTHANDLER_REGISTER(ifnet_departure_event, 1487 tcp_rl_ifnet_departure, 1488 NULL, EVENTHANDLER_PRI_ANY); 1489 rl_ifnet_arrives = EVENTHANDLER_REGISTER(ifnet_link_event, 1490 tcp_rl_ifnet_link, 1491 NULL, EVENTHANDLER_PRI_ANY); 1492 rl_shutdown_start = EVENTHANDLER_REGISTER(shutdown_pre_sync, 1493 tcp_rl_shutdown, NULL, 1494 SHUTDOWN_PRI_FIRST); 1495 printf("TCP_ratelimit: Is now initialized\n"); 1496 } 1497 1498 SYSINIT(tcp_rl_init, SI_SUB_SMP + 1, SI_ORDER_ANY, tcp_rs_init, NULL); 1499 #endif 1500