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
34 * This module implements IP dummynet, a bandwidth limiter/delay emulator
35 * used in conjunction with the ipfw package.
36 * Description of the data structures used is in ip_dummynet.h
37 * Here you mainly find the following blocks of code:
38 * + variable declarations;
39 * + heap management functions;
40 * + scheduler and dummynet functions;
41 * + configuration and initialization.
43 * NOTA BENE: critical sections are protected by splimp()/splx()
44 * pairs. One would think that splnet() is enough as for most of
45 * the netinet code, but it is not so because when used with
46 * bridging, dummynet is invoked at splimp().
48 * Most important Changes:
51 * 010124: Fixed WF2Q behaviour
52 * 010122: Fixed spl protection.
53 * 000601: WF2Q support
54 * 000106: large rewrite, use heaps to handle very many pipes.
55 * 980513: initial release
57 * include files marked with XXX are probably not needed
60 #include <sys/param.h>
61 #include <sys/systm.h>
62 #include <sys/malloc.h>
64 #include <sys/kernel.h>
65 #include <sys/module.h>
67 #include <sys/socket.h>
68 #include <sys/socketvar.h>
70 #include <sys/sysctl.h>
72 #include <net/route.h>
73 #include <netinet/in.h>
74 #include <netinet/in_systm.h>
75 #include <netinet/in_var.h>
76 #include <netinet/ip.h>
77 #include <netinet/ip_fw.h>
78 #include <netinet/ip_dummynet.h>
79 #include <netinet/ip_var.h>
81 #include <netinet/if_ether.h> /* for struct arpcom */
82 #include <net/bridge.h>
85 * We keep a private variable for the simulation time, but we could
86 * probably use an existing one ("softticks" in sys/kern/kern_timer.c)
88 static dn_key curr_time = 0 ; /* current simulation time */
90 static int dn_hash_size = 64 ; /* default hash size */
92 /* statistics on number of queue searches and search steps */
93 static int searches, search_steps ;
94 static int pipe_expire = 1 ; /* expire queue if empty */
95 static int dn_max_ratio = 16 ; /* max queues/buckets ratio */
97 static int red_lookup_depth = 256; /* RED - default lookup table depth */
98 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
99 static int red_max_pkt_size = 1500; /* RED - default max packet size */
102 * Three heaps contain queues and pipes that the scheduler handles:
104 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
106 * wfq_ready_heap contains the pipes associated with WF2Q flows
108 * extract_heap contains pipes associated with delay lines.
112 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
114 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
116 static int heap_init(struct dn_heap *h, int size) ;
117 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
118 static void heap_extract(struct dn_heap *h, void *obj);
120 static void transmit_event(struct dn_pipe *pipe);
121 static void ready_event(struct dn_flow_queue *q);
123 static struct dn_pipe *all_pipes = NULL ; /* list of all pipes */
124 static struct dn_flow_set *all_flow_sets = NULL ;/* list of all flow_sets */
126 static struct callout_handle dn_timeout;
129 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
130 CTLFLAG_RW, 0, "Dummynet");
131 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
132 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
133 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time,
134 CTLFLAG_RD, &curr_time, 0, "Current tick");
135 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
136 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
137 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
138 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
139 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches,
140 CTLFLAG_RD, &searches, 0, "Number of queue searches");
141 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps,
142 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
143 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
144 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
145 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
146 CTLFLAG_RW, &dn_max_ratio, 0,
147 "Max ratio between dynamic queues and buckets");
148 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
149 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
150 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
151 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
152 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
153 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
156 static int config_pipe(struct dn_pipe *p);
157 static int ip_dn_ctl(struct sockopt *sopt);
159 static void rt_unref(struct rtentry *);
160 static void dummynet(void *);
161 static void dummynet_flush(void);
162 void dummynet_drain(void);
163 static ip_dn_io_t dummynet_io;
164 static void dn_rule_delete(void *);
166 int if_tx_rdy(struct ifnet *ifp);
169 rt_unref(struct rtentry *rt)
173 if (rt->rt_refcnt <= 0)
174 printf("-- warning, refcnt now %ld, decreasing\n", rt->rt_refcnt);
179 * Heap management functions.
181 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
182 * Some macros help finding parent/children so we can optimize them.
184 * heap_init() is called to expand the heap when needed.
185 * Increment size in blocks of 16 entries.
186 * XXX failure to allocate a new element is a pretty bad failure
187 * as we basically stall a whole queue forever!!
188 * Returns 1 on error, 0 on success
190 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
191 #define HEAP_LEFT(x) ( 2*(x) + 1 )
192 #define HEAP_IS_LEFT(x) ( (x) & 1 )
193 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
194 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
195 #define HEAP_INCREMENT 15
198 heap_init(struct dn_heap *h, int new_size)
200 struct dn_heap_entry *p;
202 if (h->size >= new_size ) {
203 printf("heap_init, Bogus call, have %d want %d\n",
207 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
208 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_DONTWAIT );
210 printf(" heap_init, resize %d failed\n", new_size );
211 return 1 ; /* error */
214 bcopy(h->p, p, h->size * sizeof(*p) );
215 free(h->p, M_DUMMYNET);
223 * Insert element in heap. Normally, p != NULL, we insert p in
224 * a new position and bubble up. If p == NULL, then the element is
225 * already in place, and key is the position where to start the
227 * Returns 1 on failure (cannot allocate new heap entry)
229 * If offset > 0 the position (index, int) of the element in the heap is
230 * also stored in the element itself at the given offset in bytes.
232 #define SET_OFFSET(heap, node) \
233 if (heap->offset > 0) \
234 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
236 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
238 #define RESET_OFFSET(heap, node) \
239 if (heap->offset > 0) \
240 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
242 heap_insert(struct dn_heap *h, dn_key key1, void *p)
244 int son = h->elements ;
246 if (p == NULL) /* data already there, set starting point */
248 else { /* insert new element at the end, possibly resize */
250 if (son == h->size) /* need resize... */
251 if (heap_init(h, h->elements+1) )
252 return 1 ; /* failure... */
253 h->p[son].object = p ;
254 h->p[son].key = key1 ;
257 while (son > 0) { /* bubble up */
258 int father = HEAP_FATHER(son) ;
259 struct dn_heap_entry tmp ;
261 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
262 break ; /* found right position */
263 /* son smaller than father, swap and repeat */
264 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
273 * remove top element from heap, or obj if obj != NULL
276 heap_extract(struct dn_heap *h, void *obj)
278 int child, father, max = h->elements - 1 ;
281 printf("warning, extract from empty heap 0x%p\n", h);
284 father = 0 ; /* default: move up smallest child */
285 if (obj != NULL) { /* extract specific element, index is at offset */
287 panic("*** heap_extract from middle not supported on this heap!!!\n");
288 father = *((int *)((char *)obj + h->offset)) ;
289 if (father < 0 || father >= h->elements) {
290 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
291 father, h->elements);
292 panic("heap_extract");
295 RESET_OFFSET(h, father);
296 child = HEAP_LEFT(father) ; /* left child */
297 while (child <= max) { /* valid entry */
298 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
299 child = child+1 ; /* take right child, otherwise left */
300 h->p[father] = h->p[child] ;
301 SET_OFFSET(h, father);
303 child = HEAP_LEFT(child) ; /* left child for next loop */
308 * Fill hole with last entry and bubble up, reusing the insert code
310 h->p[father] = h->p[max] ;
311 heap_insert(h, father, NULL); /* this one cannot fail */
317 * change object position and update references
318 * XXX this one is never used!
321 heap_move(struct dn_heap *h, dn_key new_key, void *object)
325 int max = h->elements-1 ;
326 struct dn_heap_entry buf ;
329 panic("cannot move items on this heap");
331 i = *((int *)((char *)object + h->offset));
332 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
333 h->p[i].key = new_key ;
334 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
335 i = temp ) { /* bubble up */
336 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
339 } else { /* must move down */
340 h->p[i].key = new_key ;
341 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
342 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
343 temp++ ; /* select child with min key */
344 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
345 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
354 #endif /* heap_move, unused */
357 * heapify() will reorganize data inside an array to maintain the
358 * heap property. It is needed when we delete a bunch of entries.
361 heapify(struct dn_heap *h)
365 for (i = 0 ; i < h->elements ; i++ )
366 heap_insert(h, i , NULL) ;
370 * cleanup the heap and free data structure
373 heap_free(struct dn_heap *h)
376 free(h->p, M_DUMMYNET);
377 bzero(h, sizeof(*h) );
381 * --- end of heap management functions ---
385 * Scheduler functions:
387 * transmit_event() is called when the delay-line needs to enter
388 * the scheduler, either because of existing pkts getting ready,
389 * or new packets entering the queue. The event handled is the delivery
390 * time of the packet.
392 * ready_event() does something similar with fixed-rate queues, and the
393 * event handled is the finish time of the head pkt.
395 * wfq_ready_event() does something similar with WF2Q queues, and the
396 * event handled is the start time of the head pkt.
398 * In all cases, we make sure that the data structures are consistent
399 * before passing pkts out, because this might trigger recursive
400 * invocations of the procedures.
403 transmit_event(struct dn_pipe *pipe)
407 while ( (pkt = pipe->head) && DN_KEY_LEQ(pkt->output_time, curr_time) ) {
409 * first unlink, then call procedures, since ip_input() can invoke
410 * ip_output() and viceversa, thus causing nested calls
412 pipe->head = DN_NEXT(pkt) ;
415 * The actual mbuf is preceded by a struct dn_pkt, resembling an mbuf
416 * (NOT A REAL one, just a small block of malloc'ed memory) with
417 * m_type = MT_TAG, m_flags = PACKET_TAG_DUMMYNET
418 * dn_m (m_next) = actual mbuf to be processed by ip_input/output
419 * and some other fields.
420 * The block IS FREED HERE because it contains parameters passed
421 * to the called routine.
423 switch (pkt->dn_dir) {
425 (void)ip_output((struct mbuf *)pkt, NULL, NULL, 0, NULL);
426 rt_unref (pkt->ro.ro_rt) ;
430 ip_input((struct mbuf *)pkt) ;
435 /* somebody unloaded the bridge module. Drop pkt */
436 printf("-- dropping bridged packet trapped in pipe--\n");
440 case DN_TO_ETH_DEMUX:
442 struct mbuf *m = (struct mbuf *)pkt ;
443 struct ether_header *eh;
445 if (pkt->dn_m->m_len < ETHER_HDR_LEN &&
446 (pkt->dn_m = m_pullup(pkt->dn_m, ETHER_HDR_LEN)) == NULL) {
447 printf("dummynet/bridge: pullup fail, dropping pkt\n");
451 * same as ether_input, make eh be a pointer into the mbuf
453 eh = mtod(pkt->dn_m, struct ether_header *);
454 m_adj(pkt->dn_m, ETHER_HDR_LEN);
456 * bdg_forward() wants a pointer to the pseudo-mbuf-header, but
457 * on return it will supply the pointer to the actual packet
458 * (originally pkt->dn_m, but could be something else now) if
459 * it has not consumed it.
461 if (pkt->dn_dir == DN_TO_BDG_FWD) {
462 m = bdg_forward_ptr(m, eh, pkt->ifp);
466 ether_demux(NULL, eh, m); /* which consumes the mbuf */
470 ether_output_frame(pkt->ifp, (struct mbuf *)pkt);
474 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
478 free(pkt, M_DUMMYNET);
480 /* if there are leftover packets, put into the heap for next event */
481 if ( (pkt = pipe->head) )
482 heap_insert(&extract_heap, pkt->output_time, pipe ) ;
483 /* XXX should check errors on heap_insert, by draining the
484 * whole pipe p and hoping in the future we are more successful
489 * the following macro computes how many ticks we have to wait
490 * before being able to transmit a packet. The credit is taken from
491 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
493 #define SET_TICKS(pkt, q, p) \
494 (pkt->dn_m->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \
498 * extract pkt from queue, compute output time (could be now)
499 * and put into delay line (p_queue)
502 move_pkt(struct dn_pkt *pkt, struct dn_flow_queue *q,
503 struct dn_pipe *p, int len)
505 q->head = DN_NEXT(pkt) ;
507 q->len_bytes -= len ;
509 pkt->output_time = curr_time + p->delay ;
514 DN_NEXT(p->tail) = pkt;
516 DN_NEXT(p->tail) = NULL;
520 * ready_event() is invoked every time the queue must enter the
521 * scheduler, either because the first packet arrives, or because
522 * a previously scheduled event fired.
523 * On invokation, drain as many pkts as possible (could be 0) and then
524 * if there are leftover packets reinsert the pkt in the scheduler.
527 ready_event(struct dn_flow_queue *q)
530 struct dn_pipe *p = q->fs->pipe ;
534 printf("ready_event- pipe is gone\n");
537 p_was_empty = (p->head == NULL) ;
540 * schedule fixed-rate queues linked to this pipe:
541 * Account for the bw accumulated since last scheduling, then
542 * drain as many pkts as allowed by q->numbytes and move to
543 * the delay line (in p) computing output time.
544 * bandwidth==0 (no limit) means we can drain the whole queue,
545 * setting len_scaled = 0 does the job.
547 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth;
548 while ( (pkt = q->head) != NULL ) {
549 int len = pkt->dn_m->m_pkthdr.len;
550 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
551 if (len_scaled > q->numbytes )
553 q->numbytes -= len_scaled ;
554 move_pkt(pkt, q, p, len);
557 * If we have more packets queued, schedule next ready event
558 * (can only occur when bandwidth != 0, otherwise we would have
559 * flushed the whole queue in the previous loop).
560 * To this purpose we record the current time and compute how many
561 * ticks to go for the finish time of the packet.
563 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */
564 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
565 q->sched_time = curr_time ;
566 heap_insert(&ready_heap, curr_time + t, (void *)q );
567 /* XXX should check errors on heap_insert, and drain the whole
568 * queue on error hoping next time we are luckier.
570 } else /* RED needs to know when the queue becomes empty */
571 q->q_time = curr_time;
573 * If the delay line was empty call transmit_event(p) now.
574 * Otherwise, the scheduler will take care of it.
581 * Called when we can transmit packets on WF2Q queues. Take pkts out of
582 * the queues at their start time, and enqueue into the delay line.
583 * Packets are drained until p->numbytes < 0. As long as
584 * len_scaled >= p->numbytes, the packet goes into the delay line
585 * with a deadline p->delay. For the last packet, if p->numbytes<0,
586 * there is an additional delay.
589 ready_event_wfq(struct dn_pipe *p)
591 int p_was_empty = (p->head == NULL) ;
592 struct dn_heap *sch = &(p->scheduler_heap);
593 struct dn_heap *neh = &(p->not_eligible_heap) ;
595 if (p->if_name[0] == 0) /* tx clock is simulated */
596 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth;
597 else { /* tx clock is for real, the ifq must be empty or this is a NOP */
598 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
601 DEB(printf("pipe %d ready from %s --\n",
602 p->pipe_nr, p->if_name);)
607 * While we have backlogged traffic AND credit, we need to do
608 * something on the queue.
610 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) {
611 if (sch->elements > 0) { /* have some eligible pkts to send out */
612 struct dn_flow_queue *q = sch->p[0].object ;
613 struct dn_pkt *pkt = q->head;
614 struct dn_flow_set *fs = q->fs;
615 u_int64_t len = pkt->dn_m->m_pkthdr.len;
616 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
618 heap_extract(sch, NULL); /* remove queue from heap */
619 p->numbytes -= len_scaled ;
620 move_pkt(pkt, q, p, len);
622 p->V += (len<<MY_M) / p->sum ; /* update V */
623 q->S = q->F ; /* update start time */
624 if (q->len == 0) { /* Flow not backlogged any more */
626 heap_insert(&(p->idle_heap), q->F, q);
627 } else { /* still backlogged */
629 * update F and position in backlogged queue, then
630 * put flow in not_eligible_heap (we will fix this later).
632 len = (q->head)->dn_m->m_pkthdr.len;
633 q->F += (len<<MY_M)/(u_int64_t) fs->weight ;
634 if (DN_KEY_LEQ(q->S, p->V))
635 heap_insert(neh, q->S, q);
637 heap_insert(sch, q->F, q);
641 * now compute V = max(V, min(S_i)). Remember that all elements in sch
642 * have by definition S_i <= V so if sch is not empty, V is surely
643 * the max and we must not update it. Conversely, if sch is empty
644 * we only need to look at neh.
646 if (sch->elements == 0 && neh->elements > 0)
647 p->V = MAX64 ( p->V, neh->p[0].key );
648 /* move from neh to sch any packets that have become eligible */
649 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) {
650 struct dn_flow_queue *q = neh->p[0].object ;
651 heap_extract(neh, NULL);
652 heap_insert(sch, q->F, q);
655 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */
656 p->numbytes = -1 ; /* mark not ready for I/O */
660 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0
661 && p->idle_heap.elements > 0) {
663 * no traffic and no events scheduled. We can get rid of idle-heap.
667 for (i = 0 ; i < p->idle_heap.elements ; i++) {
668 struct dn_flow_queue *q = p->idle_heap.p[i].object ;
675 p->idle_heap.elements = 0 ;
678 * If we are getting clocks from dummynet (not a real interface) and
679 * If we are under credit, schedule the next ready event.
680 * Also fix the delivery time of the last packet.
682 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */
683 dn_key t=0 ; /* number of ticks i have to wait */
685 if (p->bandwidth > 0)
686 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ;
687 p->tail->output_time += t ;
688 p->sched_time = curr_time ;
689 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
690 /* XXX should check errors on heap_insert, and drain the whole
691 * queue on error hoping next time we are luckier.
695 * If the delay line was empty call transmit_event(p) now.
696 * Otherwise, the scheduler will take care of it.
703 * This is called once per tick, or HZ times per second. It is used to
704 * increment the current tick counter and schedule expired events.
707 dummynet(void * __unused unused)
709 void *p ; /* generic parameter to handler */
712 struct dn_heap *heaps[3];
716 heaps[0] = &ready_heap ; /* fixed-rate queues */
717 heaps[1] = &wfq_ready_heap ; /* wfq queues */
718 heaps[2] = &extract_heap ; /* delay line */
719 s = splimp(); /* see note on top, splnet() is not enough */
721 for (i=0; i < 3 ; i++) {
723 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time) ) {
724 DDB(if (h->p[0].key > curr_time)
725 printf("-- dummynet: warning, heap %d is %d ticks late\n",
726 i, (int)(curr_time - h->p[0].key));)
727 p = h->p[0].object ; /* store a copy before heap_extract */
728 heap_extract(h, NULL); /* need to extract before processing */
732 struct dn_pipe *pipe = p;
733 if (pipe->if_name[0] != '\0')
734 printf("*** bad ready_event_wfq for pipe %s\n",
742 /* sweep pipes trying to expire idle flow_queues */
743 for (pe = all_pipes; pe ; pe = pe->next )
744 if (pe->idle_heap.elements > 0 &&
745 DN_KEY_LT(pe->idle_heap.p[0].key, pe->V) ) {
746 struct dn_flow_queue *q = pe->idle_heap.p[0].object ;
748 heap_extract(&(pe->idle_heap), NULL);
749 q->S = q->F + 1 ; /* mark timestamp as invalid */
750 pe->sum -= q->fs->weight ;
753 dn_timeout = timeout(dummynet, NULL, 1);
757 * called by an interface when tx_rdy occurs.
760 if_tx_rdy(struct ifnet *ifp)
764 for (p = all_pipes; p ; p = p->next )
769 sprintf(buf, "%s%d",ifp->if_name, ifp->if_unit);
770 for (p = all_pipes; p ; p = p->next )
771 if (!strcmp(p->if_name, buf) ) {
773 DEB(printf("++ tx rdy from %s (now found)\n", buf);)
778 DEB(printf("++ tx rdy from %s%d - qlen %d\n", ifp->if_name,
779 ifp->if_unit, ifp->if_snd.ifq_len);)
780 p->numbytes = 0 ; /* mark ready for I/O */
787 * Unconditionally expire empty queues in case of shortage.
788 * Returns the number of queues freed.
791 expire_queues(struct dn_flow_set *fs)
793 struct dn_flow_queue *q, *prev ;
794 int i, initial_elements = fs->rq_elements ;
796 if (fs->last_expired == time_second)
798 fs->last_expired = time_second ;
799 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
800 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
801 if (q->head != NULL || q->S != q->F+1) {
804 } else { /* entry is idle, expire it */
805 struct dn_flow_queue *old_q = q ;
808 prev->next = q = q->next ;
810 fs->rq[i] = q = q->next ;
812 free(old_q, M_DUMMYNET);
814 return initial_elements - fs->rq_elements ;
818 * If room, create a new queue and put at head of slot i;
819 * otherwise, create or use the default queue.
821 static struct dn_flow_queue *
822 create_queue(struct dn_flow_set *fs, int i)
824 struct dn_flow_queue *q ;
826 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
827 expire_queues(fs) == 0) {
829 * No way to get room, use or create overflow queue.
832 if ( fs->rq[i] != NULL )
835 q = malloc(sizeof(*q), M_DUMMYNET, M_DONTWAIT | M_ZERO);
837 printf("sorry, cannot allocate queue for new flow\n");
842 q->next = fs->rq[i] ;
843 q->S = q->F + 1; /* hack - mark timestamp as invalid */
850 * Given a flow_set and a pkt in last_pkt, find a matching queue
851 * after appropriate masking. The queue is moved to front
852 * so that further searches take less time.
854 static struct dn_flow_queue *
855 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
857 int i = 0 ; /* we need i and q for new allocations */
858 struct dn_flow_queue *q, *prev;
860 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
863 /* first, do the masking */
864 id->dst_ip &= fs->flow_mask.dst_ip ;
865 id->src_ip &= fs->flow_mask.src_ip ;
866 id->dst_port &= fs->flow_mask.dst_port ;
867 id->src_port &= fs->flow_mask.src_port ;
868 id->proto &= fs->flow_mask.proto ;
869 id->flags = 0 ; /* we don't care about this one */
870 /* then, hash function */
871 i = ( (id->dst_ip) & 0xffff ) ^
872 ( (id->dst_ip >> 15) & 0xffff ) ^
873 ( (id->src_ip << 1) & 0xffff ) ^
874 ( (id->src_ip >> 16 ) & 0xffff ) ^
875 (id->dst_port << 1) ^ (id->src_port) ^
877 i = i % fs->rq_size ;
878 /* finally, scan the current list for a match */
880 for (prev=NULL, q = fs->rq[i] ; q ; ) {
882 if (bcmp(id, &(q->id), sizeof(q->id) ) == 0)
884 else if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
885 /* entry is idle and not in any heap, expire it */
886 struct dn_flow_queue *old_q = q ;
889 prev->next = q = q->next ;
891 fs->rq[i] = q = q->next ;
893 free(old_q, M_DUMMYNET);
899 if (q && prev != NULL) { /* found and not in front */
900 prev->next = q->next ;
901 q->next = fs->rq[i] ;
905 if (q == NULL) { /* no match, need to allocate a new entry */
906 q = create_queue(fs, i);
914 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
919 * RED calculates the average queue size (avg) using a low-pass filter
920 * with an exponential weighted (w_q) moving average:
921 * avg <- (1-w_q) * avg + w_q * q_size
922 * where q_size is the queue length (measured in bytes or * packets).
924 * If q_size == 0, we compute the idle time for the link, and set
925 * avg = (1 - w_q)^(idle/s)
926 * where s is the time needed for transmitting a medium-sized packet.
928 * Now, if avg < min_th the packet is enqueued.
929 * If avg > max_th the packet is dropped. Otherwise, the packet is
930 * dropped with probability P function of avg.
935 /* queue in bytes or packets ? */
936 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len;
938 DEB(printf("\n%d q: %2u ", (int) curr_time, q_size);)
940 /* average queue size estimation */
943 * queue is not empty, avg <- avg + (q_size - avg) * w_q
945 int diff = SCALE(q_size) - q->avg;
946 int64_t v = SCALE_MUL((int64_t) diff, (int64_t) fs->w_q);
951 * queue is empty, find for how long the queue has been
952 * empty and use a lookup table for computing
953 * (1 - * w_q)^(idle_time/s) where s is the time to send a
958 u_int t = (curr_time - q->q_time) / fs->lookup_step;
960 q->avg = (t < fs->lookup_depth) ?
961 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
964 DEB(printf("avg: %u ", SCALE_VAL(q->avg));)
966 /* should i drop ? */
968 if (q->avg < fs->min_th) {
970 return 0; /* accept packet ; */
972 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
973 if (fs->flags_fs & DN_IS_GENTLE_RED) {
975 * According to Gentle-RED, if avg is greater than max_th the
976 * packet is dropped with a probability
977 * p_b = c_3 * avg - c_4
978 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
980 p_b = SCALE_MUL((int64_t) fs->c_3, (int64_t) q->avg) - fs->c_4;
986 } else if (q->avg > fs->min_th) {
988 * we compute p_b using the linear dropping function p_b = c_1 *
989 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
990 * max_p * min_th / (max_th - min_th)
992 p_b = SCALE_MUL((int64_t) fs->c_1, (int64_t) q->avg) - fs->c_2;
994 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
995 p_b = (p_b * len) / fs->max_pkt_size;
997 q->random = random() & 0xffff;
1000 * q->count counts packets arrived since last drop, so a greater
1001 * value of q->count means a greater packet drop probability.
1003 if (SCALE_MUL(p_b, SCALE((int64_t) q->count)) > q->random) {
1005 DEB(printf("- red drop");)
1006 /* after a drop we calculate a new random value */
1007 q->random = random() & 0xffff;
1008 return 1; /* drop */
1011 /* end of RED algorithm */
1012 return 0 ; /* accept */
1016 struct dn_flow_set *
1017 locate_flowset(int pipe_nr, struct ip_fw *rule)
1020 struct dn_flow_set *fs;
1021 ipfw_insn *cmd = rule->cmd + rule->act_ofs;
1023 if (cmd->opcode == O_LOG)
1025 fs = ((ipfw_insn_pipe *)cmd)->pipe_ptr;
1030 if (cmd->opcode == O_QUEUE)
1032 struct dn_flow_set *fs = NULL ;
1034 if ( (rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_QUEUE )
1036 for (fs=all_flow_sets; fs && fs->fs_nr != pipe_nr; fs=fs->next)
1040 for (p1 = all_pipes; p1 && p1->pipe_nr != pipe_nr; p1 = p1->next)
1045 /* record for the future */
1047 ((ipfw_insn_pipe *)cmd)->pipe_ptr = fs;
1050 rule->pipe_ptr = fs;
1056 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1057 * depending on whether WF2Q or fixed bw is used.
1059 * pipe_nr pipe or queue the packet is destined for.
1060 * dir where shall we send the packet after dummynet.
1061 * m the mbuf with the packet
1062 * ifp the 'ifp' parameter from the caller.
1063 * NULL in ip_input, destination interface in ip_output,
1064 * real_dst in bdg_forward
1065 * ro route parameter (only used in ip_output, NULL otherwise)
1066 * dst destination address, only used by ip_output
1067 * rule matching rule, in case of multiple passes
1068 * flags flags from the caller, only used in ip_output
1072 dummynet_io(struct mbuf *m, int pipe_nr, int dir, struct ip_fw_args *fwa)
1075 struct dn_flow_set *fs;
1076 struct dn_pipe *pipe ;
1077 u_int64_t len = m->m_pkthdr.len ;
1078 struct dn_flow_queue *q = NULL ;
1082 ipfw_insn *cmd = fwa->rule->cmd + fwa->rule->act_ofs;
1084 if (cmd->opcode == O_LOG)
1086 is_pipe = (cmd->opcode == O_PIPE);
1088 is_pipe = (fwa->rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_PIPE;
1094 * this is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule
1096 fs = locate_flowset(pipe_nr, fwa->rule);
1098 goto dropit ; /* this queue/pipe does not exist! */
1100 if (pipe == NULL) { /* must be a queue, try find a matching pipe */
1101 for (pipe = all_pipes; pipe && pipe->pipe_nr != fs->parent_nr;
1107 printf("No pipe %d for queue %d, drop pkt\n",
1108 fs->parent_nr, fs->fs_nr);
1112 q = find_queue(fs, &(fwa->f_id));
1114 goto dropit ; /* cannot allocate queue */
1116 * update statistics, then check reasons to drop pkt
1118 q->tot_bytes += len ;
1120 if ( fs->plr && random() < fs->plr )
1121 goto dropit ; /* random pkt drop */
1122 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) {
1123 if (q->len_bytes > fs->qsize)
1124 goto dropit ; /* queue size overflow */
1126 if (q->len >= fs->qsize)
1127 goto dropit ; /* queue count overflow */
1129 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) )
1132 /* XXX expensive to zero, see if we can remove it*/
1133 pkt = (struct dn_pkt *)malloc(sizeof (*pkt), M_DUMMYNET, M_NOWAIT|M_ZERO);
1135 goto dropit ; /* cannot allocate packet header */
1136 /* ok, i can handle the pkt now... */
1137 /* build and enqueue packet + parameters */
1138 pkt->hdr.mh_type = MT_TAG;
1139 pkt->hdr.mh_flags = PACKET_TAG_DUMMYNET;
1140 pkt->rule = fwa->rule ;
1141 DN_NEXT(pkt) = NULL;
1145 pkt->ifp = fwa->oif;
1146 if (dir == DN_TO_IP_OUT) {
1148 * We need to copy *ro because for ICMP pkts (and maybe others)
1149 * the caller passed a pointer into the stack; dst might also be
1150 * a pointer into *ro so it needs to be updated.
1152 pkt->ro = *(fwa->ro);
1154 fwa->ro->ro_rt->rt_refcnt++ ;
1155 if (fwa->dst == (struct sockaddr_in *)&fwa->ro->ro_dst) /* dst points into ro */
1156 fwa->dst = (struct sockaddr_in *)&(pkt->ro.ro_dst) ;
1158 pkt->dn_dst = fwa->dst;
1159 pkt->flags = fwa->flags;
1161 if (q->head == NULL)
1164 DN_NEXT(q->tail) = pkt;
1167 q->len_bytes += len ;
1169 if ( q->head != pkt ) /* flow was not idle, we are done */
1172 * If we reach this point the flow was previously idle, so we need
1173 * to schedule it. This involves different actions for fixed-rate or
1178 * Fixed-rate queue: just insert into the ready_heap.
1181 if (pipe->bandwidth)
1182 t = SET_TICKS(pkt, q, pipe);
1183 q->sched_time = curr_time ;
1184 if (t == 0) /* must process it now */
1187 heap_insert(&ready_heap, curr_time + t , q );
1190 * WF2Q. First, compute start time S: if the flow was idle (S=F+1)
1191 * set S to the virtual time V for the controlling pipe, and update
1192 * the sum of weights for the pipe; otherwise, remove flow from
1193 * idle_heap and set S to max(F,V).
1194 * Second, compute finish time F = S + len/weight.
1195 * Third, if pipe was idle, update V=max(S, V).
1196 * Fourth, count one more backlogged flow.
1198 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */
1200 pipe->sum += fs->weight ; /* add weight of new queue */
1202 heap_extract(&(pipe->idle_heap), q);
1203 q->S = MAX64(q->F, pipe->V ) ;
1205 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight;
1207 if (pipe->not_eligible_heap.elements == 0 &&
1208 pipe->scheduler_heap.elements == 0)
1209 pipe->V = MAX64 ( q->S, pipe->V );
1212 * Look at eligibility. A flow is not eligibile if S>V (when
1213 * this happens, it means that there is some other flow already
1214 * scheduled for the same pipe, so the scheduler_heap cannot be
1215 * empty). If the flow is not eligible we just store it in the
1216 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1217 * and possibly invoke ready_event_wfq() right now if there is
1219 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1220 * and for all flows in not_eligible_heap (NEH), S_i > V .
1221 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH,
1222 * we only need to look into NEH.
1224 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */
1225 if (pipe->scheduler_heap.elements == 0)
1226 printf("++ ouch! not eligible but empty scheduler!\n");
1227 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1229 heap_insert(&(pipe->scheduler_heap), q->F, q);
1230 if (pipe->numbytes >= 0) { /* pipe is idle */
1231 if (pipe->scheduler_heap.elements != 1)
1232 printf("*** OUCH! pipe should have been idle!\n");
1233 DEB(printf("Waking up pipe %d at %d\n",
1234 pipe->pipe_nr, (int)(q->F >> MY_M)); )
1235 pipe->sched_time = curr_time ;
1236 ready_event_wfq(pipe);
1253 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1254 * Doing this would probably save us the initial bzero of dn_pkt
1256 #define DN_FREE_PKT(pkt) { \
1257 struct dn_pkt *n = pkt ; \
1258 rt_unref ( n->ro.ro_rt ) ; \
1260 pkt = DN_NEXT(n) ; \
1261 free(n, M_DUMMYNET) ; }
1264 * Dispose all packets and flow_queues on a flow_set.
1265 * If all=1, also remove red lookup table and other storage,
1266 * including the descriptor itself.
1267 * For the one in dn_pipe MUST also cleanup ready_heap...
1270 purge_flow_set(struct dn_flow_set *fs, int all)
1272 struct dn_pkt *pkt ;
1273 struct dn_flow_queue *q, *qn ;
1276 for (i = 0 ; i <= fs->rq_size ; i++ ) {
1277 for (q = fs->rq[i] ; q ; q = qn ) {
1278 for (pkt = q->head ; pkt ; )
1281 free(q, M_DUMMYNET);
1285 fs->rq_elements = 0 ;
1287 /* RED - free lookup table */
1289 free(fs->w_q_lookup, M_DUMMYNET);
1291 free(fs->rq, M_DUMMYNET);
1292 /* if this fs is not part of a pipe, free it */
1293 if (fs->pipe && fs != &(fs->pipe->fs) )
1294 free(fs, M_DUMMYNET);
1299 * Dispose all packets queued on a pipe (not a flow_set).
1300 * Also free all resources associated to a pipe, which is about
1304 purge_pipe(struct dn_pipe *pipe)
1306 struct dn_pkt *pkt ;
1308 purge_flow_set( &(pipe->fs), 1 );
1310 for (pkt = pipe->head ; pkt ; )
1313 heap_free( &(pipe->scheduler_heap) );
1314 heap_free( &(pipe->not_eligible_heap) );
1315 heap_free( &(pipe->idle_heap) );
1319 * Delete all pipes and heaps returning memory. Must also
1320 * remove references from all ipfw rules to all pipes.
1325 struct dn_pipe *curr_p, *p ;
1326 struct dn_flow_set *fs, *curr_fs;
1331 /* remove all references to pipes ...*/
1332 flush_pipe_ptrs(NULL);
1333 /* prevent future matches... */
1336 fs = all_flow_sets ;
1337 all_flow_sets = NULL ;
1338 /* and free heaps so we don't have unwanted events */
1339 heap_free(&ready_heap);
1340 heap_free(&wfq_ready_heap);
1341 heap_free(&extract_heap);
1344 * Now purge all queued pkts and delete all pipes
1346 /* scan and purge all flow_sets. */
1350 purge_flow_set(curr_fs, 1);
1356 free(curr_p, M_DUMMYNET);
1361 extern struct ip_fw *ip_fw_default_rule ;
1363 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1366 struct dn_flow_queue *q ;
1367 struct dn_pkt *pkt ;
1369 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1370 for (q = fs->rq[i] ; q ; q = q->next )
1371 for (pkt = q->head ; pkt ; pkt = DN_NEXT(pkt) )
1373 pkt->rule = ip_fw_default_rule ;
1376 * when a firewall rule is deleted, scan all queues and remove the flow-id
1377 * from packets matching this rule.
1380 dn_rule_delete(void *r)
1383 struct dn_pkt *pkt ;
1384 struct dn_flow_set *fs ;
1387 * If the rule references a queue (dn_flow_set), then scan
1388 * the flow set, otherwise scan pipes. Should do either, but doing
1389 * both does not harm.
1391 for ( fs = all_flow_sets ; fs ; fs = fs->next )
1392 dn_rule_delete_fs(fs, r);
1393 for ( p = all_pipes ; p ; p = p->next ) {
1395 dn_rule_delete_fs(fs, r);
1396 for (pkt = p->head ; pkt ; pkt = DN_NEXT(pkt) )
1398 pkt->rule = ip_fw_default_rule ;
1403 * setup RED parameters
1406 config_red(struct dn_flow_set *p, struct dn_flow_set * x)
1411 x->min_th = SCALE(p->min_th);
1412 x->max_th = SCALE(p->max_th);
1413 x->max_p = p->max_p;
1415 x->c_1 = p->max_p / (p->max_th - p->min_th);
1416 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1417 if (x->flags_fs & DN_IS_GENTLE_RED) {
1418 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1419 x->c_4 = (SCALE(1) - 2 * p->max_p);
1422 /* if the lookup table already exist, free and create it again */
1423 if (x->w_q_lookup) {
1424 free(x->w_q_lookup, M_DUMMYNET);
1425 x->w_q_lookup = NULL ;
1427 if (red_lookup_depth == 0) {
1428 printf("\nnet.inet.ip.dummynet.red_lookup_depth must be > 0");
1429 free(x, M_DUMMYNET);
1432 x->lookup_depth = red_lookup_depth;
1433 x->w_q_lookup = (u_int *) malloc(x->lookup_depth * sizeof(int),
1434 M_DUMMYNET, M_DONTWAIT);
1435 if (x->w_q_lookup == NULL) {
1436 printf("sorry, cannot allocate red lookup table\n");
1437 free(x, M_DUMMYNET);
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;
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_DONTWAIT | M_ZERO);
1473 if (x->rq == NULL) {
1474 printf("sorry, cannot allocate queue\n");
1482 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1484 x->flags_fs = src->flags_fs;
1485 x->qsize = src->qsize;
1487 x->flow_mask = src->flow_mask;
1488 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1489 if (x->qsize > 1024*1024)
1490 x->qsize = 1024*1024 ;
1497 /* configuring RED */
1498 if ( x->flags_fs & DN_IS_RED )
1499 config_red(src, x) ; /* XXX should check errors */
1503 * setup pipe or queue parameters.
1507 config_pipe(struct dn_pipe *p)
1510 struct dn_flow_set *pfs = &(p->fs);
1513 * The config program passes parameters as follows:
1514 * bw = bits/second (0 means no limits),
1515 * delay = ms, must be translated into ticks.
1516 * qsize = slots/bytes
1518 p->delay = ( p->delay * hz ) / 1000 ;
1519 /* We need either a pipe number or a flow_set number */
1520 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1522 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1524 if (p->pipe_nr != 0) { /* this is a pipe */
1525 struct dn_pipe *x, *a, *b;
1527 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1528 a = b , b = b->next) ;
1530 if (b == NULL || b->pipe_nr != p->pipe_nr) { /* new pipe */
1531 x = malloc(sizeof(struct dn_pipe), M_DUMMYNET, M_DONTWAIT | M_ZERO);
1533 printf("ip_dummynet.c: no memory for new pipe\n");
1536 x->pipe_nr = p->pipe_nr;
1538 /* idle_heap is the only one from which we extract from the middle.
1540 x->idle_heap.size = x->idle_heap.elements = 0 ;
1541 x->idle_heap.offset=OFFSET_OF(struct dn_flow_queue, heap_pos);
1545 x->bandwidth = p->bandwidth ;
1546 x->numbytes = 0; /* just in case... */
1547 bcopy(p->if_name, x->if_name, sizeof(p->if_name) );
1548 x->ifp = NULL ; /* reset interface ptr */
1549 x->delay = p->delay ;
1550 set_fs_parms(&(x->fs), pfs);
1553 if ( x->fs.rq == NULL ) { /* a new pipe */
1554 s = alloc_hash(&(x->fs), pfs) ;
1556 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_DONTWAIT|M_ZERO);
1579 printf("ip_dummynet.c: no memory for new flow_set\n");
1582 x->fs_nr = pfs->fs_nr;
1583 x->parent_nr = pfs->parent_nr;
1584 x->weight = pfs->weight ;
1587 else if (x->weight > 100)
1590 /* Change parent pipe not allowed; must delete and recreate */
1591 if (pfs->parent_nr != 0 && b->parent_nr != pfs->parent_nr)
1595 set_fs_parms(x, pfs);
1597 if ( x->rq == NULL ) { /* a new flow_set */
1598 s = alloc_hash(x, pfs) ;
1600 free(x, M_DUMMYNET);
1616 * Helper function to remove from a heap queues which are linked to
1617 * a flow_set about to be deleted.
1620 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1622 int i = 0, found = 0 ;
1623 for (; i < h->elements ;)
1624 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1626 h->p[i] = h->p[h->elements] ;
1635 * helper function to remove a pipe from a heap (can be there at most once)
1638 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1640 if (h->elements > 0) {
1642 for (i=0; i < h->elements ; i++ ) {
1643 if (h->p[i].object == p) { /* found it */
1645 h->p[i] = h->p[h->elements] ;
1654 * drain all queues. Called in case of severe mbuf shortage.
1659 struct dn_flow_set *fs;
1663 heap_free(&ready_heap);
1664 heap_free(&wfq_ready_heap);
1665 heap_free(&extract_heap);
1666 /* remove all references to this pipe from flow_sets */
1667 for (fs = all_flow_sets; fs; fs= fs->next )
1668 purge_flow_set(fs, 0);
1670 for (p = all_pipes; p; p= p->next ) {
1671 purge_flow_set(&(p->fs), 0);
1672 for (pkt = p->head ; pkt ; )
1674 p->head = p->tail = NULL ;
1679 * Fully delete a pipe or a queue, cleaning up associated info.
1682 delete_pipe(struct dn_pipe *p)
1686 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1688 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1690 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1691 struct dn_pipe *a, *b;
1692 struct dn_flow_set *fs;
1695 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1696 a = b , b = b->next) ;
1697 if (b == NULL || (b->pipe_nr != p->pipe_nr) )
1698 return EINVAL ; /* not found */
1702 /* unlink from list of pipes */
1704 all_pipes = b->next ;
1707 /* remove references to this pipe from the ip_fw rules. */
1708 flush_pipe_ptrs(&(b->fs));
1710 /* remove all references to this pipe from flow_sets */
1711 for (fs = all_flow_sets; fs; fs= fs->next )
1712 if (fs->pipe == b) {
1713 printf("++ ref to pipe %d from fs %d\n",
1714 p->pipe_nr, fs->fs_nr);
1716 purge_flow_set(fs, 0);
1718 fs_remove_from_heap(&ready_heap, &(b->fs));
1719 purge_pipe(b); /* remove all data associated to this pipe */
1720 /* remove reference to here from extract_heap and wfq_ready_heap */
1721 pipe_remove_from_heap(&extract_heap, b);
1722 pipe_remove_from_heap(&wfq_ready_heap, b);
1724 free(b, M_DUMMYNET);
1725 } else { /* this is a WF2Q queue (dn_flow_set) */
1726 struct dn_flow_set *a, *b;
1729 for (a = NULL, b = all_flow_sets ; b && b->fs_nr < p->fs.fs_nr ;
1730 a = b , b = b->next) ;
1731 if (b == NULL || (b->fs_nr != p->fs.fs_nr) )
1732 return EINVAL ; /* not found */
1736 all_flow_sets = b->next ;
1739 /* remove references to this flow_set from the ip_fw rules. */
1742 if (b->pipe != NULL) {
1743 /* Update total weight on parent pipe and cleanup parent heaps */
1744 b->pipe->sum -= b->weight * b->backlogged ;
1745 fs_remove_from_heap(&(b->pipe->not_eligible_heap), b);
1746 fs_remove_from_heap(&(b->pipe->scheduler_heap), b);
1747 #if 1 /* XXX should i remove from idle_heap as well ? */
1748 fs_remove_from_heap(&(b->pipe->idle_heap), b);
1751 purge_flow_set(b, 1);
1758 * helper function used to copy data from kernel in DUMMYNET_GET
1761 dn_copy_set(struct dn_flow_set *set, char *bp)
1764 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1766 for (i = 0 ; i <= set->rq_size ; i++)
1767 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
1768 if (q->hash_slot != i)
1769 printf("++ at %d: wrong slot (have %d, "
1770 "should be %d)\n", copied, q->hash_slot, i);
1772 printf("++ at %d: wrong fs ptr (have %p, should be %p)\n",
1775 bcopy(q, qp, sizeof( *q ) );
1776 /* cleanup pointers */
1778 qp->head = qp->tail = NULL ;
1781 if (copied != set->rq_elements)
1782 printf("++ wrong count, have %d should be %d\n",
1783 copied, set->rq_elements);
1788 dummynet_get(struct sockopt *sopt)
1790 char *buf, *bp ; /* bp is the "copy-pointer" */
1792 struct dn_flow_set *set ;
1798 * compute size of data structures: list of pipes and flow_sets.
1800 for (p = all_pipes, size = 0 ; p ; p = p->next )
1801 size += sizeof( *p ) +
1802 p->fs.rq_elements * sizeof(struct dn_flow_queue);
1803 for (set = all_flow_sets ; set ; set = set->next )
1804 size += sizeof ( *set ) +
1805 set->rq_elements * sizeof(struct dn_flow_queue);
1806 buf = malloc(size, M_TEMP, M_DONTWAIT);
1811 for (p = all_pipes, bp = buf ; p ; p = p->next ) {
1812 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp ;
1815 * copy pipe descriptor into *bp, convert delay back to ms,
1816 * then copy the flow_set descriptor(s) one at a time.
1817 * After each flow_set, copy the queue descriptor it owns.
1819 bcopy(p, bp, sizeof( *p ) );
1820 pipe_bp->delay = (pipe_bp->delay * 1000) / hz ;
1822 * XXX the following is a hack based on ->next being the
1823 * first field in dn_pipe and dn_flow_set. The correct
1824 * solution would be to move the dn_flow_set to the beginning
1825 * of struct dn_pipe.
1827 pipe_bp->next = (struct dn_pipe *)DN_IS_PIPE ;
1828 /* clean pointers */
1829 pipe_bp->head = pipe_bp->tail = NULL ;
1830 pipe_bp->fs.next = NULL ;
1831 pipe_bp->fs.pipe = NULL ;
1832 pipe_bp->fs.rq = NULL ;
1834 bp += sizeof( *p ) ;
1835 bp = dn_copy_set( &(p->fs), bp );
1837 for (set = all_flow_sets ; set ; set = set->next ) {
1838 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp ;
1839 bcopy(set, bp, sizeof( *set ) );
1840 /* XXX same hack as above */
1841 fs_bp->next = (struct dn_flow_set *)DN_IS_QUEUE ;
1842 fs_bp->pipe = NULL ;
1844 bp += sizeof( *set ) ;
1845 bp = dn_copy_set( set, bp );
1848 error = sooptcopyout(sopt, buf, size);
1854 * Handler for the various dummynet socket options (get, flush, config, del)
1857 ip_dn_ctl(struct sockopt *sopt)
1860 struct dn_pipe *p, tmp_pipe;
1862 /* Disallow sets in really-really secure mode. */
1863 if (sopt->sopt_dir == SOPT_SET) {
1864 #if __FreeBSD_version >= 500034
1865 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
1869 if (securelevel >= 3)
1874 switch (sopt->sopt_name) {
1876 printf("ip_dn_ctl -- unknown option %d", sopt->sopt_name);
1879 case IP_DUMMYNET_GET :
1880 error = dummynet_get(sopt);
1883 case IP_DUMMYNET_FLUSH :
1887 case IP_DUMMYNET_CONFIGURE :
1889 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
1892 error = config_pipe(p);
1895 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
1897 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
1901 error = delete_pipe(p);
1910 printf("DUMMYNET initialized (011031)\n");
1912 all_flow_sets = NULL ;
1913 ready_heap.size = ready_heap.elements = 0 ;
1914 ready_heap.offset = 0 ;
1916 wfq_ready_heap.size = wfq_ready_heap.elements = 0 ;
1917 wfq_ready_heap.offset = 0 ;
1919 extract_heap.size = extract_heap.elements = 0 ;
1920 extract_heap.offset = 0 ;
1921 ip_dn_ctl_ptr = ip_dn_ctl;
1922 ip_dn_io_ptr = dummynet_io;
1923 ip_dn_ruledel_ptr = dn_rule_delete;
1924 bzero(&dn_timeout, sizeof(struct callout_handle));
1925 dn_timeout = timeout(dummynet, NULL, 1);
1929 dummynet_modevent(module_t mod, int type, void *data)
1935 if (DUMMYNET_LOADED) {
1937 printf("DUMMYNET already loaded\n");
1945 #if !defined(KLD_MODULE)
1946 printf("dummynet statically compiled, cannot unload\n");
1950 untimeout(dummynet, NULL, dn_timeout);
1952 ip_dn_ctl_ptr = NULL;
1953 ip_dn_io_ptr = NULL;
1954 ip_dn_ruledel_ptr = NULL;
1964 static moduledata_t dummynet_mod = {
1969 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);
1970 MODULE_DEPEND(dummynet, ipfw, 1, 1, 1);
1971 MODULE_VERSION(dummynet, 1);