ip_dn_io.c revision 301231
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: stable/10/sys/netpfil/ipfw/ip_dn_io.c 301231 2016-06-03 00:48:50Z truckman $"); 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/priv.h> 43#include <sys/proc.h> 44#include <sys/rwlock.h> 45#include <sys/socket.h> 46#include <sys/time.h> 47#include <sys/sysctl.h> 48 49#include <net/if.h> /* IFNAMSIZ, struct ifaddr, ifq head, lock.h mutex.h */ 50#include <net/netisr.h> 51#include <net/vnet.h> 52 53#include <netinet/in.h> 54#include <netinet/ip.h> /* ip_len, ip_off */ 55#include <netinet/ip_var.h> /* ip_output(), IP_FORWARDING */ 56#include <netinet/ip_fw.h> 57#include <netinet/ip_dummynet.h> 58#include <netinet/if_ether.h> /* various ether_* routines */ 59#include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */ 60#include <netinet6/ip6_var.h> 61 62#include <netpfil/ipfw/ip_fw_private.h> 63#include <netpfil/ipfw/dn_heap.h> 64#include <netpfil/ipfw/ip_dn_private.h> 65#include <netpfil/ipfw/dn_sched.h> 66 67/* 68 * We keep a private variable for the simulation time, but we could 69 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c) 70 * instead of dn_cfg.curr_time 71 */ 72 73struct dn_parms dn_cfg; 74//VNET_DEFINE(struct dn_parms, _base_dn_cfg); 75 76static long tick_last; /* Last tick duration (usec). */ 77static long tick_delta; /* Last vs standard tick diff (usec). */ 78static long tick_delta_sum; /* Accumulated tick difference (usec).*/ 79static long tick_adjustment; /* Tick adjustments done. */ 80static long tick_lost; /* Lost(coalesced) ticks number. */ 81/* Adjusted vs non-adjusted curr_time difference (ticks). */ 82static long tick_diff; 83 84static unsigned long io_pkt; 85static unsigned long io_pkt_fast; 86static unsigned long io_pkt_drop; 87 88/* 89 * We use a heap to store entities for which we have pending timer events. 90 * The heap is checked at every tick and all entities with expired events 91 * are extracted. 92 */ 93 94MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap"); 95 96extern void (*bridge_dn_p)(struct mbuf *, struct ifnet *); 97 98#ifdef SYSCTL_NODE 99 100/* 101 * Because of the way the SYSBEGIN/SYSEND macros work on other 102 * platforms, there should not be functions between them. 103 * So keep the handlers outside the block. 104 */ 105static int 106sysctl_hash_size(SYSCTL_HANDLER_ARGS) 107{ 108 int error, value; 109 110 value = dn_cfg.hash_size; 111 error = sysctl_handle_int(oidp, &value, 0, req); 112 if (error != 0 || req->newptr == NULL) 113 return (error); 114 if (value < 16 || value > 65536) 115 return (EINVAL); 116 dn_cfg.hash_size = value; 117 return (0); 118} 119 120static int 121sysctl_limits(SYSCTL_HANDLER_ARGS) 122{ 123 int error; 124 long value; 125 126 if (arg2 != 0) 127 value = dn_cfg.slot_limit; 128 else 129 value = dn_cfg.byte_limit; 130 error = sysctl_handle_long(oidp, &value, 0, req); 131 132 if (error != 0 || req->newptr == NULL) 133 return (error); 134 if (arg2 != 0) { 135 if (value < 1) 136 return (EINVAL); 137 dn_cfg.slot_limit = value; 138 } else { 139 if (value < 1500) 140 return (EINVAL); 141 dn_cfg.byte_limit = value; 142 } 143 return (0); 144} 145 146SYSBEGIN(f4) 147 148SYSCTL_DECL(_net_inet); 149SYSCTL_DECL(_net_inet_ip); 150static SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW, 0, "Dummynet"); 151 152/* wrapper to pass dn_cfg fields to SYSCTL_* */ 153//#define DC(x) (&(VNET_NAME(_base_dn_cfg).x)) 154#define DC(x) (&(dn_cfg.x)) 155/* parameters */ 156 157 158SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hash_size, 159 CTLTYPE_INT | CTLFLAG_RW, 0, 0, sysctl_hash_size, 160 "I", "Default hash table size"); 161 162 163SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit, 164 CTLTYPE_LONG | CTLFLAG_RW, 0, 1, sysctl_limits, 165 "L", "Upper limit in slots for pipe queue."); 166SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit, 167 CTLTYPE_LONG | CTLFLAG_RW, 0, 0, sysctl_limits, 168 "L", "Upper limit in bytes for pipe queue."); 169SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast, 170 CTLFLAG_RW, DC(io_fast), 0, "Enable fast dummynet io."); 171SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, 172 CTLFLAG_RW, DC(debug), 0, "Dummynet debug level"); 173 174/* RED parameters */ 175SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth, 176 CTLFLAG_RD, DC(red_lookup_depth), 0, "Depth of RED lookup table"); 177SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size, 178 CTLFLAG_RD, DC(red_avg_pkt_size), 0, "RED Medium packet size"); 179SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size, 180 CTLFLAG_RD, DC(red_max_pkt_size), 0, "RED Max packet size"); 181 182/* time adjustment */ 183SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta, 184 CTLFLAG_RD, &tick_delta, 0, "Last vs standard tick difference (usec)."); 185SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum, 186 CTLFLAG_RD, &tick_delta_sum, 0, "Accumulated tick difference (usec)."); 187SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment, 188 CTLFLAG_RD, &tick_adjustment, 0, "Tick adjustments done."); 189SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff, 190 CTLFLAG_RD, &tick_diff, 0, 191 "Adjusted vs non-adjusted curr_time difference (ticks)."); 192SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost, 193 CTLFLAG_RD, &tick_lost, 0, 194 "Number of ticks coalesced by dummynet taskqueue."); 195 196/* Drain parameters */ 197SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire, 198 CTLFLAG_RW, DC(expire), 0, "Expire empty queues/pipes"); 199SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_cycle, 200 CTLFLAG_RD, DC(expire_cycle), 0, "Expire cycle for queues/pipes"); 201 202/* statistics */ 203SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, schk_count, 204 CTLFLAG_RD, DC(schk_count), 0, "Number of schedulers"); 205SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, si_count, 206 CTLFLAG_RD, DC(si_count), 0, "Number of scheduler instances"); 207SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, fsk_count, 208 CTLFLAG_RD, DC(fsk_count), 0, "Number of flowsets"); 209SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, queue_count, 210 CTLFLAG_RD, DC(queue_count), 0, "Number of queues"); 211SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt, 212 CTLFLAG_RD, &io_pkt, 0, 213 "Number of packets passed to dummynet."); 214SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast, 215 CTLFLAG_RD, &io_pkt_fast, 0, 216 "Number of packets bypassed dummynet scheduler."); 217SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop, 218 CTLFLAG_RD, &io_pkt_drop, 0, 219 "Number of packets dropped by dummynet."); 220#undef DC 221SYSEND 222 223#endif 224 225static void dummynet_send(struct mbuf *); 226 227/* 228 * Packets processed by dummynet have an mbuf tag associated with 229 * them that carries their dummynet state. 230 * Outside dummynet, only the 'rule' field is relevant, and it must 231 * be at the beginning of the structure. 232 */ 233struct dn_pkt_tag { 234 struct ipfw_rule_ref rule; /* matching rule */ 235 236 /* second part, dummynet specific */ 237 int dn_dir; /* action when packet comes out.*/ 238 /* see ip_fw_private.h */ 239 uint64_t output_time; /* when the pkt is due for delivery*/ 240 struct ifnet *ifp; /* interface, for ip_output */ 241 struct _ip6dn_args ip6opt; /* XXX ipv6 options */ 242}; 243 244/* 245 * Return the mbuf tag holding the dummynet state (it should 246 * be the first one on the list). 247 */ 248static struct dn_pkt_tag * 249dn_tag_get(struct mbuf *m) 250{ 251 struct m_tag *mtag = m_tag_first(m); 252 KASSERT(mtag != NULL && 253 mtag->m_tag_cookie == MTAG_ABI_COMPAT && 254 mtag->m_tag_id == PACKET_TAG_DUMMYNET, 255 ("packet on dummynet queue w/o dummynet tag!")); 256 return (struct dn_pkt_tag *)(mtag+1); 257} 258 259static inline void 260mq_append(struct mq *q, struct mbuf *m) 261{ 262 if (q->head == NULL) 263 q->head = m; 264 else 265 q->tail->m_nextpkt = m; 266 q->tail = m; 267 m->m_nextpkt = NULL; 268} 269 270/* 271 * Dispose a list of packet. Use a functions so if we need to do 272 * more work, this is a central point to do it. 273 */ 274void dn_free_pkts(struct mbuf *mnext) 275{ 276 struct mbuf *m; 277 278 while ((m = mnext) != NULL) { 279 mnext = m->m_nextpkt; 280 FREE_PKT(m); 281 } 282} 283 284static int 285red_drops (struct dn_queue *q, int len) 286{ 287 /* 288 * RED algorithm 289 * 290 * RED calculates the average queue size (avg) using a low-pass filter 291 * with an exponential weighted (w_q) moving average: 292 * avg <- (1-w_q) * avg + w_q * q_size 293 * where q_size is the queue length (measured in bytes or * packets). 294 * 295 * If q_size == 0, we compute the idle time for the link, and set 296 * avg = (1 - w_q)^(idle/s) 297 * where s is the time needed for transmitting a medium-sized packet. 298 * 299 * Now, if avg < min_th the packet is enqueued. 300 * If avg > max_th the packet is dropped. Otherwise, the packet is 301 * dropped with probability P function of avg. 302 */ 303 304 struct dn_fsk *fs = q->fs; 305 int64_t p_b = 0; 306 307 /* Queue in bytes or packets? */ 308 uint32_t q_size = (fs->fs.flags & DN_QSIZE_BYTES) ? 309 q->ni.len_bytes : q->ni.length; 310 311 /* Average queue size estimation. */ 312 if (q_size != 0) { 313 /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */ 314 int diff = SCALE(q_size) - q->avg; 315 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q); 316 317 q->avg += (int)v; 318 } else { 319 /* 320 * Queue is empty, find for how long the queue has been 321 * empty and use a lookup table for computing 322 * (1 - * w_q)^(idle_time/s) where s is the time to send a 323 * (small) packet. 324 * XXX check wraps... 325 */ 326 if (q->avg) { 327 u_int t = div64((dn_cfg.curr_time - q->q_time), fs->lookup_step); 328 329 q->avg = (t < fs->lookup_depth) ? 330 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0; 331 } 332 } 333 334 /* Should i drop? */ 335 if (q->avg < fs->min_th) { 336 q->count = -1; 337 return (0); /* accept packet */ 338 } 339 if (q->avg >= fs->max_th) { /* average queue >= max threshold */ 340 if (fs->fs.flags & DN_IS_ECN) 341 return (1); 342 if (fs->fs.flags & DN_IS_GENTLE_RED) { 343 /* 344 * According to Gentle-RED, if avg is greater than 345 * max_th the packet is dropped with a probability 346 * p_b = c_3 * avg - c_4 347 * where c_3 = (1 - max_p) / max_th 348 * c_4 = 1 - 2 * max_p 349 */ 350 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) - 351 fs->c_4; 352 } else { 353 q->count = -1; 354 return (1); 355 } 356 } else if (q->avg > fs->min_th) { 357 if (fs->fs.flags & DN_IS_ECN) 358 return (1); 359 /* 360 * We compute p_b using the linear dropping function 361 * p_b = c_1 * avg - c_2 362 * where c_1 = max_p / (max_th - min_th) 363 * c_2 = max_p * min_th / (max_th - min_th) 364 */ 365 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2; 366 } 367 368 if (fs->fs.flags & DN_QSIZE_BYTES) 369 p_b = div64((p_b * len) , fs->max_pkt_size); 370 if (++q->count == 0) 371 q->random = random() & 0xffff; 372 else { 373 /* 374 * q->count counts packets arrived since last drop, so a greater 375 * value of q->count means a greater packet drop probability. 376 */ 377 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) { 378 q->count = 0; 379 /* After a drop we calculate a new random value. */ 380 q->random = random() & 0xffff; 381 return (1); /* drop */ 382 } 383 } 384 /* End of RED algorithm. */ 385 386 return (0); /* accept */ 387 388} 389 390/* 391 * ECN/ECT Processing (partially adopted from altq) 392 */ 393static int 394ecn_mark(struct mbuf* m) 395{ 396 struct ip *ip; 397 ip = mtod(m, struct ip *); 398 399 switch (ip->ip_v) { 400 case IPVERSION: 401 { 402 u_int8_t otos; 403 int sum; 404 405 if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_NOTECT) 406 return (0); /* not-ECT */ 407 if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_CE) 408 return (1); /* already marked */ 409 410 /* 411 * ecn-capable but not marked, 412 * mark CE and update checksum 413 */ 414 otos = ip->ip_tos; 415 ip->ip_tos |= IPTOS_ECN_CE; 416 /* 417 * update checksum (from RFC1624) 418 * HC' = ~(~HC + ~m + m') 419 */ 420 sum = ~ntohs(ip->ip_sum) & 0xffff; 421 sum += (~otos & 0xffff) + ip->ip_tos; 422 sum = (sum >> 16) + (sum & 0xffff); 423 sum += (sum >> 16); /* add carry */ 424 ip->ip_sum = htons(~sum & 0xffff); 425 return (1); 426 } 427#ifdef INET6 428 case (IPV6_VERSION >> 4): 429 { 430 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 431 u_int32_t flowlabel; 432 433 flowlabel = ntohl(ip6->ip6_flow); 434 if ((flowlabel >> 28) != 6) 435 return (0); /* version mismatch! */ 436 if ((flowlabel & (IPTOS_ECN_MASK << 20)) == 437 (IPTOS_ECN_NOTECT << 20)) 438 return (0); /* not-ECT */ 439 if ((flowlabel & (IPTOS_ECN_MASK << 20)) == 440 (IPTOS_ECN_CE << 20)) 441 return (1); /* already marked */ 442 /* 443 * ecn-capable but not marked, mark CE 444 */ 445 flowlabel |= (IPTOS_ECN_CE << 20); 446 ip6->ip6_flow = htonl(flowlabel); 447 return (1); 448 } 449#endif 450 } 451 return (0); 452} 453 454/* 455 * Enqueue a packet in q, subject to space and queue management policy 456 * (whose parameters are in q->fs). 457 * Update stats for the queue and the scheduler. 458 * Return 0 on success, 1 on drop. The packet is consumed anyways. 459 */ 460int 461dn_enqueue(struct dn_queue *q, struct mbuf* m, int drop) 462{ 463 struct dn_fs *f; 464 struct dn_flow *ni; /* stats for scheduler instance */ 465 uint64_t len; 466 467 if (q->fs == NULL || q->_si == NULL) { 468 printf("%s fs %p si %p, dropping\n", 469 __FUNCTION__, q->fs, q->_si); 470 FREE_PKT(m); 471 return 1; 472 } 473 f = &(q->fs->fs); 474 ni = &q->_si->ni; 475 len = m->m_pkthdr.len; 476 /* Update statistics, then check reasons to drop pkt. */ 477 q->ni.tot_bytes += len; 478 q->ni.tot_pkts++; 479 ni->tot_bytes += len; 480 ni->tot_pkts++; 481 if (drop) 482 goto drop; 483 if (f->plr && random() < f->plr) 484 goto drop; 485 if (f->flags & DN_IS_RED && red_drops(q, m->m_pkthdr.len)) { 486 if (!(f->flags & DN_IS_ECN) || !ecn_mark(m)) 487 goto drop; 488 } 489 if (f->flags & DN_QSIZE_BYTES) { 490 if (q->ni.len_bytes > f->qsize) 491 goto drop; 492 } else if (q->ni.length >= f->qsize) { 493 goto drop; 494 } 495 mq_append(&q->mq, m); 496 q->ni.length++; 497 q->ni.len_bytes += len; 498 ni->length++; 499 ni->len_bytes += len; 500 return (0); 501 502drop: 503 io_pkt_drop++; 504 q->ni.drops++; 505 ni->drops++; 506 FREE_PKT(m); 507 return (1); 508} 509 510/* 511 * Fetch packets from the delay line which are due now. If there are 512 * leftover packets, reinsert the delay line in the heap. 513 * Runs under scheduler lock. 514 */ 515static void 516transmit_event(struct mq *q, struct delay_line *dline, uint64_t now) 517{ 518 struct mbuf *m; 519 struct dn_pkt_tag *pkt = NULL; 520 521 dline->oid.subtype = 0; /* not in heap */ 522 while ((m = dline->mq.head) != NULL) { 523 pkt = dn_tag_get(m); 524 if (!DN_KEY_LEQ(pkt->output_time, now)) 525 break; 526 dline->mq.head = m->m_nextpkt; 527 mq_append(q, m); 528 } 529 if (m != NULL) { 530 dline->oid.subtype = 1; /* in heap */ 531 heap_insert(&dn_cfg.evheap, pkt->output_time, dline); 532 } 533} 534 535/* 536 * Convert the additional MAC overheads/delays into an equivalent 537 * number of bits for the given data rate. The samples are 538 * in milliseconds so we need to divide by 1000. 539 */ 540static uint64_t 541extra_bits(struct mbuf *m, struct dn_schk *s) 542{ 543 int index; 544 uint64_t bits; 545 struct dn_profile *pf = s->profile; 546 547 if (!pf || pf->samples_no == 0) 548 return 0; 549 index = random() % pf->samples_no; 550 bits = div64((uint64_t)pf->samples[index] * s->link.bandwidth, 1000); 551 if (index >= pf->loss_level) { 552 struct dn_pkt_tag *dt = dn_tag_get(m); 553 if (dt) 554 dt->dn_dir = DIR_DROP; 555 } 556 return bits; 557} 558 559/* 560 * Send traffic from a scheduler instance due by 'now'. 561 * Return a pointer to the head of the queue. 562 */ 563static struct mbuf * 564serve_sched(struct mq *q, struct dn_sch_inst *si, uint64_t now) 565{ 566 struct mq def_q; 567 struct dn_schk *s = si->sched; 568 struct mbuf *m = NULL; 569 int delay_line_idle = (si->dline.mq.head == NULL); 570 int done, bw; 571 572 if (q == NULL) { 573 q = &def_q; 574 q->head = NULL; 575 } 576 577 bw = s->link.bandwidth; 578 si->kflags &= ~DN_ACTIVE; 579 580 if (bw > 0) 581 si->credit += (now - si->sched_time) * bw; 582 else 583 si->credit = 0; 584 si->sched_time = now; 585 done = 0; 586 while (si->credit >= 0 && (m = s->fp->dequeue(si)) != NULL) { 587 uint64_t len_scaled; 588 589 done++; 590 len_scaled = (bw == 0) ? 0 : hz * 591 (m->m_pkthdr.len * 8 + extra_bits(m, s)); 592 si->credit -= len_scaled; 593 /* Move packet in the delay line */ 594 dn_tag_get(m)->output_time = dn_cfg.curr_time + s->link.delay ; 595 mq_append(&si->dline.mq, m); 596 } 597 598 /* 599 * If credit >= 0 the instance is idle, mark time. 600 * Otherwise put back in the heap, and adjust the output 601 * time of the last inserted packet, m, which was too early. 602 */ 603 if (si->credit >= 0) { 604 si->idle_time = now; 605 } else { 606 uint64_t t; 607 KASSERT (bw > 0, ("bw=0 and credit<0 ?")); 608 t = div64(bw - 1 - si->credit, bw); 609 if (m) 610 dn_tag_get(m)->output_time += t; 611 si->kflags |= DN_ACTIVE; 612 heap_insert(&dn_cfg.evheap, now + t, si); 613 } 614 if (delay_line_idle && done) 615 transmit_event(q, &si->dline, now); 616 return q->head; 617} 618 619/* 620 * The timer handler for dummynet. Time is computed in ticks, but 621 * but the code is tolerant to the actual rate at which this is called. 622 * Once complete, the function reschedules itself for the next tick. 623 */ 624void 625dummynet_task(void *context, int pending) 626{ 627 struct timeval t; 628 struct mq q = { NULL, NULL }; /* queue to accumulate results */ 629 630 CURVNET_SET((struct vnet *)context); 631 632 DN_BH_WLOCK(); 633 634 /* Update number of lost(coalesced) ticks. */ 635 tick_lost += pending - 1; 636 637 getmicrouptime(&t); 638 /* Last tick duration (usec). */ 639 tick_last = (t.tv_sec - dn_cfg.prev_t.tv_sec) * 1000000 + 640 (t.tv_usec - dn_cfg.prev_t.tv_usec); 641 /* Last tick vs standard tick difference (usec). */ 642 tick_delta = (tick_last * hz - 1000000) / hz; 643 /* Accumulated tick difference (usec). */ 644 tick_delta_sum += tick_delta; 645 646 dn_cfg.prev_t = t; 647 648 /* 649 * Adjust curr_time if the accumulated tick difference is 650 * greater than the 'standard' tick. Since curr_time should 651 * be monotonically increasing, we do positive adjustments 652 * as required, and throttle curr_time in case of negative 653 * adjustment. 654 */ 655 dn_cfg.curr_time++; 656 if (tick_delta_sum - tick >= 0) { 657 int diff = tick_delta_sum / tick; 658 659 dn_cfg.curr_time += diff; 660 tick_diff += diff; 661 tick_delta_sum %= tick; 662 tick_adjustment++; 663 } else if (tick_delta_sum + tick <= 0) { 664 dn_cfg.curr_time--; 665 tick_diff--; 666 tick_delta_sum += tick; 667 tick_adjustment++; 668 } 669 670 /* serve pending events, accumulate in q */ 671 for (;;) { 672 struct dn_id *p; /* generic parameter to handler */ 673 674 if (dn_cfg.evheap.elements == 0 || 675 DN_KEY_LT(dn_cfg.curr_time, HEAP_TOP(&dn_cfg.evheap)->key)) 676 break; 677 p = HEAP_TOP(&dn_cfg.evheap)->object; 678 heap_extract(&dn_cfg.evheap, NULL); 679 680 if (p->type == DN_SCH_I) { 681 serve_sched(&q, (struct dn_sch_inst *)p, dn_cfg.curr_time); 682 } else { /* extracted a delay line */ 683 transmit_event(&q, (struct delay_line *)p, dn_cfg.curr_time); 684 } 685 } 686 if (dn_cfg.expire && ++dn_cfg.expire_cycle >= dn_cfg.expire) { 687 dn_cfg.expire_cycle = 0; 688 dn_drain_scheduler(); 689 dn_drain_queue(); 690 } 691 692 dn_reschedule(); 693 DN_BH_WUNLOCK(); 694 if (q.head != NULL) 695 dummynet_send(q.head); 696 CURVNET_RESTORE(); 697} 698 699/* 700 * forward a chain of packets to the proper destination. 701 * This runs outside the dummynet lock. 702 */ 703static void 704dummynet_send(struct mbuf *m) 705{ 706 struct mbuf *n; 707 708 for (; m != NULL; m = n) { 709 struct ifnet *ifp = NULL; /* gcc 3.4.6 complains */ 710 struct m_tag *tag; 711 int dst; 712 713 n = m->m_nextpkt; 714 m->m_nextpkt = NULL; 715 tag = m_tag_first(m); 716 if (tag == NULL) { /* should not happen */ 717 dst = DIR_DROP; 718 } else { 719 struct dn_pkt_tag *pkt = dn_tag_get(m); 720 /* extract the dummynet info, rename the tag 721 * to carry reinject info. 722 */ 723 dst = pkt->dn_dir; 724 ifp = pkt->ifp; 725 tag->m_tag_cookie = MTAG_IPFW_RULE; 726 tag->m_tag_id = 0; 727 } 728 729 switch (dst) { 730 case DIR_OUT: 731 ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL); 732 break ; 733 734 case DIR_IN : 735 netisr_dispatch(NETISR_IP, m); 736 break; 737 738#ifdef INET6 739 case DIR_IN | PROTO_IPV6: 740 netisr_dispatch(NETISR_IPV6, m); 741 break; 742 743 case DIR_OUT | PROTO_IPV6: 744 ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL); 745 break; 746#endif 747 748 case DIR_FWD | PROTO_IFB: /* DN_TO_IFB_FWD: */ 749 if (bridge_dn_p != NULL) 750 ((*bridge_dn_p)(m, ifp)); 751 else 752 printf("dummynet: if_bridge not loaded\n"); 753 754 break; 755 756 case DIR_IN | PROTO_LAYER2: /* DN_TO_ETH_DEMUX: */ 757 /* 758 * The Ethernet code assumes the Ethernet header is 759 * contiguous in the first mbuf header. 760 * Insure this is true. 761 */ 762 if (m->m_len < ETHER_HDR_LEN && 763 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) { 764 printf("dummynet/ether: pullup failed, " 765 "dropping packet\n"); 766 break; 767 } 768 ether_demux(m->m_pkthdr.rcvif, m); 769 break; 770 771 case DIR_OUT | PROTO_LAYER2: /* N_TO_ETH_OUT: */ 772 ether_output_frame(ifp, m); 773 break; 774 775 case DIR_DROP: 776 /* drop the packet after some time */ 777 FREE_PKT(m); 778 break; 779 780 default: 781 printf("dummynet: bad switch %d!\n", dst); 782 FREE_PKT(m); 783 break; 784 } 785 } 786} 787 788static inline int 789tag_mbuf(struct mbuf *m, int dir, struct ip_fw_args *fwa) 790{ 791 struct dn_pkt_tag *dt; 792 struct m_tag *mtag; 793 794 mtag = m_tag_get(PACKET_TAG_DUMMYNET, 795 sizeof(*dt), M_NOWAIT | M_ZERO); 796 if (mtag == NULL) 797 return 1; /* Cannot allocate packet header. */ 798 m_tag_prepend(m, mtag); /* Attach to mbuf chain. */ 799 dt = (struct dn_pkt_tag *)(mtag + 1); 800 dt->rule = fwa->rule; 801 dt->rule.info &= IPFW_ONEPASS; /* only keep this info */ 802 dt->dn_dir = dir; 803 dt->ifp = fwa->oif; 804 /* dt->output tame is updated as we move through */ 805 dt->output_time = dn_cfg.curr_time; 806 return 0; 807} 808 809 810/* 811 * dummynet hook for packets. 812 * We use the argument to locate the flowset fs and the sched_set sch 813 * associated to it. The we apply flow_mask and sched_mask to 814 * determine the queue and scheduler instances. 815 * 816 * dir where shall we send the packet after dummynet. 817 * *m0 the mbuf with the packet 818 * ifp the 'ifp' parameter from the caller. 819 * NULL in ip_input, destination interface in ip_output, 820 */ 821int 822dummynet_io(struct mbuf **m0, int dir, struct ip_fw_args *fwa) 823{ 824 struct mbuf *m = *m0; 825 struct dn_fsk *fs = NULL; 826 struct dn_sch_inst *si; 827 struct dn_queue *q = NULL; /* default */ 828 829 int fs_id = (fwa->rule.info & IPFW_INFO_MASK) + 830 ((fwa->rule.info & IPFW_IS_PIPE) ? 2*DN_MAX_ID : 0); 831 DN_BH_WLOCK(); 832 io_pkt++; 833 /* we could actually tag outside the lock, but who cares... */ 834 if (tag_mbuf(m, dir, fwa)) 835 goto dropit; 836 if (dn_cfg.busy) { 837 /* if the upper half is busy doing something expensive, 838 * lets queue the packet and move forward 839 */ 840 mq_append(&dn_cfg.pending, m); 841 m = *m0 = NULL; /* consumed */ 842 goto done; /* already active, nothing to do */ 843 } 844 /* XXX locate_flowset could be optimised with a direct ref. */ 845 fs = dn_ht_find(dn_cfg.fshash, fs_id, 0, NULL); 846 if (fs == NULL) 847 goto dropit; /* This queue/pipe does not exist! */ 848 if (fs->sched == NULL) /* should not happen */ 849 goto dropit; 850 /* find scheduler instance, possibly applying sched_mask */ 851 si = ipdn_si_find(fs->sched, &(fwa->f_id)); 852 if (si == NULL) 853 goto dropit; 854 /* 855 * If the scheduler supports multiple queues, find the right one 856 * (otherwise it will be ignored by enqueue). 857 */ 858 if (fs->sched->fp->flags & DN_MULTIQUEUE) { 859 q = ipdn_q_find(fs, si, &(fwa->f_id)); 860 if (q == NULL) 861 goto dropit; 862 } 863 if (fs->sched->fp->enqueue(si, q, m)) { 864 /* packet was dropped by enqueue() */ 865 m = *m0 = NULL; 866 goto dropit; 867 } 868 869 if (si->kflags & DN_ACTIVE) { 870 m = *m0 = NULL; /* consumed */ 871 goto done; /* already active, nothing to do */ 872 } 873 874 /* compute the initial allowance */ 875 if (si->idle_time < dn_cfg.curr_time) { 876 /* Do this only on the first packet on an idle pipe */ 877 struct dn_link *p = &fs->sched->link; 878 879 si->sched_time = dn_cfg.curr_time; 880 si->credit = dn_cfg.io_fast ? p->bandwidth : 0; 881 if (p->burst) { 882 uint64_t burst = (dn_cfg.curr_time - si->idle_time) * p->bandwidth; 883 if (burst > p->burst) 884 burst = p->burst; 885 si->credit += burst; 886 } 887 } 888 /* pass through scheduler and delay line */ 889 m = serve_sched(NULL, si, dn_cfg.curr_time); 890 891 /* optimization -- pass it back to ipfw for immediate send */ 892 /* XXX Don't call dummynet_send() if scheduler return the packet 893 * just enqueued. This avoid a lock order reversal. 894 * 895 */ 896 if (/*dn_cfg.io_fast &&*/ m == *m0 && (dir & PROTO_LAYER2) == 0 ) { 897 /* fast io, rename the tag * to carry reinject info. */ 898 struct m_tag *tag = m_tag_first(m); 899 900 tag->m_tag_cookie = MTAG_IPFW_RULE; 901 tag->m_tag_id = 0; 902 io_pkt_fast++; 903 if (m->m_nextpkt != NULL) { 904 printf("dummynet: fast io: pkt chain detected!\n"); 905 m->m_nextpkt = NULL; 906 } 907 m = NULL; 908 } else { 909 *m0 = NULL; 910 } 911done: 912 DN_BH_WUNLOCK(); 913 if (m) 914 dummynet_send(m); 915 return 0; 916 917dropit: 918 io_pkt_drop++; 919 DN_BH_WUNLOCK(); 920 if (m) 921 FREE_PKT(m); 922 *m0 = NULL; 923 return (fs && (fs->fs.flags & DN_NOERROR)) ? 0 : ENOBUFS; 924} 925