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 1.21 2006/09/05 03:48:12 dillon 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 */
90 * We keep a private variable for the simulation time, but we could
91 * probably use an existing one ("softticks" in sys/kern/kern_timer.c)
93 static dn_key curr_time = 0 ; /* current simulation time */
95 static int dn_hash_size = 64 ; /* default hash size */
97 /* statistics on number of queue searches and search steps */
98 static int searches, search_steps ;
99 static int pipe_expire = 1 ; /* expire queue if empty */
100 static int dn_max_ratio = 16 ; /* max queues/buckets ratio */
102 static int red_lookup_depth = 256; /* RED - default lookup table depth */
103 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
104 static int red_max_pkt_size = 1500; /* RED - default max packet size */
107 * Three heaps contain queues and pipes that the scheduler handles:
109 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
111 * wfq_ready_heap contains the pipes associated with WF2Q flows
113 * extract_heap contains pipes associated with delay lines.
117 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
119 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
121 static int heap_init(struct dn_heap *h, int size) ;
122 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
123 static void heap_extract(struct dn_heap *h, void *obj);
125 static void transmit_event(struct dn_pipe *pipe);
126 static void ready_event(struct dn_flow_queue *q);
128 static struct dn_pipe *all_pipes = NULL ; /* list of all pipes */
129 static struct dn_flow_set *all_flow_sets = NULL ;/* list of all flow_sets */
131 static struct callout dn_timeout;
134 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
135 CTLFLAG_RW, 0, "Dummynet");
136 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
137 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
138 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time,
139 CTLFLAG_RD, &curr_time, 0, "Current tick");
140 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
141 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
142 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
143 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
144 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches,
145 CTLFLAG_RD, &searches, 0, "Number of queue searches");
146 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps,
147 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
148 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
149 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
150 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
151 CTLFLAG_RW, &dn_max_ratio, 0,
152 "Max ratio between dynamic queues and buckets");
153 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
154 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
155 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
156 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
157 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
158 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
161 static int config_pipe(struct dn_pipe *p);
162 static int ip_dn_ctl(struct sockopt *sopt);
164 static void rt_unref(struct rtentry *);
165 static void dummynet(void *);
166 static void dummynet_flush(void);
167 void dummynet_drain(void);
168 static ip_dn_io_t dummynet_io;
169 static void dn_rule_delete(void *);
171 int if_tx_rdy(struct ifnet *ifp);
174 rt_unref(struct rtentry *rt)
178 if (rt->rt_refcnt <= 0)
179 printf("-- warning, refcnt now %ld, decreasing\n", rt->rt_refcnt);
184 * Heap management functions.
186 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
187 * Some macros help finding parent/children so we can optimize them.
189 * heap_init() is called to expand the heap when needed.
190 * Increment size in blocks of 16 entries.
191 * XXX failure to allocate a new element is a pretty bad failure
192 * as we basically stall a whole queue forever!!
193 * Returns 1 on error, 0 on success
195 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
196 #define HEAP_LEFT(x) ( 2*(x) + 1 )
197 #define HEAP_IS_LEFT(x) ( (x) & 1 )
198 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
199 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
200 #define HEAP_INCREMENT 15
203 heap_init(struct dn_heap *h, int new_size)
205 struct dn_heap_entry *p;
207 if (h->size >= new_size ) {
208 printf("heap_init, Bogus call, have %d want %d\n",
212 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
213 p = kmalloc(new_size * sizeof(*p), M_DUMMYNET, M_WAITOK | M_ZERO);
215 bcopy(h->p, p, h->size * sizeof(*p) );
216 kfree(h->p, M_DUMMYNET);
224 * Insert element in heap. Normally, p != NULL, we insert p in
225 * a new position and bubble up. If p == NULL, then the element is
226 * already in place, and key is the position where to start the
228 * Returns 1 on failure (cannot allocate new heap entry)
230 * If offset > 0 the position (index, int) of the element in the heap is
231 * also stored in the element itself at the given offset in bytes.
233 #define SET_OFFSET(heap, node) \
234 if (heap->offset > 0) \
235 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
237 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
239 #define RESET_OFFSET(heap, node) \
240 if (heap->offset > 0) \
241 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
243 heap_insert(struct dn_heap *h, dn_key key1, void *p)
245 int son = h->elements ;
247 if (p == NULL) /* data already there, set starting point */
249 else { /* insert new element at the end, possibly resize */
251 if (son == h->size) /* need resize... */
252 if (heap_init(h, h->elements+1) )
253 return 1 ; /* failure... */
254 h->p[son].object = p ;
255 h->p[son].key = key1 ;
258 while (son > 0) { /* bubble up */
259 int father = HEAP_FATHER(son) ;
260 struct dn_heap_entry tmp ;
262 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
263 break ; /* found right position */
264 /* son smaller than father, swap and repeat */
265 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
274 * remove top element from heap, or obj if obj != NULL
277 heap_extract(struct dn_heap *h, void *obj)
279 int child, father, max = h->elements - 1 ;
282 printf("warning, extract from empty heap 0x%p\n", h);
285 father = 0 ; /* default: move up smallest child */
286 if (obj != NULL) { /* extract specific element, index is at offset */
288 panic("*** heap_extract from middle not supported on this heap!!!\n");
289 father = *((int *)((char *)obj + h->offset)) ;
290 if (father < 0 || father >= h->elements) {
291 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
292 father, h->elements);
293 panic("heap_extract");
296 RESET_OFFSET(h, father);
297 child = HEAP_LEFT(father) ; /* left child */
298 while (child <= max) { /* valid entry */
299 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
300 child = child+1 ; /* take right child, otherwise left */
301 h->p[father] = h->p[child] ;
302 SET_OFFSET(h, father);
304 child = HEAP_LEFT(child) ; /* left child for next loop */
309 * Fill hole with last entry and bubble up, reusing the insert code
311 h->p[father] = h->p[max] ;
312 heap_insert(h, father, NULL); /* this one cannot fail */
318 * change object position and update references
319 * XXX this one is never used!
322 heap_move(struct dn_heap *h, dn_key new_key, void *object)
326 int max = h->elements-1 ;
327 struct dn_heap_entry buf ;
330 panic("cannot move items on this heap");
332 i = *((int *)((char *)object + h->offset));
333 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
334 h->p[i].key = new_key ;
335 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
336 i = temp ) { /* bubble up */
337 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
340 } else { /* must move down */
341 h->p[i].key = new_key ;
342 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
343 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
344 temp++ ; /* select child with min key */
345 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
346 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
355 #endif /* heap_move, unused */
358 * heapify() will reorganize data inside an array to maintain the
359 * heap property. It is needed when we delete a bunch of entries.
362 heapify(struct dn_heap *h)
366 for (i = 0 ; i < h->elements ; i++ )
367 heap_insert(h, i , NULL) ;
371 * cleanup the heap and free data structure
374 heap_free(struct dn_heap *h)
377 kfree(h->p, M_DUMMYNET);
378 bzero(h, sizeof(*h) );
382 * --- end of heap management functions ---
386 * Scheduler functions:
388 * transmit_event() is called when the delay-line needs to enter
389 * the scheduler, either because of existing pkts getting ready,
390 * or new packets entering the queue. The event handled is the delivery
391 * time of the packet.
393 * ready_event() does something similar with fixed-rate queues, and the
394 * event handled is the finish time of the head pkt.
396 * wfq_ready_event() does something similar with WF2Q queues, and the
397 * event handled is the start time of the head pkt.
399 * In all cases, we make sure that the data structures are consistent
400 * before passing pkts out, because this might trigger recursive
401 * invocations of the procedures.
404 transmit_event(struct dn_pipe *pipe)
408 while ( (pkt = pipe->head) && DN_KEY_LEQ(pkt->output_time, curr_time) ) {
410 * first unlink, then call procedures, since ip_input() can invoke
411 * ip_output() and viceversa, thus causing nested calls
413 pipe->head = DN_NEXT(pkt) ;
416 * The actual mbuf is preceded by a struct dn_pkt, resembling an mbuf
417 * (NOT A REAL one, just a small block of malloc'ed memory) with
418 * m_type = MT_TAG, m_flags = PACKET_TAG_DUMMYNET
419 * dn_m (m_next) = actual mbuf to be processed by ip_input/output
420 * and some other fields.
421 * The block IS FREED HERE because it contains parameters passed
422 * to the called routine.
424 switch (pkt->dn_dir) {
426 ip_output((struct mbuf *)pkt, NULL, NULL, 0, NULL, NULL);
427 rt_unref (pkt->ro.ro_rt) ;
431 ip_input((struct mbuf *)pkt) ;
434 case DN_TO_ETH_DEMUX:
436 struct mbuf *m = (struct mbuf *)pkt ;
437 struct ether_header *eh;
439 if (pkt->dn_m->m_len < ETHER_HDR_LEN &&
440 (pkt->dn_m = m_pullup(pkt->dn_m, ETHER_HDR_LEN)) == NULL) {
441 printf("dummynet: pullup fail, dropping pkt\n");
445 * same as ether_input, make eh be a pointer into the mbuf
447 eh = mtod(pkt->dn_m, struct ether_header *);
448 m_adj(pkt->dn_m, ETHER_HDR_LEN);
449 /* which consumes the mbuf */
450 ether_demux(NULL, eh, m);
454 ether_output_frame(pkt->ifp, (struct mbuf *)pkt);
458 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
462 kfree(pkt, M_DUMMYNET);
464 /* if there are leftover packets, put into the heap for next event */
465 if ( (pkt = pipe->head) )
466 heap_insert(&extract_heap, pkt->output_time, pipe ) ;
467 /* XXX should check errors on heap_insert, by draining the
468 * whole pipe p and hoping in the future we are more successful
473 * the following macro computes how many ticks we have to wait
474 * before being able to transmit a packet. The credit is taken from
475 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
477 #define SET_TICKS(pkt, q, p) \
478 (pkt->dn_m->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \
482 * extract pkt from queue, compute output time (could be now)
483 * and put into delay line (p_queue)
486 move_pkt(struct dn_pkt *pkt, struct dn_flow_queue *q,
487 struct dn_pipe *p, int len)
489 q->head = DN_NEXT(pkt) ;
491 q->len_bytes -= len ;
493 pkt->output_time = curr_time + p->delay ;
498 DN_NEXT_NC(p->tail) = (struct mbuf *)pkt;
500 DN_NEXT_NC(p->tail) = NULL;
504 * ready_event() is invoked every time the queue must enter the
505 * scheduler, either because the first packet arrives, or because
506 * a previously scheduled event fired.
507 * On invokation, drain as many pkts as possible (could be 0) and then
508 * if there are leftover packets reinsert the pkt in the scheduler.
511 ready_event(struct dn_flow_queue *q)
514 struct dn_pipe *p = q->fs->pipe ;
518 printf("ready_event- pipe is gone\n");
521 p_was_empty = (p->head == NULL) ;
524 * schedule fixed-rate queues linked to this pipe:
525 * Account for the bw accumulated since last scheduling, then
526 * drain as many pkts as allowed by q->numbytes and move to
527 * the delay line (in p) computing output time.
528 * bandwidth==0 (no limit) means we can drain the whole queue,
529 * setting len_scaled = 0 does the job.
531 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth;
532 while ( (pkt = q->head) != NULL ) {
533 int len = pkt->dn_m->m_pkthdr.len;
534 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
535 if (len_scaled > q->numbytes )
537 q->numbytes -= len_scaled ;
538 move_pkt(pkt, q, p, len);
541 * If we have more packets queued, schedule next ready event
542 * (can only occur when bandwidth != 0, otherwise we would have
543 * flushed the whole queue in the previous loop).
544 * To this purpose we record the current time and compute how many
545 * ticks to go for the finish time of the packet.
547 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */
548 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
549 q->sched_time = curr_time ;
550 heap_insert(&ready_heap, curr_time + t, (void *)q );
551 /* XXX should check errors on heap_insert, and drain the whole
552 * queue on error hoping next time we are luckier.
554 } else { /* RED needs to know when the queue becomes empty */
555 q->q_time = curr_time;
559 * If the delay line was empty call transmit_event(p) now.
560 * Otherwise, the scheduler will take care of it.
567 * Called when we can transmit packets on WF2Q queues. Take pkts out of
568 * the queues at their start time, and enqueue into the delay line.
569 * Packets are drained until p->numbytes < 0. As long as
570 * len_scaled >= p->numbytes, the packet goes into the delay line
571 * with a deadline p->delay. For the last packet, if p->numbytes<0,
572 * there is an additional delay.
575 ready_event_wfq(struct dn_pipe *p)
577 int p_was_empty = (p->head == NULL) ;
578 struct dn_heap *sch = &(p->scheduler_heap);
579 struct dn_heap *neh = &(p->not_eligible_heap) ;
581 if (p->if_name[0] == 0) /* tx clock is simulated */
582 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth;
583 else { /* tx clock is for real, the ifq must be empty or this is a NOP */
584 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
587 DEB(printf("pipe %d ready from %s --\n",
588 p->pipe_nr, p->if_name);)
593 * While we have backlogged traffic AND credit, we need to do
594 * something on the queue.
596 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) {
597 if (sch->elements > 0) { /* have some eligible pkts to send out */
598 struct dn_flow_queue *q = sch->p[0].object ;
599 struct dn_pkt *pkt = q->head;
600 struct dn_flow_set *fs = q->fs;
601 u_int64_t len = pkt->dn_m->m_pkthdr.len;
602 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
604 heap_extract(sch, NULL); /* remove queue from heap */
605 p->numbytes -= len_scaled ;
606 move_pkt(pkt, q, p, len);
608 p->V += (len<<MY_M) / p->sum ; /* update V */
609 q->S = q->F ; /* update start time */
610 if (q->len == 0) { /* Flow not backlogged any more */
612 heap_insert(&(p->idle_heap), q->F, q);
613 } else { /* still backlogged */
615 * update F and position in backlogged queue, then
616 * put flow in not_eligible_heap (we will fix this later).
618 len = (q->head)->dn_m->m_pkthdr.len;
619 q->F += (len<<MY_M)/(u_int64_t) fs->weight ;
620 if (DN_KEY_LEQ(q->S, p->V))
621 heap_insert(neh, q->S, q);
623 heap_insert(sch, q->F, q);
627 * now compute V = max(V, min(S_i)). Remember that all elements in sch
628 * have by definition S_i <= V so if sch is not empty, V is surely
629 * the max and we must not update it. Conversely, if sch is empty
630 * we only need to look at neh.
632 if (sch->elements == 0 && neh->elements > 0)
633 p->V = MAX64 ( p->V, neh->p[0].key );
634 /* move from neh to sch any packets that have become eligible */
635 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) {
636 struct dn_flow_queue *q = neh->p[0].object ;
637 heap_extract(neh, NULL);
638 heap_insert(sch, q->F, q);
641 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */
642 p->numbytes = -1 ; /* mark not ready for I/O */
646 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0
647 && p->idle_heap.elements > 0) {
649 * no traffic and no events scheduled. We can get rid of idle-heap.
653 for (i = 0 ; i < p->idle_heap.elements ; i++) {
654 struct dn_flow_queue *q = p->idle_heap.p[i].object ;
661 p->idle_heap.elements = 0 ;
664 * If we are getting clocks from dummynet (not a real interface) and
665 * If we are under credit, schedule the next ready event.
666 * Also fix the delivery time of the last packet.
668 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */
669 dn_key t=0 ; /* number of ticks i have to wait */
671 if (p->bandwidth > 0)
672 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ;
673 p->tail->output_time += t ;
674 p->sched_time = curr_time ;
675 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
676 /* XXX should check errors on heap_insert, and drain the whole
677 * queue on error hoping next time we are luckier.
681 * If the delay line was empty call transmit_event(p) now.
682 * Otherwise, the scheduler will take care of it.
689 * This is called once per tick, or HZ times per second. It is used to
690 * increment the current tick counter and schedule expired events.
693 dummynet(void * __unused unused)
695 void *p ; /* generic parameter to handler */
697 struct dn_heap *heaps[3];
701 heaps[0] = &ready_heap ; /* fixed-rate queues */
702 heaps[1] = &wfq_ready_heap ; /* wfq queues */
703 heaps[2] = &extract_heap ; /* delay line */
704 crit_enter(); /* see note on top, splnet() is not enough */
706 for (i=0; i < 3 ; i++) {
708 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time) ) {
709 DDB(if (h->p[0].key > curr_time)
710 printf("-- dummynet: warning, heap %d is %d ticks late\n",
711 i, (int)(curr_time - h->p[0].key));)
712 p = h->p[0].object ; /* store a copy before heap_extract */
713 heap_extract(h, NULL); /* need to extract before processing */
717 struct dn_pipe *pipe = p;
718 if (pipe->if_name[0] != '\0')
719 printf("*** bad ready_event_wfq for pipe %s\n",
727 /* sweep pipes trying to expire idle flow_queues */
728 for (pe = all_pipes; pe ; pe = pe->next )
729 if (pe->idle_heap.elements > 0 &&
730 DN_KEY_LT(pe->idle_heap.p[0].key, pe->V) ) {
731 struct dn_flow_queue *q = pe->idle_heap.p[0].object ;
733 heap_extract(&(pe->idle_heap), NULL);
734 q->S = q->F + 1 ; /* mark timestamp as invalid */
735 pe->sum -= q->fs->weight ;
738 callout_reset(&dn_timeout, 1, dummynet, NULL);
742 * called by an interface when tx_rdy occurs.
745 if_tx_rdy(struct ifnet *ifp)
749 for (p = all_pipes; p ; p = p->next )
753 for (p = all_pipes; p ; p = p->next )
754 if (!strcmp(p->if_name, ifp->if_xname) ) {
756 DEB(printf("++ tx rdy from %s (now found)\n", ifp->if_xname);)
761 DEB(printf("++ tx rdy from %s - qlen %d\n", ifp->if_xname,
762 ifp->if_snd.ifq_len);)
763 p->numbytes = 0 ; /* mark ready for I/O */
770 * Unconditionally expire empty queues in case of shortage.
771 * Returns the number of queues freed.
774 expire_queues(struct dn_flow_set *fs)
776 struct dn_flow_queue *q, *prev ;
777 int i, initial_elements = fs->rq_elements ;
779 if (fs->last_expired == time_second)
781 fs->last_expired = time_second ;
782 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
783 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
784 if (q->head != NULL || q->S != q->F+1) {
787 } else { /* entry is idle, expire it */
788 struct dn_flow_queue *old_q = q ;
791 prev->next = q = q->next ;
793 fs->rq[i] = q = q->next ;
795 kfree(old_q, M_DUMMYNET);
797 return initial_elements - fs->rq_elements ;
801 * If room, create a new queue and put at head of slot i;
802 * otherwise, create or use the default queue.
804 static struct dn_flow_queue *
805 create_queue(struct dn_flow_set *fs, int i)
807 struct dn_flow_queue *q ;
809 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
810 expire_queues(fs) == 0) {
812 * No way to get room, use or create overflow queue.
815 if ( fs->rq[i] != NULL )
818 q = kmalloc(sizeof(*q), M_DUMMYNET, M_WAITOK | M_ZERO);
821 q->next = fs->rq[i] ;
822 q->S = q->F + 1; /* hack - mark timestamp as invalid */
829 * Given a flow_set and a pkt in last_pkt, find a matching queue
830 * after appropriate masking. The queue is moved to front
831 * so that further searches take less time.
833 static struct dn_flow_queue *
834 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
836 int i = 0 ; /* we need i and q for new allocations */
837 struct dn_flow_queue *q, *prev;
839 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
842 /* first, do the masking */
843 id->dst_ip &= fs->flow_mask.dst_ip ;
844 id->src_ip &= fs->flow_mask.src_ip ;
845 id->dst_port &= fs->flow_mask.dst_port ;
846 id->src_port &= fs->flow_mask.src_port ;
847 id->proto &= fs->flow_mask.proto ;
848 id->flags = 0 ; /* we don't care about this one */
849 /* then, hash function */
850 i = ( (id->dst_ip) & 0xffff ) ^
851 ( (id->dst_ip >> 15) & 0xffff ) ^
852 ( (id->src_ip << 1) & 0xffff ) ^
853 ( (id->src_ip >> 16 ) & 0xffff ) ^
854 (id->dst_port << 1) ^ (id->src_port) ^
856 i = i % fs->rq_size ;
857 /* finally, scan the current list for a match */
859 for (prev=NULL, q = fs->rq[i] ; q ; ) {
861 if (id->dst_ip == q->id.dst_ip &&
862 id->src_ip == q->id.src_ip &&
863 id->dst_port == q->id.dst_port &&
864 id->src_port == q->id.src_port &&
865 id->proto == q->id.proto &&
866 id->flags == q->id.flags)
868 else if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
869 /* entry is idle and not in any heap, expire it */
870 struct dn_flow_queue *old_q = q ;
873 prev->next = q = q->next ;
875 fs->rq[i] = q = q->next ;
877 kfree(old_q, M_DUMMYNET);
883 if (q && prev != NULL) { /* found and not in front */
884 prev->next = q->next ;
885 q->next = fs->rq[i] ;
889 if (q == NULL) { /* no match, need to allocate a new entry */
890 q = create_queue(fs, i);
898 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
903 * RED calculates the average queue size (avg) using a low-pass filter
904 * with an exponential weighted (w_q) moving average:
905 * avg <- (1-w_q) * avg + w_q * q_size
906 * where q_size is the queue length (measured in bytes or * packets).
908 * If q_size == 0, we compute the idle time for the link, and set
909 * avg = (1 - w_q)^(idle/s)
910 * where s is the time needed for transmitting a medium-sized packet.
912 * Now, if avg < min_th the packet is enqueued.
913 * If avg > max_th the packet is dropped. Otherwise, the packet is
914 * dropped with probability P function of avg.
919 /* queue in bytes or packets ? */
920 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len;
922 DEB(printf("\n%d q: %2u ", (int) curr_time, q_size);)
924 /* average queue size estimation */
927 * queue is not empty, avg <- avg + (q_size - avg) * w_q
929 int diff = SCALE(q_size) - q->avg;
930 int64_t v = SCALE_MUL((int64_t) diff, (int64_t) fs->w_q);
935 * queue is empty, find for how long the queue has been
936 * empty and use a lookup table for computing
937 * (1 - * w_q)^(idle_time/s) where s is the time to send a
942 u_int t = (curr_time - q->q_time) / fs->lookup_step;
944 q->avg = (t < fs->lookup_depth) ?
945 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
948 DEB(printf("avg: %u ", SCALE_VAL(q->avg));)
950 /* should i drop ? */
952 if (q->avg < fs->min_th) {
954 return 0; /* accept packet ; */
956 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
957 if (fs->flags_fs & DN_IS_GENTLE_RED) {
959 * According to Gentle-RED, if avg is greater than max_th the
960 * packet is dropped with a probability
961 * p_b = c_3 * avg - c_4
962 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
964 p_b = SCALE_MUL((int64_t) fs->c_3, (int64_t) q->avg) - fs->c_4;
970 } else if (q->avg > fs->min_th) {
972 * we compute p_b using the linear dropping function p_b = c_1 *
973 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
974 * max_p * min_th / (max_th - min_th)
976 p_b = SCALE_MUL((int64_t) fs->c_1, (int64_t) q->avg) - fs->c_2;
978 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
979 p_b = (p_b * len) / fs->max_pkt_size;
981 q->random = krandom() & 0xffff;
984 * q->count counts packets arrived since last drop, so a greater
985 * value of q->count means a greater packet drop probability.
987 if (SCALE_MUL(p_b, SCALE((int64_t) q->count)) > q->random) {
989 DEB(printf("- red drop");)
990 /* after a drop we calculate a new random value */
991 q->random = krandom() & 0xffff;
995 /* end of RED algorithm */
996 return 0 ; /* accept */
1000 struct dn_flow_set *
1001 locate_flowset(int pipe_nr, struct ip_fw *rule)
1004 struct dn_flow_set *fs;
1005 ipfw_insn *cmd = rule->cmd + rule->act_ofs;
1007 if (cmd->opcode == O_LOG)
1009 fs = ((ipfw_insn_pipe *)cmd)->pipe_ptr;
1014 if (cmd->opcode == O_QUEUE)
1016 struct dn_flow_set *fs = NULL ;
1018 if ( (rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_QUEUE )
1020 for (fs=all_flow_sets; fs && fs->fs_nr != pipe_nr; fs=fs->next)
1024 for (p1 = all_pipes; p1 && p1->pipe_nr != pipe_nr; p1 = p1->next)
1029 /* record for the future */
1031 ((ipfw_insn_pipe *)cmd)->pipe_ptr = fs;
1034 rule->pipe_ptr = fs;
1040 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1041 * depending on whether WF2Q or fixed bw is used.
1043 * pipe_nr pipe or queue the packet is destined for.
1044 * dir where shall we send the packet after dummynet.
1045 * m the mbuf with the packet
1046 * ifp the 'ifp' parameter from the caller.
1047 * NULL in ip_input, destination interface in ip_output
1048 * ro route parameter (only used in ip_output, NULL otherwise)
1049 * dst destination address, only used by ip_output
1050 * rule matching rule, in case of multiple passes
1051 * flags flags from the caller, only used in ip_output
1055 dummynet_io(struct mbuf *m, int pipe_nr, int dir, struct ip_fw_args *fwa)
1058 struct dn_flow_set *fs;
1059 struct dn_pipe *pipe ;
1060 u_int64_t len = m->m_pkthdr.len ;
1061 struct dn_flow_queue *q = NULL ;
1066 ipfw_insn *cmd = fwa->rule->cmd + fwa->rule->act_ofs;
1068 if (cmd->opcode == O_LOG)
1070 is_pipe = (cmd->opcode == O_PIPE);
1072 is_pipe = (fwa->rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_PIPE;
1078 * this is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule
1080 fs = locate_flowset(pipe_nr, fwa->rule);
1082 goto dropit ; /* this queue/pipe does not exist! */
1084 if (pipe == NULL) { /* must be a queue, try find a matching pipe */
1085 for (pipe = all_pipes; pipe && pipe->pipe_nr != fs->parent_nr;
1091 printf("No pipe %d for queue %d, drop pkt\n",
1092 fs->parent_nr, fs->fs_nr);
1096 q = find_queue(fs, &(fwa->f_id));
1098 goto dropit ; /* cannot allocate queue */
1100 * update statistics, then check reasons to drop pkt
1102 q->tot_bytes += len ;
1104 if ( fs->plr && krandom() < fs->plr )
1105 goto dropit ; /* random pkt drop */
1106 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) {
1107 if (q->len_bytes > fs->qsize)
1108 goto dropit ; /* queue size overflow */
1110 if (q->len >= fs->qsize)
1111 goto dropit ; /* queue count overflow */
1113 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) )
1116 /* XXX expensive to zero, see if we can remove it*/
1117 pkt = kmalloc(sizeof (*pkt), M_DUMMYNET, M_INTWAIT | M_ZERO | M_NULLOK);
1119 goto dropit; /* cannot allocate packet header */
1121 /* ok, i can handle the pkt now... */
1122 /* build and enqueue packet + parameters */
1123 pkt->hdr.mh_type = MT_TAG;
1124 pkt->hdr.mh_flags = PACKET_TAG_DUMMYNET;
1125 pkt->rule = fwa->rule ;
1126 DN_NEXT_NC(pkt) = NULL;
1130 pkt->ifp = fwa->oif;
1131 if (dir == DN_TO_IP_OUT) {
1133 * We need to copy *ro because for ICMP pkts (and maybe others)
1134 * the caller passed a pointer into the stack; dst might also be
1135 * a pointer into *ro so it needs to be updated.
1137 pkt->ro = *(fwa->ro);
1139 fwa->ro->ro_rt->rt_refcnt++ ;
1140 if (fwa->dst == (struct sockaddr_in *)&fwa->ro->ro_dst) /* dst points into ro */
1141 fwa->dst = (struct sockaddr_in *)&(pkt->ro.ro_dst) ;
1143 pkt->dn_dst = fwa->dst;
1144 pkt->flags = fwa->flags;
1146 if (q->head == NULL)
1149 DN_NEXT_NC(q->tail) = (struct mbuf *)pkt;
1152 q->len_bytes += len ;
1154 if ( q->head != pkt ) /* flow was not idle, we are done */
1157 * If we reach this point the flow was previously idle, so we need
1158 * to schedule it. This involves different actions for fixed-rate or
1163 * Fixed-rate queue: just insert into the ready_heap.
1166 if (pipe->bandwidth)
1167 t = SET_TICKS(pkt, q, pipe);
1168 q->sched_time = curr_time ;
1169 if (t == 0) /* must process it now */
1172 heap_insert(&ready_heap, curr_time + t , q );
1175 * WF2Q. First, compute start time S: if the flow was idle (S=F+1)
1176 * set S to the virtual time V for the controlling pipe, and update
1177 * the sum of weights for the pipe; otherwise, remove flow from
1178 * idle_heap and set S to max(F,V).
1179 * Second, compute finish time F = S + len/weight.
1180 * Third, if pipe was idle, update V=max(S, V).
1181 * Fourth, count one more backlogged flow.
1183 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */
1185 pipe->sum += fs->weight ; /* add weight of new queue */
1187 heap_extract(&(pipe->idle_heap), q);
1188 q->S = MAX64(q->F, pipe->V ) ;
1190 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight;
1192 if (pipe->not_eligible_heap.elements == 0 &&
1193 pipe->scheduler_heap.elements == 0)
1194 pipe->V = MAX64 ( q->S, pipe->V );
1197 * Look at eligibility. A flow is not eligibile if S>V (when
1198 * this happens, it means that there is some other flow already
1199 * scheduled for the same pipe, so the scheduler_heap cannot be
1200 * empty). If the flow is not eligible we just store it in the
1201 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1202 * and possibly invoke ready_event_wfq() right now if there is
1204 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1205 * and for all flows in not_eligible_heap (NEH), S_i > V .
1206 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH,
1207 * we only need to look into NEH.
1209 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */
1210 if (pipe->scheduler_heap.elements == 0)
1211 printf("++ ouch! not eligible but empty scheduler!\n");
1212 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1214 heap_insert(&(pipe->scheduler_heap), q->F, q);
1215 if (pipe->numbytes >= 0) { /* pipe is idle */
1216 if (pipe->scheduler_heap.elements != 1)
1217 printf("*** OUCH! pipe should have been idle!\n");
1218 DEB(printf("Waking up pipe %d at %d\n",
1219 pipe->pipe_nr, (int)(q->F >> MY_M)); )
1220 pipe->sched_time = curr_time ;
1221 ready_event_wfq(pipe);
1234 return ( (fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1238 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1239 * Doing this would probably save us the initial bzero of dn_pkt
1241 #define DN_FREE_PKT(pkt) { \
1242 struct dn_pkt *n = pkt ; \
1243 rt_unref ( n->ro.ro_rt ) ; \
1245 pkt = DN_NEXT(n) ; \
1246 kfree(n, M_DUMMYNET) ; }
1249 * Dispose all packets and flow_queues on a flow_set.
1250 * If all=1, also remove red lookup table and other storage,
1251 * including the descriptor itself.
1252 * For the one in dn_pipe MUST also cleanup ready_heap...
1255 purge_flow_set(struct dn_flow_set *fs, int all)
1257 struct dn_pkt *pkt ;
1258 struct dn_flow_queue *q, *qn ;
1261 for (i = 0 ; i <= fs->rq_size ; i++ ) {
1262 for (q = fs->rq[i] ; q ; q = qn ) {
1263 for (pkt = q->head ; pkt ; )
1266 kfree(q, M_DUMMYNET);
1270 fs->rq_elements = 0 ;
1272 /* RED - free lookup table */
1274 kfree(fs->w_q_lookup, M_DUMMYNET);
1276 kfree(fs->rq, M_DUMMYNET);
1277 /* if this fs is not part of a pipe, free it */
1278 if (fs->pipe && fs != &(fs->pipe->fs) )
1279 kfree(fs, M_DUMMYNET);
1284 * Dispose all packets queued on a pipe (not a flow_set).
1285 * Also free all resources associated to a pipe, which is about
1289 purge_pipe(struct dn_pipe *pipe)
1291 struct dn_pkt *pkt ;
1293 purge_flow_set( &(pipe->fs), 1 );
1295 for (pkt = pipe->head ; pkt ; )
1298 heap_free( &(pipe->scheduler_heap) );
1299 heap_free( &(pipe->not_eligible_heap) );
1300 heap_free( &(pipe->idle_heap) );
1304 * Delete all pipes and heaps returning memory. Must also
1305 * remove references from all ipfw rules to all pipes.
1308 dummynet_flush(void)
1310 struct dn_pipe *curr_p, *p ;
1311 struct dn_flow_set *fs, *curr_fs;
1315 /* remove all references to pipes ...*/
1316 flush_pipe_ptrs(NULL);
1317 /* prevent future matches... */
1320 fs = all_flow_sets ;
1321 all_flow_sets = NULL ;
1322 /* and free heaps so we don't have unwanted events */
1323 heap_free(&ready_heap);
1324 heap_free(&wfq_ready_heap);
1325 heap_free(&extract_heap);
1328 * Now purge all queued pkts and delete all pipes
1330 /* scan and purge all flow_sets. */
1334 purge_flow_set(curr_fs, 1);
1340 kfree(curr_p, M_DUMMYNET);
1345 extern struct ip_fw *ip_fw_default_rule ;
1347 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1350 struct dn_flow_queue *q ;
1351 struct dn_pkt *pkt ;
1353 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1354 for (q = fs->rq[i] ; q ; q = q->next )
1355 for (pkt = q->head ; pkt ; pkt = DN_NEXT(pkt) )
1357 pkt->rule = ip_fw_default_rule ;
1360 * when a firewall rule is deleted, scan all queues and remove the flow-id
1361 * from packets matching this rule.
1364 dn_rule_delete(void *r)
1367 struct dn_pkt *pkt ;
1368 struct dn_flow_set *fs ;
1371 * If the rule references a queue (dn_flow_set), then scan
1372 * the flow set, otherwise scan pipes. Should do either, but doing
1373 * both does not harm.
1375 for ( fs = all_flow_sets ; fs ; fs = fs->next )
1376 dn_rule_delete_fs(fs, r);
1377 for ( p = all_pipes ; p ; p = p->next ) {
1379 dn_rule_delete_fs(fs, r);
1380 for (pkt = p->head ; pkt ; pkt = DN_NEXT(pkt) )
1382 pkt->rule = ip_fw_default_rule ;
1387 * setup RED parameters
1390 config_red(struct dn_flow_set *p, struct dn_flow_set * x)
1395 x->min_th = SCALE(p->min_th);
1396 x->max_th = SCALE(p->max_th);
1397 x->max_p = p->max_p;
1399 x->c_1 = p->max_p / (p->max_th - p->min_th);
1400 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1401 if (x->flags_fs & DN_IS_GENTLE_RED) {
1402 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1403 x->c_4 = (SCALE(1) - 2 * p->max_p);
1406 /* if the lookup table already exist, free and create it again */
1407 if (x->w_q_lookup) {
1408 kfree(x->w_q_lookup, M_DUMMYNET);
1409 x->w_q_lookup = NULL ;
1411 if (red_lookup_depth == 0) {
1412 printf("\nnet.inet.ip.dummynet.red_lookup_depth must be > 0");
1413 kfree(x, M_DUMMYNET);
1416 x->lookup_depth = red_lookup_depth;
1417 x->w_q_lookup = kmalloc(x->lookup_depth * sizeof(int),
1418 M_DUMMYNET, M_WAITOK);
1420 /* fill the lookup table with (1 - w_q)^x */
1421 x->lookup_step = p->lookup_step ;
1422 x->lookup_weight = p->lookup_weight ;
1423 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1424 for (i = 1; i < x->lookup_depth; i++)
1425 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1426 if (red_avg_pkt_size < 1)
1427 red_avg_pkt_size = 512 ;
1428 x->avg_pkt_size = red_avg_pkt_size ;
1429 if (red_max_pkt_size < 1)
1430 red_max_pkt_size = 1500 ;
1431 x->max_pkt_size = red_max_pkt_size ;
1436 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1438 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1439 int l = pfs->rq_size;
1445 else if (l > DN_MAX_HASH_SIZE)
1446 l = DN_MAX_HASH_SIZE;
1448 } else /* one is enough for null mask */
1450 x->rq = kmalloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1451 M_DUMMYNET, M_WAITOK | M_ZERO);
1457 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1459 x->flags_fs = src->flags_fs;
1460 x->qsize = src->qsize;
1462 x->flow_mask = src->flow_mask;
1463 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1464 if (x->qsize > 1024*1024)
1465 x->qsize = 1024*1024 ;
1472 /* configuring RED */
1473 if ( x->flags_fs & DN_IS_RED )
1474 config_red(src, x) ; /* XXX should check errors */
1478 * setup pipe or queue parameters.
1482 config_pipe(struct dn_pipe *p)
1485 struct dn_flow_set *pfs = &(p->fs);
1486 struct dn_flow_queue *q;
1489 * The config program passes parameters as follows:
1490 * bw = bits/second (0 means no limits),
1491 * delay = ms, must be translated into ticks.
1492 * qsize = slots/bytes
1494 p->delay = ( p->delay * hz ) / 1000 ;
1495 /* We need either a pipe number or a flow_set number */
1496 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1498 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1500 if (p->pipe_nr != 0) { /* this is a pipe */
1501 struct dn_pipe *x, *a, *b;
1503 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1504 a = b , b = b->next) ;
1506 if (b == NULL || b->pipe_nr != p->pipe_nr) { /* new pipe */
1507 x = kmalloc(sizeof(struct dn_pipe), M_DUMMYNET, M_WAITOK | M_ZERO);
1508 x->pipe_nr = p->pipe_nr;
1510 /* idle_heap is the only one from which we extract from the middle.
1512 x->idle_heap.size = x->idle_heap.elements = 0 ;
1513 x->idle_heap.offset=OFFSET_OF(struct dn_flow_queue, heap_pos);
1517 /* Flush accumulated credit for all queues */
1518 for (i = 0; i <= x->fs.rq_size; i++)
1519 for (q = x->fs.rq[i]; q; q = q->next)
1525 x->bandwidth = p->bandwidth ;
1526 x->numbytes = 0; /* just in case... */
1527 bcopy(p->if_name, x->if_name, sizeof(p->if_name) );
1528 x->ifp = NULL ; /* reset interface ptr */
1529 x->delay = p->delay ;
1530 set_fs_parms(&(x->fs), pfs);
1533 if ( x->fs.rq == NULL ) { /* a new pipe */
1534 s = alloc_hash(&(x->fs), pfs) ;
1536 kfree(x, M_DUMMYNET);
1546 } else { /* config queue */
1547 struct dn_flow_set *x, *a, *b ;
1549 /* locate flow_set */
1550 for (a=NULL, b=all_flow_sets ; b && b->fs_nr < pfs->fs_nr ;
1551 a = b , b = b->next) ;
1553 if (b == NULL || b->fs_nr != pfs->fs_nr) { /* new */
1554 if (pfs->parent_nr == 0) /* need link to a pipe */
1556 x = kmalloc(sizeof(struct dn_flow_set), M_DUMMYNET, M_WAITOK|M_ZERO);
1557 x->fs_nr = pfs->fs_nr;
1558 x->parent_nr = pfs->parent_nr;
1559 x->weight = pfs->weight ;
1562 else if (x->weight > 100)
1565 /* Change parent pipe not allowed; must delete and recreate */
1566 if (pfs->parent_nr != 0 && b->parent_nr != pfs->parent_nr)
1571 set_fs_parms(x, pfs);
1573 if ( x->rq == NULL ) { /* a new flow_set */
1574 s = alloc_hash(x, pfs) ;
1576 kfree(x, M_DUMMYNET);
1591 * Helper function to remove from a heap queues which are linked to
1592 * a flow_set about to be deleted.
1595 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1597 int i = 0, found = 0 ;
1598 for (; i < h->elements ;)
1599 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1601 h->p[i] = h->p[h->elements] ;
1610 * helper function to remove a pipe from a heap (can be there at most once)
1613 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1615 if (h->elements > 0) {
1617 for (i=0; i < h->elements ; i++ ) {
1618 if (h->p[i].object == p) { /* found it */
1620 h->p[i] = h->p[h->elements] ;
1629 * drain all queues. Called in case of severe mbuf shortage.
1632 dummynet_drain(void)
1634 struct dn_flow_set *fs;
1638 heap_free(&ready_heap);
1639 heap_free(&wfq_ready_heap);
1640 heap_free(&extract_heap);
1641 /* remove all references to this pipe from flow_sets */
1642 for (fs = all_flow_sets; fs; fs= fs->next )
1643 purge_flow_set(fs, 0);
1645 for (p = all_pipes; p; p= p->next ) {
1646 purge_flow_set(&(p->fs), 0);
1647 for (pkt = p->head ; pkt ; )
1649 p->head = p->tail = NULL ;
1654 * Fully delete a pipe or a queue, cleaning up associated info.
1657 delete_pipe(struct dn_pipe *p)
1659 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1661 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1663 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1664 struct dn_pipe *a, *b;
1665 struct dn_flow_set *fs;
1668 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1669 a = b , b = b->next) ;
1670 if (b == NULL || (b->pipe_nr != p->pipe_nr) )
1671 return EINVAL ; /* not found */
1675 /* unlink from list of pipes */
1677 all_pipes = b->next ;
1680 /* remove references to this pipe from the ip_fw rules. */
1681 flush_pipe_ptrs(&(b->fs));
1683 /* remove all references to this pipe from flow_sets */
1684 for (fs = all_flow_sets; fs; fs= fs->next )
1685 if (fs->pipe == b) {
1686 printf("++ ref to pipe %d from fs %d\n",
1687 p->pipe_nr, fs->fs_nr);
1689 purge_flow_set(fs, 0);
1691 fs_remove_from_heap(&ready_heap, &(b->fs));
1692 purge_pipe(b); /* remove all data associated to this pipe */
1693 /* remove reference to here from extract_heap and wfq_ready_heap */
1694 pipe_remove_from_heap(&extract_heap, b);
1695 pipe_remove_from_heap(&wfq_ready_heap, b);
1697 kfree(b, M_DUMMYNET);
1698 } else { /* this is a WF2Q queue (dn_flow_set) */
1699 struct dn_flow_set *a, *b;
1702 for (a = NULL, b = all_flow_sets ; b && b->fs_nr < p->fs.fs_nr ;
1703 a = b , b = b->next) ;
1704 if (b == NULL || (b->fs_nr != p->fs.fs_nr) )
1705 return EINVAL ; /* not found */
1709 all_flow_sets = b->next ;
1712 /* remove references to this flow_set from the ip_fw rules. */
1715 if (b->pipe != NULL) {
1716 /* Update total weight on parent pipe and cleanup parent heaps */
1717 b->pipe->sum -= b->weight * b->backlogged ;
1718 fs_remove_from_heap(&(b->pipe->not_eligible_heap), b);
1719 fs_remove_from_heap(&(b->pipe->scheduler_heap), b);
1720 #if 1 /* XXX should i remove from idle_heap as well ? */
1721 fs_remove_from_heap(&(b->pipe->idle_heap), b);
1724 purge_flow_set(b, 1);
1731 * helper function used to copy data from kernel in DUMMYNET_GET
1734 dn_copy_set(struct dn_flow_set *set, char *bp)
1737 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1739 for (i = 0 ; i <= set->rq_size ; i++)
1740 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
1741 if (q->hash_slot != i)
1742 printf("++ at %d: wrong slot (have %d, "
1743 "should be %d)\n", copied, q->hash_slot, i);
1745 printf("++ at %d: wrong fs ptr (have %p, should be %p)\n",
1748 bcopy(q, qp, sizeof( *q ) );
1749 /* cleanup pointers */
1751 qp->head = qp->tail = NULL ;
1754 if (copied != set->rq_elements)
1755 printf("++ wrong count, have %d should be %d\n",
1756 copied, set->rq_elements);
1761 dummynet_get(struct sockopt *sopt)
1763 char *buf, *bp ; /* bp is the "copy-pointer" */
1765 struct dn_flow_set *set ;
1771 * compute size of data structures: list of pipes and flow_sets.
1773 for (p = all_pipes, size = 0 ; p ; p = p->next )
1774 size += sizeof( *p ) +
1775 p->fs.rq_elements * sizeof(struct dn_flow_queue);
1776 for (set = all_flow_sets ; set ; set = set->next )
1777 size += sizeof ( *set ) +
1778 set->rq_elements * sizeof(struct dn_flow_queue);
1779 buf = kmalloc(size, M_TEMP, M_WAITOK);
1780 for (p = all_pipes, bp = buf ; p ; p = p->next ) {
1781 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp ;
1784 * copy pipe descriptor into *bp, convert delay back to ms,
1785 * then copy the flow_set descriptor(s) one at a time.
1786 * After each flow_set, copy the queue descriptor it owns.
1788 bcopy(p, bp, sizeof( *p ) );
1789 pipe_bp->delay = (pipe_bp->delay * 1000) / hz ;
1791 * XXX the following is a hack based on ->next being the
1792 * first field in dn_pipe and dn_flow_set. The correct
1793 * solution would be to move the dn_flow_set to the beginning
1794 * of struct dn_pipe.
1796 pipe_bp->next = (struct dn_pipe *)DN_IS_PIPE ;
1797 /* clean pointers */
1798 pipe_bp->head = pipe_bp->tail = NULL ;
1799 pipe_bp->fs.next = NULL ;
1800 pipe_bp->fs.pipe = NULL ;
1801 pipe_bp->fs.rq = NULL ;
1803 bp += sizeof( *p ) ;
1804 bp = dn_copy_set( &(p->fs), bp );
1806 for (set = all_flow_sets ; set ; set = set->next ) {
1807 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp ;
1808 bcopy(set, bp, sizeof( *set ) );
1809 /* XXX same hack as above */
1810 fs_bp->next = (struct dn_flow_set *)DN_IS_QUEUE ;
1811 fs_bp->pipe = NULL ;
1813 bp += sizeof( *set ) ;
1814 bp = dn_copy_set( set, bp );
1817 error = sooptcopyout(sopt, buf, size);
1823 * Handler for the various dummynet socket options (get, flush, config, del)
1826 ip_dn_ctl(struct sockopt *sopt)
1829 struct dn_pipe *p, tmp_pipe;
1831 /* Disallow sets in really-really secure mode. */
1832 if (sopt->sopt_dir == SOPT_SET) {
1833 #if defined(__FreeBSD__) && __FreeBSD_version >= 500034
1834 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
1838 if (securelevel >= 3)
1843 switch (sopt->sopt_name) {
1845 printf("ip_dn_ctl -- unknown option %d", sopt->sopt_name);
1848 case IP_DUMMYNET_GET :
1849 error = dummynet_get(sopt);
1852 case IP_DUMMYNET_FLUSH :
1856 case IP_DUMMYNET_CONFIGURE :
1858 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
1861 error = config_pipe(p);
1864 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
1866 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
1870 error = delete_pipe(p);
1879 printf("DUMMYNET initialized (011031)\n");
1881 all_flow_sets = NULL ;
1882 ready_heap.size = ready_heap.elements = 0 ;
1883 ready_heap.offset = 0 ;
1885 wfq_ready_heap.size = wfq_ready_heap.elements = 0 ;
1886 wfq_ready_heap.offset = 0 ;
1888 extract_heap.size = extract_heap.elements = 0 ;
1889 extract_heap.offset = 0 ;
1890 ip_dn_ctl_ptr = ip_dn_ctl;
1891 ip_dn_io_ptr = dummynet_io;
1892 ip_dn_ruledel_ptr = dn_rule_delete;
1893 callout_init(&dn_timeout);
1894 callout_reset(&dn_timeout, 1, dummynet, NULL);
1898 dummynet_modevent(module_t mod, int type, void *data)
1903 if (DUMMYNET_LOADED) {
1905 printf("DUMMYNET already loaded\n");
1913 #if !defined(KLD_MODULE)
1914 printf("dummynet statically compiled, cannot unload\n");
1918 callout_stop(&dn_timeout);
1920 ip_dn_ctl_ptr = NULL;
1921 ip_dn_io_ptr = NULL;
1922 ip_dn_ruledel_ptr = NULL;
1932 static moduledata_t dummynet_mod = {
1937 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);
1938 MODULE_DEPEND(dummynet, ipfw, 1, 1, 1);
1939 MODULE_VERSION(dummynet, 1);