1/*- 2 * SPDX-License-Identifier: BSD-2-Clause 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#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/if_var.h> /* NET_EPOCH_... */ 52#include <net/if_private.h> 53#include <net/netisr.h> 54#include <net/vnet.h> 55 56#include <netinet/in.h> 57#include <netinet/ip.h> /* ip_len, ip_off */ 58#include <netinet/ip_var.h> /* ip_output(), IP_FORWARDING */ 59#include <netinet/ip_fw.h> 60#include <netinet/ip_dummynet.h> 61#include <netinet/if_ether.h> /* various ether_* routines */ 62#include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */ 63#include <netinet6/ip6_var.h> 64 65#include <netpfil/ipfw/ip_fw_private.h> 66#include <netpfil/ipfw/dn_heap.h> 67#include <netpfil/ipfw/ip_dn_private.h> 68#ifdef NEW_AQM 69#include <netpfil/ipfw/dn_aqm.h> 70#endif 71#include <netpfil/ipfw/dn_sched.h> 72 73/* 74 * We keep a private variable for the simulation time, but we could 75 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c) 76 * instead of V_dn_cfg.curr_time 77 */ 78VNET_DEFINE(struct dn_parms, dn_cfg); 79#define V_dn_cfg VNET(dn_cfg) 80 81/* 82 * We use a heap to store entities for which we have pending timer events. 83 * The heap is checked at every tick and all entities with expired events 84 * are extracted. 85 */ 86 87MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap"); 88 89extern void (*bridge_dn_p)(struct mbuf *, struct ifnet *); 90 91#ifdef SYSCTL_NODE 92 93/* 94 * Because of the way the SYSBEGIN/SYSEND macros work on other 95 * platforms, there should not be functions between them. 96 * So keep the handlers outside the block. 97 */ 98static int 99sysctl_hash_size(SYSCTL_HANDLER_ARGS) 100{ 101 int error, value; 102 103 value = V_dn_cfg.hash_size; 104 error = sysctl_handle_int(oidp, &value, 0, req); 105 if (error != 0 || req->newptr == NULL) 106 return (error); 107 if (value < 16 || value > 65536) 108 return (EINVAL); 109 V_dn_cfg.hash_size = value; 110 return (0); 111} 112 113static int 114sysctl_limits(SYSCTL_HANDLER_ARGS) 115{ 116 int error; 117 long value; 118 119 if (arg2 != 0) 120 value = V_dn_cfg.slot_limit; 121 else 122 value = V_dn_cfg.byte_limit; 123 error = sysctl_handle_long(oidp, &value, 0, req); 124 125 if (error != 0 || req->newptr == NULL) 126 return (error); 127 if (arg2 != 0) { 128 if (value < 1) 129 return (EINVAL); 130 V_dn_cfg.slot_limit = value; 131 } else { 132 if (value < 1500) 133 return (EINVAL); 134 V_dn_cfg.byte_limit = value; 135 } 136 return (0); 137} 138 139SYSBEGIN(f4) 140 141SYSCTL_DECL(_net_inet); 142SYSCTL_DECL(_net_inet_ip); 143#ifdef NEW_AQM 144SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 145 "Dummynet"); 146#else 147static SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, 148 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 149 "Dummynet"); 150#endif 151 152/* wrapper to pass V_dn_cfg fields to SYSCTL_* */ 153#define DC(x) (&(VNET_NAME(dn_cfg).x)) 154 155/* parameters */ 156 157SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hash_size, 158 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 159 0, 0, sysctl_hash_size, "I", 160 "Default hash table size"); 161 162SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit, 163 CTLTYPE_LONG | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 164 0, 1, sysctl_limits, "L", 165 "Upper limit in slots for pipe queue."); 166SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit, 167 CTLTYPE_LONG | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 168 0, 0, sysctl_limits, "L", 169 "Upper limit in bytes for pipe queue."); 170SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast, 171 CTLFLAG_RW | CTLFLAG_VNET, DC(io_fast), 0, "Enable fast dummynet io."); 172SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, 173 CTLFLAG_RW | CTLFLAG_VNET, DC(debug), 0, "Dummynet debug level"); 174 175/* RED parameters */ 176SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth, 177 CTLFLAG_RD | CTLFLAG_VNET, DC(red_lookup_depth), 0, "Depth of RED lookup table"); 178SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size, 179 CTLFLAG_RD | CTLFLAG_VNET, DC(red_avg_pkt_size), 0, "RED Medium packet size"); 180SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size, 181 CTLFLAG_RD | CTLFLAG_VNET, DC(red_max_pkt_size), 0, "RED Max packet size"); 182 183/* time adjustment */ 184SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta, 185 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_delta), 0, "Last vs standard tick difference (usec)."); 186SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum, 187 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_delta_sum), 0, "Accumulated tick difference (usec)."); 188SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment, 189 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_adjustment), 0, "Tick adjustments done."); 190SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff, 191 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_diff), 0, 192 "Adjusted vs non-adjusted curr_time difference (ticks)."); 193SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost, 194 CTLFLAG_RD | CTLFLAG_VNET, DC(tick_lost), 0, 195 "Number of ticks coalesced by dummynet taskqueue."); 196 197/* Drain parameters */ 198SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire, 199 CTLFLAG_RW | CTLFLAG_VNET, DC(expire), 0, "Expire empty queues/pipes"); 200SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_cycle, 201 CTLFLAG_RD | CTLFLAG_VNET, DC(expire_cycle), 0, "Expire cycle for queues/pipes"); 202 203/* statistics */ 204SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, schk_count, 205 CTLFLAG_RD | CTLFLAG_VNET, DC(schk_count), 0, "Number of schedulers"); 206SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, si_count, 207 CTLFLAG_RD | CTLFLAG_VNET, DC(si_count), 0, "Number of scheduler instances"); 208SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, fsk_count, 209 CTLFLAG_RD | CTLFLAG_VNET, DC(fsk_count), 0, "Number of flowsets"); 210SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, queue_count, 211 CTLFLAG_RD | CTLFLAG_VNET, DC(queue_count), 0, "Number of queues"); 212SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt, 213 CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt), 0, 214 "Number of packets passed to dummynet."); 215SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast, 216 CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt_fast), 0, 217 "Number of packets bypassed dummynet scheduler."); 218SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop, 219 CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt_drop), 0, 220 "Number of packets dropped by dummynet."); 221#undef DC 222SYSEND 223 224#endif 225 226static void dummynet_send(struct mbuf *); 227 228/* 229 * Return the mbuf tag holding the dummynet state (it should 230 * be the first one on the list). 231 */ 232struct dn_pkt_tag * 233dn_tag_get(struct mbuf *m) 234{ 235 struct m_tag *mtag = m_tag_first(m); 236#ifdef NEW_AQM 237 /* XXX: to skip ts m_tag. For Debugging only*/ 238 if (mtag != NULL && mtag->m_tag_id == DN_AQM_MTAG_TS) { 239 m_tag_delete(m,mtag); 240 mtag = m_tag_first(m); 241 D("skip TS tag"); 242 } 243#endif 244 KASSERT(mtag != NULL && 245 mtag->m_tag_cookie == MTAG_ABI_COMPAT && 246 mtag->m_tag_id == PACKET_TAG_DUMMYNET, 247 ("packet on dummynet queue w/o dummynet tag!")); 248 return (struct dn_pkt_tag *)(mtag+1); 249} 250 251#ifndef NEW_AQM 252static inline void 253mq_append(struct mq *q, struct mbuf *m) 254{ 255#ifdef USERSPACE 256 // buffers from netmap need to be copied 257 // XXX note that the routine is not expected to fail 258 ND("append %p to %p", m, q); 259 if (m->m_flags & M_STACK) { 260 struct mbuf *m_new; 261 void *p; 262 int l, ofs; 263 264 ofs = m->m_data - m->__m_extbuf; 265 // XXX allocate 266 MGETHDR(m_new, M_NOWAIT, MT_DATA); 267 ND("*** WARNING, volatile buf %p ext %p %d dofs %d m_new %p", 268 m, m->__m_extbuf, m->__m_extlen, ofs, m_new); 269 p = m_new->__m_extbuf; /* new pointer */ 270 l = m_new->__m_extlen; /* new len */ 271 if (l <= m->__m_extlen) { 272 panic("extlen too large"); 273 } 274 275 *m_new = *m; // copy 276 m_new->m_flags &= ~M_STACK; 277 m_new->__m_extbuf = p; // point to new buffer 278 _pkt_copy(m->__m_extbuf, p, m->__m_extlen); 279 m_new->m_data = p + ofs; 280 m = m_new; 281 } 282#endif /* USERSPACE */ 283 if (q->head == NULL) 284 q->head = m; 285 else 286 q->tail->m_nextpkt = m; 287 q->count++; 288 q->tail = m; 289 m->m_nextpkt = NULL; 290} 291#endif 292 293/* 294 * Dispose a list of packet. Use a functions so if we need to do 295 * more work, this is a central point to do it. 296 */ 297void dn_free_pkts(struct mbuf *mnext) 298{ 299 struct mbuf *m; 300 301 while ((m = mnext) != NULL) { 302 mnext = m->m_nextpkt; 303 FREE_PKT(m); 304 } 305} 306 307static int 308red_drops (struct dn_queue *q, int len) 309{ 310 /* 311 * RED algorithm 312 * 313 * RED calculates the average queue size (avg) using a low-pass filter 314 * with an exponential weighted (w_q) moving average: 315 * avg <- (1-w_q) * avg + w_q * q_size 316 * where q_size is the queue length (measured in bytes or * packets). 317 * 318 * If q_size == 0, we compute the idle time for the link, and set 319 * avg = (1 - w_q)^(idle/s) 320 * where s is the time needed for transmitting a medium-sized packet. 321 * 322 * Now, if avg < min_th the packet is enqueued. 323 * If avg > max_th the packet is dropped. Otherwise, the packet is 324 * dropped with probability P function of avg. 325 */ 326 327 struct dn_fsk *fs = q->fs; 328 int64_t p_b = 0; 329 330 /* Queue in bytes or packets? */ 331 uint32_t q_size = (fs->fs.flags & DN_QSIZE_BYTES) ? 332 q->ni.len_bytes : q->ni.length; 333 334 /* Average queue size estimation. */ 335 if (q_size != 0) { 336 /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */ 337 int diff = SCALE(q_size) - q->avg; 338 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q); 339 340 q->avg += (int)v; 341 } else { 342 /* 343 * Queue is empty, find for how long the queue has been 344 * empty and use a lookup table for computing 345 * (1 - * w_q)^(idle_time/s) where s is the time to send a 346 * (small) packet. 347 * XXX check wraps... 348 */ 349 if (q->avg) { 350 u_int t = div64((V_dn_cfg.curr_time - q->q_time), fs->lookup_step); 351 352 q->avg = (t < fs->lookup_depth) ? 353 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0; 354 } 355 } 356 357 /* Should i drop? */ 358 if (q->avg < fs->min_th) { 359 q->count = -1; 360 return (0); /* accept packet */ 361 } 362 if (q->avg >= fs->max_th) { /* average queue >= max threshold */ 363 if (fs->fs.flags & DN_IS_ECN) 364 return (1); 365 if (fs->fs.flags & DN_IS_GENTLE_RED) { 366 /* 367 * According to Gentle-RED, if avg is greater than 368 * max_th the packet is dropped with a probability 369 * p_b = c_3 * avg - c_4 370 * where c_3 = (1 - max_p) / max_th 371 * c_4 = 1 - 2 * max_p 372 */ 373 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) - 374 fs->c_4; 375 } else { 376 q->count = -1; 377 return (1); 378 } 379 } else if (q->avg > fs->min_th) { 380 if (fs->fs.flags & DN_IS_ECN) 381 return (1); 382 /* 383 * We compute p_b using the linear dropping function 384 * p_b = c_1 * avg - c_2 385 * where c_1 = max_p / (max_th - min_th) 386 * c_2 = max_p * min_th / (max_th - min_th) 387 */ 388 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2; 389 } 390 391 if (fs->fs.flags & DN_QSIZE_BYTES) 392 p_b = div64((p_b * len) , fs->max_pkt_size); 393 if (++q->count == 0) 394 q->random = random() & 0xffff; 395 else { 396 /* 397 * q->count counts packets arrived since last drop, so a greater 398 * value of q->count means a greater packet drop probability. 399 */ 400 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) { 401 q->count = 0; 402 /* After a drop we calculate a new random value. */ 403 q->random = random() & 0xffff; 404 return (1); /* drop */ 405 } 406 } 407 /* End of RED algorithm. */ 408 409 return (0); /* accept */ 410 411} 412 413/* 414 * ECN/ECT Processing (partially adopted from altq) 415 */ 416#ifndef NEW_AQM 417static 418#endif 419int 420ecn_mark(struct mbuf* m) 421{ 422 struct ip *ip; 423 ip = (struct ip *)mtodo(m, dn_tag_get(m)->iphdr_off); 424 425 switch (ip->ip_v) { 426 case IPVERSION: 427 { 428 uint16_t old; 429 430 if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_NOTECT) 431 return (0); /* not-ECT */ 432 if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_CE) 433 return (1); /* already marked */ 434 435 /* 436 * ecn-capable but not marked, 437 * mark CE and update checksum 438 */ 439 old = *(uint16_t *)ip; 440 ip->ip_tos |= IPTOS_ECN_CE; 441 ip->ip_sum = cksum_adjust(ip->ip_sum, old, *(uint16_t *)ip); 442 return (1); 443 } 444#ifdef INET6 445 case (IPV6_VERSION >> 4): 446 { 447 struct ip6_hdr *ip6 = (struct ip6_hdr *)ip; 448 u_int32_t flowlabel; 449 450 flowlabel = ntohl(ip6->ip6_flow); 451 if ((flowlabel >> 28) != 6) 452 return (0); /* version mismatch! */ 453 if ((flowlabel & (IPTOS_ECN_MASK << 20)) == 454 (IPTOS_ECN_NOTECT << 20)) 455 return (0); /* not-ECT */ 456 if ((flowlabel & (IPTOS_ECN_MASK << 20)) == 457 (IPTOS_ECN_CE << 20)) 458 return (1); /* already marked */ 459 /* 460 * ecn-capable but not marked, mark CE 461 */ 462 flowlabel |= (IPTOS_ECN_CE << 20); 463 ip6->ip6_flow = htonl(flowlabel); 464 return (1); 465 } 466#endif 467 } 468 return (0); 469} 470 471/* 472 * Enqueue a packet in q, subject to space and queue management policy 473 * (whose parameters are in q->fs). 474 * Update stats for the queue and the scheduler. 475 * Return 0 on success, 1 on drop. The packet is consumed anyways. 476 */ 477int 478dn_enqueue(struct dn_queue *q, struct mbuf* m, int drop) 479{ 480 struct dn_fs *f; 481 struct dn_flow *ni; /* stats for scheduler instance */ 482 uint64_t len; 483 484 if (q->fs == NULL || q->_si == NULL) { 485 printf("%s fs %p si %p, dropping\n", 486 __FUNCTION__, q->fs, q->_si); 487 FREE_PKT(m); 488 return 1; 489 } 490 f = &(q->fs->fs); 491 ni = &q->_si->ni; 492 len = m->m_pkthdr.len; 493 /* Update statistics, then check reasons to drop pkt. */ 494 q->ni.tot_bytes += len; 495 q->ni.tot_pkts++; 496 ni->tot_bytes += len; 497 ni->tot_pkts++; 498 if (drop) 499 goto drop; 500 if (f->plr[0] || f->plr[1]) { 501 if (__predict_true(f->plr[1] == 0)) { 502 if (random() < f->plr[0]) 503 goto drop; 504 } else { 505 switch (f->pl_state) { 506 case PLR_STATE_B: 507 if (random() < f->plr[3]) 508 f->pl_state = PLR_STATE_G; 509 if (random() < f->plr[2]) 510 goto drop; 511 break; 512 case PLR_STATE_G: /* FALLTHROUGH */ 513 default: 514 if (random() < f->plr[1]) 515 f->pl_state = PLR_STATE_B; 516 if (random() < f->plr[0]) 517 goto drop; 518 break; 519 } 520 } 521 } 522 if (m->m_pkthdr.rcvif != NULL) 523 m_rcvif_serialize(m); 524#ifdef NEW_AQM 525 /* Call AQM enqueue function */ 526 if (q->fs->aqmfp) 527 return q->fs->aqmfp->enqueue(q ,m); 528#endif 529 if (f->flags & DN_IS_RED && red_drops(q, m->m_pkthdr.len)) { 530 if (!(f->flags & DN_IS_ECN) || !ecn_mark(m)) 531 goto drop; 532 } 533 if (f->flags & DN_QSIZE_BYTES) { 534 if (q->ni.len_bytes > f->qsize) 535 goto drop; 536 } else if (q->ni.length >= f->qsize) { 537 goto drop; 538 } 539 mq_append(&q->mq, m); 540 q->ni.length++; 541 q->ni.len_bytes += len; 542 ni->length++; 543 ni->len_bytes += len; 544 return (0); 545 546drop: 547 V_dn_cfg.io_pkt_drop++; 548 q->ni.drops++; 549 ni->drops++; 550 FREE_PKT(m); 551 return (1); 552} 553 554/* 555 * Fetch packets from the delay line which are due now. If there are 556 * leftover packets, reinsert the delay line in the heap. 557 * Runs under scheduler lock. 558 */ 559static void 560transmit_event(struct mq *q, struct delay_line *dline, uint64_t now) 561{ 562 struct mbuf *m; 563 struct dn_pkt_tag *pkt = NULL; 564 565 dline->oid.subtype = 0; /* not in heap */ 566 while ((m = dline->mq.head) != NULL) { 567 pkt = dn_tag_get(m); 568 if (!DN_KEY_LEQ(pkt->output_time, now)) 569 break; 570 dline->mq.head = m->m_nextpkt; 571 dline->mq.count--; 572 if (m->m_pkthdr.rcvif != NULL && 573 __predict_false(m_rcvif_restore(m) == NULL)) 574 m_freem(m); 575 else 576 mq_append(q, m); 577 } 578 if (m != NULL) { 579 dline->oid.subtype = 1; /* in heap */ 580 heap_insert(&V_dn_cfg.evheap, pkt->output_time, dline); 581 } 582} 583 584/* 585 * Convert the additional MAC overheads/delays into an equivalent 586 * number of bits for the given data rate. The samples are 587 * in milliseconds so we need to divide by 1000. 588 */ 589static uint64_t 590extra_bits(struct mbuf *m, struct dn_schk *s) 591{ 592 int index; 593 uint64_t bits; 594 struct dn_profile *pf = s->profile; 595 596 if (!pf || pf->samples_no == 0) 597 return 0; 598 index = random() % pf->samples_no; 599 bits = div64((uint64_t)pf->samples[index] * s->link.bandwidth, 1000); 600 if (index >= pf->loss_level) { 601 struct dn_pkt_tag *dt = dn_tag_get(m); 602 if (dt) 603 dt->dn_dir = DIR_DROP; 604 } 605 return bits; 606} 607 608/* 609 * Send traffic from a scheduler instance due by 'now'. 610 * Return a pointer to the head of the queue. 611 */ 612static struct mbuf * 613serve_sched(struct mq *q, struct dn_sch_inst *si, uint64_t now) 614{ 615 struct mq def_q; 616 struct dn_schk *s = si->sched; 617 struct mbuf *m = NULL; 618 int delay_line_idle = (si->dline.mq.head == NULL); 619 int done; 620 uint32_t bw; 621 622 if (q == NULL) { 623 q = &def_q; 624 q->head = NULL; 625 } 626 627 bw = s->link.bandwidth; 628 si->kflags &= ~DN_ACTIVE; 629 630 if (bw > 0) 631 si->credit += (now - si->sched_time) * bw; 632 else 633 si->credit = 0; 634 si->sched_time = now; 635 done = 0; 636 while (si->credit >= 0 && (m = s->fp->dequeue(si)) != NULL) { 637 uint64_t len_scaled; 638 639 done++; 640 len_scaled = (bw == 0) ? 0 : hz * 641 (m->m_pkthdr.len * 8 + extra_bits(m, s)); 642 si->credit -= len_scaled; 643 /* Move packet in the delay line */ 644 dn_tag_get(m)->output_time = V_dn_cfg.curr_time + s->link.delay ; 645 if (m->m_pkthdr.rcvif != NULL) 646 m_rcvif_serialize(m); 647 mq_append(&si->dline.mq, m); 648 } 649 650 /* 651 * If credit >= 0 the instance is idle, mark time. 652 * Otherwise put back in the heap, and adjust the output 653 * time of the last inserted packet, m, which was too early. 654 */ 655 if (si->credit >= 0) { 656 si->idle_time = now; 657 } else { 658 uint64_t t; 659 KASSERT (bw > 0, ("bw=0 and credit<0 ?")); 660 t = div64(bw - 1 - si->credit, bw); 661 if (m) 662 dn_tag_get(m)->output_time += t; 663 si->kflags |= DN_ACTIVE; 664 heap_insert(&V_dn_cfg.evheap, now + t, si); 665 } 666 if (delay_line_idle && done) 667 transmit_event(q, &si->dline, now); 668 return q->head; 669} 670 671/* 672 * The timer handler for dummynet. Time is computed in ticks, but 673 * but the code is tolerant to the actual rate at which this is called. 674 * Once complete, the function reschedules itself for the next tick. 675 */ 676void 677dummynet_task(void *context, int pending) 678{ 679 struct timeval t; 680 struct mq q = { NULL, NULL }; /* queue to accumulate results */ 681 struct epoch_tracker et; 682 683 VNET_ITERATOR_DECL(vnet_iter); 684 VNET_LIST_RLOCK(); 685 NET_EPOCH_ENTER(et); 686 687 VNET_FOREACH(vnet_iter) { 688 memset(&q, 0, sizeof(struct mq)); 689 CURVNET_SET(vnet_iter); 690 691 if (! V_dn_cfg.init_done) { 692 CURVNET_RESTORE(); 693 continue; 694 } 695 696 DN_BH_WLOCK(); 697 698 /* Update number of lost(coalesced) ticks. */ 699 V_dn_cfg.tick_lost += pending - 1; 700 701 getmicrouptime(&t); 702 /* Last tick duration (usec). */ 703 V_dn_cfg.tick_last = (t.tv_sec - V_dn_cfg.prev_t.tv_sec) * 1000000 + 704 (t.tv_usec - V_dn_cfg.prev_t.tv_usec); 705 /* Last tick vs standard tick difference (usec). */ 706 V_dn_cfg.tick_delta = (V_dn_cfg.tick_last * hz - 1000000) / hz; 707 /* Accumulated tick difference (usec). */ 708 V_dn_cfg.tick_delta_sum += V_dn_cfg.tick_delta; 709 710 V_dn_cfg.prev_t = t; 711 712 /* 713 * Adjust curr_time if the accumulated tick difference is 714 * greater than the 'standard' tick. Since curr_time should 715 * be monotonically increasing, we do positive adjustments 716 * as required, and throttle curr_time in case of negative 717 * adjustment. 718 */ 719 V_dn_cfg.curr_time++; 720 if (V_dn_cfg.tick_delta_sum - tick >= 0) { 721 int diff = V_dn_cfg.tick_delta_sum / tick; 722 723 V_dn_cfg.curr_time += diff; 724 V_dn_cfg.tick_diff += diff; 725 V_dn_cfg.tick_delta_sum %= tick; 726 V_dn_cfg.tick_adjustment++; 727 } else if (V_dn_cfg.tick_delta_sum + tick <= 0) { 728 V_dn_cfg.curr_time--; 729 V_dn_cfg.tick_diff--; 730 V_dn_cfg.tick_delta_sum += tick; 731 V_dn_cfg.tick_adjustment++; 732 } 733 734 /* serve pending events, accumulate in q */ 735 for (;;) { 736 struct dn_id *p; /* generic parameter to handler */ 737 738 if (V_dn_cfg.evheap.elements == 0 || 739 DN_KEY_LT(V_dn_cfg.curr_time, HEAP_TOP(&V_dn_cfg.evheap)->key)) 740 break; 741 p = HEAP_TOP(&V_dn_cfg.evheap)->object; 742 heap_extract(&V_dn_cfg.evheap, NULL); 743 if (p->type == DN_SCH_I) { 744 serve_sched(&q, (struct dn_sch_inst *)p, V_dn_cfg.curr_time); 745 } else { /* extracted a delay line */ 746 transmit_event(&q, (struct delay_line *)p, V_dn_cfg.curr_time); 747 } 748 } 749 if (V_dn_cfg.expire && ++V_dn_cfg.expire_cycle >= V_dn_cfg.expire) { 750 V_dn_cfg.expire_cycle = 0; 751 dn_drain_scheduler(); 752 dn_drain_queue(); 753 } 754 DN_BH_WUNLOCK(); 755 if (q.head != NULL) 756 dummynet_send(q.head); 757 758 CURVNET_RESTORE(); 759 } 760 NET_EPOCH_EXIT(et); 761 VNET_LIST_RUNLOCK(); 762 763 /* Schedule our next run. */ 764 dn_reschedule(); 765} 766 767/* 768 * forward a chain of packets to the proper destination. 769 * This runs outside the dummynet lock. 770 */ 771static void 772dummynet_send(struct mbuf *m) 773{ 774 struct mbuf *n; 775 776 NET_EPOCH_ASSERT(); 777 778 for (; m != NULL; m = n) { 779 struct ifnet *ifp = NULL; /* gcc 3.4.6 complains */ 780 struct m_tag *tag; 781 int dst; 782 783 n = m->m_nextpkt; 784 m->m_nextpkt = NULL; 785 tag = m_tag_first(m); 786 if (tag == NULL) { /* should not happen */ 787 dst = DIR_DROP; 788 } else { 789 struct dn_pkt_tag *pkt = dn_tag_get(m); 790 /* extract the dummynet info, rename the tag 791 * to carry reinject info. 792 */ 793 ifp = ifnet_byindexgen(pkt->if_index, pkt->if_idxgen); 794 if (((pkt->dn_dir == (DIR_OUT | PROTO_LAYER2)) || 795 (pkt->dn_dir == (DIR_OUT | PROTO_LAYER2 | PROTO_IPV6))) && 796 ifp == NULL) { 797 dst = DIR_DROP; 798 } else { 799 dst = pkt->dn_dir; 800 tag->m_tag_cookie = MTAG_IPFW_RULE; 801 tag->m_tag_id = 0; 802 } 803 } 804 805 switch (dst) { 806 case DIR_OUT: 807 ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL); 808 break ; 809 810 case DIR_IN : 811 netisr_dispatch(NETISR_IP, m); 812 break; 813 814#ifdef INET6 815 case DIR_IN | PROTO_IPV6: 816 netisr_dispatch(NETISR_IPV6, m); 817 break; 818 819 case DIR_OUT | PROTO_IPV6: 820 ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL); 821 break; 822#endif 823 824 case DIR_FWD | PROTO_IFB: /* DN_TO_IFB_FWD: */ 825 if (bridge_dn_p != NULL) 826 ((*bridge_dn_p)(m, ifp)); 827 else 828 printf("dummynet: if_bridge not loaded\n"); 829 830 break; 831 832 case DIR_IN | PROTO_LAYER2 | PROTO_IPV6: 833 case DIR_IN | PROTO_LAYER2: /* DN_TO_ETH_DEMUX: */ 834 /* 835 * The Ethernet code assumes the Ethernet header is 836 * contiguous in the first mbuf header. 837 * Insure this is true. 838 */ 839 if (m->m_len < ETHER_HDR_LEN && 840 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) { 841 printf("dummynet/ether: pullup failed, " 842 "dropping packet\n"); 843 break; 844 } 845 ether_demux(m->m_pkthdr.rcvif, m); 846 break; 847 848 case DIR_OUT | PROTO_LAYER2 | PROTO_IPV6: 849 case DIR_OUT | PROTO_LAYER2: /* DN_TO_ETH_OUT: */ 850 MPASS(ifp != NULL); 851 ether_output_frame(ifp, m); 852 break; 853 854 case DIR_DROP: 855 /* drop the packet after some time */ 856 FREE_PKT(m); 857 break; 858 859 default: 860 printf("dummynet: bad switch %d!\n", dst); 861 FREE_PKT(m); 862 break; 863 } 864 } 865} 866 867static inline int 868tag_mbuf(struct mbuf *m, int dir, struct ip_fw_args *fwa) 869{ 870 struct dn_pkt_tag *dt; 871 struct m_tag *mtag; 872 873 mtag = m_tag_get(PACKET_TAG_DUMMYNET, 874 sizeof(*dt), M_NOWAIT | M_ZERO); 875 if (mtag == NULL) 876 return 1; /* Cannot allocate packet header. */ 877 m_tag_prepend(m, mtag); /* Attach to mbuf chain. */ 878 dt = (struct dn_pkt_tag *)(mtag + 1); 879 dt->rule = fwa->rule; 880 /* only keep this info */ 881 dt->rule.info &= (IPFW_ONEPASS | IPFW_IS_DUMMYNET); 882 dt->dn_dir = dir; 883 if (fwa->flags & IPFW_ARGS_OUT && fwa->ifp != NULL) { 884 NET_EPOCH_ASSERT(); 885 dt->if_index = fwa->ifp->if_index; 886 dt->if_idxgen = fwa->ifp->if_idxgen; 887 } 888 /* dt->output_time is updated as we move through */ 889 dt->output_time = V_dn_cfg.curr_time; 890 dt->iphdr_off = (dir & PROTO_LAYER2) ? ETHER_HDR_LEN : 0; 891 return 0; 892} 893 894/* 895 * dummynet hook for packets. 896 * We use the argument to locate the flowset fs and the sched_set sch 897 * associated to it. The we apply flow_mask and sched_mask to 898 * determine the queue and scheduler instances. 899 */ 900int 901dummynet_io(struct mbuf **m0, struct ip_fw_args *fwa) 902{ 903 struct mbuf *m = *m0; 904 struct dn_fsk *fs = NULL; 905 struct dn_sch_inst *si; 906 struct dn_queue *q = NULL; /* default */ 907 int fs_id, dir; 908 909 fs_id = (fwa->rule.info & IPFW_INFO_MASK) + 910 ((fwa->rule.info & IPFW_IS_PIPE) ? 2*DN_MAX_ID : 0); 911 /* XXXGL: convert args to dir */ 912 if (fwa->flags & IPFW_ARGS_IN) 913 dir = DIR_IN; 914 else 915 dir = DIR_OUT; 916 if (fwa->flags & IPFW_ARGS_ETHER) 917 dir |= PROTO_LAYER2; 918 else if (fwa->flags & IPFW_ARGS_IP6) 919 dir |= PROTO_IPV6; 920 DN_BH_WLOCK(); 921 V_dn_cfg.io_pkt++; 922 /* we could actually tag outside the lock, but who cares... */ 923 if (tag_mbuf(m, dir, fwa)) 924 goto dropit; 925 /* XXX locate_flowset could be optimised with a direct ref. */ 926 fs = dn_ht_find(V_dn_cfg.fshash, fs_id, 0, NULL); 927 if (fs == NULL) 928 goto dropit; /* This queue/pipe does not exist! */ 929 if (fs->sched == NULL) /* should not happen */ 930 goto dropit; 931 /* find scheduler instance, possibly applying sched_mask */ 932 si = ipdn_si_find(fs->sched, &(fwa->f_id)); 933 if (si == NULL) 934 goto dropit; 935 /* 936 * If the scheduler supports multiple queues, find the right one 937 * (otherwise it will be ignored by enqueue). 938 */ 939 if (fs->sched->fp->flags & DN_MULTIQUEUE) { 940 q = ipdn_q_find(fs, si, &(fwa->f_id)); 941 if (q == NULL) 942 goto dropit; 943 } 944 if (fs->sched->fp->enqueue(si, q, m)) { 945 /* packet was dropped by enqueue() */ 946 m = *m0 = NULL; 947 948 /* dn_enqueue already increases io_pkt_drop */ 949 V_dn_cfg.io_pkt_drop--; 950 951 goto dropit; 952 } 953 954 if (si->kflags & DN_ACTIVE) { 955 m = *m0 = NULL; /* consumed */ 956 goto done; /* already active, nothing to do */ 957 } 958 959 /* compute the initial allowance */ 960 if (si->idle_time < V_dn_cfg.curr_time) { 961 /* Do this only on the first packet on an idle pipe */ 962 struct dn_link *p = &fs->sched->link; 963 964 si->sched_time = V_dn_cfg.curr_time; 965 si->credit = V_dn_cfg.io_fast ? p->bandwidth : 0; 966 if (p->burst) { 967 uint64_t burst = (V_dn_cfg.curr_time - si->idle_time) * p->bandwidth; 968 if (burst > p->burst) 969 burst = p->burst; 970 si->credit += burst; 971 } 972 } 973 /* pass through scheduler and delay line */ 974 m = serve_sched(NULL, si, V_dn_cfg.curr_time); 975 976 /* optimization -- pass it back to ipfw for immediate send */ 977 /* XXX Don't call dummynet_send() if scheduler return the packet 978 * just enqueued. This avoid a lock order reversal. 979 * 980 */ 981 if (/*V_dn_cfg.io_fast &&*/ m == *m0 && (dir & PROTO_LAYER2) == 0 ) { 982 /* fast io, rename the tag * to carry reinject info. */ 983 struct m_tag *tag = m_tag_first(m); 984 985 tag->m_tag_cookie = MTAG_IPFW_RULE; 986 tag->m_tag_id = 0; 987 V_dn_cfg.io_pkt_fast++; 988 if (m->m_nextpkt != NULL) { 989 printf("dummynet: fast io: pkt chain detected!\n"); 990 m->m_nextpkt = NULL; 991 } 992 m = NULL; 993 } else { 994 *m0 = NULL; 995 } 996done: 997 DN_BH_WUNLOCK(); 998 if (m) 999 dummynet_send(m); 1000 return 0; 1001 1002dropit: 1003 V_dn_cfg.io_pkt_drop++; 1004 DN_BH_WUNLOCK(); 1005 if (m) 1006 FREE_PKT(m); 1007 *m0 = NULL; 1008 return (fs && (fs->fs.flags & DN_NOERROR)) ? 0 : ENOBUFS; 1009} 1010