2 * Copyright (c) 1998-2002 Luigi Rizzo, Universita` di Pisa
3 * Portions Copyright (c) 2000 Akamba Corp.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * $FreeBSD: src/sys/netinet/ip_dummynet.c,v 1.24.2.22 2003/05/13 09:31:06 maxim Exp $
28 * $DragonFly: src/sys/net/dummynet/ip_dummynet.c,v 1.16 2005/12/19 00:07:02 corecode Exp $
31 #if !defined(KLD_MODULE)
32 #include "opt_ipfw.h" /* for IPFW2 definition */
39 * This module implements IP dummynet, a bandwidth limiter/delay emulator
40 * used in conjunction with the ipfw package.
41 * Description of the data structures used is in ip_dummynet.h
42 * Here you mainly find the following blocks of code:
43 * + variable declarations;
44 * + heap management functions;
45 * + scheduler and dummynet functions;
46 * + configuration and initialization.
48 * NOTA BENE: critical sections are protected by splimp()/splx()
49 * pairs. One would think that splnet() is enough as for most of
50 * the netinet code, but it is not so because when used with
51 * bridging, dummynet is invoked at splimp().
53 * Most important Changes:
56 * 010124: Fixed WF2Q behaviour
57 * 010122: Fixed spl protection.
58 * 000601: WF2Q support
59 * 000106: large rewrite, use heaps to handle very many pipes.
60 * 980513: initial release
62 * include files marked with XXX are probably not needed
65 #include <sys/param.h>
66 #include <sys/systm.h>
67 #include <sys/malloc.h>
69 #include <sys/kernel.h>
70 #include <sys/module.h>
72 #include <sys/socket.h>
73 #include <sys/socketvar.h>
75 #include <sys/sysctl.h>
76 #include <sys/thread2.h>
78 #include <net/route.h>
79 #include <netinet/in.h>
80 #include <netinet/in_systm.h>
81 #include <netinet/in_var.h>
82 #include <netinet/ip.h>
83 #include <net/ipfw/ip_fw.h>
84 #include "ip_dummynet.h"
85 #include <netinet/ip_var.h>
87 #include <netinet/if_ether.h> /* for struct arpcom */
88 #include <net/oldbridge/bridge.h>
91 * We keep a private variable for the simulation time, but we could
92 * probably use an existing one ("softticks" in sys/kern/kern_timer.c)
94 static dn_key curr_time = 0 ; /* current simulation time */
96 static int dn_hash_size = 64 ; /* default hash size */
98 /* statistics on number of queue searches and search steps */
99 static int searches, search_steps ;
100 static int pipe_expire = 1 ; /* expire queue if empty */
101 static int dn_max_ratio = 16 ; /* max queues/buckets ratio */
103 static int red_lookup_depth = 256; /* RED - default lookup table depth */
104 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
105 static int red_max_pkt_size = 1500; /* RED - default max packet size */
108 * Three heaps contain queues and pipes that the scheduler handles:
110 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
112 * wfq_ready_heap contains the pipes associated with WF2Q flows
114 * extract_heap contains pipes associated with delay lines.
118 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
120 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
122 static int heap_init(struct dn_heap *h, int size) ;
123 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
124 static void heap_extract(struct dn_heap *h, void *obj);
126 static void transmit_event(struct dn_pipe *pipe);
127 static void ready_event(struct dn_flow_queue *q);
129 static struct dn_pipe *all_pipes = NULL ; /* list of all pipes */
130 static struct dn_flow_set *all_flow_sets = NULL ;/* list of all flow_sets */
132 static struct callout dn_timeout;
135 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
136 CTLFLAG_RW, 0, "Dummynet");
137 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
138 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
139 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time,
140 CTLFLAG_RD, &curr_time, 0, "Current tick");
141 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
142 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
143 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
144 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
145 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches,
146 CTLFLAG_RD, &searches, 0, "Number of queue searches");
147 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps,
148 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
149 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
150 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
151 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
152 CTLFLAG_RW, &dn_max_ratio, 0,
153 "Max ratio between dynamic queues and buckets");
154 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
155 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
156 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
157 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
158 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
159 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
162 static int config_pipe(struct dn_pipe *p);
163 static int ip_dn_ctl(struct sockopt *sopt);
165 static void rt_unref(struct rtentry *);
166 static void dummynet(void *);
167 static void dummynet_flush(void);
168 void dummynet_drain(void);
169 static ip_dn_io_t dummynet_io;
170 static void dn_rule_delete(void *);
172 int if_tx_rdy(struct ifnet *ifp);
175 rt_unref(struct rtentry *rt)
179 if (rt->rt_refcnt <= 0)
180 printf("-- warning, refcnt now %ld, decreasing\n", rt->rt_refcnt);
185 * Heap management functions.
187 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
188 * Some macros help finding parent/children so we can optimize them.
190 * heap_init() is called to expand the heap when needed.
191 * Increment size in blocks of 16 entries.
192 * XXX failure to allocate a new element is a pretty bad failure
193 * as we basically stall a whole queue forever!!
194 * Returns 1 on error, 0 on success
196 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
197 #define HEAP_LEFT(x) ( 2*(x) + 1 )
198 #define HEAP_IS_LEFT(x) ( (x) & 1 )
199 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
200 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
201 #define HEAP_INCREMENT 15
204 heap_init(struct dn_heap *h, int new_size)
206 struct dn_heap_entry *p;
208 if (h->size >= new_size ) {
209 printf("heap_init, Bogus call, have %d want %d\n",
213 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
214 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_WAITOK | M_ZERO);
216 bcopy(h->p, p, h->size * sizeof(*p) );
217 free(h->p, M_DUMMYNET);
225 * Insert element in heap. Normally, p != NULL, we insert p in
226 * a new position and bubble up. If p == NULL, then the element is
227 * already in place, and key is the position where to start the
229 * Returns 1 on failure (cannot allocate new heap entry)
231 * If offset > 0 the position (index, int) of the element in the heap is
232 * also stored in the element itself at the given offset in bytes.
234 #define SET_OFFSET(heap, node) \
235 if (heap->offset > 0) \
236 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
238 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
240 #define RESET_OFFSET(heap, node) \
241 if (heap->offset > 0) \
242 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
244 heap_insert(struct dn_heap *h, dn_key key1, void *p)
246 int son = h->elements ;
248 if (p == NULL) /* data already there, set starting point */
250 else { /* insert new element at the end, possibly resize */
252 if (son == h->size) /* need resize... */
253 if (heap_init(h, h->elements+1) )
254 return 1 ; /* failure... */
255 h->p[son].object = p ;
256 h->p[son].key = key1 ;
259 while (son > 0) { /* bubble up */
260 int father = HEAP_FATHER(son) ;
261 struct dn_heap_entry tmp ;
263 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
264 break ; /* found right position */
265 /* son smaller than father, swap and repeat */
266 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
275 * remove top element from heap, or obj if obj != NULL
278 heap_extract(struct dn_heap *h, void *obj)
280 int child, father, max = h->elements - 1 ;
283 printf("warning, extract from empty heap 0x%p\n", h);
286 father = 0 ; /* default: move up smallest child */
287 if (obj != NULL) { /* extract specific element, index is at offset */
289 panic("*** heap_extract from middle not supported on this heap!!!\n");
290 father = *((int *)((char *)obj + h->offset)) ;
291 if (father < 0 || father >= h->elements) {
292 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
293 father, h->elements);
294 panic("heap_extract");
297 RESET_OFFSET(h, father);
298 child = HEAP_LEFT(father) ; /* left child */
299 while (child <= max) { /* valid entry */
300 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
301 child = child+1 ; /* take right child, otherwise left */
302 h->p[father] = h->p[child] ;
303 SET_OFFSET(h, father);
305 child = HEAP_LEFT(child) ; /* left child for next loop */
310 * Fill hole with last entry and bubble up, reusing the insert code
312 h->p[father] = h->p[max] ;
313 heap_insert(h, father, NULL); /* this one cannot fail */
319 * change object position and update references
320 * XXX this one is never used!
323 heap_move(struct dn_heap *h, dn_key new_key, void *object)
327 int max = h->elements-1 ;
328 struct dn_heap_entry buf ;
331 panic("cannot move items on this heap");
333 i = *((int *)((char *)object + h->offset));
334 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
335 h->p[i].key = new_key ;
336 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
337 i = temp ) { /* bubble up */
338 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
341 } else { /* must move down */
342 h->p[i].key = new_key ;
343 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
344 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
345 temp++ ; /* select child with min key */
346 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
347 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
356 #endif /* heap_move, unused */
359 * heapify() will reorganize data inside an array to maintain the
360 * heap property. It is needed when we delete a bunch of entries.
363 heapify(struct dn_heap *h)
367 for (i = 0 ; i < h->elements ; i++ )
368 heap_insert(h, i , NULL) ;
372 * cleanup the heap and free data structure
375 heap_free(struct dn_heap *h)
378 free(h->p, M_DUMMYNET);
379 bzero(h, sizeof(*h) );
383 * --- end of heap management functions ---
387 * Scheduler functions:
389 * transmit_event() is called when the delay-line needs to enter
390 * the scheduler, either because of existing pkts getting ready,
391 * or new packets entering the queue. The event handled is the delivery
392 * time of the packet.
394 * ready_event() does something similar with fixed-rate queues, and the
395 * event handled is the finish time of the head pkt.
397 * wfq_ready_event() does something similar with WF2Q queues, and the
398 * event handled is the start time of the head pkt.
400 * In all cases, we make sure that the data structures are consistent
401 * before passing pkts out, because this might trigger recursive
402 * invocations of the procedures.
405 transmit_event(struct dn_pipe *pipe)
409 while ( (pkt = pipe->head) && DN_KEY_LEQ(pkt->output_time, curr_time) ) {
411 * first unlink, then call procedures, since ip_input() can invoke
412 * ip_output() and viceversa, thus causing nested calls
414 pipe->head = DN_NEXT(pkt) ;
417 * The actual mbuf is preceded by a struct dn_pkt, resembling an mbuf
418 * (NOT A REAL one, just a small block of malloc'ed memory) with
419 * m_type = MT_TAG, m_flags = PACKET_TAG_DUMMYNET
420 * dn_m (m_next) = actual mbuf to be processed by ip_input/output
421 * and some other fields.
422 * The block IS FREED HERE because it contains parameters passed
423 * to the called routine.
425 switch (pkt->dn_dir) {
427 (void)ip_output((struct mbuf *)pkt, NULL, NULL, 0, NULL, NULL);
428 rt_unref (pkt->ro.ro_rt) ;
432 ip_input((struct mbuf *)pkt) ;
437 /* somebody unloaded the bridge module. Drop pkt */
438 printf("-- dropping bridged packet trapped in pipe--\n");
442 case DN_TO_ETH_DEMUX:
444 struct mbuf *m = (struct mbuf *)pkt ;
445 struct ether_header *eh;
447 if (pkt->dn_m->m_len < ETHER_HDR_LEN &&
448 (pkt->dn_m = m_pullup(pkt->dn_m, ETHER_HDR_LEN)) == NULL) {
449 printf("dummynet/bridge: pullup fail, dropping pkt\n");
453 * same as ether_input, make eh be a pointer into the mbuf
455 eh = mtod(pkt->dn_m, struct ether_header *);
456 m_adj(pkt->dn_m, ETHER_HDR_LEN);
458 * bdg_forward() wants a pointer to the pseudo-mbuf-header, but
459 * on return it will supply the pointer to the actual packet
460 * (originally pkt->dn_m, but could be something else now) if
461 * it has not consumed it.
463 if (pkt->dn_dir == DN_TO_BDG_FWD) {
464 m = bdg_forward_ptr(m, eh, pkt->ifp);
468 /* which consumes the mbuf */
469 ether_demux(NULL, eh, m);
474 ether_output_frame(pkt->ifp, (struct mbuf *)pkt);
478 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
482 free(pkt, M_DUMMYNET);
484 /* if there are leftover packets, put into the heap for next event */
485 if ( (pkt = pipe->head) )
486 heap_insert(&extract_heap, pkt->output_time, pipe ) ;
487 /* XXX should check errors on heap_insert, by draining the
488 * whole pipe p and hoping in the future we are more successful
493 * the following macro computes how many ticks we have to wait
494 * before being able to transmit a packet. The credit is taken from
495 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
497 #define SET_TICKS(pkt, q, p) \
498 (pkt->dn_m->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \
502 * extract pkt from queue, compute output time (could be now)
503 * and put into delay line (p_queue)
506 move_pkt(struct dn_pkt *pkt, struct dn_flow_queue *q,
507 struct dn_pipe *p, int len)
509 q->head = DN_NEXT(pkt) ;
511 q->len_bytes -= len ;
513 pkt->output_time = curr_time + p->delay ;
518 DN_NEXT_NC(p->tail) = (struct mbuf *)pkt;
520 DN_NEXT_NC(p->tail) = NULL;
524 * ready_event() is invoked every time the queue must enter the
525 * scheduler, either because the first packet arrives, or because
526 * a previously scheduled event fired.
527 * On invokation, drain as many pkts as possible (could be 0) and then
528 * if there are leftover packets reinsert the pkt in the scheduler.
531 ready_event(struct dn_flow_queue *q)
534 struct dn_pipe *p = q->fs->pipe ;
538 printf("ready_event- pipe is gone\n");
541 p_was_empty = (p->head == NULL) ;
544 * schedule fixed-rate queues linked to this pipe:
545 * Account for the bw accumulated since last scheduling, then
546 * drain as many pkts as allowed by q->numbytes and move to
547 * the delay line (in p) computing output time.
548 * bandwidth==0 (no limit) means we can drain the whole queue,
549 * setting len_scaled = 0 does the job.
551 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth;
552 while ( (pkt = q->head) != NULL ) {
553 int len = pkt->dn_m->m_pkthdr.len;
554 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
555 if (len_scaled > q->numbytes )
557 q->numbytes -= len_scaled ;
558 move_pkt(pkt, q, p, len);
561 * If we have more packets queued, schedule next ready event
562 * (can only occur when bandwidth != 0, otherwise we would have
563 * flushed the whole queue in the previous loop).
564 * To this purpose we record the current time and compute how many
565 * ticks to go for the finish time of the packet.
567 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */
568 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
569 q->sched_time = curr_time ;
570 heap_insert(&ready_heap, curr_time + t, (void *)q );
571 /* XXX should check errors on heap_insert, and drain the whole
572 * queue on error hoping next time we are luckier.
574 } else { /* RED needs to know when the queue becomes empty */
575 q->q_time = curr_time;
579 * If the delay line was empty call transmit_event(p) now.
580 * Otherwise, the scheduler will take care of it.
587 * Called when we can transmit packets on WF2Q queues. Take pkts out of
588 * the queues at their start time, and enqueue into the delay line.
589 * Packets are drained until p->numbytes < 0. As long as
590 * len_scaled >= p->numbytes, the packet goes into the delay line
591 * with a deadline p->delay. For the last packet, if p->numbytes<0,
592 * there is an additional delay.
595 ready_event_wfq(struct dn_pipe *p)
597 int p_was_empty = (p->head == NULL) ;
598 struct dn_heap *sch = &(p->scheduler_heap);
599 struct dn_heap *neh = &(p->not_eligible_heap) ;
601 if (p->if_name[0] == 0) /* tx clock is simulated */
602 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth;
603 else { /* tx clock is for real, the ifq must be empty or this is a NOP */
604 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
607 DEB(printf("pipe %d ready from %s --\n",
608 p->pipe_nr, p->if_name);)
613 * While we have backlogged traffic AND credit, we need to do
614 * something on the queue.
616 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) {
617 if (sch->elements > 0) { /* have some eligible pkts to send out */
618 struct dn_flow_queue *q = sch->p[0].object ;
619 struct dn_pkt *pkt = q->head;
620 struct dn_flow_set *fs = q->fs;
621 u_int64_t len = pkt->dn_m->m_pkthdr.len;
622 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
624 heap_extract(sch, NULL); /* remove queue from heap */
625 p->numbytes -= len_scaled ;
626 move_pkt(pkt, q, p, len);
628 p->V += (len<<MY_M) / p->sum ; /* update V */
629 q->S = q->F ; /* update start time */
630 if (q->len == 0) { /* Flow not backlogged any more */
632 heap_insert(&(p->idle_heap), q->F, q);
633 } else { /* still backlogged */
635 * update F and position in backlogged queue, then
636 * put flow in not_eligible_heap (we will fix this later).
638 len = (q->head)->dn_m->m_pkthdr.len;
639 q->F += (len<<MY_M)/(u_int64_t) fs->weight ;
640 if (DN_KEY_LEQ(q->S, p->V))
641 heap_insert(neh, q->S, q);
643 heap_insert(sch, q->F, q);
647 * now compute V = max(V, min(S_i)). Remember that all elements in sch
648 * have by definition S_i <= V so if sch is not empty, V is surely
649 * the max and we must not update it. Conversely, if sch is empty
650 * we only need to look at neh.
652 if (sch->elements == 0 && neh->elements > 0)
653 p->V = MAX64 ( p->V, neh->p[0].key );
654 /* move from neh to sch any packets that have become eligible */
655 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) {
656 struct dn_flow_queue *q = neh->p[0].object ;
657 heap_extract(neh, NULL);
658 heap_insert(sch, q->F, q);
661 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */
662 p->numbytes = -1 ; /* mark not ready for I/O */
666 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0
667 && p->idle_heap.elements > 0) {
669 * no traffic and no events scheduled. We can get rid of idle-heap.
673 for (i = 0 ; i < p->idle_heap.elements ; i++) {
674 struct dn_flow_queue *q = p->idle_heap.p[i].object ;
681 p->idle_heap.elements = 0 ;
684 * If we are getting clocks from dummynet (not a real interface) and
685 * If we are under credit, schedule the next ready event.
686 * Also fix the delivery time of the last packet.
688 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */
689 dn_key t=0 ; /* number of ticks i have to wait */
691 if (p->bandwidth > 0)
692 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ;
693 p->tail->output_time += t ;
694 p->sched_time = curr_time ;
695 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
696 /* XXX should check errors on heap_insert, and drain the whole
697 * queue on error hoping next time we are luckier.
701 * If the delay line was empty call transmit_event(p) now.
702 * Otherwise, the scheduler will take care of it.
709 * This is called once per tick, or HZ times per second. It is used to
710 * increment the current tick counter and schedule expired events.
713 dummynet(void * __unused unused)
715 void *p ; /* generic parameter to handler */
717 struct dn_heap *heaps[3];
721 heaps[0] = &ready_heap ; /* fixed-rate queues */
722 heaps[1] = &wfq_ready_heap ; /* wfq queues */
723 heaps[2] = &extract_heap ; /* delay line */
724 crit_enter(); /* see note on top, splnet() is not enough */
726 for (i=0; i < 3 ; i++) {
728 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time) ) {
729 DDB(if (h->p[0].key > curr_time)
730 printf("-- dummynet: warning, heap %d is %d ticks late\n",
731 i, (int)(curr_time - h->p[0].key));)
732 p = h->p[0].object ; /* store a copy before heap_extract */
733 heap_extract(h, NULL); /* need to extract before processing */
737 struct dn_pipe *pipe = p;
738 if (pipe->if_name[0] != '\0')
739 printf("*** bad ready_event_wfq for pipe %s\n",
747 /* sweep pipes trying to expire idle flow_queues */
748 for (pe = all_pipes; pe ; pe = pe->next )
749 if (pe->idle_heap.elements > 0 &&
750 DN_KEY_LT(pe->idle_heap.p[0].key, pe->V) ) {
751 struct dn_flow_queue *q = pe->idle_heap.p[0].object ;
753 heap_extract(&(pe->idle_heap), NULL);
754 q->S = q->F + 1 ; /* mark timestamp as invalid */
755 pe->sum -= q->fs->weight ;
758 callout_reset(&dn_timeout, 1, dummynet, NULL);
762 * called by an interface when tx_rdy occurs.
765 if_tx_rdy(struct ifnet *ifp)
769 for (p = all_pipes; p ; p = p->next )
773 for (p = all_pipes; p ; p = p->next )
774 if (!strcmp(p->if_name, ifp->if_xname) ) {
776 DEB(printf("++ tx rdy from %s (now found)\n", ifp->if_xname);)
781 DEB(printf("++ tx rdy from %s - qlen %d\n", ifp->if_xname,
782 ifp->if_snd.ifq_len);)
783 p->numbytes = 0 ; /* mark ready for I/O */
790 * Unconditionally expire empty queues in case of shortage.
791 * Returns the number of queues freed.
794 expire_queues(struct dn_flow_set *fs)
796 struct dn_flow_queue *q, *prev ;
797 int i, initial_elements = fs->rq_elements ;
799 if (fs->last_expired == time_second)
801 fs->last_expired = time_second ;
802 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
803 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
804 if (q->head != NULL || q->S != q->F+1) {
807 } else { /* entry is idle, expire it */
808 struct dn_flow_queue *old_q = q ;
811 prev->next = q = q->next ;
813 fs->rq[i] = q = q->next ;
815 free(old_q, M_DUMMYNET);
817 return initial_elements - fs->rq_elements ;
821 * If room, create a new queue and put at head of slot i;
822 * otherwise, create or use the default queue.
824 static struct dn_flow_queue *
825 create_queue(struct dn_flow_set *fs, int i)
827 struct dn_flow_queue *q ;
829 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
830 expire_queues(fs) == 0) {
832 * No way to get room, use or create overflow queue.
835 if ( fs->rq[i] != NULL )
838 q = malloc(sizeof(*q), M_DUMMYNET, M_WAITOK | M_ZERO);
841 q->next = fs->rq[i] ;
842 q->S = q->F + 1; /* hack - mark timestamp as invalid */
849 * Given a flow_set and a pkt in last_pkt, find a matching queue
850 * after appropriate masking. The queue is moved to front
851 * so that further searches take less time.
853 static struct dn_flow_queue *
854 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
856 int i = 0 ; /* we need i and q for new allocations */
857 struct dn_flow_queue *q, *prev;
859 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
862 /* first, do the masking */
863 id->dst_ip &= fs->flow_mask.dst_ip ;
864 id->src_ip &= fs->flow_mask.src_ip ;
865 id->dst_port &= fs->flow_mask.dst_port ;
866 id->src_port &= fs->flow_mask.src_port ;
867 id->proto &= fs->flow_mask.proto ;
868 id->flags = 0 ; /* we don't care about this one */
869 /* then, hash function */
870 i = ( (id->dst_ip) & 0xffff ) ^
871 ( (id->dst_ip >> 15) & 0xffff ) ^
872 ( (id->src_ip << 1) & 0xffff ) ^
873 ( (id->src_ip >> 16 ) & 0xffff ) ^
874 (id->dst_port << 1) ^ (id->src_port) ^
876 i = i % fs->rq_size ;
877 /* finally, scan the current list for a match */
879 for (prev=NULL, q = fs->rq[i] ; q ; ) {
881 if (id->dst_ip == q->id.dst_ip &&
882 id->src_ip == q->id.src_ip &&
883 id->dst_port == q->id.dst_port &&
884 id->src_port == q->id.src_port &&
885 id->proto == q->id.proto &&
886 id->flags == q->id.flags)
888 else if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
889 /* entry is idle and not in any heap, expire it */
890 struct dn_flow_queue *old_q = q ;
893 prev->next = q = q->next ;
895 fs->rq[i] = q = q->next ;
897 free(old_q, M_DUMMYNET);
903 if (q && prev != NULL) { /* found and not in front */
904 prev->next = q->next ;
905 q->next = fs->rq[i] ;
909 if (q == NULL) { /* no match, need to allocate a new entry */
910 q = create_queue(fs, i);
918 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
923 * RED calculates the average queue size (avg) using a low-pass filter
924 * with an exponential weighted (w_q) moving average:
925 * avg <- (1-w_q) * avg + w_q * q_size
926 * where q_size is the queue length (measured in bytes or * packets).
928 * If q_size == 0, we compute the idle time for the link, and set
929 * avg = (1 - w_q)^(idle/s)
930 * where s is the time needed for transmitting a medium-sized packet.
932 * Now, if avg < min_th the packet is enqueued.
933 * If avg > max_th the packet is dropped. Otherwise, the packet is
934 * dropped with probability P function of avg.
939 /* queue in bytes or packets ? */
940 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len;
942 DEB(printf("\n%d q: %2u ", (int) curr_time, q_size);)
944 /* average queue size estimation */
947 * queue is not empty, avg <- avg + (q_size - avg) * w_q
949 int diff = SCALE(q_size) - q->avg;
950 int64_t v = SCALE_MUL((int64_t) diff, (int64_t) fs->w_q);
955 * queue is empty, find for how long the queue has been
956 * empty and use a lookup table for computing
957 * (1 - * w_q)^(idle_time/s) where s is the time to send a
962 u_int t = (curr_time - q->q_time) / fs->lookup_step;
964 q->avg = (t < fs->lookup_depth) ?
965 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
968 DEB(printf("avg: %u ", SCALE_VAL(q->avg));)
970 /* should i drop ? */
972 if (q->avg < fs->min_th) {
974 return 0; /* accept packet ; */
976 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
977 if (fs->flags_fs & DN_IS_GENTLE_RED) {
979 * According to Gentle-RED, if avg is greater than max_th the
980 * packet is dropped with a probability
981 * p_b = c_3 * avg - c_4
982 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
984 p_b = SCALE_MUL((int64_t) fs->c_3, (int64_t) q->avg) - fs->c_4;
990 } else if (q->avg > fs->min_th) {
992 * we compute p_b using the linear dropping function p_b = c_1 *
993 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
994 * max_p * min_th / (max_th - min_th)
996 p_b = SCALE_MUL((int64_t) fs->c_1, (int64_t) q->avg) - fs->c_2;
998 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
999 p_b = (p_b * len) / fs->max_pkt_size;
1000 if (++q->count == 0)
1001 q->random = random() & 0xffff;
1004 * q->count counts packets arrived since last drop, so a greater
1005 * value of q->count means a greater packet drop probability.
1007 if (SCALE_MUL(p_b, SCALE((int64_t) q->count)) > q->random) {
1009 DEB(printf("- red drop");)
1010 /* after a drop we calculate a new random value */
1011 q->random = random() & 0xffff;
1012 return 1; /* drop */
1015 /* end of RED algorithm */
1016 return 0 ; /* accept */
1020 struct dn_flow_set *
1021 locate_flowset(int pipe_nr, struct ip_fw *rule)
1024 struct dn_flow_set *fs;
1025 ipfw_insn *cmd = rule->cmd + rule->act_ofs;
1027 if (cmd->opcode == O_LOG)
1029 fs = ((ipfw_insn_pipe *)cmd)->pipe_ptr;
1034 if (cmd->opcode == O_QUEUE)
1036 struct dn_flow_set *fs = NULL ;
1038 if ( (rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_QUEUE )
1040 for (fs=all_flow_sets; fs && fs->fs_nr != pipe_nr; fs=fs->next)
1044 for (p1 = all_pipes; p1 && p1->pipe_nr != pipe_nr; p1 = p1->next)
1049 /* record for the future */
1051 ((ipfw_insn_pipe *)cmd)->pipe_ptr = fs;
1054 rule->pipe_ptr = fs;
1060 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1061 * depending on whether WF2Q or fixed bw is used.
1063 * pipe_nr pipe or queue the packet is destined for.
1064 * dir where shall we send the packet after dummynet.
1065 * m the mbuf with the packet
1066 * ifp the 'ifp' parameter from the caller.
1067 * NULL in ip_input, destination interface in ip_output,
1068 * real_dst in bdg_forward
1069 * ro route parameter (only used in ip_output, NULL otherwise)
1070 * dst destination address, only used by ip_output
1071 * rule matching rule, in case of multiple passes
1072 * flags flags from the caller, only used in ip_output
1076 dummynet_io(struct mbuf *m, int pipe_nr, int dir, struct ip_fw_args *fwa)
1079 struct dn_flow_set *fs;
1080 struct dn_pipe *pipe ;
1081 u_int64_t len = m->m_pkthdr.len ;
1082 struct dn_flow_queue *q = NULL ;
1087 ipfw_insn *cmd = fwa->rule->cmd + fwa->rule->act_ofs;
1089 if (cmd->opcode == O_LOG)
1091 is_pipe = (cmd->opcode == O_PIPE);
1093 is_pipe = (fwa->rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_PIPE;
1099 * this is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule
1101 fs = locate_flowset(pipe_nr, fwa->rule);
1103 goto dropit ; /* this queue/pipe does not exist! */
1105 if (pipe == NULL) { /* must be a queue, try find a matching pipe */
1106 for (pipe = all_pipes; pipe && pipe->pipe_nr != fs->parent_nr;
1112 printf("No pipe %d for queue %d, drop pkt\n",
1113 fs->parent_nr, fs->fs_nr);
1117 q = find_queue(fs, &(fwa->f_id));
1119 goto dropit ; /* cannot allocate queue */
1121 * update statistics, then check reasons to drop pkt
1123 q->tot_bytes += len ;
1125 if ( fs->plr && random() < fs->plr )
1126 goto dropit ; /* random pkt drop */
1127 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) {
1128 if (q->len_bytes > fs->qsize)
1129 goto dropit ; /* queue size overflow */
1131 if (q->len >= fs->qsize)
1132 goto dropit ; /* queue count overflow */
1134 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) )
1137 /* XXX expensive to zero, see if we can remove it*/
1138 pkt = malloc(sizeof (*pkt), M_DUMMYNET, M_INTWAIT | M_ZERO | M_NULLOK);
1140 goto dropit; /* cannot allocate packet header */
1142 /* ok, i can handle the pkt now... */
1143 /* build and enqueue packet + parameters */
1144 pkt->hdr.mh_type = MT_TAG;
1145 pkt->hdr.mh_flags = PACKET_TAG_DUMMYNET;
1146 pkt->rule = fwa->rule ;
1147 DN_NEXT_NC(pkt) = NULL;
1151 pkt->ifp = fwa->oif;
1152 if (dir == DN_TO_IP_OUT) {
1154 * We need to copy *ro because for ICMP pkts (and maybe others)
1155 * the caller passed a pointer into the stack; dst might also be
1156 * a pointer into *ro so it needs to be updated.
1158 pkt->ro = *(fwa->ro);
1160 fwa->ro->ro_rt->rt_refcnt++ ;
1161 if (fwa->dst == (struct sockaddr_in *)&fwa->ro->ro_dst) /* dst points into ro */
1162 fwa->dst = (struct sockaddr_in *)&(pkt->ro.ro_dst) ;
1164 pkt->dn_dst = fwa->dst;
1165 pkt->flags = fwa->flags;
1167 if (q->head == NULL)
1170 DN_NEXT_NC(q->tail) = (struct mbuf *)pkt;
1173 q->len_bytes += len ;
1175 if ( q->head != pkt ) /* flow was not idle, we are done */
1178 * If we reach this point the flow was previously idle, so we need
1179 * to schedule it. This involves different actions for fixed-rate or
1184 * Fixed-rate queue: just insert into the ready_heap.
1187 if (pipe->bandwidth)
1188 t = SET_TICKS(pkt, q, pipe);
1189 q->sched_time = curr_time ;
1190 if (t == 0) /* must process it now */
1193 heap_insert(&ready_heap, curr_time + t , q );
1196 * WF2Q. First, compute start time S: if the flow was idle (S=F+1)
1197 * set S to the virtual time V for the controlling pipe, and update
1198 * the sum of weights for the pipe; otherwise, remove flow from
1199 * idle_heap and set S to max(F,V).
1200 * Second, compute finish time F = S + len/weight.
1201 * Third, if pipe was idle, update V=max(S, V).
1202 * Fourth, count one more backlogged flow.
1204 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */
1206 pipe->sum += fs->weight ; /* add weight of new queue */
1208 heap_extract(&(pipe->idle_heap), q);
1209 q->S = MAX64(q->F, pipe->V ) ;
1211 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight;
1213 if (pipe->not_eligible_heap.elements == 0 &&
1214 pipe->scheduler_heap.elements == 0)
1215 pipe->V = MAX64 ( q->S, pipe->V );
1218 * Look at eligibility. A flow is not eligibile if S>V (when
1219 * this happens, it means that there is some other flow already
1220 * scheduled for the same pipe, so the scheduler_heap cannot be
1221 * empty). If the flow is not eligible we just store it in the
1222 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1223 * and possibly invoke ready_event_wfq() right now if there is
1225 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1226 * and for all flows in not_eligible_heap (NEH), S_i > V .
1227 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH,
1228 * we only need to look into NEH.
1230 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */
1231 if (pipe->scheduler_heap.elements == 0)
1232 printf("++ ouch! not eligible but empty scheduler!\n");
1233 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1235 heap_insert(&(pipe->scheduler_heap), q->F, q);
1236 if (pipe->numbytes >= 0) { /* pipe is idle */
1237 if (pipe->scheduler_heap.elements != 1)
1238 printf("*** OUCH! pipe should have been idle!\n");
1239 DEB(printf("Waking up pipe %d at %d\n",
1240 pipe->pipe_nr, (int)(q->F >> MY_M)); )
1241 pipe->sched_time = curr_time ;
1242 ready_event_wfq(pipe);
1255 return ( (fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1259 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1260 * Doing this would probably save us the initial bzero of dn_pkt
1262 #define DN_FREE_PKT(pkt) { \
1263 struct dn_pkt *n = pkt ; \
1264 rt_unref ( n->ro.ro_rt ) ; \
1266 pkt = DN_NEXT(n) ; \
1267 free(n, M_DUMMYNET) ; }
1270 * Dispose all packets and flow_queues on a flow_set.
1271 * If all=1, also remove red lookup table and other storage,
1272 * including the descriptor itself.
1273 * For the one in dn_pipe MUST also cleanup ready_heap...
1276 purge_flow_set(struct dn_flow_set *fs, int all)
1278 struct dn_pkt *pkt ;
1279 struct dn_flow_queue *q, *qn ;
1282 for (i = 0 ; i <= fs->rq_size ; i++ ) {
1283 for (q = fs->rq[i] ; q ; q = qn ) {
1284 for (pkt = q->head ; pkt ; )
1287 free(q, M_DUMMYNET);
1291 fs->rq_elements = 0 ;
1293 /* RED - free lookup table */
1295 free(fs->w_q_lookup, M_DUMMYNET);
1297 free(fs->rq, M_DUMMYNET);
1298 /* if this fs is not part of a pipe, free it */
1299 if (fs->pipe && fs != &(fs->pipe->fs) )
1300 free(fs, M_DUMMYNET);
1305 * Dispose all packets queued on a pipe (not a flow_set).
1306 * Also free all resources associated to a pipe, which is about
1310 purge_pipe(struct dn_pipe *pipe)
1312 struct dn_pkt *pkt ;
1314 purge_flow_set( &(pipe->fs), 1 );
1316 for (pkt = pipe->head ; pkt ; )
1319 heap_free( &(pipe->scheduler_heap) );
1320 heap_free( &(pipe->not_eligible_heap) );
1321 heap_free( &(pipe->idle_heap) );
1325 * Delete all pipes and heaps returning memory. Must also
1326 * remove references from all ipfw rules to all pipes.
1329 dummynet_flush(void)
1331 struct dn_pipe *curr_p, *p ;
1332 struct dn_flow_set *fs, *curr_fs;
1336 /* remove all references to pipes ...*/
1337 flush_pipe_ptrs(NULL);
1338 /* prevent future matches... */
1341 fs = all_flow_sets ;
1342 all_flow_sets = NULL ;
1343 /* and free heaps so we don't have unwanted events */
1344 heap_free(&ready_heap);
1345 heap_free(&wfq_ready_heap);
1346 heap_free(&extract_heap);
1349 * Now purge all queued pkts and delete all pipes
1351 /* scan and purge all flow_sets. */
1355 purge_flow_set(curr_fs, 1);
1361 free(curr_p, M_DUMMYNET);
1366 extern struct ip_fw *ip_fw_default_rule ;
1368 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1371 struct dn_flow_queue *q ;
1372 struct dn_pkt *pkt ;
1374 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1375 for (q = fs->rq[i] ; q ; q = q->next )
1376 for (pkt = q->head ; pkt ; pkt = DN_NEXT(pkt) )
1378 pkt->rule = ip_fw_default_rule ;
1381 * when a firewall rule is deleted, scan all queues and remove the flow-id
1382 * from packets matching this rule.
1385 dn_rule_delete(void *r)
1388 struct dn_pkt *pkt ;
1389 struct dn_flow_set *fs ;
1392 * If the rule references a queue (dn_flow_set), then scan
1393 * the flow set, otherwise scan pipes. Should do either, but doing
1394 * both does not harm.
1396 for ( fs = all_flow_sets ; fs ; fs = fs->next )
1397 dn_rule_delete_fs(fs, r);
1398 for ( p = all_pipes ; p ; p = p->next ) {
1400 dn_rule_delete_fs(fs, r);
1401 for (pkt = p->head ; pkt ; pkt = DN_NEXT(pkt) )
1403 pkt->rule = ip_fw_default_rule ;
1408 * setup RED parameters
1411 config_red(struct dn_flow_set *p, struct dn_flow_set * x)
1416 x->min_th = SCALE(p->min_th);
1417 x->max_th = SCALE(p->max_th);
1418 x->max_p = p->max_p;
1420 x->c_1 = p->max_p / (p->max_th - p->min_th);
1421 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1422 if (x->flags_fs & DN_IS_GENTLE_RED) {
1423 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1424 x->c_4 = (SCALE(1) - 2 * p->max_p);
1427 /* if the lookup table already exist, free and create it again */
1428 if (x->w_q_lookup) {
1429 free(x->w_q_lookup, M_DUMMYNET);
1430 x->w_q_lookup = NULL ;
1432 if (red_lookup_depth == 0) {
1433 printf("\nnet.inet.ip.dummynet.red_lookup_depth must be > 0");
1434 free(x, M_DUMMYNET);
1437 x->lookup_depth = red_lookup_depth;
1438 x->w_q_lookup = malloc(x->lookup_depth * sizeof(int),
1439 M_DUMMYNET, M_WAITOK);
1441 /* fill the lookup table with (1 - w_q)^x */
1442 x->lookup_step = p->lookup_step ;
1443 x->lookup_weight = p->lookup_weight ;
1444 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1445 for (i = 1; i < x->lookup_depth; i++)
1446 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1447 if (red_avg_pkt_size < 1)
1448 red_avg_pkt_size = 512 ;
1449 x->avg_pkt_size = red_avg_pkt_size ;
1450 if (red_max_pkt_size < 1)
1451 red_max_pkt_size = 1500 ;
1452 x->max_pkt_size = red_max_pkt_size ;
1457 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1459 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1460 int l = pfs->rq_size;
1466 else if (l > DN_MAX_HASH_SIZE)
1467 l = DN_MAX_HASH_SIZE;
1469 } else /* one is enough for null mask */
1471 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1472 M_DUMMYNET, M_WAITOK | M_ZERO);
1478 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1480 x->flags_fs = src->flags_fs;
1481 x->qsize = src->qsize;
1483 x->flow_mask = src->flow_mask;
1484 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1485 if (x->qsize > 1024*1024)
1486 x->qsize = 1024*1024 ;
1493 /* configuring RED */
1494 if ( x->flags_fs & DN_IS_RED )
1495 config_red(src, x) ; /* XXX should check errors */
1499 * setup pipe or queue parameters.
1503 config_pipe(struct dn_pipe *p)
1506 struct dn_flow_set *pfs = &(p->fs);
1507 struct dn_flow_queue *q;
1510 * The config program passes parameters as follows:
1511 * bw = bits/second (0 means no limits),
1512 * delay = ms, must be translated into ticks.
1513 * qsize = slots/bytes
1515 p->delay = ( p->delay * hz ) / 1000 ;
1516 /* We need either a pipe number or a flow_set number */
1517 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1519 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1521 if (p->pipe_nr != 0) { /* this is a pipe */
1522 struct dn_pipe *x, *a, *b;
1524 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1525 a = b , b = b->next) ;
1527 if (b == NULL || b->pipe_nr != p->pipe_nr) { /* new pipe */
1528 x = malloc(sizeof(struct dn_pipe), M_DUMMYNET, M_WAITOK | M_ZERO);
1529 x->pipe_nr = p->pipe_nr;
1531 /* idle_heap is the only one from which we extract from the middle.
1533 x->idle_heap.size = x->idle_heap.elements = 0 ;
1534 x->idle_heap.offset=OFFSET_OF(struct dn_flow_queue, heap_pos);
1538 /* Flush accumulated credit for all queues */
1539 for (i = 0; i <= x->fs.rq_size; i++)
1540 for (q = x->fs.rq[i]; q; q = q->next)
1546 x->bandwidth = p->bandwidth ;
1547 x->numbytes = 0; /* just in case... */
1548 bcopy(p->if_name, x->if_name, sizeof(p->if_name) );
1549 x->ifp = NULL ; /* reset interface ptr */
1550 x->delay = p->delay ;
1551 set_fs_parms(&(x->fs), pfs);
1554 if ( x->fs.rq == NULL ) { /* a new pipe */
1555 s = alloc_hash(&(x->fs), pfs) ;
1557 free(x, M_DUMMYNET);
1567 } else { /* config queue */
1568 struct dn_flow_set *x, *a, *b ;
1570 /* locate flow_set */
1571 for (a=NULL, b=all_flow_sets ; b && b->fs_nr < pfs->fs_nr ;
1572 a = b , b = b->next) ;
1574 if (b == NULL || b->fs_nr != pfs->fs_nr) { /* new */
1575 if (pfs->parent_nr == 0) /* need link to a pipe */
1577 x = malloc(sizeof(struct dn_flow_set), M_DUMMYNET, M_WAITOK|M_ZERO);
1578 x->fs_nr = pfs->fs_nr;
1579 x->parent_nr = pfs->parent_nr;
1580 x->weight = pfs->weight ;
1583 else if (x->weight > 100)
1586 /* Change parent pipe not allowed; must delete and recreate */
1587 if (pfs->parent_nr != 0 && b->parent_nr != pfs->parent_nr)
1592 set_fs_parms(x, pfs);
1594 if ( x->rq == NULL ) { /* a new flow_set */
1595 s = alloc_hash(x, pfs) ;
1597 free(x, M_DUMMYNET);
1612 * Helper function to remove from a heap queues which are linked to
1613 * a flow_set about to be deleted.
1616 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1618 int i = 0, found = 0 ;
1619 for (; i < h->elements ;)
1620 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1622 h->p[i] = h->p[h->elements] ;
1631 * helper function to remove a pipe from a heap (can be there at most once)
1634 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1636 if (h->elements > 0) {
1638 for (i=0; i < h->elements ; i++ ) {
1639 if (h->p[i].object == p) { /* found it */
1641 h->p[i] = h->p[h->elements] ;
1650 * drain all queues. Called in case of severe mbuf shortage.
1653 dummynet_drain(void)
1655 struct dn_flow_set *fs;
1659 heap_free(&ready_heap);
1660 heap_free(&wfq_ready_heap);
1661 heap_free(&extract_heap);
1662 /* remove all references to this pipe from flow_sets */
1663 for (fs = all_flow_sets; fs; fs= fs->next )
1664 purge_flow_set(fs, 0);
1666 for (p = all_pipes; p; p= p->next ) {
1667 purge_flow_set(&(p->fs), 0);
1668 for (pkt = p->head ; pkt ; )
1670 p->head = p->tail = NULL ;
1675 * Fully delete a pipe or a queue, cleaning up associated info.
1678 delete_pipe(struct dn_pipe *p)
1680 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1682 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1684 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1685 struct dn_pipe *a, *b;
1686 struct dn_flow_set *fs;
1689 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1690 a = b , b = b->next) ;
1691 if (b == NULL || (b->pipe_nr != p->pipe_nr) )
1692 return EINVAL ; /* not found */
1696 /* unlink from list of pipes */
1698 all_pipes = b->next ;
1701 /* remove references to this pipe from the ip_fw rules. */
1702 flush_pipe_ptrs(&(b->fs));
1704 /* remove all references to this pipe from flow_sets */
1705 for (fs = all_flow_sets; fs; fs= fs->next )
1706 if (fs->pipe == b) {
1707 printf("++ ref to pipe %d from fs %d\n",
1708 p->pipe_nr, fs->fs_nr);
1710 purge_flow_set(fs, 0);
1712 fs_remove_from_heap(&ready_heap, &(b->fs));
1713 purge_pipe(b); /* remove all data associated to this pipe */
1714 /* remove reference to here from extract_heap and wfq_ready_heap */
1715 pipe_remove_from_heap(&extract_heap, b);
1716 pipe_remove_from_heap(&wfq_ready_heap, b);
1718 free(b, M_DUMMYNET);
1719 } else { /* this is a WF2Q queue (dn_flow_set) */
1720 struct dn_flow_set *a, *b;
1723 for (a = NULL, b = all_flow_sets ; b && b->fs_nr < p->fs.fs_nr ;
1724 a = b , b = b->next) ;
1725 if (b == NULL || (b->fs_nr != p->fs.fs_nr) )
1726 return EINVAL ; /* not found */
1730 all_flow_sets = b->next ;
1733 /* remove references to this flow_set from the ip_fw rules. */
1736 if (b->pipe != NULL) {
1737 /* Update total weight on parent pipe and cleanup parent heaps */
1738 b->pipe->sum -= b->weight * b->backlogged ;
1739 fs_remove_from_heap(&(b->pipe->not_eligible_heap), b);
1740 fs_remove_from_heap(&(b->pipe->scheduler_heap), b);
1741 #if 1 /* XXX should i remove from idle_heap as well ? */
1742 fs_remove_from_heap(&(b->pipe->idle_heap), b);
1745 purge_flow_set(b, 1);
1752 * helper function used to copy data from kernel in DUMMYNET_GET
1755 dn_copy_set(struct dn_flow_set *set, char *bp)
1758 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1760 for (i = 0 ; i <= set->rq_size ; i++)
1761 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
1762 if (q->hash_slot != i)
1763 printf("++ at %d: wrong slot (have %d, "
1764 "should be %d)\n", copied, q->hash_slot, i);
1766 printf("++ at %d: wrong fs ptr (have %p, should be %p)\n",
1769 bcopy(q, qp, sizeof( *q ) );
1770 /* cleanup pointers */
1772 qp->head = qp->tail = NULL ;
1775 if (copied != set->rq_elements)
1776 printf("++ wrong count, have %d should be %d\n",
1777 copied, set->rq_elements);
1782 dummynet_get(struct sockopt *sopt)
1784 char *buf, *bp ; /* bp is the "copy-pointer" */
1786 struct dn_flow_set *set ;
1792 * compute size of data structures: list of pipes and flow_sets.
1794 for (p = all_pipes, size = 0 ; p ; p = p->next )
1795 size += sizeof( *p ) +
1796 p->fs.rq_elements * sizeof(struct dn_flow_queue);
1797 for (set = all_flow_sets ; set ; set = set->next )
1798 size += sizeof ( *set ) +
1799 set->rq_elements * sizeof(struct dn_flow_queue);
1800 buf = malloc(size, M_TEMP, M_WAITOK);
1801 for (p = all_pipes, bp = buf ; p ; p = p->next ) {
1802 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp ;
1805 * copy pipe descriptor into *bp, convert delay back to ms,
1806 * then copy the flow_set descriptor(s) one at a time.
1807 * After each flow_set, copy the queue descriptor it owns.
1809 bcopy(p, bp, sizeof( *p ) );
1810 pipe_bp->delay = (pipe_bp->delay * 1000) / hz ;
1812 * XXX the following is a hack based on ->next being the
1813 * first field in dn_pipe and dn_flow_set. The correct
1814 * solution would be to move the dn_flow_set to the beginning
1815 * of struct dn_pipe.
1817 pipe_bp->next = (struct dn_pipe *)DN_IS_PIPE ;
1818 /* clean pointers */
1819 pipe_bp->head = pipe_bp->tail = NULL ;
1820 pipe_bp->fs.next = NULL ;
1821 pipe_bp->fs.pipe = NULL ;
1822 pipe_bp->fs.rq = NULL ;
1824 bp += sizeof( *p ) ;
1825 bp = dn_copy_set( &(p->fs), bp );
1827 for (set = all_flow_sets ; set ; set = set->next ) {
1828 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp ;
1829 bcopy(set, bp, sizeof( *set ) );
1830 /* XXX same hack as above */
1831 fs_bp->next = (struct dn_flow_set *)DN_IS_QUEUE ;
1832 fs_bp->pipe = NULL ;
1834 bp += sizeof( *set ) ;
1835 bp = dn_copy_set( set, bp );
1838 error = sooptcopyout(sopt, buf, size);
1844 * Handler for the various dummynet socket options (get, flush, config, del)
1847 ip_dn_ctl(struct sockopt *sopt)
1850 struct dn_pipe *p, tmp_pipe;
1852 /* Disallow sets in really-really secure mode. */
1853 if (sopt->sopt_dir == SOPT_SET) {
1854 #if defined(__FreeBSD__) && __FreeBSD_version >= 500034
1855 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
1859 if (securelevel >= 3)
1864 switch (sopt->sopt_name) {
1866 printf("ip_dn_ctl -- unknown option %d", sopt->sopt_name);
1869 case IP_DUMMYNET_GET :
1870 error = dummynet_get(sopt);
1873 case IP_DUMMYNET_FLUSH :
1877 case IP_DUMMYNET_CONFIGURE :
1879 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
1882 error = config_pipe(p);
1885 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
1887 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
1891 error = delete_pipe(p);
1900 printf("DUMMYNET initialized (011031)\n");
1902 all_flow_sets = NULL ;
1903 ready_heap.size = ready_heap.elements = 0 ;
1904 ready_heap.offset = 0 ;
1906 wfq_ready_heap.size = wfq_ready_heap.elements = 0 ;
1907 wfq_ready_heap.offset = 0 ;
1909 extract_heap.size = extract_heap.elements = 0 ;
1910 extract_heap.offset = 0 ;
1911 ip_dn_ctl_ptr = ip_dn_ctl;
1912 ip_dn_io_ptr = dummynet_io;
1913 ip_dn_ruledel_ptr = dn_rule_delete;
1914 callout_init(&dn_timeout);
1915 callout_reset(&dn_timeout, 1, dummynet, NULL);
1919 dummynet_modevent(module_t mod, int type, void *data)
1924 if (DUMMYNET_LOADED) {
1926 printf("DUMMYNET already loaded\n");
1934 #if !defined(KLD_MODULE)
1935 printf("dummynet statically compiled, cannot unload\n");
1939 callout_stop(&dn_timeout);
1941 ip_dn_ctl_ptr = NULL;
1942 ip_dn_io_ptr = NULL;
1943 ip_dn_ruledel_ptr = NULL;
1953 static moduledata_t dummynet_mod = {
1958 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);
1959 MODULE_DEPEND(dummynet, ipfw, 1, 1, 1);
1960 MODULE_VERSION(dummynet, 1);