1204591Sluigi/*- 2204591Sluigi * Copyright (c) 2010 Luigi Rizzo, Riccardo Panicucci, Universita` di Pisa 3204591Sluigi * All rights reserved 4204591Sluigi * 5204591Sluigi * Redistribution and use in source and binary forms, with or without 6204591Sluigi * modification, are permitted provided that the following conditions 7204591Sluigi * are met: 8204591Sluigi * 1. Redistributions of source code must retain the above copyright 9204591Sluigi * notice, this list of conditions and the following disclaimer. 10204591Sluigi * 2. Redistributions in binary form must reproduce the above copyright 11204591Sluigi * notice, this list of conditions and the following disclaimer in the 12204591Sluigi * documentation and/or other materials provided with the distribution. 13204591Sluigi * 14204591Sluigi * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15204591Sluigi * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16204591Sluigi * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17204591Sluigi * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18204591Sluigi * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19204591Sluigi * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20204591Sluigi * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21204591Sluigi * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22204591Sluigi * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23204591Sluigi * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24204591Sluigi * SUCH DAMAGE. 25204591Sluigi */ 26204591Sluigi 27204591Sluigi/* 28204591Sluigi * Dummynet portions related to packet handling. 29204591Sluigi */ 30204591Sluigi#include <sys/cdefs.h> 31204591Sluigi__FBSDID("$FreeBSD$"); 32204591Sluigi 33204591Sluigi#include "opt_inet6.h" 34204591Sluigi 35204591Sluigi#include <sys/param.h> 36204591Sluigi#include <sys/systm.h> 37204591Sluigi#include <sys/malloc.h> 38204591Sluigi#include <sys/mbuf.h> 39204591Sluigi#include <sys/kernel.h> 40204591Sluigi#include <sys/lock.h> 41204591Sluigi#include <sys/module.h> 42204591Sluigi#include <sys/priv.h> 43204591Sluigi#include <sys/proc.h> 44204591Sluigi#include <sys/rwlock.h> 45204591Sluigi#include <sys/socket.h> 46204591Sluigi#include <sys/time.h> 47204591Sluigi#include <sys/sysctl.h> 48206461Sbz 49204591Sluigi#include <net/if.h> /* IFNAMSIZ, struct ifaddr, ifq head, lock.h mutex.h */ 50204591Sluigi#include <net/netisr.h> 51206461Sbz#include <net/vnet.h> 52206461Sbz 53204591Sluigi#include <netinet/in.h> 54204591Sluigi#include <netinet/ip.h> /* ip_len, ip_off */ 55204591Sluigi#include <netinet/ip_var.h> /* ip_output(), IP_FORWARDING */ 56204591Sluigi#include <netinet/ip_fw.h> 57204591Sluigi#include <netinet/ip_dummynet.h> 58204591Sluigi#include <netinet/if_ether.h> /* various ether_* routines */ 59204591Sluigi#include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */ 60204591Sluigi#include <netinet6/ip6_var.h> 61204591Sluigi 62240494Sglebius#include <netpfil/ipfw/ip_fw_private.h> 63240494Sglebius#include <netpfil/ipfw/dn_heap.h> 64240494Sglebius#include <netpfil/ipfw/ip_dn_private.h> 65240494Sglebius#include <netpfil/ipfw/dn_sched.h> 66240494Sglebius 67204591Sluigi/* 68204591Sluigi * We keep a private variable for the simulation time, but we could 69204591Sluigi * probably use an existing one ("softticks" in sys/kern/kern_timeout.c) 70204591Sluigi * instead of dn_cfg.curr_time 71204591Sluigi */ 72204591Sluigi 73204591Sluigistruct dn_parms dn_cfg; 74206428Sluigi//VNET_DEFINE(struct dn_parms, _base_dn_cfg); 75204591Sluigi 76204591Sluigistatic long tick_last; /* Last tick duration (usec). */ 77204591Sluigistatic long tick_delta; /* Last vs standard tick diff (usec). */ 78204591Sluigistatic long tick_delta_sum; /* Accumulated tick difference (usec).*/ 79204591Sluigistatic long tick_adjustment; /* Tick adjustments done. */ 80204591Sluigistatic long tick_lost; /* Lost(coalesced) ticks number. */ 81204591Sluigi/* Adjusted vs non-adjusted curr_time difference (ticks). */ 82204591Sluigistatic long tick_diff; 83204591Sluigi 84204591Sluigistatic unsigned long io_pkt; 85204591Sluigistatic unsigned long io_pkt_fast; 86204591Sluigistatic unsigned long io_pkt_drop; 87204591Sluigi 88204591Sluigi/* 89204591Sluigi * We use a heap to store entities for which we have pending timer events. 90204591Sluigi * The heap is checked at every tick and all entities with expired events 91204591Sluigi * are extracted. 92204591Sluigi */ 93204591Sluigi 94204591SluigiMALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap"); 95204591Sluigi 96204591Sluigiextern void (*bridge_dn_p)(struct mbuf *, struct ifnet *); 97204591Sluigi 98204591Sluigi#ifdef SYSCTL_NODE 99204591Sluigi 100239093Sluigi/* 101239093Sluigi * Because of the way the SYSBEGIN/SYSEND macros work on other 102239093Sluigi * platforms, there should not be functions between them. 103239093Sluigi * So keep the handlers outside the block. 104239093Sluigi */ 105220832Saestatic int 106220832Saesysctl_hash_size(SYSCTL_HANDLER_ARGS) 107220832Sae{ 108220832Sae int error, value; 109220832Sae 110220832Sae value = dn_cfg.hash_size; 111220832Sae error = sysctl_handle_int(oidp, &value, 0, req); 112220832Sae if (error != 0 || req->newptr == NULL) 113220832Sae return (error); 114220832Sae if (value < 16 || value > 65536) 115220832Sae return (EINVAL); 116220832Sae dn_cfg.hash_size = value; 117220832Sae return (0); 118220832Sae} 119220832Sae 120220832Saestatic int 121220832Saesysctl_limits(SYSCTL_HANDLER_ARGS) 122220832Sae{ 123220832Sae int error; 124220832Sae long value; 125220832Sae 126220832Sae if (arg2 != 0) 127220832Sae value = dn_cfg.slot_limit; 128220832Sae else 129220832Sae value = dn_cfg.byte_limit; 130220832Sae error = sysctl_handle_long(oidp, &value, 0, req); 131220832Sae 132220832Sae if (error != 0 || req->newptr == NULL) 133220832Sae return (error); 134220832Sae if (arg2 != 0) { 135220832Sae if (value < 1) 136220832Sae return (EINVAL); 137220832Sae dn_cfg.slot_limit = value; 138220832Sae } else { 139220832Sae if (value < 1500) 140220832Sae return (EINVAL); 141220832Sae dn_cfg.byte_limit = value; 142220832Sae } 143220832Sae return (0); 144220832Sae} 145220832Sae 146239093SluigiSYSBEGIN(f4) 147239093Sluigi 148239093SluigiSYSCTL_DECL(_net_inet); 149239093SluigiSYSCTL_DECL(_net_inet_ip); 150239093Sluigistatic SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW, 0, "Dummynet"); 151239093Sluigi 152239093Sluigi/* wrapper to pass dn_cfg fields to SYSCTL_* */ 153239093Sluigi//#define DC(x) (&(VNET_NAME(_base_dn_cfg).x)) 154239093Sluigi#define DC(x) (&(dn_cfg.x)) 155239093Sluigi/* parameters */ 156239093Sluigi 157239093Sluigi 158239093SluigiSYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hash_size, 159239093Sluigi CTLTYPE_INT | CTLFLAG_RW, 0, 0, sysctl_hash_size, 160239093Sluigi "I", "Default hash table size"); 161239093Sluigi 162239093Sluigi 163220832SaeSYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit, 164220832Sae CTLTYPE_LONG | CTLFLAG_RW, 0, 1, sysctl_limits, 165220832Sae "L", "Upper limit in slots for pipe queue."); 166220832SaeSYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit, 167220832Sae CTLTYPE_LONG | CTLFLAG_RW, 0, 0, sysctl_limits, 168220832Sae "L", "Upper limit in bytes for pipe queue."); 169204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast, 170206428Sluigi CTLFLAG_RW, DC(io_fast), 0, "Enable fast dummynet io."); 171204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, 172206428Sluigi CTLFLAG_RW, DC(debug), 0, "Dummynet debug level"); 173204591Sluigi 174204591Sluigi/* RED parameters */ 175204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth, 176206428Sluigi CTLFLAG_RD, DC(red_lookup_depth), 0, "Depth of RED lookup table"); 177204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size, 178206428Sluigi CTLFLAG_RD, DC(red_avg_pkt_size), 0, "RED Medium packet size"); 179204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size, 180206428Sluigi CTLFLAG_RD, DC(red_max_pkt_size), 0, "RED Max packet size"); 181204591Sluigi 182204591Sluigi/* time adjustment */ 183204591SluigiSYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta, 184204591Sluigi CTLFLAG_RD, &tick_delta, 0, "Last vs standard tick difference (usec)."); 185204591SluigiSYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum, 186204591Sluigi CTLFLAG_RD, &tick_delta_sum, 0, "Accumulated tick difference (usec)."); 187204591SluigiSYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment, 188204591Sluigi CTLFLAG_RD, &tick_adjustment, 0, "Tick adjustments done."); 189204591SluigiSYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff, 190204591Sluigi CTLFLAG_RD, &tick_diff, 0, 191204591Sluigi "Adjusted vs non-adjusted curr_time difference (ticks)."); 192204591SluigiSYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost, 193204591Sluigi CTLFLAG_RD, &tick_lost, 0, 194204591Sluigi "Number of ticks coalesced by dummynet taskqueue."); 195204591Sluigi 196213267Sluigi/* Drain parameters */ 197217322SmdfSYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire, 198213267Sluigi CTLFLAG_RW, DC(expire), 0, "Expire empty queues/pipes"); 199217322SmdfSYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_cycle, 200213267Sluigi CTLFLAG_RD, DC(expire_cycle), 0, "Expire cycle for queues/pipes"); 201213267Sluigi 202204591Sluigi/* statistics */ 203204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, schk_count, 204206428Sluigi CTLFLAG_RD, DC(schk_count), 0, "Number of schedulers"); 205204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, si_count, 206206428Sluigi CTLFLAG_RD, DC(si_count), 0, "Number of scheduler instances"); 207204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, fsk_count, 208206428Sluigi CTLFLAG_RD, DC(fsk_count), 0, "Number of flowsets"); 209204591SluigiSYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, queue_count, 210206428Sluigi CTLFLAG_RD, DC(queue_count), 0, "Number of queues"); 211204591SluigiSYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt, 212204591Sluigi CTLFLAG_RD, &io_pkt, 0, 213204591Sluigi "Number of packets passed to dummynet."); 214204591SluigiSYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast, 215204591Sluigi CTLFLAG_RD, &io_pkt_fast, 0, 216204591Sluigi "Number of packets bypassed dummynet scheduler."); 217204591SluigiSYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop, 218204591Sluigi CTLFLAG_RD, &io_pkt_drop, 0, 219204591Sluigi "Number of packets dropped by dummynet."); 220206428Sluigi#undef DC 221204591SluigiSYSEND 222204591Sluigi 223204591Sluigi#endif 224204591Sluigi 225204591Sluigistatic void dummynet_send(struct mbuf *); 226204591Sluigi 227204591Sluigi/* 228204591Sluigi * Packets processed by dummynet have an mbuf tag associated with 229204591Sluigi * them that carries their dummynet state. 230204591Sluigi * Outside dummynet, only the 'rule' field is relevant, and it must 231204591Sluigi * be at the beginning of the structure. 232204591Sluigi */ 233204591Sluigistruct dn_pkt_tag { 234204591Sluigi struct ipfw_rule_ref rule; /* matching rule */ 235204591Sluigi 236204591Sluigi /* second part, dummynet specific */ 237204591Sluigi int dn_dir; /* action when packet comes out.*/ 238204591Sluigi /* see ip_fw_private.h */ 239204591Sluigi uint64_t output_time; /* when the pkt is due for delivery*/ 240204591Sluigi struct ifnet *ifp; /* interface, for ip_output */ 241204591Sluigi struct _ip6dn_args ip6opt; /* XXX ipv6 options */ 242204591Sluigi}; 243204591Sluigi 244204591Sluigi/* 245204591Sluigi * Return the mbuf tag holding the dummynet state (it should 246204591Sluigi * be the first one on the list). 247204591Sluigi */ 248204591Sluigistatic struct dn_pkt_tag * 249204591Sluigidn_tag_get(struct mbuf *m) 250204591Sluigi{ 251204591Sluigi struct m_tag *mtag = m_tag_first(m); 252204591Sluigi KASSERT(mtag != NULL && 253204591Sluigi mtag->m_tag_cookie == MTAG_ABI_COMPAT && 254204591Sluigi mtag->m_tag_id == PACKET_TAG_DUMMYNET, 255204591Sluigi ("packet on dummynet queue w/o dummynet tag!")); 256204591Sluigi return (struct dn_pkt_tag *)(mtag+1); 257204591Sluigi} 258204591Sluigi 259204591Sluigistatic inline void 260204591Sluigimq_append(struct mq *q, struct mbuf *m) 261204591Sluigi{ 262204591Sluigi if (q->head == NULL) 263204591Sluigi q->head = m; 264204591Sluigi else 265204591Sluigi q->tail->m_nextpkt = m; 266204591Sluigi q->tail = m; 267204591Sluigi m->m_nextpkt = NULL; 268204591Sluigi} 269204591Sluigi 270204591Sluigi/* 271204591Sluigi * Dispose a list of packet. Use a functions so if we need to do 272204591Sluigi * more work, this is a central point to do it. 273204591Sluigi */ 274204591Sluigivoid dn_free_pkts(struct mbuf *mnext) 275204591Sluigi{ 276204591Sluigi struct mbuf *m; 277204591Sluigi 278204591Sluigi while ((m = mnext) != NULL) { 279204591Sluigi mnext = m->m_nextpkt; 280204591Sluigi FREE_PKT(m); 281204591Sluigi } 282204591Sluigi} 283204591Sluigi 284204591Sluigistatic int 285204591Sluigired_drops (struct dn_queue *q, int len) 286204591Sluigi{ 287204591Sluigi /* 288204591Sluigi * RED algorithm 289204591Sluigi * 290204591Sluigi * RED calculates the average queue size (avg) using a low-pass filter 291204591Sluigi * with an exponential weighted (w_q) moving average: 292204591Sluigi * avg <- (1-w_q) * avg + w_q * q_size 293204591Sluigi * where q_size is the queue length (measured in bytes or * packets). 294204591Sluigi * 295204591Sluigi * If q_size == 0, we compute the idle time for the link, and set 296204591Sluigi * avg = (1 - w_q)^(idle/s) 297204591Sluigi * where s is the time needed for transmitting a medium-sized packet. 298204591Sluigi * 299204591Sluigi * Now, if avg < min_th the packet is enqueued. 300204591Sluigi * If avg > max_th the packet is dropped. Otherwise, the packet is 301204591Sluigi * dropped with probability P function of avg. 302204591Sluigi */ 303204591Sluigi 304204591Sluigi struct dn_fsk *fs = q->fs; 305204591Sluigi int64_t p_b = 0; 306204591Sluigi 307204591Sluigi /* Queue in bytes or packets? */ 308204591Sluigi uint32_t q_size = (fs->fs.flags & DN_QSIZE_BYTES) ? 309204591Sluigi q->ni.len_bytes : q->ni.length; 310204591Sluigi 311204591Sluigi /* Average queue size estimation. */ 312204591Sluigi if (q_size != 0) { 313204591Sluigi /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */ 314204591Sluigi int diff = SCALE(q_size) - q->avg; 315204591Sluigi int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q); 316204591Sluigi 317204591Sluigi q->avg += (int)v; 318204591Sluigi } else { 319204591Sluigi /* 320204591Sluigi * Queue is empty, find for how long the queue has been 321204591Sluigi * empty and use a lookup table for computing 322204591Sluigi * (1 - * w_q)^(idle_time/s) where s is the time to send a 323204591Sluigi * (small) packet. 324204591Sluigi * XXX check wraps... 325204591Sluigi */ 326204591Sluigi if (q->avg) { 327204591Sluigi u_int t = div64((dn_cfg.curr_time - q->q_time), fs->lookup_step); 328204591Sluigi 329204591Sluigi q->avg = (t < fs->lookup_depth) ? 330204591Sluigi SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0; 331204591Sluigi } 332204591Sluigi } 333204591Sluigi 334204591Sluigi /* Should i drop? */ 335204591Sluigi if (q->avg < fs->min_th) { 336204591Sluigi q->count = -1; 337204591Sluigi return (0); /* accept packet */ 338204591Sluigi } 339204591Sluigi if (q->avg >= fs->max_th) { /* average queue >= max threshold */ 340204591Sluigi if (fs->fs.flags & DN_IS_GENTLE_RED) { 341204591Sluigi /* 342204591Sluigi * According to Gentle-RED, if avg is greater than 343204591Sluigi * max_th the packet is dropped with a probability 344204591Sluigi * p_b = c_3 * avg - c_4 345204591Sluigi * where c_3 = (1 - max_p) / max_th 346204591Sluigi * c_4 = 1 - 2 * max_p 347204591Sluigi */ 348204591Sluigi p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) - 349204591Sluigi fs->c_4; 350204591Sluigi } else { 351204591Sluigi q->count = -1; 352204591Sluigi return (1); 353204591Sluigi } 354204591Sluigi } else if (q->avg > fs->min_th) { 355204591Sluigi /* 356204591Sluigi * We compute p_b using the linear dropping function 357204591Sluigi * p_b = c_1 * avg - c_2 358204591Sluigi * where c_1 = max_p / (max_th - min_th) 359204591Sluigi * c_2 = max_p * min_th / (max_th - min_th) 360204591Sluigi */ 361204591Sluigi p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2; 362204591Sluigi } 363204591Sluigi 364204591Sluigi if (fs->fs.flags & DN_QSIZE_BYTES) 365204591Sluigi p_b = div64((p_b * len) , fs->max_pkt_size); 366204591Sluigi if (++q->count == 0) 367204591Sluigi q->random = random() & 0xffff; 368204591Sluigi else { 369204591Sluigi /* 370204591Sluigi * q->count counts packets arrived since last drop, so a greater 371204591Sluigi * value of q->count means a greater packet drop probability. 372204591Sluigi */ 373204591Sluigi if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) { 374204591Sluigi q->count = 0; 375204591Sluigi /* After a drop we calculate a new random value. */ 376204591Sluigi q->random = random() & 0xffff; 377204591Sluigi return (1); /* drop */ 378204591Sluigi } 379204591Sluigi } 380204591Sluigi /* End of RED algorithm. */ 381204591Sluigi 382204591Sluigi return (0); /* accept */ 383204591Sluigi 384204591Sluigi} 385204591Sluigi 386204591Sluigi/* 387204591Sluigi * Enqueue a packet in q, subject to space and queue management policy 388204591Sluigi * (whose parameters are in q->fs). 389204591Sluigi * Update stats for the queue and the scheduler. 390204591Sluigi * Return 0 on success, 1 on drop. The packet is consumed anyways. 391204591Sluigi */ 392204591Sluigiint 393204591Sluigidn_enqueue(struct dn_queue *q, struct mbuf* m, int drop) 394204591Sluigi{ 395204591Sluigi struct dn_fs *f; 396204591Sluigi struct dn_flow *ni; /* stats for scheduler instance */ 397204591Sluigi uint64_t len; 398204591Sluigi 399204591Sluigi if (q->fs == NULL || q->_si == NULL) { 400204591Sluigi printf("%s fs %p si %p, dropping\n", 401204591Sluigi __FUNCTION__, q->fs, q->_si); 402204591Sluigi FREE_PKT(m); 403204591Sluigi return 1; 404204591Sluigi } 405204591Sluigi f = &(q->fs->fs); 406204591Sluigi ni = &q->_si->ni; 407204591Sluigi len = m->m_pkthdr.len; 408204591Sluigi /* Update statistics, then check reasons to drop pkt. */ 409204591Sluigi q->ni.tot_bytes += len; 410204591Sluigi q->ni.tot_pkts++; 411204591Sluigi ni->tot_bytes += len; 412204591Sluigi ni->tot_pkts++; 413204591Sluigi if (drop) 414204591Sluigi goto drop; 415204591Sluigi if (f->plr && random() < f->plr) 416204591Sluigi goto drop; 417204591Sluigi if (f->flags & DN_IS_RED && red_drops(q, m->m_pkthdr.len)) 418204591Sluigi goto drop; 419204591Sluigi if (f->flags & DN_QSIZE_BYTES) { 420204591Sluigi if (q->ni.len_bytes > f->qsize) 421204591Sluigi goto drop; 422204591Sluigi } else if (q->ni.length >= f->qsize) { 423204591Sluigi goto drop; 424204591Sluigi } 425204591Sluigi mq_append(&q->mq, m); 426204591Sluigi q->ni.length++; 427204591Sluigi q->ni.len_bytes += len; 428204591Sluigi ni->length++; 429204591Sluigi ni->len_bytes += len; 430204591Sluigi return 0; 431204591Sluigi 432204591Sluigidrop: 433204591Sluigi io_pkt_drop++; 434204591Sluigi q->ni.drops++; 435204591Sluigi ni->drops++; 436204591Sluigi FREE_PKT(m); 437204591Sluigi return 1; 438204591Sluigi} 439204591Sluigi 440204591Sluigi/* 441204591Sluigi * Fetch packets from the delay line which are due now. If there are 442204591Sluigi * leftover packets, reinsert the delay line in the heap. 443204591Sluigi * Runs under scheduler lock. 444204591Sluigi */ 445204591Sluigistatic void 446204591Sluigitransmit_event(struct mq *q, struct delay_line *dline, uint64_t now) 447204591Sluigi{ 448204591Sluigi struct mbuf *m; 449204591Sluigi struct dn_pkt_tag *pkt = NULL; 450204591Sluigi 451204591Sluigi dline->oid.subtype = 0; /* not in heap */ 452204591Sluigi while ((m = dline->mq.head) != NULL) { 453204591Sluigi pkt = dn_tag_get(m); 454204591Sluigi if (!DN_KEY_LEQ(pkt->output_time, now)) 455204591Sluigi break; 456204591Sluigi dline->mq.head = m->m_nextpkt; 457204591Sluigi mq_append(q, m); 458204591Sluigi } 459204591Sluigi if (m != NULL) { 460204591Sluigi dline->oid.subtype = 1; /* in heap */ 461204591Sluigi heap_insert(&dn_cfg.evheap, pkt->output_time, dline); 462204591Sluigi } 463204591Sluigi} 464204591Sluigi 465204591Sluigi/* 466204591Sluigi * Convert the additional MAC overheads/delays into an equivalent 467204591Sluigi * number of bits for the given data rate. The samples are 468204591Sluigi * in milliseconds so we need to divide by 1000. 469204591Sluigi */ 470204591Sluigistatic uint64_t 471204591Sluigiextra_bits(struct mbuf *m, struct dn_schk *s) 472204591Sluigi{ 473204591Sluigi int index; 474204591Sluigi uint64_t bits; 475204591Sluigi struct dn_profile *pf = s->profile; 476204591Sluigi 477204591Sluigi if (!pf || pf->samples_no == 0) 478204591Sluigi return 0; 479204591Sluigi index = random() % pf->samples_no; 480204591Sluigi bits = div64((uint64_t)pf->samples[index] * s->link.bandwidth, 1000); 481204591Sluigi if (index >= pf->loss_level) { 482204591Sluigi struct dn_pkt_tag *dt = dn_tag_get(m); 483204591Sluigi if (dt) 484204591Sluigi dt->dn_dir = DIR_DROP; 485204591Sluigi } 486204591Sluigi return bits; 487204591Sluigi} 488204591Sluigi 489204591Sluigi/* 490204591Sluigi * Send traffic from a scheduler instance due by 'now'. 491204591Sluigi * Return a pointer to the head of the queue. 492204591Sluigi */ 493204591Sluigistatic struct mbuf * 494204591Sluigiserve_sched(struct mq *q, struct dn_sch_inst *si, uint64_t now) 495204591Sluigi{ 496204591Sluigi struct mq def_q; 497204591Sluigi struct dn_schk *s = si->sched; 498204591Sluigi struct mbuf *m = NULL; 499204591Sluigi int delay_line_idle = (si->dline.mq.head == NULL); 500204591Sluigi int done, bw; 501204591Sluigi 502204591Sluigi if (q == NULL) { 503204591Sluigi q = &def_q; 504204591Sluigi q->head = NULL; 505204591Sluigi } 506204591Sluigi 507204591Sluigi bw = s->link.bandwidth; 508204591Sluigi si->kflags &= ~DN_ACTIVE; 509204591Sluigi 510204591Sluigi if (bw > 0) 511204591Sluigi si->credit += (now - si->sched_time) * bw; 512204591Sluigi else 513204591Sluigi si->credit = 0; 514204591Sluigi si->sched_time = now; 515204591Sluigi done = 0; 516204591Sluigi while (si->credit >= 0 && (m = s->fp->dequeue(si)) != NULL) { 517204591Sluigi uint64_t len_scaled; 518213267Sluigi 519204591Sluigi done++; 520204591Sluigi len_scaled = (bw == 0) ? 0 : hz * 521213267Sluigi (m->m_pkthdr.len * 8 + extra_bits(m, s)); 522204591Sluigi si->credit -= len_scaled; 523204591Sluigi /* Move packet in the delay line */ 524218360Sluigi dn_tag_get(m)->output_time = dn_cfg.curr_time + s->link.delay ; 525204591Sluigi mq_append(&si->dline.mq, m); 526204591Sluigi } 527213267Sluigi 528204591Sluigi /* 529204591Sluigi * If credit >= 0 the instance is idle, mark time. 530204591Sluigi * Otherwise put back in the heap, and adjust the output 531204591Sluigi * time of the last inserted packet, m, which was too early. 532204591Sluigi */ 533204591Sluigi if (si->credit >= 0) { 534204591Sluigi si->idle_time = now; 535204591Sluigi } else { 536204591Sluigi uint64_t t; 537204591Sluigi KASSERT (bw > 0, ("bw=0 and credit<0 ?")); 538204591Sluigi t = div64(bw - 1 - si->credit, bw); 539204591Sluigi if (m) 540204591Sluigi dn_tag_get(m)->output_time += t; 541204591Sluigi si->kflags |= DN_ACTIVE; 542204591Sluigi heap_insert(&dn_cfg.evheap, now + t, si); 543204591Sluigi } 544204591Sluigi if (delay_line_idle && done) 545204591Sluigi transmit_event(q, &si->dline, now); 546204591Sluigi return q->head; 547204591Sluigi} 548204591Sluigi 549204591Sluigi/* 550204591Sluigi * The timer handler for dummynet. Time is computed in ticks, but 551204591Sluigi * but the code is tolerant to the actual rate at which this is called. 552204591Sluigi * Once complete, the function reschedules itself for the next tick. 553204591Sluigi */ 554204591Sluigivoid 555204591Sluigidummynet_task(void *context, int pending) 556204591Sluigi{ 557204591Sluigi struct timeval t; 558204591Sluigi struct mq q = { NULL, NULL }; /* queue to accumulate results */ 559204591Sluigi 560206461Sbz CURVNET_SET((struct vnet *)context); 561206428Sluigi 562204591Sluigi DN_BH_WLOCK(); 563204591Sluigi 564204591Sluigi /* Update number of lost(coalesced) ticks. */ 565204591Sluigi tick_lost += pending - 1; 566204591Sluigi 567204591Sluigi getmicrouptime(&t); 568204591Sluigi /* Last tick duration (usec). */ 569204591Sluigi tick_last = (t.tv_sec - dn_cfg.prev_t.tv_sec) * 1000000 + 570204591Sluigi (t.tv_usec - dn_cfg.prev_t.tv_usec); 571204591Sluigi /* Last tick vs standard tick difference (usec). */ 572204591Sluigi tick_delta = (tick_last * hz - 1000000) / hz; 573204591Sluigi /* Accumulated tick difference (usec). */ 574204591Sluigi tick_delta_sum += tick_delta; 575204591Sluigi 576204591Sluigi dn_cfg.prev_t = t; 577204591Sluigi 578204591Sluigi /* 579204591Sluigi * Adjust curr_time if the accumulated tick difference is 580204591Sluigi * greater than the 'standard' tick. Since curr_time should 581204591Sluigi * be monotonically increasing, we do positive adjustments 582204591Sluigi * as required, and throttle curr_time in case of negative 583204591Sluigi * adjustment. 584204591Sluigi */ 585204591Sluigi dn_cfg.curr_time++; 586204591Sluigi if (tick_delta_sum - tick >= 0) { 587204591Sluigi int diff = tick_delta_sum / tick; 588204591Sluigi 589204591Sluigi dn_cfg.curr_time += diff; 590204591Sluigi tick_diff += diff; 591204591Sluigi tick_delta_sum %= tick; 592204591Sluigi tick_adjustment++; 593204591Sluigi } else if (tick_delta_sum + tick <= 0) { 594204591Sluigi dn_cfg.curr_time--; 595204591Sluigi tick_diff--; 596204591Sluigi tick_delta_sum += tick; 597204591Sluigi tick_adjustment++; 598204591Sluigi } 599204591Sluigi 600204591Sluigi /* serve pending events, accumulate in q */ 601204591Sluigi for (;;) { 602204591Sluigi struct dn_id *p; /* generic parameter to handler */ 603204591Sluigi 604204591Sluigi if (dn_cfg.evheap.elements == 0 || 605204591Sluigi DN_KEY_LT(dn_cfg.curr_time, HEAP_TOP(&dn_cfg.evheap)->key)) 606204591Sluigi break; 607204591Sluigi p = HEAP_TOP(&dn_cfg.evheap)->object; 608204591Sluigi heap_extract(&dn_cfg.evheap, NULL); 609204591Sluigi 610204591Sluigi if (p->type == DN_SCH_I) { 611204591Sluigi serve_sched(&q, (struct dn_sch_inst *)p, dn_cfg.curr_time); 612204591Sluigi } else { /* extracted a delay line */ 613204591Sluigi transmit_event(&q, (struct delay_line *)p, dn_cfg.curr_time); 614204591Sluigi } 615204591Sluigi } 616205173Sluigi if (dn_cfg.expire && ++dn_cfg.expire_cycle >= dn_cfg.expire) { 617205173Sluigi dn_cfg.expire_cycle = 0; 618205173Sluigi dn_drain_scheduler(); 619205173Sluigi dn_drain_queue(); 620205173Sluigi } 621204591Sluigi 622204591Sluigi DN_BH_WUNLOCK(); 623204591Sluigi dn_reschedule(); 624204591Sluigi if (q.head != NULL) 625204591Sluigi dummynet_send(q.head); 626206428Sluigi CURVNET_RESTORE(); 627204591Sluigi} 628204591Sluigi 629204591Sluigi/* 630204591Sluigi * forward a chain of packets to the proper destination. 631204591Sluigi * This runs outside the dummynet lock. 632204591Sluigi */ 633204591Sluigistatic void 634204591Sluigidummynet_send(struct mbuf *m) 635204591Sluigi{ 636204591Sluigi struct mbuf *n; 637204591Sluigi 638204591Sluigi for (; m != NULL; m = n) { 639204591Sluigi struct ifnet *ifp = NULL; /* gcc 3.4.6 complains */ 640204591Sluigi struct m_tag *tag; 641204591Sluigi int dst; 642204591Sluigi 643204591Sluigi n = m->m_nextpkt; 644204591Sluigi m->m_nextpkt = NULL; 645204591Sluigi tag = m_tag_first(m); 646204591Sluigi if (tag == NULL) { /* should not happen */ 647204591Sluigi dst = DIR_DROP; 648204591Sluigi } else { 649204591Sluigi struct dn_pkt_tag *pkt = dn_tag_get(m); 650204591Sluigi /* extract the dummynet info, rename the tag 651204591Sluigi * to carry reinject info. 652204591Sluigi */ 653204591Sluigi dst = pkt->dn_dir; 654204591Sluigi ifp = pkt->ifp; 655204591Sluigi tag->m_tag_cookie = MTAG_IPFW_RULE; 656204591Sluigi tag->m_tag_id = 0; 657204591Sluigi } 658204591Sluigi 659204591Sluigi switch (dst) { 660204591Sluigi case DIR_OUT: 661204591Sluigi ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL); 662204591Sluigi break ; 663204591Sluigi 664204591Sluigi case DIR_IN : 665204591Sluigi netisr_dispatch(NETISR_IP, m); 666204591Sluigi break; 667204591Sluigi 668204591Sluigi#ifdef INET6 669204591Sluigi case DIR_IN | PROTO_IPV6: 670204591Sluigi netisr_dispatch(NETISR_IPV6, m); 671204591Sluigi break; 672204591Sluigi 673204591Sluigi case DIR_OUT | PROTO_IPV6: 674204591Sluigi ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL); 675204591Sluigi break; 676204591Sluigi#endif 677204591Sluigi 678204591Sluigi case DIR_FWD | PROTO_IFB: /* DN_TO_IFB_FWD: */ 679204591Sluigi if (bridge_dn_p != NULL) 680204591Sluigi ((*bridge_dn_p)(m, ifp)); 681204591Sluigi else 682204591Sluigi printf("dummynet: if_bridge not loaded\n"); 683204591Sluigi 684204591Sluigi break; 685204591Sluigi 686204591Sluigi case DIR_IN | PROTO_LAYER2: /* DN_TO_ETH_DEMUX: */ 687204591Sluigi /* 688204591Sluigi * The Ethernet code assumes the Ethernet header is 689204591Sluigi * contiguous in the first mbuf header. 690204591Sluigi * Insure this is true. 691204591Sluigi */ 692204591Sluigi if (m->m_len < ETHER_HDR_LEN && 693204591Sluigi (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) { 694204591Sluigi printf("dummynet/ether: pullup failed, " 695204591Sluigi "dropping packet\n"); 696204591Sluigi break; 697204591Sluigi } 698204591Sluigi ether_demux(m->m_pkthdr.rcvif, m); 699204591Sluigi break; 700204591Sluigi 701204591Sluigi case DIR_OUT | PROTO_LAYER2: /* N_TO_ETH_OUT: */ 702204591Sluigi ether_output_frame(ifp, m); 703204591Sluigi break; 704204591Sluigi 705204591Sluigi case DIR_DROP: 706204591Sluigi /* drop the packet after some time */ 707204591Sluigi FREE_PKT(m); 708204591Sluigi break; 709204591Sluigi 710204591Sluigi default: 711204591Sluigi printf("dummynet: bad switch %d!\n", dst); 712204591Sluigi FREE_PKT(m); 713204591Sluigi break; 714204591Sluigi } 715204591Sluigi } 716204591Sluigi} 717204591Sluigi 718204591Sluigistatic inline int 719204591Sluigitag_mbuf(struct mbuf *m, int dir, struct ip_fw_args *fwa) 720204591Sluigi{ 721204591Sluigi struct dn_pkt_tag *dt; 722204591Sluigi struct m_tag *mtag; 723204591Sluigi 724204591Sluigi mtag = m_tag_get(PACKET_TAG_DUMMYNET, 725204591Sluigi sizeof(*dt), M_NOWAIT | M_ZERO); 726204591Sluigi if (mtag == NULL) 727204591Sluigi return 1; /* Cannot allocate packet header. */ 728204591Sluigi m_tag_prepend(m, mtag); /* Attach to mbuf chain. */ 729204591Sluigi dt = (struct dn_pkt_tag *)(mtag + 1); 730204591Sluigi dt->rule = fwa->rule; 731204591Sluigi dt->rule.info &= IPFW_ONEPASS; /* only keep this info */ 732204591Sluigi dt->dn_dir = dir; 733204591Sluigi dt->ifp = fwa->oif; 734204591Sluigi /* dt->output tame is updated as we move through */ 735204591Sluigi dt->output_time = dn_cfg.curr_time; 736204591Sluigi return 0; 737204591Sluigi} 738204591Sluigi 739204591Sluigi 740204591Sluigi/* 741204591Sluigi * dummynet hook for packets. 742204591Sluigi * We use the argument to locate the flowset fs and the sched_set sch 743204591Sluigi * associated to it. The we apply flow_mask and sched_mask to 744204591Sluigi * determine the queue and scheduler instances. 745204591Sluigi * 746204591Sluigi * dir where shall we send the packet after dummynet. 747204591Sluigi * *m0 the mbuf with the packet 748204591Sluigi * ifp the 'ifp' parameter from the caller. 749204591Sluigi * NULL in ip_input, destination interface in ip_output, 750204591Sluigi */ 751204591Sluigiint 752204591Sluigidummynet_io(struct mbuf **m0, int dir, struct ip_fw_args *fwa) 753204591Sluigi{ 754204591Sluigi struct mbuf *m = *m0; 755204591Sluigi struct dn_fsk *fs = NULL; 756204591Sluigi struct dn_sch_inst *si; 757204591Sluigi struct dn_queue *q = NULL; /* default */ 758204591Sluigi 759204591Sluigi int fs_id = (fwa->rule.info & IPFW_INFO_MASK) + 760204591Sluigi ((fwa->rule.info & IPFW_IS_PIPE) ? 2*DN_MAX_ID : 0); 761204591Sluigi DN_BH_WLOCK(); 762204591Sluigi io_pkt++; 763204591Sluigi /* we could actually tag outside the lock, but who cares... */ 764204591Sluigi if (tag_mbuf(m, dir, fwa)) 765204591Sluigi goto dropit; 766204591Sluigi if (dn_cfg.busy) { 767204591Sluigi /* if the upper half is busy doing something expensive, 768204591Sluigi * lets queue the packet and move forward 769204591Sluigi */ 770204591Sluigi mq_append(&dn_cfg.pending, m); 771204591Sluigi m = *m0 = NULL; /* consumed */ 772204591Sluigi goto done; /* already active, nothing to do */ 773204591Sluigi } 774204591Sluigi /* XXX locate_flowset could be optimised with a direct ref. */ 775204591Sluigi fs = dn_ht_find(dn_cfg.fshash, fs_id, 0, NULL); 776204591Sluigi if (fs == NULL) 777204591Sluigi goto dropit; /* This queue/pipe does not exist! */ 778204591Sluigi if (fs->sched == NULL) /* should not happen */ 779204591Sluigi goto dropit; 780204591Sluigi /* find scheduler instance, possibly applying sched_mask */ 781204591Sluigi si = ipdn_si_find(fs->sched, &(fwa->f_id)); 782204591Sluigi if (si == NULL) 783204591Sluigi goto dropit; 784204591Sluigi /* 785204591Sluigi * If the scheduler supports multiple queues, find the right one 786204591Sluigi * (otherwise it will be ignored by enqueue). 787204591Sluigi */ 788204591Sluigi if (fs->sched->fp->flags & DN_MULTIQUEUE) { 789204591Sluigi q = ipdn_q_find(fs, si, &(fwa->f_id)); 790204591Sluigi if (q == NULL) 791204591Sluigi goto dropit; 792204591Sluigi } 793204591Sluigi if (fs->sched->fp->enqueue(si, q, m)) { 794204591Sluigi /* packet was dropped by enqueue() */ 795204591Sluigi m = *m0 = NULL; 796204591Sluigi goto dropit; 797204591Sluigi } 798204591Sluigi 799204591Sluigi if (si->kflags & DN_ACTIVE) { 800204591Sluigi m = *m0 = NULL; /* consumed */ 801204591Sluigi goto done; /* already active, nothing to do */ 802204591Sluigi } 803204591Sluigi 804204591Sluigi /* compute the initial allowance */ 805213265Sluigi if (si->idle_time < dn_cfg.curr_time) { 806213265Sluigi /* Do this only on the first packet on an idle pipe */ 807204591Sluigi struct dn_link *p = &fs->sched->link; 808213265Sluigi 809213329Sluigi si->sched_time = dn_cfg.curr_time; 810204591Sluigi si->credit = dn_cfg.io_fast ? p->bandwidth : 0; 811204591Sluigi if (p->burst) { 812204591Sluigi uint64_t burst = (dn_cfg.curr_time - si->idle_time) * p->bandwidth; 813204591Sluigi if (burst > p->burst) 814204591Sluigi burst = p->burst; 815204591Sluigi si->credit += burst; 816204591Sluigi } 817204591Sluigi } 818204591Sluigi /* pass through scheduler and delay line */ 819204591Sluigi m = serve_sched(NULL, si, dn_cfg.curr_time); 820204591Sluigi 821204591Sluigi /* optimization -- pass it back to ipfw for immediate send */ 822204591Sluigi /* XXX Don't call dummynet_send() if scheduler return the packet 823204591Sluigi * just enqueued. This avoid a lock order reversal. 824204591Sluigi * 825204591Sluigi */ 826204591Sluigi if (/*dn_cfg.io_fast &&*/ m == *m0 && (dir & PROTO_LAYER2) == 0 ) { 827205602Sluigi /* fast io, rename the tag * to carry reinject info. */ 828205602Sluigi struct m_tag *tag = m_tag_first(m); 829205602Sluigi 830205602Sluigi tag->m_tag_cookie = MTAG_IPFW_RULE; 831205602Sluigi tag->m_tag_id = 0; 832204591Sluigi io_pkt_fast++; 833204591Sluigi if (m->m_nextpkt != NULL) { 834204591Sluigi printf("dummynet: fast io: pkt chain detected!\n"); 835204591Sluigi m->m_nextpkt = NULL; 836204591Sluigi } 837204591Sluigi m = NULL; 838204591Sluigi } else { 839204591Sluigi *m0 = NULL; 840204591Sluigi } 841204591Sluigidone: 842204591Sluigi DN_BH_WUNLOCK(); 843204591Sluigi if (m) 844204591Sluigi dummynet_send(m); 845204591Sluigi return 0; 846204591Sluigi 847204591Sluigidropit: 848204591Sluigi io_pkt_drop++; 849204591Sluigi DN_BH_WUNLOCK(); 850204591Sluigi if (m) 851204591Sluigi FREE_PKT(m); 852204591Sluigi *m0 = NULL; 853204591Sluigi return (fs && (fs->fs.flags & DN_NOERROR)) ? 0 : ENOBUFS; 854204591Sluigi} 855