1 /*- 2 * Copyright (c) 2010 Luigi Rizzo, Riccardo Panicucci, Universita` di Pisa 3 * All rights reserved 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 */ 26 27 /* 28 * Dummynet portions related to packet handling. 29 */ 30 #include <sys/cdefs.h> 31 __FBSDID("$FreeBSD$"); 32 33 #include "opt_inet6.h" 34 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/malloc.h> 38 #include <sys/mbuf.h> 39 #include <sys/kernel.h> 40 #include <sys/lock.h> 41 #include <sys/module.h> 42 #include <sys/mutex.h> 43 #include <sys/priv.h> 44 #include <sys/proc.h> 45 #include <sys/rwlock.h> 46 #include <sys/socket.h> 47 #include <sys/time.h> 48 #include <sys/sysctl.h> 49 50 #include <net/if.h> /* IFNAMSIZ, struct ifaddr, ifq head, lock.h mutex.h */ 51 #include <net/netisr.h> 52 #include <net/vnet.h> 53 54 #include <netinet/in.h> 55 #include <netinet/ip.h> /* ip_len, ip_off */ 56 #include <netinet/ip_var.h> /* ip_output(), IP_FORWARDING */ 57 #include <netinet/ip_fw.h> 58 #include <netinet/ip_dummynet.h> 59 #include <netinet/if_ether.h> /* various ether_* routines */ 60 #include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */ 61 #include <netinet6/ip6_var.h> 62 63 #include <netpfil/ipfw/ip_fw_private.h> 64 #include <netpfil/ipfw/dn_heap.h> 65 #include <netpfil/ipfw/ip_dn_private.h> 66 #include <netpfil/ipfw/dn_sched.h> 67 68 /* 69 * We keep a private variable for the simulation time, but we could 70 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c) 71 * instead of dn_cfg.curr_time 72 */ 73 74 struct dn_parms dn_cfg; 75 //VNET_DEFINE(struct dn_parms, _base_dn_cfg); 76 77 static long tick_last; /* Last tick duration (usec). */ 78 static long tick_delta; /* Last vs standard tick diff (usec). */ 79 static long tick_delta_sum; /* Accumulated tick difference (usec).*/ 80 static long tick_adjustment; /* Tick adjustments done. */ 81 static long tick_lost; /* Lost(coalesced) ticks number. */ 82 /* Adjusted vs non-adjusted curr_time difference (ticks). */ 83 static long tick_diff; 84 85 static unsigned long io_pkt; 86 static unsigned long io_pkt_fast; 87 static unsigned long io_pkt_drop; 88 89 /* 90 * We use a heap to store entities for which we have pending timer events. 91 * The heap is checked at every tick and all entities with expired events 92 * are extracted. 93 */ 94 95 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap"); 96 97 extern void (*bridge_dn_p)(struct mbuf *, struct ifnet *); 98 99 #ifdef SYSCTL_NODE 100 101 /* 102 * Because of the way the SYSBEGIN/SYSEND macros work on other 103 * platforms, there should not be functions between them. 104 * So keep the handlers outside the block. 105 */ 106 static int 107 sysctl_hash_size(SYSCTL_HANDLER_ARGS) 108 { 109 int error, value; 110 111 value = dn_cfg.hash_size; 112 error = sysctl_handle_int(oidp, &value, 0, req); 113 if (error != 0 || req->newptr == NULL) 114 return (error); 115 if (value < 16 || value > 65536) 116 return (EINVAL); 117 dn_cfg.hash_size = value; 118 return (0); 119 } 120 121 static int 122 sysctl_limits(SYSCTL_HANDLER_ARGS) 123 { 124 int error; 125 long value; 126 127 if (arg2 != 0) 128 value = dn_cfg.slot_limit; 129 else 130 value = dn_cfg.byte_limit; 131 error = sysctl_handle_long(oidp, &value, 0, req); 132 133 if (error != 0 || req->newptr == NULL) 134 return (error); 135 if (arg2 != 0) { 136 if (value < 1) 137 return (EINVAL); 138 dn_cfg.slot_limit = value; 139 } else { 140 if (value < 1500) 141 return (EINVAL); 142 dn_cfg.byte_limit = value; 143 } 144 return (0); 145 } 146 147 SYSBEGIN(f4) 148 149 SYSCTL_DECL(_net_inet); 150 SYSCTL_DECL(_net_inet_ip); 151 static SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW, 0, "Dummynet"); 152 153 /* wrapper to pass dn_cfg fields to SYSCTL_* */ 154 //#define DC(x) (&(VNET_NAME(_base_dn_cfg).x)) 155 #define DC(x) (&(dn_cfg.x)) 156 /* parameters */ 157 158 159 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hash_size, 160 CTLTYPE_INT | CTLFLAG_RW, 0, 0, sysctl_hash_size, 161 "I", "Default hash table size"); 162 163 164 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit, 165 CTLTYPE_LONG | CTLFLAG_RW, 0, 1, sysctl_limits, 166 "L", "Upper limit in slots for pipe queue."); 167 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit, 168 CTLTYPE_LONG | CTLFLAG_RW, 0, 0, sysctl_limits, 169 "L", "Upper limit in bytes for pipe queue."); 170 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast, 171 CTLFLAG_RW, DC(io_fast), 0, "Enable fast dummynet io."); 172 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, 173 CTLFLAG_RW, DC(debug), 0, "Dummynet debug level"); 174 175 /* RED parameters */ 176 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth, 177 CTLFLAG_RD, DC(red_lookup_depth), 0, "Depth of RED lookup table"); 178 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size, 179 CTLFLAG_RD, DC(red_avg_pkt_size), 0, "RED Medium packet size"); 180 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size, 181 CTLFLAG_RD, DC(red_max_pkt_size), 0, "RED Max packet size"); 182 183 /* time adjustment */ 184 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta, 185 CTLFLAG_RD, &tick_delta, 0, "Last vs standard tick difference (usec)."); 186 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum, 187 CTLFLAG_RD, &tick_delta_sum, 0, "Accumulated tick difference (usec)."); 188 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment, 189 CTLFLAG_RD, &tick_adjustment, 0, "Tick adjustments done."); 190 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff, 191 CTLFLAG_RD, &tick_diff, 0, 192 "Adjusted vs non-adjusted curr_time difference (ticks)."); 193 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost, 194 CTLFLAG_RD, &tick_lost, 0, 195 "Number of ticks coalesced by dummynet taskqueue."); 196 197 /* Drain parameters */ 198 SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire, 199 CTLFLAG_RW, DC(expire), 0, "Expire empty queues/pipes"); 200 SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_cycle, 201 CTLFLAG_RD, DC(expire_cycle), 0, "Expire cycle for queues/pipes"); 202 203 /* statistics */ 204 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, schk_count, 205 CTLFLAG_RD, DC(schk_count), 0, "Number of schedulers"); 206 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, si_count, 207 CTLFLAG_RD, DC(si_count), 0, "Number of scheduler instances"); 208 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, fsk_count, 209 CTLFLAG_RD, DC(fsk_count), 0, "Number of flowsets"); 210 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, queue_count, 211 CTLFLAG_RD, DC(queue_count), 0, "Number of queues"); 212 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt, 213 CTLFLAG_RD, &io_pkt, 0, 214 "Number of packets passed to dummynet."); 215 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast, 216 CTLFLAG_RD, &io_pkt_fast, 0, 217 "Number of packets bypassed dummynet scheduler."); 218 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop, 219 CTLFLAG_RD, &io_pkt_drop, 0, 220 "Number of packets dropped by dummynet."); 221 #undef DC 222 SYSEND 223 224 #endif 225 226 static void dummynet_send(struct mbuf *); 227 228 /* 229 * Packets processed by dummynet have an mbuf tag associated with 230 * them that carries their dummynet state. 231 * Outside dummynet, only the 'rule' field is relevant, and it must 232 * be at the beginning of the structure. 233 */ 234 struct dn_pkt_tag { 235 struct ipfw_rule_ref rule; /* matching rule */ 236 237 /* second part, dummynet specific */ 238 int dn_dir; /* action when packet comes out.*/ 239 /* see ip_fw_private.h */ 240 uint64_t output_time; /* when the pkt is due for delivery*/ 241 struct ifnet *ifp; /* interface, for ip_output */ 242 struct _ip6dn_args ip6opt; /* XXX ipv6 options */ 243 }; 244 245 /* 246 * Return the mbuf tag holding the dummynet state (it should 247 * be the first one on the list). 248 */ 249 static struct dn_pkt_tag * 250 dn_tag_get(struct mbuf *m) 251 { 252 struct m_tag *mtag = m_tag_first(m); 253 KASSERT(mtag != NULL && 254 mtag->m_tag_cookie == MTAG_ABI_COMPAT && 255 mtag->m_tag_id == PACKET_TAG_DUMMYNET, 256 ("packet on dummynet queue w/o dummynet tag!")); 257 return (struct dn_pkt_tag *)(mtag+1); 258 } 259 260 static inline void 261 mq_append(struct mq *q, struct mbuf *m) 262 { 263 #ifdef USERSPACE 264 // buffers from netmap need to be copied 265 // XXX note that the routine is not expected to fail 266 ND("append %p to %p", m, q); 267 if (m->m_flags & M_STACK) { 268 struct mbuf *m_new; 269 void *p; 270 int l, ofs; 271 272 ofs = m->m_data - m->__m_extbuf; 273 // XXX allocate 274 MGETHDR(m_new, M_NOWAIT, MT_DATA); 275 ND("*** WARNING, volatile buf %p ext %p %d dofs %d m_new %p", 276 m, m->__m_extbuf, m->__m_extlen, ofs, m_new); 277 p = m_new->__m_extbuf; /* new pointer */ 278 l = m_new->__m_extlen; /* new len */ 279 if (l <= m->__m_extlen) { 280 panic("extlen too large"); 281 } 282 283 *m_new = *m; // copy 284 m_new->m_flags &= ~M_STACK; 285 m_new->__m_extbuf = p; // point to new buffer 286 _pkt_copy(m->__m_extbuf, p, m->__m_extlen); 287 m_new->m_data = p + ofs; 288 m = m_new; 289 } 290 #endif /* USERSPACE */ 291 if (q->head == NULL) 292 q->head = m; 293 else 294 q->tail->m_nextpkt = m; 295 q->count++; 296 q->tail = m; 297 m->m_nextpkt = NULL; 298 } 299 300 /* 301 * Dispose a list of packet. Use a functions so if we need to do 302 * more work, this is a central point to do it. 303 */ 304 void dn_free_pkts(struct mbuf *mnext) 305 { 306 struct mbuf *m; 307 308 while ((m = mnext) != NULL) { 309 mnext = m->m_nextpkt; 310 FREE_PKT(m); 311 } 312 } 313 314 static int 315 red_drops (struct dn_queue *q, int len) 316 { 317 /* 318 * RED algorithm 319 * 320 * RED calculates the average queue size (avg) using a low-pass filter 321 * with an exponential weighted (w_q) moving average: 322 * avg <- (1-w_q) * avg + w_q * q_size 323 * where q_size is the queue length (measured in bytes or * packets). 324 * 325 * If q_size == 0, we compute the idle time for the link, and set 326 * avg = (1 - w_q)^(idle/s) 327 * where s is the time needed for transmitting a medium-sized packet. 328 * 329 * Now, if avg < min_th the packet is enqueued. 330 * If avg > max_th the packet is dropped. Otherwise, the packet is 331 * dropped with probability P function of avg. 332 */ 333 334 struct dn_fsk *fs = q->fs; 335 int64_t p_b = 0; 336 337 /* Queue in bytes or packets? */ 338 uint32_t q_size = (fs->fs.flags & DN_QSIZE_BYTES) ? 339 q->ni.len_bytes : q->ni.length; 340 341 /* Average queue size estimation. */ 342 if (q_size != 0) { 343 /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */ 344 int diff = SCALE(q_size) - q->avg; 345 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q); 346 347 q->avg += (int)v; 348 } else { 349 /* 350 * Queue is empty, find for how long the queue has been 351 * empty and use a lookup table for computing 352 * (1 - * w_q)^(idle_time/s) where s is the time to send a 353 * (small) packet. 354 * XXX check wraps... 355 */ 356 if (q->avg) { 357 u_int t = div64((dn_cfg.curr_time - q->q_time), fs->lookup_step); 358 359 q->avg = (t < fs->lookup_depth) ? 360 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0; 361 } 362 } 363 364 /* Should i drop? */ 365 if (q->avg < fs->min_th) { 366 q->count = -1; 367 return (0); /* accept packet */ 368 } 369 if (q->avg >= fs->max_th) { /* average queue >= max threshold */ 370 if (fs->fs.flags & DN_IS_ECN) 371 return (1); 372 if (fs->fs.flags & DN_IS_GENTLE_RED) { 373 /* 374 * According to Gentle-RED, if avg is greater than 375 * max_th the packet is dropped with a probability 376 * p_b = c_3 * avg - c_4 377 * where c_3 = (1 - max_p) / max_th 378 * c_4 = 1 - 2 * max_p 379 */ 380 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) - 381 fs->c_4; 382 } else { 383 q->count = -1; 384 return (1); 385 } 386 } else if (q->avg > fs->min_th) { 387 if (fs->fs.flags & DN_IS_ECN) 388 return (1); 389 /* 390 * We compute p_b using the linear dropping function 391 * p_b = c_1 * avg - c_2 392 * where c_1 = max_p / (max_th - min_th) 393 * c_2 = max_p * min_th / (max_th - min_th) 394 */ 395 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2; 396 } 397 398 if (fs->fs.flags & DN_QSIZE_BYTES) 399 p_b = div64((p_b * len) , fs->max_pkt_size); 400 if (++q->count == 0) 401 q->random = random() & 0xffff; 402 else { 403 /* 404 * q->count counts packets arrived since last drop, so a greater 405 * value of q->count means a greater packet drop probability. 406 */ 407 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) { 408 q->count = 0; 409 /* After a drop we calculate a new random value. */ 410 q->random = random() & 0xffff; 411 return (1); /* drop */ 412 } 413 } 414 /* End of RED algorithm. */ 415 416 return (0); /* accept */ 417 418 } 419 420 /* 421 * ECN/ECT Processing (partially adopted from altq) 422 */ 423 static int 424 ecn_mark(struct mbuf* m) 425 { 426 struct ip *ip; 427 ip = mtod(m, struct ip *); 428 429 switch (ip->ip_v) { 430 case IPVERSION: 431 { 432 uint16_t old; 433 434 if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_NOTECT) 435 return (0); /* not-ECT */ 436 if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_CE) 437 return (1); /* already marked */ 438 439 /* 440 * ecn-capable but not marked, 441 * mark CE and update checksum 442 */ 443 old = *(uint16_t *)ip; 444 ip->ip_tos |= IPTOS_ECN_CE; 445 ip->ip_sum = cksum_adjust(ip->ip_sum, old, *(uint16_t *)ip); 446 return (1); 447 } 448 #ifdef INET6 449 case (IPV6_VERSION >> 4): 450 { 451 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 452 u_int32_t flowlabel; 453 454 flowlabel = ntohl(ip6->ip6_flow); 455 if ((flowlabel >> 28) != 6) 456 return (0); /* version mismatch! */ 457 if ((flowlabel & (IPTOS_ECN_MASK << 20)) == 458 (IPTOS_ECN_NOTECT << 20)) 459 return (0); /* not-ECT */ 460 if ((flowlabel & (IPTOS_ECN_MASK << 20)) == 461 (IPTOS_ECN_CE << 20)) 462 return (1); /* already marked */ 463 /* 464 * ecn-capable but not marked, mark CE 465 */ 466 flowlabel |= (IPTOS_ECN_CE << 20); 467 ip6->ip6_flow = htonl(flowlabel); 468 return (1); 469 } 470 #endif 471 } 472 return (0); 473 } 474 475 /* 476 * Enqueue a packet in q, subject to space and queue management policy 477 * (whose parameters are in q->fs). 478 * Update stats for the queue and the scheduler. 479 * Return 0 on success, 1 on drop. The packet is consumed anyways. 480 */ 481 int 482 dn_enqueue(struct dn_queue *q, struct mbuf* m, int drop) 483 { 484 struct dn_fs *f; 485 struct dn_flow *ni; /* stats for scheduler instance */ 486 uint64_t len; 487 488 if (q->fs == NULL || q->_si == NULL) { 489 printf("%s fs %p si %p, dropping\n", 490 __FUNCTION__, q->fs, q->_si); 491 FREE_PKT(m); 492 return 1; 493 } 494 f = &(q->fs->fs); 495 ni = &q->_si->ni; 496 len = m->m_pkthdr.len; 497 /* Update statistics, then check reasons to drop pkt. */ 498 q->ni.tot_bytes += len; 499 q->ni.tot_pkts++; 500 ni->tot_bytes += len; 501 ni->tot_pkts++; 502 if (drop) 503 goto drop; 504 if (f->plr && random() < f->plr) 505 goto drop; 506 if (f->flags & DN_IS_RED && red_drops(q, m->m_pkthdr.len)) { 507 if (!(f->flags & DN_IS_ECN) || !ecn_mark(m)) 508 goto drop; 509 } 510 if (f->flags & DN_QSIZE_BYTES) { 511 if (q->ni.len_bytes > f->qsize) 512 goto drop; 513 } else if (q->ni.length >= f->qsize) { 514 goto drop; 515 } 516 mq_append(&q->mq, m); 517 q->ni.length++; 518 q->ni.len_bytes += len; 519 ni->length++; 520 ni->len_bytes += len; 521 return (0); 522 523 drop: 524 io_pkt_drop++; 525 q->ni.drops++; 526 ni->drops++; 527 FREE_PKT(m); 528 return (1); 529 } 530 531 /* 532 * Fetch packets from the delay line which are due now. If there are 533 * leftover packets, reinsert the delay line in the heap. 534 * Runs under scheduler lock. 535 */ 536 static void 537 transmit_event(struct mq *q, struct delay_line *dline, uint64_t now) 538 { 539 struct mbuf *m; 540 struct dn_pkt_tag *pkt = NULL; 541 542 dline->oid.subtype = 0; /* not in heap */ 543 while ((m = dline->mq.head) != NULL) { 544 pkt = dn_tag_get(m); 545 if (!DN_KEY_LEQ(pkt->output_time, now)) 546 break; 547 dline->mq.head = m->m_nextpkt; 548 dline->mq.count--; 549 mq_append(q, m); 550 } 551 if (m != NULL) { 552 dline->oid.subtype = 1; /* in heap */ 553 heap_insert(&dn_cfg.evheap, pkt->output_time, dline); 554 } 555 } 556 557 /* 558 * Convert the additional MAC overheads/delays into an equivalent 559 * number of bits for the given data rate. The samples are 560 * in milliseconds so we need to divide by 1000. 561 */ 562 static uint64_t 563 extra_bits(struct mbuf *m, struct dn_schk *s) 564 { 565 int index; 566 uint64_t bits; 567 struct dn_profile *pf = s->profile; 568 569 if (!pf || pf->samples_no == 0) 570 return 0; 571 index = random() % pf->samples_no; 572 bits = div64((uint64_t)pf->samples[index] * s->link.bandwidth, 1000); 573 if (index >= pf->loss_level) { 574 struct dn_pkt_tag *dt = dn_tag_get(m); 575 if (dt) 576 dt->dn_dir = DIR_DROP; 577 } 578 return bits; 579 } 580 581 /* 582 * Send traffic from a scheduler instance due by 'now'. 583 * Return a pointer to the head of the queue. 584 */ 585 static struct mbuf * 586 serve_sched(struct mq *q, struct dn_sch_inst *si, uint64_t now) 587 { 588 struct mq def_q; 589 struct dn_schk *s = si->sched; 590 struct mbuf *m = NULL; 591 int delay_line_idle = (si->dline.mq.head == NULL); 592 int done, bw; 593 594 if (q == NULL) { 595 q = &def_q; 596 q->head = NULL; 597 } 598 599 bw = s->link.bandwidth; 600 si->kflags &= ~DN_ACTIVE; 601 602 if (bw > 0) 603 si->credit += (now - si->sched_time) * bw; 604 else 605 si->credit = 0; 606 si->sched_time = now; 607 done = 0; 608 while (si->credit >= 0 && (m = s->fp->dequeue(si)) != NULL) { 609 uint64_t len_scaled; 610 611 done++; 612 len_scaled = (bw == 0) ? 0 : hz * 613 (m->m_pkthdr.len * 8 + extra_bits(m, s)); 614 si->credit -= len_scaled; 615 /* Move packet in the delay line */ 616 dn_tag_get(m)->output_time = dn_cfg.curr_time + s->link.delay ; 617 mq_append(&si->dline.mq, m); 618 } 619 620 /* 621 * If credit >= 0 the instance is idle, mark time. 622 * Otherwise put back in the heap, and adjust the output 623 * time of the last inserted packet, m, which was too early. 624 */ 625 if (si->credit >= 0) { 626 si->idle_time = now; 627 } else { 628 uint64_t t; 629 KASSERT (bw > 0, ("bw=0 and credit<0 ?")); 630 t = div64(bw - 1 - si->credit, bw); 631 if (m) 632 dn_tag_get(m)->output_time += t; 633 si->kflags |= DN_ACTIVE; 634 heap_insert(&dn_cfg.evheap, now + t, si); 635 } 636 if (delay_line_idle && done) 637 transmit_event(q, &si->dline, now); 638 return q->head; 639 } 640 641 /* 642 * The timer handler for dummynet. Time is computed in ticks, but 643 * but the code is tolerant to the actual rate at which this is called. 644 * Once complete, the function reschedules itself for the next tick. 645 */ 646 void 647 dummynet_task(void *context, int pending) 648 { 649 struct timeval t; 650 struct mq q = { NULL, NULL }; /* queue to accumulate results */ 651 652 CURVNET_SET((struct vnet *)context); 653 654 DN_BH_WLOCK(); 655 656 /* Update number of lost(coalesced) ticks. */ 657 tick_lost += pending - 1; 658 659 getmicrouptime(&t); 660 /* Last tick duration (usec). */ 661 tick_last = (t.tv_sec - dn_cfg.prev_t.tv_sec) * 1000000 + 662 (t.tv_usec - dn_cfg.prev_t.tv_usec); 663 /* Last tick vs standard tick difference (usec). */ 664 tick_delta = (tick_last * hz - 1000000) / hz; 665 /* Accumulated tick difference (usec). */ 666 tick_delta_sum += tick_delta; 667 668 dn_cfg.prev_t = t; 669 670 /* 671 * Adjust curr_time if the accumulated tick difference is 672 * greater than the 'standard' tick. Since curr_time should 673 * be monotonically increasing, we do positive adjustments 674 * as required, and throttle curr_time in case of negative 675 * adjustment. 676 */ 677 dn_cfg.curr_time++; 678 if (tick_delta_sum - tick >= 0) { 679 int diff = tick_delta_sum / tick; 680 681 dn_cfg.curr_time += diff; 682 tick_diff += diff; 683 tick_delta_sum %= tick; 684 tick_adjustment++; 685 } else if (tick_delta_sum + tick <= 0) { 686 dn_cfg.curr_time--; 687 tick_diff--; 688 tick_delta_sum += tick; 689 tick_adjustment++; 690 } 691 692 /* serve pending events, accumulate in q */ 693 for (;;) { 694 struct dn_id *p; /* generic parameter to handler */ 695 696 if (dn_cfg.evheap.elements == 0 || 697 DN_KEY_LT(dn_cfg.curr_time, HEAP_TOP(&dn_cfg.evheap)->key)) 698 break; 699 p = HEAP_TOP(&dn_cfg.evheap)->object; 700 heap_extract(&dn_cfg.evheap, NULL); 701 702 if (p->type == DN_SCH_I) { 703 serve_sched(&q, (struct dn_sch_inst *)p, dn_cfg.curr_time); 704 } else { /* extracted a delay line */ 705 transmit_event(&q, (struct delay_line *)p, dn_cfg.curr_time); 706 } 707 } 708 if (dn_cfg.expire && ++dn_cfg.expire_cycle >= dn_cfg.expire) { 709 dn_cfg.expire_cycle = 0; 710 dn_drain_scheduler(); 711 dn_drain_queue(); 712 } 713 714 DN_BH_WUNLOCK(); 715 dn_reschedule(); 716 if (q.head != NULL) 717 dummynet_send(q.head); 718 CURVNET_RESTORE(); 719 } 720 721 /* 722 * forward a chain of packets to the proper destination. 723 * This runs outside the dummynet lock. 724 */ 725 static void 726 dummynet_send(struct mbuf *m) 727 { 728 struct mbuf *n; 729 730 for (; m != NULL; m = n) { 731 struct ifnet *ifp = NULL; /* gcc 3.4.6 complains */ 732 struct m_tag *tag; 733 int dst; 734 735 n = m->m_nextpkt; 736 m->m_nextpkt = NULL; 737 tag = m_tag_first(m); 738 if (tag == NULL) { /* should not happen */ 739 dst = DIR_DROP; 740 } else { 741 struct dn_pkt_tag *pkt = dn_tag_get(m); 742 /* extract the dummynet info, rename the tag 743 * to carry reinject info. 744 */ 745 if (pkt->dn_dir == (DIR_OUT | PROTO_LAYER2) && 746 pkt->ifp == NULL) { 747 dst = DIR_DROP; 748 } else { 749 dst = pkt->dn_dir; 750 ifp = pkt->ifp; 751 tag->m_tag_cookie = MTAG_IPFW_RULE; 752 tag->m_tag_id = 0; 753 } 754 } 755 756 switch (dst) { 757 case DIR_OUT: 758 ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL); 759 break ; 760 761 case DIR_IN : 762 netisr_dispatch(NETISR_IP, m); 763 break; 764 765 #ifdef INET6 766 case DIR_IN | PROTO_IPV6: 767 netisr_dispatch(NETISR_IPV6, m); 768 break; 769 770 case DIR_OUT | PROTO_IPV6: 771 ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL); 772 break; 773 #endif 774 775 case DIR_FWD | PROTO_IFB: /* DN_TO_IFB_FWD: */ 776 if (bridge_dn_p != NULL) 777 ((*bridge_dn_p)(m, ifp)); 778 else 779 printf("dummynet: if_bridge not loaded\n"); 780 781 break; 782 783 case DIR_IN | PROTO_LAYER2: /* DN_TO_ETH_DEMUX: */ 784 /* 785 * The Ethernet code assumes the Ethernet header is 786 * contiguous in the first mbuf header. 787 * Insure this is true. 788 */ 789 if (m->m_len < ETHER_HDR_LEN && 790 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) { 791 printf("dummynet/ether: pullup failed, " 792 "dropping packet\n"); 793 break; 794 } 795 ether_demux(m->m_pkthdr.rcvif, m); 796 break; 797 798 case DIR_OUT | PROTO_LAYER2: /* N_TO_ETH_OUT: */ 799 ether_output_frame(ifp, m); 800 break; 801 802 case DIR_DROP: 803 /* drop the packet after some time */ 804 FREE_PKT(m); 805 break; 806 807 default: 808 printf("dummynet: bad switch %d!\n", dst); 809 FREE_PKT(m); 810 break; 811 } 812 } 813 } 814 815 static inline int 816 tag_mbuf(struct mbuf *m, int dir, struct ip_fw_args *fwa) 817 { 818 struct dn_pkt_tag *dt; 819 struct m_tag *mtag; 820 821 mtag = m_tag_get(PACKET_TAG_DUMMYNET, 822 sizeof(*dt), M_NOWAIT | M_ZERO); 823 if (mtag == NULL) 824 return 1; /* Cannot allocate packet header. */ 825 m_tag_prepend(m, mtag); /* Attach to mbuf chain. */ 826 dt = (struct dn_pkt_tag *)(mtag + 1); 827 dt->rule = fwa->rule; 828 dt->rule.info &= IPFW_ONEPASS; /* only keep this info */ 829 dt->dn_dir = dir; 830 dt->ifp = fwa->oif; 831 /* dt->output tame is updated as we move through */ 832 dt->output_time = dn_cfg.curr_time; 833 return 0; 834 } 835 836 837 /* 838 * dummynet hook for packets. 839 * We use the argument to locate the flowset fs and the sched_set sch 840 * associated to it. The we apply flow_mask and sched_mask to 841 * determine the queue and scheduler instances. 842 * 843 * dir where shall we send the packet after dummynet. 844 * *m0 the mbuf with the packet 845 * ifp the 'ifp' parameter from the caller. 846 * NULL in ip_input, destination interface in ip_output, 847 */ 848 int 849 dummynet_io(struct mbuf **m0, int dir, struct ip_fw_args *fwa) 850 { 851 struct mbuf *m = *m0; 852 struct dn_fsk *fs = NULL; 853 struct dn_sch_inst *si; 854 struct dn_queue *q = NULL; /* default */ 855 856 int fs_id = (fwa->rule.info & IPFW_INFO_MASK) + 857 ((fwa->rule.info & IPFW_IS_PIPE) ? 2*DN_MAX_ID : 0); 858 DN_BH_WLOCK(); 859 io_pkt++; 860 /* we could actually tag outside the lock, but who cares... */ 861 if (tag_mbuf(m, dir, fwa)) 862 goto dropit; 863 if (dn_cfg.busy) { 864 /* if the upper half is busy doing something expensive, 865 * lets queue the packet and move forward 866 */ 867 mq_append(&dn_cfg.pending, m); 868 m = *m0 = NULL; /* consumed */ 869 goto done; /* already active, nothing to do */ 870 } 871 /* XXX locate_flowset could be optimised with a direct ref. */ 872 fs = dn_ht_find(dn_cfg.fshash, fs_id, 0, NULL); 873 if (fs == NULL) 874 goto dropit; /* This queue/pipe does not exist! */ 875 if (fs->sched == NULL) /* should not happen */ 876 goto dropit; 877 /* find scheduler instance, possibly applying sched_mask */ 878 si = ipdn_si_find(fs->sched, &(fwa->f_id)); 879 if (si == NULL) 880 goto dropit; 881 /* 882 * If the scheduler supports multiple queues, find the right one 883 * (otherwise it will be ignored by enqueue). 884 */ 885 if (fs->sched->fp->flags & DN_MULTIQUEUE) { 886 q = ipdn_q_find(fs, si, &(fwa->f_id)); 887 if (q == NULL) 888 goto dropit; 889 } 890 if (fs->sched->fp->enqueue(si, q, m)) { 891 /* packet was dropped by enqueue() */ 892 m = *m0 = NULL; 893 goto dropit; 894 } 895 896 if (si->kflags & DN_ACTIVE) { 897 m = *m0 = NULL; /* consumed */ 898 goto done; /* already active, nothing to do */ 899 } 900 901 /* compute the initial allowance */ 902 if (si->idle_time < dn_cfg.curr_time) { 903 /* Do this only on the first packet on an idle pipe */ 904 struct dn_link *p = &fs->sched->link; 905 906 si->sched_time = dn_cfg.curr_time; 907 si->credit = dn_cfg.io_fast ? p->bandwidth : 0; 908 if (p->burst) { 909 uint64_t burst = (dn_cfg.curr_time - si->idle_time) * p->bandwidth; 910 if (burst > p->burst) 911 burst = p->burst; 912 si->credit += burst; 913 } 914 } 915 /* pass through scheduler and delay line */ 916 m = serve_sched(NULL, si, dn_cfg.curr_time); 917 918 /* optimization -- pass it back to ipfw for immediate send */ 919 /* XXX Don't call dummynet_send() if scheduler return the packet 920 * just enqueued. This avoid a lock order reversal. 921 * 922 */ 923 if (/*dn_cfg.io_fast &&*/ m == *m0 && (dir & PROTO_LAYER2) == 0 ) { 924 /* fast io, rename the tag * to carry reinject info. */ 925 struct m_tag *tag = m_tag_first(m); 926 927 tag->m_tag_cookie = MTAG_IPFW_RULE; 928 tag->m_tag_id = 0; 929 io_pkt_fast++; 930 if (m->m_nextpkt != NULL) { 931 printf("dummynet: fast io: pkt chain detected!\n"); 932 m->m_nextpkt = NULL; 933 } 934 m = NULL; 935 } else { 936 *m0 = NULL; 937 } 938 done: 939 DN_BH_WUNLOCK(); 940 if (m) 941 dummynet_send(m); 942 return 0; 943 944 dropit: 945 io_pkt_drop++; 946 DN_BH_WUNLOCK(); 947 if (m) 948 FREE_PKT(m); 949 *m0 = NULL; 950 return (fs && (fs->fs.flags & DN_NOERROR)) ? 0 : ENOBUFS; 951 } 952