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