1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2010 Luigi Rizzo, Riccardo Panicucci, Universita` di Pisa 5 * All rights reserved 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 /* 30 * Dummynet portions related to packet handling. 31 */ 32 #include <sys/cdefs.h> 33 __FBSDID("$FreeBSD$"); 34 35 #include "opt_inet6.h" 36 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/malloc.h> 40 #include <sys/mbuf.h> 41 #include <sys/kernel.h> 42 #include <sys/lock.h> 43 #include <sys/module.h> 44 #include <sys/mutex.h> 45 #include <sys/priv.h> 46 #include <sys/proc.h> 47 #include <sys/rwlock.h> 48 #include <sys/socket.h> 49 #include <sys/time.h> 50 #include <sys/sysctl.h> 51 52 #include <net/if.h> /* IFNAMSIZ, struct ifaddr, ifq head, lock.h mutex.h */ 53 #include <net/if_var.h> /* NET_EPOCH_... */ 54 #include <net/netisr.h> 55 #include <net/vnet.h> 56 57 #include <netinet/in.h> 58 #include <netinet/ip.h> /* ip_len, ip_off */ 59 #include <netinet/ip_var.h> /* ip_output(), IP_FORWARDING */ 60 #include <netinet/ip_fw.h> 61 #include <netinet/ip_dummynet.h> 62 #include <netinet/if_ether.h> /* various ether_* routines */ 63 #include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */ 64 #include <netinet6/ip6_var.h> 65 66 #include <netpfil/ipfw/ip_fw_private.h> 67 #include <netpfil/ipfw/dn_heap.h> 68 #include <netpfil/ipfw/ip_dn_private.h> 69 #ifdef NEW_AQM 70 #include <netpfil/ipfw/dn_aqm.h> 71 #endif 72 #include <netpfil/ipfw/dn_sched.h> 73 74 /* 75 * We keep a private variable for the simulation time, but we could 76 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c) 77 * instead of V_dn_cfg.curr_time 78 */ 79 VNET_DEFINE(struct dn_parms, dn_cfg); 80 #define V_dn_cfg VNET(dn_cfg) 81 82 /* 83 * We use a heap to store entities for which we have pending timer events. 84 * The heap is checked at every tick and all entities with expired events 85 * are extracted. 86 */ 87 88 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap"); 89 90 extern void (*bridge_dn_p)(struct mbuf *, struct ifnet *); 91 92 #ifdef SYSCTL_NODE 93 94 /* 95 * Because of the way the SYSBEGIN/SYSEND macros work on other 96 * platforms, there should not be functions between them. 97 * So keep the handlers outside the block. 98 */ 99 static int 100 sysctl_hash_size(SYSCTL_HANDLER_ARGS) 101 { 102 int error, value; 103 104 value = V_dn_cfg.hash_size; 105 error = sysctl_handle_int(oidp, &value, 0, req); 106 if (error != 0 || req->newptr == NULL) 107 return (error); 108 if (value < 16 || value > 65536) 109 return (EINVAL); 110 V_dn_cfg.hash_size = value; 111 return (0); 112 } 113 114 static int 115 sysctl_limits(SYSCTL_HANDLER_ARGS) 116 { 117 int error; 118 long value; 119 120 if (arg2 != 0) 121 value = V_dn_cfg.slot_limit; 122 else 123 value = V_dn_cfg.byte_limit; 124 error = sysctl_handle_long(oidp, &value, 0, req); 125 126 if (error != 0 || req->newptr == NULL) 127 return (error); 128 if (arg2 != 0) { 129 if (value < 1) 130 return (EINVAL); 131 V_dn_cfg.slot_limit = value; 132 } else { 133 if (value < 1500) 134 return (EINVAL); 135 V_dn_cfg.byte_limit = value; 136 } 137 return (0); 138 } 139 140 SYSBEGIN(f4) 141 142 SYSCTL_DECL(_net_inet); 143 SYSCTL_DECL(_net_inet_ip); 144 #ifdef NEW_AQM 145 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 146 "Dummynet"); 147 #else 148 static SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, 149 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 150 "Dummynet"); 151 #endif 152 153 /* wrapper to pass V_dn_cfg fields to SYSCTL_* */ 154 #define DC(x) (&(VNET_NAME(dn_cfg).x)) 155 156 /* parameters */ 157 158 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hash_size, 159 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 160 0, 0, sysctl_hash_size, "I", 161 "Default hash table size"); 162 163 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit, 164 CTLTYPE_LONG | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 165 0, 1, sysctl_limits, "L", 166 "Upper limit in slots for pipe queue."); 167 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit, 168 CTLTYPE_LONG | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 169 0, 0, sysctl_limits, "L", 170 "Upper limit in bytes for pipe queue."); 171 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast, 172 CTLFLAG_RW | CTLFLAG_VNET, DC(io_fast), 0, "Enable fast dummynet io."); 173 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, 174 CTLFLAG_RW | CTLFLAG_VNET, DC(debug), 0, "Dummynet debug level"); 175 176 /* RED parameters */ 177 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth, 178 CTLFLAG_RD | CTLFLAG_VNET, DC(red_lookup_depth), 0, "Depth of RED lookup table"); 179 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size, 180 CTLFLAG_RD | CTLFLAG_VNET, DC(red_avg_pkt_size), 0, "RED Medium packet size"); 181 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size, 182 CTLFLAG_RD | CTLFLAG_VNET, DC(red_max_pkt_size), 0, "RED Max packet size"); 183 184 /* time adjustment */ 185 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta, 186 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_delta), 0, "Last vs standard tick difference (usec)."); 187 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum, 188 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_delta_sum), 0, "Accumulated tick difference (usec)."); 189 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment, 190 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_adjustment), 0, "Tick adjustments done."); 191 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff, 192 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_diff), 0, 193 "Adjusted vs non-adjusted curr_time difference (ticks)."); 194 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost, 195 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_lost), 0, 196 "Number of ticks coalesced by dummynet taskqueue."); 197 198 /* Drain parameters */ 199 SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire, 200 CTLFLAG_RW | CTLFLAG_VNET, DC(expire), 0, "Expire empty queues/pipes"); 201 SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_cycle, 202 CTLFLAG_RD | CTLFLAG_VNET, DC(expire_cycle), 0, "Expire cycle for queues/pipes"); 203 204 /* statistics */ 205 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, schk_count, 206 CTLFLAG_RD | CTLFLAG_VNET, DC(schk_count), 0, "Number of schedulers"); 207 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, si_count, 208 CTLFLAG_RD | CTLFLAG_VNET, DC(si_count), 0, "Number of scheduler instances"); 209 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, fsk_count, 210 CTLFLAG_RD | CTLFLAG_VNET, DC(fsk_count), 0, "Number of flowsets"); 211 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, queue_count, 212 CTLFLAG_RD | CTLFLAG_VNET, DC(queue_count), 0, "Number of queues"); 213 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt, 214 CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt), 0, 215 "Number of packets passed to dummynet."); 216 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast, 217 CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt_fast), 0, 218 "Number of packets bypassed dummynet scheduler."); 219 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop, 220 CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt_drop), 0, 221 "Number of packets dropped by dummynet."); 222 #undef DC 223 SYSEND 224 225 #endif 226 227 static void dummynet_send(struct mbuf *); 228 229 /* 230 * Return the mbuf tag holding the dummynet state (it should 231 * be the first one on the list). 232 */ 233 struct dn_pkt_tag * 234 dn_tag_get(struct mbuf *m) 235 { 236 struct m_tag *mtag = m_tag_first(m); 237 #ifdef NEW_AQM 238 /* XXX: to skip ts m_tag. For Debugging only*/ 239 if (mtag != NULL && mtag->m_tag_id == DN_AQM_MTAG_TS) { 240 m_tag_delete(m,mtag); 241 mtag = m_tag_first(m); 242 D("skip TS tag"); 243 } 244 #endif 245 KASSERT(mtag != NULL && 246 mtag->m_tag_cookie == MTAG_ABI_COMPAT && 247 mtag->m_tag_id == PACKET_TAG_DUMMYNET, 248 ("packet on dummynet queue w/o dummynet tag!")); 249 return (struct dn_pkt_tag *)(mtag+1); 250 } 251 252 #ifndef NEW_AQM 253 static inline void 254 mq_append(struct mq *q, struct mbuf *m) 255 { 256 #ifdef USERSPACE 257 // buffers from netmap need to be copied 258 // XXX note that the routine is not expected to fail 259 ND("append %p to %p", m, q); 260 if (m->m_flags & M_STACK) { 261 struct mbuf *m_new; 262 void *p; 263 int l, ofs; 264 265 ofs = m->m_data - m->__m_extbuf; 266 // XXX allocate 267 MGETHDR(m_new, M_NOWAIT, MT_DATA); 268 ND("*** WARNING, volatile buf %p ext %p %d dofs %d m_new %p", 269 m, m->__m_extbuf, m->__m_extlen, ofs, m_new); 270 p = m_new->__m_extbuf; /* new pointer */ 271 l = m_new->__m_extlen; /* new len */ 272 if (l <= m->__m_extlen) { 273 panic("extlen too large"); 274 } 275 276 *m_new = *m; // copy 277 m_new->m_flags &= ~M_STACK; 278 m_new->__m_extbuf = p; // point to new buffer 279 _pkt_copy(m->__m_extbuf, p, m->__m_extlen); 280 m_new->m_data = p + ofs; 281 m = m_new; 282 } 283 #endif /* USERSPACE */ 284 if (q->head == NULL) 285 q->head = m; 286 else 287 q->tail->m_nextpkt = m; 288 q->count++; 289 q->tail = m; 290 m->m_nextpkt = NULL; 291 } 292 #endif 293 294 /* 295 * Dispose a list of packet. Use a functions so if we need to do 296 * more work, this is a central point to do it. 297 */ 298 void dn_free_pkts(struct mbuf *mnext) 299 { 300 struct mbuf *m; 301 302 while ((m = mnext) != NULL) { 303 mnext = m->m_nextpkt; 304 FREE_PKT(m); 305 } 306 } 307 308 static int 309 red_drops (struct dn_queue *q, int len) 310 { 311 /* 312 * RED algorithm 313 * 314 * RED calculates the average queue size (avg) using a low-pass filter 315 * with an exponential weighted (w_q) moving average: 316 * avg <- (1-w_q) * avg + w_q * q_size 317 * where q_size is the queue length (measured in bytes or * packets). 318 * 319 * If q_size == 0, we compute the idle time for the link, and set 320 * avg = (1 - w_q)^(idle/s) 321 * where s is the time needed for transmitting a medium-sized packet. 322 * 323 * Now, if avg < min_th the packet is enqueued. 324 * If avg > max_th the packet is dropped. Otherwise, the packet is 325 * dropped with probability P function of avg. 326 */ 327 328 struct dn_fsk *fs = q->fs; 329 int64_t p_b = 0; 330 331 /* Queue in bytes or packets? */ 332 uint32_t q_size = (fs->fs.flags & DN_QSIZE_BYTES) ? 333 q->ni.len_bytes : q->ni.length; 334 335 /* Average queue size estimation. */ 336 if (q_size != 0) { 337 /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */ 338 int diff = SCALE(q_size) - q->avg; 339 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q); 340 341 q->avg += (int)v; 342 } else { 343 /* 344 * Queue is empty, find for how long the queue has been 345 * empty and use a lookup table for computing 346 * (1 - * w_q)^(idle_time/s) where s is the time to send a 347 * (small) packet. 348 * XXX check wraps... 349 */ 350 if (q->avg) { 351 u_int t = div64((V_dn_cfg.curr_time - q->q_time), fs->lookup_step); 352 353 q->avg = (t < fs->lookup_depth) ? 354 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0; 355 } 356 } 357 358 /* Should i drop? */ 359 if (q->avg < fs->min_th) { 360 q->count = -1; 361 return (0); /* accept packet */ 362 } 363 if (q->avg >= fs->max_th) { /* average queue >= max threshold */ 364 if (fs->fs.flags & DN_IS_ECN) 365 return (1); 366 if (fs->fs.flags & DN_IS_GENTLE_RED) { 367 /* 368 * According to Gentle-RED, if avg is greater than 369 * max_th the packet is dropped with a probability 370 * p_b = c_3 * avg - c_4 371 * where c_3 = (1 - max_p) / max_th 372 * c_4 = 1 - 2 * max_p 373 */ 374 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) - 375 fs->c_4; 376 } else { 377 q->count = -1; 378 return (1); 379 } 380 } else if (q->avg > fs->min_th) { 381 if (fs->fs.flags & DN_IS_ECN) 382 return (1); 383 /* 384 * We compute p_b using the linear dropping function 385 * p_b = c_1 * avg - c_2 386 * where c_1 = max_p / (max_th - min_th) 387 * c_2 = max_p * min_th / (max_th - min_th) 388 */ 389 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2; 390 } 391 392 if (fs->fs.flags & DN_QSIZE_BYTES) 393 p_b = div64((p_b * len) , fs->max_pkt_size); 394 if (++q->count == 0) 395 q->random = random() & 0xffff; 396 else { 397 /* 398 * q->count counts packets arrived since last drop, so a greater 399 * value of q->count means a greater packet drop probability. 400 */ 401 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) { 402 q->count = 0; 403 /* After a drop we calculate a new random value. */ 404 q->random = random() & 0xffff; 405 return (1); /* drop */ 406 } 407 } 408 /* End of RED algorithm. */ 409 410 return (0); /* accept */ 411 412 } 413 414 /* 415 * ECN/ECT Processing (partially adopted from altq) 416 */ 417 #ifndef NEW_AQM 418 static 419 #endif 420 int 421 ecn_mark(struct mbuf* m) 422 { 423 struct ip *ip; 424 ip = (struct ip *)mtodo(m, dn_tag_get(m)->iphdr_off); 425 426 switch (ip->ip_v) { 427 case IPVERSION: 428 { 429 uint16_t old; 430 431 if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_NOTECT) 432 return (0); /* not-ECT */ 433 if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_CE) 434 return (1); /* already marked */ 435 436 /* 437 * ecn-capable but not marked, 438 * mark CE and update checksum 439 */ 440 old = *(uint16_t *)ip; 441 ip->ip_tos |= IPTOS_ECN_CE; 442 ip->ip_sum = cksum_adjust(ip->ip_sum, old, *(uint16_t *)ip); 443 return (1); 444 } 445 #ifdef INET6 446 case (IPV6_VERSION >> 4): 447 { 448 struct ip6_hdr *ip6 = (struct ip6_hdr *)ip; 449 u_int32_t flowlabel; 450 451 flowlabel = ntohl(ip6->ip6_flow); 452 if ((flowlabel >> 28) != 6) 453 return (0); /* version mismatch! */ 454 if ((flowlabel & (IPTOS_ECN_MASK << 20)) == 455 (IPTOS_ECN_NOTECT << 20)) 456 return (0); /* not-ECT */ 457 if ((flowlabel & (IPTOS_ECN_MASK << 20)) == 458 (IPTOS_ECN_CE << 20)) 459 return (1); /* already marked */ 460 /* 461 * ecn-capable but not marked, mark CE 462 */ 463 flowlabel |= (IPTOS_ECN_CE << 20); 464 ip6->ip6_flow = htonl(flowlabel); 465 return (1); 466 } 467 #endif 468 } 469 return (0); 470 } 471 472 /* 473 * Enqueue a packet in q, subject to space and queue management policy 474 * (whose parameters are in q->fs). 475 * Update stats for the queue and the scheduler. 476 * Return 0 on success, 1 on drop. The packet is consumed anyways. 477 */ 478 int 479 dn_enqueue(struct dn_queue *q, struct mbuf* m, int drop) 480 { 481 struct dn_fs *f; 482 struct dn_flow *ni; /* stats for scheduler instance */ 483 uint64_t len; 484 485 if (q->fs == NULL || q->_si == NULL) { 486 printf("%s fs %p si %p, dropping\n", 487 __FUNCTION__, q->fs, q->_si); 488 FREE_PKT(m); 489 return 1; 490 } 491 f = &(q->fs->fs); 492 ni = &q->_si->ni; 493 len = m->m_pkthdr.len; 494 /* Update statistics, then check reasons to drop pkt. */ 495 q->ni.tot_bytes += len; 496 q->ni.tot_pkts++; 497 ni->tot_bytes += len; 498 ni->tot_pkts++; 499 if (drop) 500 goto drop; 501 if (f->plr && random() < f->plr) 502 goto drop; 503 if (m->m_pkthdr.rcvif != NULL) 504 m_rcvif_serialize(m); 505 #ifdef NEW_AQM 506 /* Call AQM enqueue function */ 507 if (q->fs->aqmfp) 508 return q->fs->aqmfp->enqueue(q ,m); 509 #endif 510 if (f->flags & DN_IS_RED && red_drops(q, m->m_pkthdr.len)) { 511 if (!(f->flags & DN_IS_ECN) || !ecn_mark(m)) 512 goto drop; 513 } 514 if (f->flags & DN_QSIZE_BYTES) { 515 if (q->ni.len_bytes > f->qsize) 516 goto drop; 517 } else if (q->ni.length >= f->qsize) { 518 goto drop; 519 } 520 mq_append(&q->mq, m); 521 q->ni.length++; 522 q->ni.len_bytes += len; 523 ni->length++; 524 ni->len_bytes += len; 525 return (0); 526 527 drop: 528 V_dn_cfg.io_pkt_drop++; 529 q->ni.drops++; 530 ni->drops++; 531 FREE_PKT(m); 532 return (1); 533 } 534 535 /* 536 * Fetch packets from the delay line which are due now. If there are 537 * leftover packets, reinsert the delay line in the heap. 538 * Runs under scheduler lock. 539 */ 540 static void 541 transmit_event(struct mq *q, struct delay_line *dline, uint64_t now) 542 { 543 struct mbuf *m; 544 struct dn_pkt_tag *pkt = NULL; 545 546 dline->oid.subtype = 0; /* not in heap */ 547 while ((m = dline->mq.head) != NULL) { 548 pkt = dn_tag_get(m); 549 if (!DN_KEY_LEQ(pkt->output_time, now)) 550 break; 551 dline->mq.head = m->m_nextpkt; 552 dline->mq.count--; 553 if (m->m_pkthdr.rcvif != NULL && 554 __predict_false(m_rcvif_restore(m) == NULL)) 555 m_freem(m); 556 else 557 mq_append(q, m); 558 } 559 if (m != NULL) { 560 dline->oid.subtype = 1; /* in heap */ 561 heap_insert(&V_dn_cfg.evheap, pkt->output_time, dline); 562 } 563 } 564 565 /* 566 * Convert the additional MAC overheads/delays into an equivalent 567 * number of bits for the given data rate. The samples are 568 * in milliseconds so we need to divide by 1000. 569 */ 570 static uint64_t 571 extra_bits(struct mbuf *m, struct dn_schk *s) 572 { 573 int index; 574 uint64_t bits; 575 struct dn_profile *pf = s->profile; 576 577 if (!pf || pf->samples_no == 0) 578 return 0; 579 index = random() % pf->samples_no; 580 bits = div64((uint64_t)pf->samples[index] * s->link.bandwidth, 1000); 581 if (index >= pf->loss_level) { 582 struct dn_pkt_tag *dt = dn_tag_get(m); 583 if (dt) 584 dt->dn_dir = DIR_DROP; 585 } 586 return bits; 587 } 588 589 /* 590 * Send traffic from a scheduler instance due by 'now'. 591 * Return a pointer to the head of the queue. 592 */ 593 static struct mbuf * 594 serve_sched(struct mq *q, struct dn_sch_inst *si, uint64_t now) 595 { 596 struct mq def_q; 597 struct dn_schk *s = si->sched; 598 struct mbuf *m = NULL; 599 int delay_line_idle = (si->dline.mq.head == NULL); 600 int done; 601 uint32_t bw; 602 603 if (q == NULL) { 604 q = &def_q; 605 q->head = NULL; 606 } 607 608 bw = s->link.bandwidth; 609 si->kflags &= ~DN_ACTIVE; 610 611 if (bw > 0) 612 si->credit += (now - si->sched_time) * bw; 613 else 614 si->credit = 0; 615 si->sched_time = now; 616 done = 0; 617 while (si->credit >= 0 && (m = s->fp->dequeue(si)) != NULL) { 618 uint64_t len_scaled; 619 620 done++; 621 len_scaled = (bw == 0) ? 0 : hz * 622 (m->m_pkthdr.len * 8 + extra_bits(m, s)); 623 si->credit -= len_scaled; 624 /* Move packet in the delay line */ 625 dn_tag_get(m)->output_time = V_dn_cfg.curr_time + s->link.delay ; 626 if (m->m_pkthdr.rcvif != NULL) 627 m_rcvif_serialize(m); 628 mq_append(&si->dline.mq, m); 629 } 630 631 /* 632 * If credit >= 0 the instance is idle, mark time. 633 * Otherwise put back in the heap, and adjust the output 634 * time of the last inserted packet, m, which was too early. 635 */ 636 if (si->credit >= 0) { 637 si->idle_time = now; 638 } else { 639 uint64_t t; 640 KASSERT (bw > 0, ("bw=0 and credit<0 ?")); 641 t = div64(bw - 1 - si->credit, bw); 642 if (m) 643 dn_tag_get(m)->output_time += t; 644 si->kflags |= DN_ACTIVE; 645 heap_insert(&V_dn_cfg.evheap, now + t, si); 646 } 647 if (delay_line_idle && done) 648 transmit_event(q, &si->dline, now); 649 return q->head; 650 } 651 652 /* 653 * The timer handler for dummynet. Time is computed in ticks, but 654 * but the code is tolerant to the actual rate at which this is called. 655 * Once complete, the function reschedules itself for the next tick. 656 */ 657 void 658 dummynet_task(void *context, int pending) 659 { 660 struct timeval t; 661 struct mq q = { NULL, NULL }; /* queue to accumulate results */ 662 struct epoch_tracker et; 663 664 VNET_ITERATOR_DECL(vnet_iter); 665 VNET_LIST_RLOCK(); 666 NET_EPOCH_ENTER(et); 667 668 VNET_FOREACH(vnet_iter) { 669 memset(&q, 0, sizeof(struct mq)); 670 CURVNET_SET(vnet_iter); 671 672 if (! V_dn_cfg.init_done) { 673 CURVNET_RESTORE(); 674 continue; 675 } 676 677 DN_BH_WLOCK(); 678 679 /* Update number of lost(coalesced) ticks. */ 680 V_dn_cfg.tick_lost += pending - 1; 681 682 getmicrouptime(&t); 683 /* Last tick duration (usec). */ 684 V_dn_cfg.tick_last = (t.tv_sec - V_dn_cfg.prev_t.tv_sec) * 1000000 + 685 (t.tv_usec - V_dn_cfg.prev_t.tv_usec); 686 /* Last tick vs standard tick difference (usec). */ 687 V_dn_cfg.tick_delta = (V_dn_cfg.tick_last * hz - 1000000) / hz; 688 /* Accumulated tick difference (usec). */ 689 V_dn_cfg.tick_delta_sum += V_dn_cfg.tick_delta; 690 691 V_dn_cfg.prev_t = t; 692 693 /* 694 * Adjust curr_time if the accumulated tick difference is 695 * greater than the 'standard' tick. Since curr_time should 696 * be monotonically increasing, we do positive adjustments 697 * as required, and throttle curr_time in case of negative 698 * adjustment. 699 */ 700 V_dn_cfg.curr_time++; 701 if (V_dn_cfg.tick_delta_sum - tick >= 0) { 702 int diff = V_dn_cfg.tick_delta_sum / tick; 703 704 V_dn_cfg.curr_time += diff; 705 V_dn_cfg.tick_diff += diff; 706 V_dn_cfg.tick_delta_sum %= tick; 707 V_dn_cfg.tick_adjustment++; 708 } else if (V_dn_cfg.tick_delta_sum + tick <= 0) { 709 V_dn_cfg.curr_time--; 710 V_dn_cfg.tick_diff--; 711 V_dn_cfg.tick_delta_sum += tick; 712 V_dn_cfg.tick_adjustment++; 713 } 714 715 /* serve pending events, accumulate in q */ 716 for (;;) { 717 struct dn_id *p; /* generic parameter to handler */ 718 719 if (V_dn_cfg.evheap.elements == 0 || 720 DN_KEY_LT(V_dn_cfg.curr_time, HEAP_TOP(&V_dn_cfg.evheap)->key)) 721 break; 722 p = HEAP_TOP(&V_dn_cfg.evheap)->object; 723 heap_extract(&V_dn_cfg.evheap, NULL); 724 if (p->type == DN_SCH_I) { 725 serve_sched(&q, (struct dn_sch_inst *)p, V_dn_cfg.curr_time); 726 } else { /* extracted a delay line */ 727 transmit_event(&q, (struct delay_line *)p, V_dn_cfg.curr_time); 728 } 729 } 730 if (V_dn_cfg.expire && ++V_dn_cfg.expire_cycle >= V_dn_cfg.expire) { 731 V_dn_cfg.expire_cycle = 0; 732 dn_drain_scheduler(); 733 dn_drain_queue(); 734 } 735 DN_BH_WUNLOCK(); 736 if (q.head != NULL) 737 dummynet_send(q.head); 738 739 CURVNET_RESTORE(); 740 } 741 NET_EPOCH_EXIT(et); 742 VNET_LIST_RUNLOCK(); 743 744 /* Schedule our next run. */ 745 dn_reschedule(); 746 } 747 748 /* 749 * forward a chain of packets to the proper destination. 750 * This runs outside the dummynet lock. 751 */ 752 static void 753 dummynet_send(struct mbuf *m) 754 { 755 struct mbuf *n; 756 757 NET_EPOCH_ASSERT(); 758 759 for (; m != NULL; m = n) { 760 struct ifnet *ifp = NULL; /* gcc 3.4.6 complains */ 761 struct m_tag *tag; 762 int dst; 763 764 n = m->m_nextpkt; 765 m->m_nextpkt = NULL; 766 tag = m_tag_first(m); 767 if (tag == NULL) { /* should not happen */ 768 dst = DIR_DROP; 769 } else { 770 struct dn_pkt_tag *pkt = dn_tag_get(m); 771 /* extract the dummynet info, rename the tag 772 * to carry reinject info. 773 */ 774 ifp = ifnet_byindexgen(pkt->if_index, pkt->if_idxgen); 775 if (((pkt->dn_dir == (DIR_OUT | PROTO_LAYER2)) || 776 (pkt->dn_dir == (DIR_OUT | PROTO_LAYER2 | PROTO_IPV6))) && 777 ifp == NULL) { 778 dst = DIR_DROP; 779 } else { 780 dst = pkt->dn_dir; 781 tag->m_tag_cookie = MTAG_IPFW_RULE; 782 tag->m_tag_id = 0; 783 } 784 } 785 786 switch (dst) { 787 case DIR_OUT: 788 ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL); 789 break ; 790 791 case DIR_IN : 792 netisr_dispatch(NETISR_IP, m); 793 break; 794 795 #ifdef INET6 796 case DIR_IN | PROTO_IPV6: 797 netisr_dispatch(NETISR_IPV6, m); 798 break; 799 800 case DIR_OUT | PROTO_IPV6: 801 ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL); 802 break; 803 #endif 804 805 case DIR_FWD | PROTO_IFB: /* DN_TO_IFB_FWD: */ 806 if (bridge_dn_p != NULL) 807 ((*bridge_dn_p)(m, ifp)); 808 else 809 printf("dummynet: if_bridge not loaded\n"); 810 811 break; 812 813 case DIR_IN | PROTO_LAYER2 | PROTO_IPV6: 814 case DIR_IN | PROTO_LAYER2: /* DN_TO_ETH_DEMUX: */ 815 /* 816 * The Ethernet code assumes the Ethernet header is 817 * contiguous in the first mbuf header. 818 * Insure this is true. 819 */ 820 if (m->m_len < ETHER_HDR_LEN && 821 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) { 822 printf("dummynet/ether: pullup failed, " 823 "dropping packet\n"); 824 break; 825 } 826 ether_demux(m->m_pkthdr.rcvif, m); 827 break; 828 829 case DIR_OUT | PROTO_LAYER2 | PROTO_IPV6: 830 case DIR_OUT | PROTO_LAYER2: /* DN_TO_ETH_OUT: */ 831 MPASS(ifp != NULL); 832 ether_output_frame(ifp, m); 833 break; 834 835 case DIR_DROP: 836 /* drop the packet after some time */ 837 FREE_PKT(m); 838 break; 839 840 default: 841 printf("dummynet: bad switch %d!\n", dst); 842 FREE_PKT(m); 843 break; 844 } 845 } 846 } 847 848 static inline int 849 tag_mbuf(struct mbuf *m, int dir, struct ip_fw_args *fwa) 850 { 851 struct dn_pkt_tag *dt; 852 struct m_tag *mtag; 853 854 mtag = m_tag_get(PACKET_TAG_DUMMYNET, 855 sizeof(*dt), M_NOWAIT | M_ZERO); 856 if (mtag == NULL) 857 return 1; /* Cannot allocate packet header. */ 858 m_tag_prepend(m, mtag); /* Attach to mbuf chain. */ 859 dt = (struct dn_pkt_tag *)(mtag + 1); 860 dt->rule = fwa->rule; 861 /* only keep this info */ 862 dt->rule.info &= (IPFW_ONEPASS | IPFW_IS_DUMMYNET); 863 dt->dn_dir = dir; 864 if (fwa->flags & IPFW_ARGS_OUT && fwa->ifp != NULL) { 865 NET_EPOCH_ASSERT(); 866 dt->if_index = fwa->ifp->if_index; 867 dt->if_idxgen = fwa->ifp->if_idxgen; 868 } 869 /* dt->output tame is updated as we move through */ 870 dt->output_time = V_dn_cfg.curr_time; 871 dt->iphdr_off = (dir & PROTO_LAYER2) ? ETHER_HDR_LEN : 0; 872 return 0; 873 } 874 875 /* 876 * dummynet hook for packets. 877 * We use the argument to locate the flowset fs and the sched_set sch 878 * associated to it. The we apply flow_mask and sched_mask to 879 * determine the queue and scheduler instances. 880 */ 881 int 882 dummynet_io(struct mbuf **m0, struct ip_fw_args *fwa) 883 { 884 struct mbuf *m = *m0; 885 struct dn_fsk *fs = NULL; 886 struct dn_sch_inst *si; 887 struct dn_queue *q = NULL; /* default */ 888 int fs_id, dir; 889 890 fs_id = (fwa->rule.info & IPFW_INFO_MASK) + 891 ((fwa->rule.info & IPFW_IS_PIPE) ? 2*DN_MAX_ID : 0); 892 /* XXXGL: convert args to dir */ 893 if (fwa->flags & IPFW_ARGS_IN) 894 dir = DIR_IN; 895 else 896 dir = DIR_OUT; 897 if (fwa->flags & IPFW_ARGS_ETHER) 898 dir |= PROTO_LAYER2; 899 else if (fwa->flags & IPFW_ARGS_IP6) 900 dir |= PROTO_IPV6; 901 DN_BH_WLOCK(); 902 V_dn_cfg.io_pkt++; 903 /* we could actually tag outside the lock, but who cares... */ 904 if (tag_mbuf(m, dir, fwa)) 905 goto dropit; 906 /* XXX locate_flowset could be optimised with a direct ref. */ 907 fs = dn_ht_find(V_dn_cfg.fshash, fs_id, 0, NULL); 908 if (fs == NULL) 909 goto dropit; /* This queue/pipe does not exist! */ 910 if (fs->sched == NULL) /* should not happen */ 911 goto dropit; 912 /* find scheduler instance, possibly applying sched_mask */ 913 si = ipdn_si_find(fs->sched, &(fwa->f_id)); 914 if (si == NULL) 915 goto dropit; 916 /* 917 * If the scheduler supports multiple queues, find the right one 918 * (otherwise it will be ignored by enqueue). 919 */ 920 if (fs->sched->fp->flags & DN_MULTIQUEUE) { 921 q = ipdn_q_find(fs, si, &(fwa->f_id)); 922 if (q == NULL) 923 goto dropit; 924 } 925 if (fs->sched->fp->enqueue(si, q, m)) { 926 /* packet was dropped by enqueue() */ 927 m = *m0 = NULL; 928 929 /* dn_enqueue already increases io_pkt_drop */ 930 V_dn_cfg.io_pkt_drop--; 931 932 goto dropit; 933 } 934 935 if (si->kflags & DN_ACTIVE) { 936 m = *m0 = NULL; /* consumed */ 937 goto done; /* already active, nothing to do */ 938 } 939 940 /* compute the initial allowance */ 941 if (si->idle_time < V_dn_cfg.curr_time) { 942 /* Do this only on the first packet on an idle pipe */ 943 struct dn_link *p = &fs->sched->link; 944 945 si->sched_time = V_dn_cfg.curr_time; 946 si->credit = V_dn_cfg.io_fast ? p->bandwidth : 0; 947 if (p->burst) { 948 uint64_t burst = (V_dn_cfg.curr_time - si->idle_time) * p->bandwidth; 949 if (burst > p->burst) 950 burst = p->burst; 951 si->credit += burst; 952 } 953 } 954 /* pass through scheduler and delay line */ 955 m = serve_sched(NULL, si, V_dn_cfg.curr_time); 956 957 /* optimization -- pass it back to ipfw for immediate send */ 958 /* XXX Don't call dummynet_send() if scheduler return the packet 959 * just enqueued. This avoid a lock order reversal. 960 * 961 */ 962 if (/*V_dn_cfg.io_fast &&*/ m == *m0 && (dir & PROTO_LAYER2) == 0 ) { 963 /* fast io, rename the tag * to carry reinject info. */ 964 struct m_tag *tag = m_tag_first(m); 965 966 tag->m_tag_cookie = MTAG_IPFW_RULE; 967 tag->m_tag_id = 0; 968 V_dn_cfg.io_pkt_fast++; 969 if (m->m_nextpkt != NULL) { 970 printf("dummynet: fast io: pkt chain detected!\n"); 971 m->m_nextpkt = NULL; 972 } 973 m = NULL; 974 } else { 975 *m0 = NULL; 976 } 977 done: 978 DN_BH_WUNLOCK(); 979 if (m) 980 dummynet_send(m); 981 return 0; 982 983 dropit: 984 V_dn_cfg.io_pkt_drop++; 985 DN_BH_WUNLOCK(); 986 if (m) 987 FREE_PKT(m); 988 *m0 = NULL; 989 return (fs && (fs->fs.flags & DN_NOERROR)) ? 0 : ENOBUFS; 990 } 991