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_hpts.h> 61 #include <netinet/tcp_log_buf.h> 62 #include <netinet/tcp_ratelimit.h> 63 #ifndef USECS_IN_SECOND 64 #define USECS_IN_SECOND 1000000 65 #endif 66 /* 67 * For the purposes of each send, what is the size 68 * of an ethernet frame. 69 */ 70 MALLOC_DEFINE(M_TCPPACE, "tcp_hwpace", "TCP Hardware pacing memory"); 71 #ifdef RATELIMIT 72 73 /* 74 * The following preferred table will seem weird to 75 * the casual viewer. Why do we not have any rates below 76 * 1Mbps? Why do we have a rate at 1.44Mbps called common? 77 * Why do the rates cluster in the 1-100Mbps range more 78 * than others? Why does the table jump around at the beginnign 79 * and then be more consistently raising? 80 * 81 * Let me try to answer those questions. A lot of 82 * this is dependant on the hardware. We have three basic 83 * supporters of rate limiting 84 * 85 * Chelsio - Supporting 16 configurable rates. 86 * Mlx - c4 supporting 13 fixed rates. 87 * Mlx - c5 & c6 supporting 127 configurable rates. 88 * 89 * The c4 is why we have a common rate that is available 90 * in all rate tables. This is a selected rate from the 91 * c4 table and we assure its available in all ratelimit 92 * tables. This way the tcp_ratelimit code has an assured 93 * rate it should always be able to get. This answers a 94 * couple of the questions above. 95 * 96 * So what about the rest, well the table is built to 97 * try to get the most out of a joint hardware/software 98 * pacing system. The software pacer will always pick 99 * a rate higher than the b/w that it is estimating 100 * 101 * on the path. This is done for two reasons. 102 * a) So we can discover more b/w 103 * and 104 * b) So we can send a block of MSS's down and then 105 * have the software timer go off after the previous 106 * send is completely out of the hardware. 107 * 108 * But when we do <b> we don't want to have the delay 109 * between the last packet sent by the hardware be 110 * excessively long (to reach our desired rate). 111 * 112 * So let me give an example for clarity. 113 * 114 * Lets assume that the tcp stack sees that 29,110,000 bps is 115 * what the bw of the path is. The stack would select the 116 * rate 31Mbps. 31Mbps means that each send that is done 117 * by the hardware will cause a 390 micro-second gap between 118 * the packets sent at that rate. For 29,110,000 bps we 119 * would need 416 micro-seconds gap between each send. 120 * 121 * Note that are calculating a complete time for pacing 122 * which includes the ethernet, IP and TCP overhead. So 123 * a full 1514 bytes is used for the above calculations. 124 * My testing has shown that both cards are also using this 125 * as their basis i.e. full payload size of the ethernet frame. 126 * The TCP stack caller needs to be aware of this and make the 127 * appropriate overhead calculations be included in its choices. 128 * 129 * Now, continuing our example, we pick a MSS size based on the 130 * delta between the two rates (416 - 390) divided into the rate 131 * we really wish to send at rounded up. That results in a MSS 132 * send of 17 mss's at once. The hardware then will 133 * run out of data in a single 17MSS send in 6,630 micro-seconds. 134 * 135 * On the other hand the software pacer will send more data 136 * in 7,072 micro-seconds. This means that we will refill 137 * the hardware 52 microseconds after it would have sent 138 * next if it had not ran out of data. This is a win since we are 139 * only sending every 7ms or so and yet all the packets are spaced on 140 * the wire with 94% of what they should be and only 141 * the last packet is delayed extra to make up for the 142 * difference. 143 * 144 * Note that the above formula has two important caveat. 145 * If we are above (b/w wise) over 100Mbps we double the result 146 * of the MSS calculation. The second caveat is if we are 500Mbps 147 * or more we just send the maximum MSS at once i.e. 45MSS. At 148 * the higher b/w's even the cards have limits to what times (timer granularity) 149 * they can insert between packets and start to send more than one 150 * packet at a time on the wire. 151 * 152 */ 153 #define COMMON_RATE 180500 154 const uint64_t desired_rates[] = { 155 122500, /* 1Mbps - rate 1 */ 156 180500, /* 1.44Mpbs - rate 2 common rate */ 157 375000, /* 3Mbps - rate 3 */ 158 625000, /* 5Mbps - rate 4 */ 159 1250000, /* 10Mbps - rate 5 */ 160 1875000, /* 15Mbps - rate 6 */ 161 2500000, /* 20Mbps - rate 7 */ 162 3125000, /* 25Mbps - rate 8 */ 163 3750000, /* 30Mbps - rate 9 */ 164 4375000, /* 35Mbps - rate 10 */ 165 5000000, /* 40Meg - rate 11 */ 166 6250000, /* 50Mbps - rate 12 */ 167 12500000, /* 100Mbps - rate 13 */ 168 25000000, /* 200Mbps - rate 14 */ 169 50000000, /* 400Mbps - rate 15 */ 170 100000000, /* 800Mbps - rate 16 */ 171 5625000, /* 45Mbps - rate 17 */ 172 6875000, /* 55Mbps - rate 19 */ 173 7500000, /* 60Mbps - rate 20 */ 174 8125000, /* 65Mbps - rate 21 */ 175 8750000, /* 70Mbps - rate 22 */ 176 9375000, /* 75Mbps - rate 23 */ 177 10000000, /* 80Mbps - rate 24 */ 178 10625000, /* 85Mbps - rate 25 */ 179 11250000, /* 90Mbps - rate 26 */ 180 11875000, /* 95Mbps - rate 27 */ 181 12500000, /* 100Mbps - rate 28 */ 182 13750000, /* 110Mbps - rate 29 */ 183 15000000, /* 120Mbps - rate 30 */ 184 16250000, /* 130Mbps - rate 31 */ 185 17500000, /* 140Mbps - rate 32 */ 186 18750000, /* 150Mbps - rate 33 */ 187 20000000, /* 160Mbps - rate 34 */ 188 21250000, /* 170Mbps - rate 35 */ 189 22500000, /* 180Mbps - rate 36 */ 190 23750000, /* 190Mbps - rate 37 */ 191 26250000, /* 210Mbps - rate 38 */ 192 27500000, /* 220Mbps - rate 39 */ 193 28750000, /* 230Mbps - rate 40 */ 194 30000000, /* 240Mbps - rate 41 */ 195 31250000, /* 250Mbps - rate 42 */ 196 34375000, /* 275Mbps - rate 43 */ 197 37500000, /* 300Mbps - rate 44 */ 198 40625000, /* 325Mbps - rate 45 */ 199 43750000, /* 350Mbps - rate 46 */ 200 46875000, /* 375Mbps - rate 47 */ 201 53125000, /* 425Mbps - rate 48 */ 202 56250000, /* 450Mbps - rate 49 */ 203 59375000, /* 475Mbps - rate 50 */ 204 62500000, /* 500Mbps - rate 51 */ 205 68750000, /* 550Mbps - rate 52 */ 206 75000000, /* 600Mbps - rate 53 */ 207 81250000, /* 650Mbps - rate 54 */ 208 87500000, /* 700Mbps - rate 55 */ 209 93750000, /* 750Mbps - rate 56 */ 210 106250000, /* 850Mbps - rate 57 */ 211 112500000, /* 900Mbps - rate 58 */ 212 125000000, /* 1Gbps - rate 59 */ 213 156250000, /* 1.25Gps - rate 60 */ 214 187500000, /* 1.5Gps - rate 61 */ 215 218750000, /* 1.75Gps - rate 62 */ 216 250000000, /* 2Gbps - rate 63 */ 217 281250000, /* 2.25Gps - rate 64 */ 218 312500000, /* 2.5Gbps - rate 65 */ 219 343750000, /* 2.75Gbps - rate 66 */ 220 375000000, /* 3Gbps - rate 67 */ 221 500000000, /* 4Gbps - rate 68 */ 222 625000000, /* 5Gbps - rate 69 */ 223 750000000, /* 6Gbps - rate 70 */ 224 875000000, /* 7Gbps - rate 71 */ 225 1000000000, /* 8Gbps - rate 72 */ 226 1125000000, /* 9Gbps - rate 73 */ 227 1250000000, /* 10Gbps - rate 74 */ 228 1875000000, /* 15Gbps - rate 75 */ 229 2500000000 /* 20Gbps - rate 76 */ 230 }; 231 232 #define MAX_HDWR_RATES (sizeof(desired_rates)/sizeof(uint64_t)) 233 #define RS_ORDERED_COUNT 16 /* 234 * Number that are in order 235 * at the beginning of the table, 236 * over this a sort is required. 237 */ 238 #define RS_NEXT_ORDER_GROUP 16 /* 239 * The point in our table where 240 * we come fill in a second ordered 241 * group (index wise means -1). 242 */ 243 #define ALL_HARDWARE_RATES 1004 /* 244 * 1Meg - 1Gig in 1 Meg steps 245 * plus 100, 200k and 500k and 246 * 10Gig 247 */ 248 249 #define RS_ONE_MEGABIT_PERSEC 1000000 250 #define RS_ONE_GIGABIT_PERSEC 1000000000 251 #define RS_TEN_GIGABIT_PERSEC 10000000000 252 253 static struct head_tcp_rate_set int_rs; 254 static struct mtx rs_mtx; 255 uint32_t rs_number_alive; 256 uint32_t rs_number_dead; 257 static uint32_t rs_floor_mss = 0; 258 static uint32_t wait_time_floor = 8000; /* 8 ms */ 259 static uint32_t rs_hw_floor_mss = 16; 260 static uint32_t num_of_waits_allowed = 1; /* How many time blocks are we willing to wait */ 261 262 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, rl, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 263 "TCP Ratelimit stats"); 264 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, alive, CTLFLAG_RW, 265 &rs_number_alive, 0, 266 "Number of interfaces initialized for ratelimiting"); 267 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, dead, CTLFLAG_RW, 268 &rs_number_dead, 0, 269 "Number of interfaces departing from ratelimiting"); 270 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, floor_mss, CTLFLAG_RW, 271 &rs_floor_mss, 0, 272 "Number of MSS that will override the normal minimums (0 means don't enforce)"); 273 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, wait_floor, CTLFLAG_RW, 274 &wait_time_floor, 2000, 275 "Has b/w increases what is the wait floor we are willing to wait at the end?"); 276 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, time_blocks, CTLFLAG_RW, 277 &num_of_waits_allowed, 1, 278 "How many time blocks on the end should software pacing be willing to wait?"); 279 280 SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, hw_floor_mss, CTLFLAG_RW, 281 &rs_hw_floor_mss, 16, 282 "Number of mss that are a minum for hardware pacing?"); 283 284 285 static void 286 rl_add_syctl_entries(struct sysctl_oid *rl_sysctl_root, struct tcp_rate_set *rs) 287 { 288 /* 289 * Add sysctl entries for thus interface. 290 */ 291 if (rs->rs_flags & RS_INTF_NO_SUP) { 292 SYSCTL_ADD_S32(&rs->sysctl_ctx, 293 SYSCTL_CHILDREN(rl_sysctl_root), 294 OID_AUTO, "disable", CTLFLAG_RD, 295 &rs->rs_disable, 0, 296 "Disable this interface from new hdwr limiting?"); 297 } else { 298 SYSCTL_ADD_S32(&rs->sysctl_ctx, 299 SYSCTL_CHILDREN(rl_sysctl_root), 300 OID_AUTO, "disable", CTLFLAG_RW, 301 &rs->rs_disable, 0, 302 "Disable this interface from new hdwr limiting?"); 303 } 304 SYSCTL_ADD_S32(&rs->sysctl_ctx, 305 SYSCTL_CHILDREN(rl_sysctl_root), 306 OID_AUTO, "minseg", CTLFLAG_RW, 307 &rs->rs_min_seg, 0, 308 "What is the minimum we need to send on this interface?"); 309 SYSCTL_ADD_U64(&rs->sysctl_ctx, 310 SYSCTL_CHILDREN(rl_sysctl_root), 311 OID_AUTO, "flow_limit", CTLFLAG_RW, 312 &rs->rs_flow_limit, 0, 313 "What is the limit for number of flows (0=unlimited)?"); 314 SYSCTL_ADD_S32(&rs->sysctl_ctx, 315 SYSCTL_CHILDREN(rl_sysctl_root), 316 OID_AUTO, "highest", CTLFLAG_RD, 317 &rs->rs_highest_valid, 0, 318 "Highest valid rate"); 319 SYSCTL_ADD_S32(&rs->sysctl_ctx, 320 SYSCTL_CHILDREN(rl_sysctl_root), 321 OID_AUTO, "lowest", CTLFLAG_RD, 322 &rs->rs_lowest_valid, 0, 323 "Lowest valid rate"); 324 SYSCTL_ADD_S32(&rs->sysctl_ctx, 325 SYSCTL_CHILDREN(rl_sysctl_root), 326 OID_AUTO, "flags", CTLFLAG_RD, 327 &rs->rs_flags, 0, 328 "What lags are on the entry?"); 329 SYSCTL_ADD_S32(&rs->sysctl_ctx, 330 SYSCTL_CHILDREN(rl_sysctl_root), 331 OID_AUTO, "numrates", CTLFLAG_RD, 332 &rs->rs_rate_cnt, 0, 333 "How many rates re there?"); 334 SYSCTL_ADD_U64(&rs->sysctl_ctx, 335 SYSCTL_CHILDREN(rl_sysctl_root), 336 OID_AUTO, "flows_using", CTLFLAG_RD, 337 &rs->rs_flows_using, 0, 338 "How many flows are using this interface now?"); 339 #ifdef DETAILED_RATELIMIT_SYSCTL 340 if (rs->rs_rlt && rs->rs_rate_cnt > 0) { 341 /* Lets display the rates */ 342 int i; 343 struct sysctl_oid *rl_rates; 344 struct sysctl_oid *rl_rate_num; 345 char rate_num[16]; 346 rl_rates = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 347 SYSCTL_CHILDREN(rl_sysctl_root), 348 OID_AUTO, 349 "rate", 350 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 351 "Ratelist"); 352 for( i = 0; i < rs->rs_rate_cnt; i++) { 353 sprintf(rate_num, "%d", i); 354 rl_rate_num = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 355 SYSCTL_CHILDREN(rl_rates), 356 OID_AUTO, 357 rate_num, 358 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 359 "Individual Rate"); 360 SYSCTL_ADD_U32(&rs->sysctl_ctx, 361 SYSCTL_CHILDREN(rl_rate_num), 362 OID_AUTO, "flags", CTLFLAG_RD, 363 &rs->rs_rlt[i].flags, 0, 364 "Flags on this rate"); 365 SYSCTL_ADD_U32(&rs->sysctl_ctx, 366 SYSCTL_CHILDREN(rl_rate_num), 367 OID_AUTO, "pacetime", CTLFLAG_RD, 368 &rs->rs_rlt[i].time_between, 0, 369 "Time hardware inserts between 1500 byte sends"); 370 SYSCTL_ADD_U64(&rs->sysctl_ctx, 371 SYSCTL_CHILDREN(rl_rate_num), 372 OID_AUTO, "rate", CTLFLAG_RD, 373 &rs->rs_rlt[i].rate, 0, 374 "Rate in bytes per second"); 375 } 376 } 377 #endif 378 } 379 380 static void 381 rs_destroy(epoch_context_t ctx) 382 { 383 struct tcp_rate_set *rs; 384 bool do_free_rs; 385 386 rs = __containerof(ctx, struct tcp_rate_set, rs_epoch_ctx); 387 388 mtx_lock(&rs_mtx); 389 rs->rs_flags &= ~RS_FUNERAL_SCHD; 390 /* 391 * In theory its possible (but unlikely) 392 * that while the delete was occuring 393 * and we were applying the DEAD flag 394 * someone slipped in and found the 395 * interface in a lookup. While we 396 * decided rs_flows_using were 0 and 397 * scheduling the epoch_call, the other 398 * thread incremented rs_flow_using. This 399 * is because users have a pointer and 400 * we only use the rs_flows_using in an 401 * atomic fashion, i.e. the other entities 402 * are not protected. To assure this did 403 * not occur, we check rs_flows_using here 404 * before deleting. 405 */ 406 do_free_rs = (rs->rs_flows_using == 0); 407 rs_number_dead--; 408 mtx_unlock(&rs_mtx); 409 410 if (do_free_rs) { 411 sysctl_ctx_free(&rs->sysctl_ctx); 412 free(rs->rs_rlt, M_TCPPACE); 413 free(rs, M_TCPPACE); 414 } 415 } 416 417 static void 418 rs_defer_destroy(struct tcp_rate_set *rs) 419 { 420 421 mtx_assert(&rs_mtx, MA_OWNED); 422 423 /* Check if already pending. */ 424 if (rs->rs_flags & RS_FUNERAL_SCHD) 425 return; 426 427 rs_number_dead++; 428 429 /* Set flag to only defer once. */ 430 rs->rs_flags |= RS_FUNERAL_SCHD; 431 NET_EPOCH_CALL(rs_destroy, &rs->rs_epoch_ctx); 432 } 433 434 #ifdef INET 435 extern counter_u64_t rate_limit_new; 436 extern counter_u64_t rate_limit_chg; 437 extern counter_u64_t rate_limit_set_ok; 438 extern counter_u64_t rate_limit_active; 439 extern counter_u64_t rate_limit_alloc_fail; 440 #endif 441 442 static int 443 rl_attach_txrtlmt(struct ifnet *ifp, 444 uint32_t flowtype, 445 int flowid, 446 uint64_t cfg_rate, 447 struct m_snd_tag **tag) 448 { 449 int error; 450 union if_snd_tag_alloc_params params = { 451 .rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT, 452 .rate_limit.hdr.flowid = flowid, 453 .rate_limit.hdr.flowtype = flowtype, 454 .rate_limit.max_rate = cfg_rate, 455 .rate_limit.flags = M_NOWAIT, 456 }; 457 458 error = m_snd_tag_alloc(ifp, ¶ms, tag); 459 #ifdef INET 460 if (error == 0) { 461 counter_u64_add(rate_limit_set_ok, 1); 462 counter_u64_add(rate_limit_active, 1); 463 } else if (error != EOPNOTSUPP) 464 counter_u64_add(rate_limit_alloc_fail, 1); 465 #endif 466 return (error); 467 } 468 469 static void 470 populate_canned_table(struct tcp_rate_set *rs, const uint64_t *rate_table_act) 471 { 472 /* 473 * The internal table is "special", it 474 * is two seperate ordered tables that 475 * must be merged. We get here when the 476 * adapter specifies a number of rates that 477 * covers both ranges in the table in some 478 * form. 479 */ 480 int i, at_low, at_high; 481 uint8_t low_disabled = 0, high_disabled = 0; 482 483 for(i = 0, at_low = 0, at_high = RS_NEXT_ORDER_GROUP; i < rs->rs_rate_cnt; i++) { 484 rs->rs_rlt[i].flags = 0; 485 rs->rs_rlt[i].time_between = 0; 486 if ((low_disabled == 0) && 487 (high_disabled || 488 (rate_table_act[at_low] < rate_table_act[at_high]))) { 489 rs->rs_rlt[i].rate = rate_table_act[at_low]; 490 at_low++; 491 if (at_low == RS_NEXT_ORDER_GROUP) 492 low_disabled = 1; 493 } else if (high_disabled == 0) { 494 rs->rs_rlt[i].rate = rate_table_act[at_high]; 495 at_high++; 496 if (at_high == MAX_HDWR_RATES) 497 high_disabled = 1; 498 } 499 } 500 } 501 502 static struct tcp_rate_set * 503 rt_setup_new_rs(struct ifnet *ifp, int *error) 504 { 505 struct tcp_rate_set *rs; 506 const uint64_t *rate_table_act; 507 uint64_t lentim, res; 508 size_t sz; 509 uint32_t hash_type; 510 int i; 511 struct if_ratelimit_query_results rl; 512 struct sysctl_oid *rl_sysctl_root; 513 struct epoch_tracker et; 514 /* 515 * We expect to enter with the 516 * mutex locked. 517 */ 518 519 if (ifp->if_ratelimit_query == NULL) { 520 /* 521 * We can do nothing if we cannot 522 * get a query back from the driver. 523 */ 524 printf("Warning:No query functions for %s:%d-- failed\n", 525 ifp->if_dname, ifp->if_dunit); 526 return (NULL); 527 } 528 rs = malloc(sizeof(struct tcp_rate_set), M_TCPPACE, M_NOWAIT | M_ZERO); 529 if (rs == NULL) { 530 if (error) 531 *error = ENOMEM; 532 printf("Warning:No memory for malloc of tcp_rate_set\n"); 533 return (NULL); 534 } 535 memset(&rl, 0, sizeof(rl)); 536 rl.flags = RT_NOSUPPORT; 537 ifp->if_ratelimit_query(ifp, &rl); 538 if (rl.flags & RT_IS_UNUSABLE) { 539 /* 540 * The interface does not really support 541 * the rate-limiting. 542 */ 543 memset(rs, 0, sizeof(struct tcp_rate_set)); 544 rs->rs_ifp = ifp; 545 rs->rs_if_dunit = ifp->if_dunit; 546 rs->rs_flags = RS_INTF_NO_SUP; 547 rs->rs_disable = 1; 548 rs_number_alive++; 549 sysctl_ctx_init(&rs->sysctl_ctx); 550 rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 551 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), 552 OID_AUTO, 553 rs->rs_ifp->if_xname, 554 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 555 ""); 556 rl_add_syctl_entries(rl_sysctl_root, rs); 557 NET_EPOCH_ENTER(et); 558 mtx_lock(&rs_mtx); 559 CK_LIST_INSERT_HEAD(&int_rs, rs, next); 560 mtx_unlock(&rs_mtx); 561 NET_EPOCH_EXIT(et); 562 return (rs); 563 } else if ((rl.flags & RT_IS_INDIRECT) == RT_IS_INDIRECT) { 564 memset(rs, 0, sizeof(struct tcp_rate_set)); 565 rs->rs_ifp = ifp; 566 rs->rs_if_dunit = ifp->if_dunit; 567 rs->rs_flags = RS_IS_DEFF; 568 rs_number_alive++; 569 sysctl_ctx_init(&rs->sysctl_ctx); 570 rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 571 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), 572 OID_AUTO, 573 rs->rs_ifp->if_xname, 574 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 575 ""); 576 rl_add_syctl_entries(rl_sysctl_root, rs); 577 NET_EPOCH_ENTER(et); 578 mtx_lock(&rs_mtx); 579 CK_LIST_INSERT_HEAD(&int_rs, rs, next); 580 mtx_unlock(&rs_mtx); 581 NET_EPOCH_EXIT(et); 582 return (rs); 583 } else if ((rl.flags & RT_IS_FIXED_TABLE) == RT_IS_FIXED_TABLE) { 584 /* Mellanox C4 likely */ 585 rs->rs_ifp = ifp; 586 rs->rs_if_dunit = ifp->if_dunit; 587 rs->rs_rate_cnt = rl.number_of_rates; 588 rs->rs_min_seg = rl.min_segment_burst; 589 rs->rs_highest_valid = 0; 590 rs->rs_flow_limit = rl.max_flows; 591 rs->rs_flags = RS_IS_INTF | RS_NO_PRE; 592 rs->rs_disable = 0; 593 rate_table_act = rl.rate_table; 594 } else if ((rl.flags & RT_IS_SELECTABLE) == RT_IS_SELECTABLE) { 595 /* Chelsio, C5 and C6 of Mellanox? */ 596 rs->rs_ifp = ifp; 597 rs->rs_if_dunit = ifp->if_dunit; 598 rs->rs_rate_cnt = rl.number_of_rates; 599 rs->rs_min_seg = rl.min_segment_burst; 600 rs->rs_disable = 0; 601 rs->rs_flow_limit = rl.max_flows; 602 rate_table_act = desired_rates; 603 if ((rs->rs_rate_cnt > MAX_HDWR_RATES) && 604 (rs->rs_rate_cnt < ALL_HARDWARE_RATES)) { 605 /* 606 * Our desired table is not big 607 * enough, do what we can. 608 */ 609 rs->rs_rate_cnt = MAX_HDWR_RATES; 610 } 611 if (rs->rs_rate_cnt <= RS_ORDERED_COUNT) 612 rs->rs_flags = RS_IS_INTF; 613 else 614 rs->rs_flags = RS_IS_INTF | RS_INT_TBL; 615 if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) 616 rs->rs_rate_cnt = ALL_HARDWARE_RATES; 617 } else { 618 free(rs, M_TCPPACE); 619 return (NULL); 620 } 621 sz = sizeof(struct tcp_hwrate_limit_table) * rs->rs_rate_cnt; 622 rs->rs_rlt = malloc(sz, M_TCPPACE, M_NOWAIT); 623 if (rs->rs_rlt == NULL) { 624 if (error) 625 *error = ENOMEM; 626 bail: 627 free(rs, M_TCPPACE); 628 return (NULL); 629 } 630 if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) { 631 /* 632 * The interface supports all 633 * the rates we could possibly want. 634 */ 635 uint64_t rat; 636 637 rs->rs_rlt[0].rate = 12500; /* 100k */ 638 rs->rs_rlt[1].rate = 25000; /* 200k */ 639 rs->rs_rlt[2].rate = 62500; /* 500k */ 640 /* Note 125000 == 1Megabit 641 * populate 1Meg - 1000meg. 642 */ 643 for(i = 3, rat = 125000; i< (ALL_HARDWARE_RATES-1); i++) { 644 rs->rs_rlt[i].rate = rat; 645 rat += 125000; 646 } 647 rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate = 1250000000; 648 } else if (rs->rs_flags & RS_INT_TBL) { 649 /* We populate this in a special way */ 650 populate_canned_table(rs, rate_table_act); 651 } else { 652 /* 653 * Just copy in the rates from 654 * the table, it is in order. 655 */ 656 for (i=0; i<rs->rs_rate_cnt; i++) { 657 rs->rs_rlt[i].rate = rate_table_act[i]; 658 rs->rs_rlt[i].time_between = 0; 659 rs->rs_rlt[i].flags = 0; 660 } 661 } 662 for (i = (rs->rs_rate_cnt - 1); i >= 0; i--) { 663 /* 664 * We go backwards through the list so that if we can't get 665 * a rate and fail to init one, we have at least a chance of 666 * getting the highest one. 667 */ 668 rs->rs_rlt[i].ptbl = rs; 669 rs->rs_rlt[i].tag = NULL; 670 /* 671 * Calculate the time between. 672 */ 673 lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND; 674 res = lentim / rs->rs_rlt[i].rate; 675 if (res > 0) 676 rs->rs_rlt[i].time_between = res; 677 else 678 rs->rs_rlt[i].time_between = 1; 679 if (rs->rs_flags & RS_NO_PRE) { 680 rs->rs_rlt[i].flags = HDWRPACE_INITED; 681 rs->rs_lowest_valid = i; 682 } else { 683 int err; 684 685 if ((rl.flags & RT_IS_SETUP_REQ) && 686 (ifp->if_ratelimit_query)) { 687 err = ifp->if_ratelimit_setup(ifp, 688 rs->rs_rlt[i].rate, i); 689 if (err) 690 goto handle_err; 691 } 692 #ifdef RSS 693 hash_type = M_HASHTYPE_RSS_TCP_IPV4; 694 #else 695 hash_type = M_HASHTYPE_OPAQUE_HASH; 696 #endif 697 err = rl_attach_txrtlmt(ifp, 698 hash_type, 699 (i + 1), 700 rs->rs_rlt[i].rate, 701 &rs->rs_rlt[i].tag); 702 if (err) { 703 handle_err: 704 if (i == (rs->rs_rate_cnt - 1)) { 705 /* 706 * Huh - first rate and we can't get 707 * it? 708 */ 709 free(rs->rs_rlt, M_TCPPACE); 710 if (error) 711 *error = err; 712 goto bail; 713 } else { 714 if (error) 715 *error = err; 716 } 717 break; 718 } else { 719 rs->rs_rlt[i].flags = HDWRPACE_INITED | HDWRPACE_TAGPRESENT; 720 rs->rs_lowest_valid = i; 721 } 722 } 723 } 724 /* Did we get at least 1 rate? */ 725 if (rs->rs_rlt[(rs->rs_rate_cnt - 1)].flags & HDWRPACE_INITED) 726 rs->rs_highest_valid = rs->rs_rate_cnt - 1; 727 else { 728 free(rs->rs_rlt, M_TCPPACE); 729 goto bail; 730 } 731 rs_number_alive++; 732 sysctl_ctx_init(&rs->sysctl_ctx); 733 rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, 734 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), 735 OID_AUTO, 736 rs->rs_ifp->if_xname, 737 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 738 ""); 739 rl_add_syctl_entries(rl_sysctl_root, rs); 740 NET_EPOCH_ENTER(et); 741 mtx_lock(&rs_mtx); 742 CK_LIST_INSERT_HEAD(&int_rs, rs, next); 743 mtx_unlock(&rs_mtx); 744 NET_EPOCH_EXIT(et); 745 return (rs); 746 } 747 748 /* 749 * For an explanation of why the argument is volatile please 750 * look at the comments around rt_setup_rate(). 751 */ 752 static const struct tcp_hwrate_limit_table * 753 tcp_int_find_suitable_rate(const volatile struct tcp_rate_set *rs, 754 uint64_t bytes_per_sec, uint32_t flags, uint64_t *lower_rate) 755 { 756 struct tcp_hwrate_limit_table *arte = NULL, *rte = NULL; 757 uint64_t mbits_per_sec, ind_calc, previous_rate = 0; 758 int i; 759 760 mbits_per_sec = (bytes_per_sec * 8); 761 if (flags & RS_PACING_LT) { 762 if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && 763 (rs->rs_lowest_valid <= 2)){ 764 /* 765 * Smaller than 1Meg, only 766 * 3 entries can match it. 767 */ 768 previous_rate = 0; 769 for(i = rs->rs_lowest_valid; i < 3; i++) { 770 if (bytes_per_sec <= rs->rs_rlt[i].rate) { 771 rte = &rs->rs_rlt[i]; 772 break; 773 } else if (rs->rs_rlt[i].flags & HDWRPACE_INITED) { 774 arte = &rs->rs_rlt[i]; 775 } 776 previous_rate = rs->rs_rlt[i].rate; 777 } 778 goto done; 779 } else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) && 780 (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){ 781 /* 782 * Larger than 1G (the majority of 783 * our table. 784 */ 785 if (mbits_per_sec < RS_TEN_GIGABIT_PERSEC) 786 rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 787 else 788 arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 789 previous_rate = rs->rs_rlt[(ALL_HARDWARE_RATES-2)].rate; 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 if (ind_calc >= 1) 811 previous_rate = rs->rs_rlt[(ind_calc-1)].rate; 812 } 813 } else if (flags & RS_PACING_EXACT_MATCH) { 814 if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && 815 (rs->rs_lowest_valid <= 2)){ 816 for(i = rs->rs_lowest_valid; i < 3; i++) { 817 if (bytes_per_sec == rs->rs_rlt[i].rate) { 818 rte = &rs->rs_rlt[i]; 819 break; 820 } 821 } 822 } else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) && 823 (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) { 824 /* > 1Gbps only one rate */ 825 if (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) { 826 /* Its 10G wow */ 827 rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 828 } 829 } else { 830 /* Ok it must be a exact meg (its between 1G and 1Meg) */ 831 ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC; 832 if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) { 833 /* its an exact Mbps */ 834 ind_calc += 2; 835 if (ind_calc > (ALL_HARDWARE_RATES-1)) { 836 /* This should not happen */ 837 ind_calc = ALL_HARDWARE_RATES-1; 838 } 839 if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) 840 rte = &rs->rs_rlt[ind_calc]; 841 } 842 } 843 } else { 844 /* we want greater than the requested rate */ 845 if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && 846 (rs->rs_lowest_valid <= 2)){ 847 arte = &rs->rs_rlt[3]; /* set alternate to 1Meg */ 848 for (i=2; i>=rs->rs_lowest_valid; i--) { 849 if (bytes_per_sec < rs->rs_rlt[i].rate) { 850 rte = &rs->rs_rlt[i]; 851 if (i >= 1) { 852 previous_rate = rs->rs_rlt[(i-1)].rate; 853 } 854 break; 855 } else if ((flags & RS_PACING_GEQ) && 856 (bytes_per_sec == rs->rs_rlt[i].rate)) { 857 rte = &rs->rs_rlt[i]; 858 if (i >= 1) { 859 previous_rate = rs->rs_rlt[(i-1)].rate; 860 } 861 break; 862 } else { 863 arte = &rs->rs_rlt[i]; /* new alternate */ 864 } 865 } 866 } else if (mbits_per_sec > RS_ONE_GIGABIT_PERSEC) { 867 if ((bytes_per_sec < rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) && 868 (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){ 869 /* Our top rate is larger than the request */ 870 rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 871 } else if ((flags & RS_PACING_GEQ) && 872 (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) && 873 (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) { 874 /* It matches our top rate */ 875 rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 876 } else if (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED) { 877 /* The top rate is an alternative */ 878 arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; 879 } 880 previous_rate = rs->rs_rlt[(ALL_HARDWARE_RATES-2)].rate; 881 } else { 882 /* Its in our range 1Meg - 1Gig */ 883 if (flags & RS_PACING_GEQ) { 884 ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC; 885 if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) { 886 if (ind_calc > (ALL_HARDWARE_RATES-1)) { 887 /* This should not happen */ 888 ind_calc = (ALL_HARDWARE_RATES-1); 889 } 890 rte = &rs->rs_rlt[ind_calc]; 891 if (ind_calc >= 1) 892 previous_rate = rs->rs_rlt[(ind_calc-1)].rate; 893 } 894 goto done; 895 } 896 ind_calc = (mbits_per_sec + (RS_ONE_MEGABIT_PERSEC-1))/RS_ONE_MEGABIT_PERSEC; 897 ind_calc += 2; 898 if (ind_calc > (ALL_HARDWARE_RATES-1)) { 899 /* This should not happen */ 900 ind_calc = ALL_HARDWARE_RATES-1; 901 } 902 if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) { 903 rte = &rs->rs_rlt[ind_calc]; 904 if (ind_calc >= 1) 905 previous_rate = rs->rs_rlt[(ind_calc-1)].rate; 906 } 907 } 908 } 909 done: 910 if ((rte == NULL) && 911 (arte != NULL) && 912 (flags & RS_PACING_SUB_OK)) { 913 /* We can use the substitute */ 914 rte = arte; 915 } 916 if (lower_rate) 917 *lower_rate = previous_rate; 918 return (rte); 919 } 920 921 /* 922 * For an explanation of why the argument is volatile please 923 * look at the comments around rt_setup_rate(). 924 */ 925 static const struct tcp_hwrate_limit_table * 926 tcp_find_suitable_rate(const volatile struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags, uint64_t *lower_rate) 927 { 928 /** 929 * Hunt the rate table with the restrictions in flags and find a 930 * suitable rate if possible. 931 * RS_PACING_EXACT_MATCH - look for an exact match to rate. 932 * RS_PACING_GT - must be greater than. 933 * RS_PACING_GEQ - must be greater than or equal. 934 * RS_PACING_LT - must be less than. 935 * RS_PACING_SUB_OK - If we don't meet criteria a 936 * substitute is ok. 937 */ 938 int i, matched; 939 struct tcp_hwrate_limit_table *rte = NULL; 940 uint64_t previous_rate = 0; 941 942 if ((rs->rs_flags & RS_INT_TBL) && 943 (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)) { 944 /* 945 * Here we don't want to paw thru 946 * a big table, we have everything 947 * from 1Meg - 1000Meg in 1Meg increments. 948 * Use an alternate method to "lookup". 949 */ 950 return (tcp_int_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate)); 951 } 952 if ((flags & RS_PACING_LT) || 953 (flags & RS_PACING_EXACT_MATCH)) { 954 /* 955 * For exact and less than we go forward through the table. 956 * This way when we find one larger we stop (exact was a 957 * toss up). 958 */ 959 for (i = rs->rs_lowest_valid, matched = 0; i <= rs->rs_highest_valid; i++) { 960 if ((flags & RS_PACING_EXACT_MATCH) && 961 (bytes_per_sec == rs->rs_rlt[i].rate)) { 962 rte = &rs->rs_rlt[i]; 963 matched = 1; 964 if (lower_rate != NULL) 965 *lower_rate = previous_rate; 966 break; 967 } else if ((flags & RS_PACING_LT) && 968 (bytes_per_sec <= rs->rs_rlt[i].rate)) { 969 rte = &rs->rs_rlt[i]; 970 matched = 1; 971 if (lower_rate != NULL) 972 *lower_rate = previous_rate; 973 break; 974 } 975 previous_rate = rs->rs_rlt[i].rate; 976 if (bytes_per_sec > rs->rs_rlt[i].rate) 977 break; 978 } 979 if ((matched == 0) && 980 (flags & RS_PACING_LT) && 981 (flags & RS_PACING_SUB_OK)) { 982 /* Kick in a substitute (the lowest) */ 983 rte = &rs->rs_rlt[rs->rs_lowest_valid]; 984 } 985 } else { 986 /* 987 * Here we go backward through the table so that we can find 988 * the one greater in theory faster (but its probably a 989 * wash). 990 */ 991 for (i = rs->rs_highest_valid, matched = 0; i >= rs->rs_lowest_valid; i--) { 992 if (rs->rs_rlt[i].rate > bytes_per_sec) { 993 /* A possible candidate */ 994 rte = &rs->rs_rlt[i]; 995 } 996 if ((flags & RS_PACING_GEQ) && 997 (bytes_per_sec == rs->rs_rlt[i].rate)) { 998 /* An exact match and we want equal */ 999 matched = 1; 1000 rte = &rs->rs_rlt[i]; 1001 break; 1002 } else if (rte) { 1003 /* 1004 * Found one that is larger than but don't 1005 * stop, there may be a more closer match. 1006 */ 1007 matched = 1; 1008 } 1009 if (rs->rs_rlt[i].rate < bytes_per_sec) { 1010 /* 1011 * We found a table entry that is smaller, 1012 * stop there will be none greater or equal. 1013 */ 1014 if (lower_rate != NULL) 1015 *lower_rate = rs->rs_rlt[i].rate; 1016 break; 1017 } 1018 } 1019 if ((matched == 0) && 1020 (flags & RS_PACING_SUB_OK)) { 1021 /* Kick in a substitute (the highest) */ 1022 rte = &rs->rs_rlt[rs->rs_highest_valid]; 1023 } 1024 } 1025 return (rte); 1026 } 1027 1028 static struct ifnet * 1029 rt_find_real_interface(struct ifnet *ifp, struct inpcb *inp, int *error) 1030 { 1031 struct ifnet *tifp; 1032 struct m_snd_tag *tag, *ntag; 1033 union if_snd_tag_alloc_params params = { 1034 .rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT, 1035 .rate_limit.hdr.flowid = inp->inp_flowid, 1036 .rate_limit.hdr.numa_domain = inp->inp_numa_domain, 1037 .rate_limit.max_rate = COMMON_RATE, 1038 .rate_limit.flags = M_NOWAIT, 1039 }; 1040 int err; 1041 #ifdef RSS 1042 params.rate_limit.hdr.flowtype = ((inp->inp_vflag & INP_IPV6) ? 1043 M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_TCP_IPV4); 1044 #else 1045 params.rate_limit.hdr.flowtype = M_HASHTYPE_OPAQUE_HASH; 1046 #endif 1047 err = m_snd_tag_alloc(ifp, ¶ms, &tag); 1048 if (err) { 1049 /* Failed to setup a tag? */ 1050 if (error) 1051 *error = err; 1052 return (NULL); 1053 } 1054 ntag = tag; 1055 while(ntag->ifp->if_next_snd_tag != NULL) { 1056 ntag = ntag->ifp->if_next_snd_tag(ntag); 1057 } 1058 tifp = ntag->ifp; 1059 m_snd_tag_rele(tag); 1060 return (tifp); 1061 } 1062 1063 static void 1064 rl_increment_using(const struct tcp_hwrate_limit_table *rte) 1065 { 1066 } 1067 1068 static void 1069 rl_decrement_using(const struct tcp_hwrate_limit_table *rte) 1070 { 1071 } 1072 1073 void 1074 tcp_rl_log_enobuf(const struct tcp_hwrate_limit_table *rte) 1075 { 1076 } 1077 1078 /* 1079 * Do NOT take the __noinline out of the 1080 * find_rs_for_ifp() function. If you do the inline 1081 * of it for the rt_setup_rate() will show you a 1082 * compiler bug. For some reason the compiler thinks 1083 * the list can never be empty. The consequence of 1084 * this will be a crash when we dereference NULL 1085 * if an ifp is removed just has a hw rate limit 1086 * is attempted. If you are working on the compiler 1087 * and want to "test" this go ahead and take the noinline 1088 * out otherwise let sleeping dogs ly until such time 1089 * as we get a compiler fix 10/2/20 -- RRS 1090 */ 1091 static __noinline struct tcp_rate_set * 1092 find_rs_for_ifp(struct ifnet *ifp) 1093 { 1094 struct tcp_rate_set *rs; 1095 1096 CK_LIST_FOREACH(rs, &int_rs, next) { 1097 if ((rs->rs_ifp == ifp) && 1098 (rs->rs_if_dunit == ifp->if_dunit)) { 1099 /* Ok we found it */ 1100 return (rs); 1101 } 1102 } 1103 return (NULL); 1104 } 1105 1106 1107 static const struct tcp_hwrate_limit_table * 1108 rt_setup_rate(struct inpcb *inp, struct ifnet *ifp, uint64_t bytes_per_sec, 1109 uint32_t flags, int *error, uint64_t *lower_rate) 1110 { 1111 /* First lets find the interface if it exists */ 1112 const struct tcp_hwrate_limit_table *rte; 1113 /* 1114 * So why is rs volatile? This is to defeat a 1115 * compiler bug where in the compiler is convinced 1116 * that rs can never be NULL (which is not true). Because 1117 * of its conviction it nicely optimizes out the if ((rs == NULL 1118 * below which means if you get a NULL back you dereference it. 1119 */ 1120 volatile struct tcp_rate_set *rs; 1121 struct epoch_tracker et; 1122 struct ifnet *oifp = ifp; 1123 int err; 1124 1125 NET_EPOCH_ENTER(et); 1126 use_real_interface: 1127 rs = find_rs_for_ifp(ifp); 1128 if ((rs == NULL) || 1129 (rs->rs_flags & RS_INTF_NO_SUP) || 1130 (rs->rs_flags & RS_IS_DEAD)) { 1131 /* 1132 * This means we got a packet *before* 1133 * the IF-UP was processed below, <or> 1134 * while or after we already received an interface 1135 * departed event. In either case we really don't 1136 * want to do anything with pacing, in 1137 * the departing case the packet is not 1138 * going to go very far. The new case 1139 * might be arguable, but its impossible 1140 * to tell from the departing case. 1141 */ 1142 if (error) 1143 *error = ENODEV; 1144 NET_EPOCH_EXIT(et); 1145 return (NULL); 1146 } 1147 1148 if ((rs == NULL) || (rs->rs_disable != 0)) { 1149 if (error) 1150 *error = ENOSPC; 1151 NET_EPOCH_EXIT(et); 1152 return (NULL); 1153 } 1154 if (rs->rs_flags & RS_IS_DEFF) { 1155 /* We need to find the real interface */ 1156 struct ifnet *tifp; 1157 1158 tifp = rt_find_real_interface(ifp, inp, error); 1159 if (tifp == NULL) { 1160 if (rs->rs_disable && error) 1161 *error = ENOTSUP; 1162 NET_EPOCH_EXIT(et); 1163 return (NULL); 1164 } 1165 KASSERT((tifp != ifp), 1166 ("Lookup failure ifp:%p inp:%p rt_find_real_interface() returns the same interface tifp:%p?\n", 1167 ifp, inp, tifp)); 1168 ifp = tifp; 1169 goto use_real_interface; 1170 } 1171 if (rs->rs_flow_limit && 1172 ((rs->rs_flows_using + 1) > rs->rs_flow_limit)) { 1173 if (error) 1174 *error = ENOSPC; 1175 NET_EPOCH_EXIT(et); 1176 return (NULL); 1177 } 1178 rte = tcp_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate); 1179 if (rte) { 1180 err = in_pcbattach_txrtlmt(inp, oifp, 1181 inp->inp_flowtype, 1182 inp->inp_flowid, 1183 rte->rate, 1184 &inp->inp_snd_tag); 1185 if (err) { 1186 /* Failed to attach */ 1187 if (error) 1188 *error = err; 1189 rte = NULL; 1190 } else { 1191 KASSERT((inp->inp_snd_tag != NULL) , 1192 ("Setup rate has no snd_tag inp:%p rte:%p rate:%llu rs:%p", 1193 inp, rte, (unsigned long long)rte->rate, rs)); 1194 #ifdef INET 1195 counter_u64_add(rate_limit_new, 1); 1196 #endif 1197 } 1198 } 1199 if (rte) { 1200 /* 1201 * We use an atomic here for accounting so we don't have to 1202 * use locks when freeing. 1203 */ 1204 atomic_add_64(&rs->rs_flows_using, 1); 1205 } 1206 NET_EPOCH_EXIT(et); 1207 return (rte); 1208 } 1209 1210 static void 1211 tcp_rl_ifnet_link(void *arg __unused, struct ifnet *ifp, int link_state) 1212 { 1213 int error; 1214 struct tcp_rate_set *rs; 1215 struct epoch_tracker et; 1216 1217 if (((ifp->if_capenable & IFCAP_TXRTLMT) == 0) || 1218 (link_state != LINK_STATE_UP)) { 1219 /* 1220 * We only care on an interface going up that is rate-limit 1221 * capable. 1222 */ 1223 return; 1224 } 1225 NET_EPOCH_ENTER(et); 1226 mtx_lock(&rs_mtx); 1227 rs = find_rs_for_ifp(ifp); 1228 if (rs) { 1229 /* We already have initialized this guy */ 1230 mtx_unlock(&rs_mtx); 1231 NET_EPOCH_EXIT(et); 1232 return; 1233 } 1234 mtx_unlock(&rs_mtx); 1235 NET_EPOCH_EXIT(et); 1236 rt_setup_new_rs(ifp, &error); 1237 } 1238 1239 static void 1240 tcp_rl_ifnet_departure(void *arg __unused, struct ifnet *ifp) 1241 { 1242 struct tcp_rate_set *rs; 1243 struct epoch_tracker et; 1244 int i; 1245 1246 NET_EPOCH_ENTER(et); 1247 mtx_lock(&rs_mtx); 1248 rs = find_rs_for_ifp(ifp); 1249 if (rs) { 1250 CK_LIST_REMOVE(rs, next); 1251 rs_number_alive--; 1252 rs->rs_flags |= RS_IS_DEAD; 1253 for (i = 0; i < rs->rs_rate_cnt; i++) { 1254 if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) { 1255 in_pcbdetach_tag(rs->rs_rlt[i].tag); 1256 rs->rs_rlt[i].tag = NULL; 1257 } 1258 rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED; 1259 } 1260 if (rs->rs_flows_using == 0) 1261 rs_defer_destroy(rs); 1262 } 1263 mtx_unlock(&rs_mtx); 1264 NET_EPOCH_EXIT(et); 1265 } 1266 1267 static void 1268 tcp_rl_shutdown(void *arg __unused, int howto __unused) 1269 { 1270 struct tcp_rate_set *rs, *nrs; 1271 struct epoch_tracker et; 1272 int i; 1273 1274 NET_EPOCH_ENTER(et); 1275 mtx_lock(&rs_mtx); 1276 CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) { 1277 CK_LIST_REMOVE(rs, next); 1278 rs_number_alive--; 1279 rs->rs_flags |= RS_IS_DEAD; 1280 for (i = 0; i < rs->rs_rate_cnt; i++) { 1281 if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) { 1282 in_pcbdetach_tag(rs->rs_rlt[i].tag); 1283 rs->rs_rlt[i].tag = NULL; 1284 } 1285 rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED; 1286 } 1287 if (rs->rs_flows_using == 0) 1288 rs_defer_destroy(rs); 1289 } 1290 mtx_unlock(&rs_mtx); 1291 NET_EPOCH_EXIT(et); 1292 } 1293 1294 const struct tcp_hwrate_limit_table * 1295 tcp_set_pacing_rate(struct tcpcb *tp, struct ifnet *ifp, 1296 uint64_t bytes_per_sec, int flags, int *error, uint64_t *lower_rate) 1297 { 1298 const struct tcp_hwrate_limit_table *rte; 1299 #ifdef KERN_TLS 1300 struct ktls_session *tls; 1301 #endif 1302 1303 INP_WLOCK_ASSERT(tp->t_inpcb); 1304 1305 if (tp->t_inpcb->inp_snd_tag == NULL) { 1306 /* 1307 * We are setting up a rate for the first time. 1308 */ 1309 if ((ifp->if_capenable & IFCAP_TXRTLMT) == 0) { 1310 /* Not supported by the egress */ 1311 if (error) 1312 *error = ENODEV; 1313 return (NULL); 1314 } 1315 #ifdef KERN_TLS 1316 tls = NULL; 1317 if (tp->t_inpcb->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) { 1318 tls = tp->t_inpcb->inp_socket->so_snd.sb_tls_info; 1319 1320 if ((ifp->if_capenable & IFCAP_TXTLS_RTLMT) == 0 || 1321 tls->mode != TCP_TLS_MODE_IFNET) { 1322 if (error) 1323 *error = ENODEV; 1324 return (NULL); 1325 } 1326 } 1327 #endif 1328 rte = rt_setup_rate(tp->t_inpcb, ifp, bytes_per_sec, flags, error, lower_rate); 1329 if (rte) 1330 rl_increment_using(rte); 1331 #ifdef KERN_TLS 1332 if (rte != NULL && tls != NULL && tls->snd_tag != NULL) { 1333 /* 1334 * Fake a route change error to reset the TLS 1335 * send tag. This will convert the existing 1336 * tag to a TLS ratelimit tag. 1337 */ 1338 MPASS(tls->snd_tag->type == IF_SND_TAG_TYPE_TLS); 1339 ktls_output_eagain(tp->t_inpcb, tls); 1340 } 1341 #endif 1342 } else { 1343 /* 1344 * We are modifying a rate, wrong interface? 1345 */ 1346 if (error) 1347 *error = EINVAL; 1348 rte = NULL; 1349 } 1350 if (rte != NULL) { 1351 tp->t_pacing_rate = rte->rate; 1352 *error = 0; 1353 } 1354 return (rte); 1355 } 1356 1357 const struct tcp_hwrate_limit_table * 1358 tcp_chg_pacing_rate(const struct tcp_hwrate_limit_table *crte, 1359 struct tcpcb *tp, struct ifnet *ifp, 1360 uint64_t bytes_per_sec, int flags, int *error, uint64_t *lower_rate) 1361 { 1362 const struct tcp_hwrate_limit_table *nrte; 1363 const struct tcp_rate_set *rs; 1364 #ifdef KERN_TLS 1365 struct ktls_session *tls = NULL; 1366 #endif 1367 int err; 1368 1369 INP_WLOCK_ASSERT(tp->t_inpcb); 1370 1371 if (crte == NULL) { 1372 /* Wrong interface */ 1373 if (error) 1374 *error = EINVAL; 1375 return (NULL); 1376 } 1377 1378 #ifdef KERN_TLS 1379 if (tp->t_inpcb->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) { 1380 tls = tp->t_inpcb->inp_socket->so_snd.sb_tls_info; 1381 MPASS(tls->mode == TCP_TLS_MODE_IFNET); 1382 if (tls->snd_tag != NULL && 1383 tls->snd_tag->type != IF_SND_TAG_TYPE_TLS_RATE_LIMIT) { 1384 /* 1385 * NIC probably doesn't support ratelimit TLS 1386 * tags if it didn't allocate one when an 1387 * existing rate was present, so ignore. 1388 */ 1389 if (error) 1390 *error = EOPNOTSUPP; 1391 return (NULL); 1392 } 1393 } 1394 #endif 1395 if (tp->t_inpcb->inp_snd_tag == NULL) { 1396 /* Wrong interface */ 1397 if (error) 1398 *error = EINVAL; 1399 return (NULL); 1400 } 1401 rs = crte->ptbl; 1402 if ((rs->rs_flags & RS_IS_DEAD) || 1403 (crte->flags & HDWRPACE_IFPDEPARTED)) { 1404 /* Release the rate, and try anew */ 1405 1406 tcp_rel_pacing_rate(crte, tp); 1407 nrte = tcp_set_pacing_rate(tp, ifp, 1408 bytes_per_sec, flags, error, lower_rate); 1409 return (nrte); 1410 } 1411 nrte = tcp_find_suitable_rate(rs, bytes_per_sec, flags, lower_rate); 1412 if (nrte == crte) { 1413 /* No change */ 1414 if (error) 1415 *error = 0; 1416 return (crte); 1417 } 1418 if (nrte == NULL) { 1419 /* Release the old rate */ 1420 if (error) 1421 *error = ENOENT; 1422 tcp_rel_pacing_rate(crte, tp); 1423 return (NULL); 1424 } 1425 rl_decrement_using(crte); 1426 rl_increment_using(nrte); 1427 /* Change rates to our new entry */ 1428 #ifdef KERN_TLS 1429 if (tls != NULL) 1430 err = ktls_modify_txrtlmt(tls, nrte->rate); 1431 else 1432 #endif 1433 err = in_pcbmodify_txrtlmt(tp->t_inpcb, nrte->rate); 1434 if (err) { 1435 rl_decrement_using(nrte); 1436 /* Do we still have a snd-tag attached? */ 1437 if (tp->t_inpcb->inp_snd_tag) 1438 in_pcbdetach_txrtlmt(tp->t_inpcb); 1439 if (error) 1440 *error = err; 1441 return (NULL); 1442 } else { 1443 #ifdef INET 1444 counter_u64_add(rate_limit_chg, 1); 1445 #endif 1446 } 1447 if (error) 1448 *error = 0; 1449 tp->t_pacing_rate = nrte->rate; 1450 return (nrte); 1451 } 1452 1453 void 1454 tcp_rel_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp) 1455 { 1456 const struct tcp_rate_set *crs; 1457 struct tcp_rate_set *rs; 1458 uint64_t pre; 1459 1460 INP_WLOCK_ASSERT(tp->t_inpcb); 1461 1462 tp->t_pacing_rate = -1; 1463 crs = crte->ptbl; 1464 /* 1465 * Now we must break the const 1466 * in order to release our refcount. 1467 */ 1468 rs = __DECONST(struct tcp_rate_set *, crs); 1469 rl_decrement_using(crte); 1470 pre = atomic_fetchadd_64(&rs->rs_flows_using, -1); 1471 if (pre == 1) { 1472 struct epoch_tracker et; 1473 1474 NET_EPOCH_ENTER(et); 1475 mtx_lock(&rs_mtx); 1476 /* 1477 * Is it dead? 1478 */ 1479 if (rs->rs_flags & RS_IS_DEAD) 1480 rs_defer_destroy(rs); 1481 mtx_unlock(&rs_mtx); 1482 NET_EPOCH_EXIT(et); 1483 } 1484 1485 /* 1486 * XXX: If this connection is using ifnet TLS, should we 1487 * switch it to using an unlimited rate, or perhaps use 1488 * ktls_output_eagain() to reset the send tag to a plain 1489 * TLS tag? 1490 */ 1491 in_pcbdetach_txrtlmt(tp->t_inpcb); 1492 } 1493 1494 #define ONE_POINT_TWO_MEG 150000 /* 1.2 megabits in bytes */ 1495 #define ONE_HUNDRED_MBPS 12500000 /* 100Mbps in bytes per second */ 1496 #define FIVE_HUNDRED_MBPS 62500000 /* 500Mbps in bytes per second */ 1497 #define MAX_MSS_SENT 43 /* 43 mss = 43 x 1500 = 64,500 bytes */ 1498 1499 static void 1500 tcp_log_pacing_size(struct tcpcb *tp, uint64_t bw, uint32_t segsiz, uint32_t new_tso, 1501 uint64_t hw_rate, uint32_t time_between, uint32_t calc_time_between, 1502 uint32_t segs, uint32_t res_div, uint16_t mult, uint8_t mod) 1503 { 1504 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 1505 union tcp_log_stackspecific log; 1506 struct timeval tv; 1507 uint32_t cts; 1508 1509 memset(&log, 0, sizeof(log)); 1510 cts = tcp_get_usecs(&tv); 1511 log.u_bbr.flex1 = segsiz; 1512 log.u_bbr.flex2 = new_tso; 1513 log.u_bbr.flex3 = time_between; 1514 log.u_bbr.flex4 = calc_time_between; 1515 log.u_bbr.flex5 = segs; 1516 log.u_bbr.flex6 = res_div; 1517 log.u_bbr.flex7 = mult; 1518 log.u_bbr.flex8 = mod; 1519 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 1520 log.u_bbr.cur_del_rate = bw; 1521 log.u_bbr.delRate = hw_rate; 1522 TCP_LOG_EVENTP(tp, NULL, 1523 &tp->t_inpcb->inp_socket->so_rcv, 1524 &tp->t_inpcb->inp_socket->so_snd, 1525 TCP_HDWR_PACE_SIZE, 0, 1526 0, &log, false, &tv); 1527 } 1528 } 1529 1530 uint32_t 1531 tcp_get_pacing_burst_size (struct tcpcb *tp, uint64_t bw, uint32_t segsiz, int can_use_1mss, 1532 const struct tcp_hwrate_limit_table *te, int *err) 1533 { 1534 /* 1535 * We use the google formula to calculate the 1536 * TSO size. I.E. 1537 * bw < 24Meg 1538 * tso = 2mss 1539 * else 1540 * tso = min(bw/1000, 64k) 1541 * 1542 * Note for these calculations we ignore the 1543 * packet overhead (enet hdr, ip hdr and tcp hdr). 1544 */ 1545 uint64_t lentim, res, bytes; 1546 uint32_t new_tso, min_tso_segs; 1547 1548 bytes = bw / 1000; 1549 if (bytes > (64 * 1000)) 1550 bytes = 64 * 1000; 1551 /* Round up */ 1552 new_tso = (bytes + segsiz - 1) / segsiz; 1553 if (can_use_1mss && (bw < ONE_POINT_TWO_MEG)) 1554 min_tso_segs = 1; 1555 else 1556 min_tso_segs = 2; 1557 if (rs_floor_mss && (new_tso < rs_floor_mss)) 1558 new_tso = rs_floor_mss; 1559 else if (new_tso < min_tso_segs) 1560 new_tso = min_tso_segs; 1561 if (new_tso > MAX_MSS_SENT) 1562 new_tso = MAX_MSS_SENT; 1563 new_tso *= segsiz; 1564 tcp_log_pacing_size(tp, bw, segsiz, new_tso, 1565 0, 0, 0, 0, 0, 0, 1); 1566 /* 1567 * If we are not doing hardware pacing 1568 * then we are done. 1569 */ 1570 if (te == NULL) { 1571 if (err) 1572 *err = 0; 1573 return(new_tso); 1574 } 1575 /* 1576 * For hardware pacing we look at the 1577 * rate you are sending at and compare 1578 * that to the rate you have in hardware. 1579 * 1580 * If the hardware rate is slower than your 1581 * software rate then you are in error and 1582 * we will build a queue in our hardware whic 1583 * is probably not desired, in such a case 1584 * just return the non-hardware TSO size. 1585 * 1586 * If the rate in hardware is faster (which 1587 * it should be) then look at how long it 1588 * takes to send one ethernet segment size at 1589 * your b/w and compare that to the time it 1590 * takes to send at the rate you had selected. 1591 * 1592 * If your time is greater (which we hope it is) 1593 * we get the delta between the two, and then 1594 * divide that into your pacing time. This tells 1595 * us how many MSS you can send down at once (rounded up). 1596 * 1597 * Note we also double this value if the b/w is over 1598 * 100Mbps. If its over 500meg we just set you to the 1599 * max (43 segments). 1600 */ 1601 if (te->rate > FIVE_HUNDRED_MBPS) 1602 goto max; 1603 if (te->rate == bw) { 1604 /* We are pacing at exactly the hdwr rate */ 1605 max: 1606 tcp_log_pacing_size(tp, bw, segsiz, new_tso, 1607 te->rate, te->time_between, (uint32_t)0, 1608 (segsiz * MAX_MSS_SENT), 0, 0, 3); 1609 return (segsiz * MAX_MSS_SENT); 1610 } 1611 lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND; 1612 res = lentim / bw; 1613 if (res > te->time_between) { 1614 uint32_t delta, segs, res_div; 1615 1616 res_div = ((res * num_of_waits_allowed) + wait_time_floor); 1617 delta = res - te->time_between; 1618 segs = (res_div + delta - 1)/delta; 1619 if (segs < min_tso_segs) 1620 segs = min_tso_segs; 1621 if (segs < rs_hw_floor_mss) 1622 segs = rs_hw_floor_mss; 1623 if (segs > MAX_MSS_SENT) 1624 segs = MAX_MSS_SENT; 1625 segs *= segsiz; 1626 tcp_log_pacing_size(tp, bw, segsiz, new_tso, 1627 te->rate, te->time_between, (uint32_t)res, 1628 segs, res_div, 1, 3); 1629 if (err) 1630 *err = 0; 1631 if (segs < new_tso) { 1632 /* unexpected ? */ 1633 return(new_tso); 1634 } else { 1635 return (segs); 1636 } 1637 } else { 1638 /* 1639 * Your time is smaller which means 1640 * we will grow a queue on our 1641 * hardware. Send back the non-hardware 1642 * rate. 1643 */ 1644 tcp_log_pacing_size(tp, bw, segsiz, new_tso, 1645 te->rate, te->time_between, (uint32_t)res, 1646 0, 0, 0, 4); 1647 if (err) 1648 *err = -1; 1649 return (new_tso); 1650 } 1651 } 1652 1653 uint64_t 1654 tcp_hw_highest_rate_ifp(struct ifnet *ifp, struct inpcb *inp) 1655 { 1656 struct epoch_tracker et; 1657 struct tcp_rate_set *rs; 1658 uint64_t rate_ret; 1659 1660 NET_EPOCH_ENTER(et); 1661 use_next_interface: 1662 rs = find_rs_for_ifp(ifp); 1663 if (rs == NULL) { 1664 /* This interface does not do ratelimiting */ 1665 rate_ret = 0; 1666 } else if (rs->rs_flags & RS_IS_DEFF) { 1667 /* We need to find the real interface */ 1668 struct ifnet *tifp; 1669 1670 tifp = rt_find_real_interface(ifp, inp, NULL); 1671 if (tifp == NULL) { 1672 NET_EPOCH_EXIT(et); 1673 return (0); 1674 } 1675 ifp = tifp; 1676 goto use_next_interface; 1677 } else { 1678 /* Lets return the highest rate this guy has */ 1679 rate_ret = rs->rs_rlt[rs->rs_highest_valid].rate; 1680 } 1681 NET_EPOCH_EXIT(et); 1682 return(rate_ret); 1683 } 1684 1685 static eventhandler_tag rl_ifnet_departs; 1686 static eventhandler_tag rl_ifnet_arrives; 1687 static eventhandler_tag rl_shutdown_start; 1688 1689 static void 1690 tcp_rs_init(void *st __unused) 1691 { 1692 CK_LIST_INIT(&int_rs); 1693 rs_number_alive = 0; 1694 rs_number_dead = 0; 1695 mtx_init(&rs_mtx, "tcp_rs_mtx", "rsmtx", MTX_DEF); 1696 rl_ifnet_departs = EVENTHANDLER_REGISTER(ifnet_departure_event, 1697 tcp_rl_ifnet_departure, 1698 NULL, EVENTHANDLER_PRI_ANY); 1699 rl_ifnet_arrives = EVENTHANDLER_REGISTER(ifnet_link_event, 1700 tcp_rl_ifnet_link, 1701 NULL, EVENTHANDLER_PRI_ANY); 1702 rl_shutdown_start = EVENTHANDLER_REGISTER(shutdown_pre_sync, 1703 tcp_rl_shutdown, NULL, 1704 SHUTDOWN_PRI_FIRST); 1705 printf("TCP_ratelimit: Is now initialized\n"); 1706 } 1707 1708 SYSINIT(tcp_rl_init, SI_SUB_SMP + 1, SI_ORDER_ANY, tcp_rs_init, NULL); 1709 #endif 1710