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 $
30 #if !defined(KLD_MODULE)
31 #include "opt_ipfw.h" /* for IPFW2 definition */
38 * This module implements IP dummynet, a bandwidth limiter/delay emulator
39 * used in conjunction with the ipfw package.
40 * Description of the data structures used is in ip_dummynet.h
41 * Here you mainly find the following blocks of code:
42 * + variable declarations;
43 * + heap management functions;
44 * + scheduler and dummynet functions;
45 * + configuration and initialization.
47 * NOTA BENE: critical sections are protected by splimp()/splx()
48 * pairs. One would think that splnet() is enough as for most of
49 * the netinet code, but it is not so because when used with
50 * bridging, dummynet is invoked at splimp().
52 * Most important Changes:
55 * 010124: Fixed WF2Q behaviour
56 * 010122: Fixed spl protection.
57 * 000601: WF2Q support
58 * 000106: large rewrite, use heaps to handle very many pipes.
59 * 980513: initial release
61 * include files marked with XXX are probably not needed
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/malloc.h>
68 #include <sys/kernel.h>
69 #include <sys/module.h>
71 #include <sys/socket.h>
72 #include <sys/socketvar.h>
74 #include <sys/sysctl.h>
76 #include <net/route.h>
77 #include <netinet/in.h>
78 #include <netinet/in_systm.h>
79 #include <netinet/in_var.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip_fw.h>
82 #include <netinet/ip_dummynet.h>
83 #include <netinet/ip_var.h>
85 #include <netinet/if_ether.h> /* for struct arpcom */
86 #include <net/bridge.h>
89 * We keep a private variable for the simulation time, but we could
90 * probably use an existing one ("softticks" in sys/kern/kern_timer.c)
92 static dn_key curr_time = 0 ; /* current simulation time */
94 static int dn_hash_size = 64 ; /* default hash size */
96 /* statistics on number of queue searches and search steps */
97 static int searches, search_steps ;
98 static int pipe_expire = 1 ; /* expire queue if empty */
99 static int dn_max_ratio = 16 ; /* max queues/buckets ratio */
101 static int red_lookup_depth = 256; /* RED - default lookup table depth */
102 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
103 static int red_max_pkt_size = 1500; /* RED - default max packet size */
106 * Three heaps contain queues and pipes that the scheduler handles:
108 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
110 * wfq_ready_heap contains the pipes associated with WF2Q flows
112 * extract_heap contains pipes associated with delay lines.
116 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
118 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
120 static int heap_init(struct dn_heap *h, int size) ;
121 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
122 static void heap_extract(struct dn_heap *h, void *obj);
124 static void transmit_event(struct dn_pipe *pipe);
125 static void ready_event(struct dn_flow_queue *q);
127 static struct dn_pipe *all_pipes = NULL ; /* list of all pipes */
128 static struct dn_flow_set *all_flow_sets = NULL ;/* list of all flow_sets */
130 static struct callout_handle dn_timeout;
133 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
134 CTLFLAG_RW, 0, "Dummynet");
135 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
136 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
137 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time,
138 CTLFLAG_RD, &curr_time, 0, "Current tick");
139 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
140 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
141 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
142 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
143 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches,
144 CTLFLAG_RD, &searches, 0, "Number of queue searches");
145 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps,
146 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
147 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
148 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
149 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
150 CTLFLAG_RW, &dn_max_ratio, 0,
151 "Max ratio between dynamic queues and buckets");
152 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
153 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
154 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
155 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
156 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
157 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
160 static int config_pipe(struct dn_pipe *p);
161 static int ip_dn_ctl(struct sockopt *sopt);
163 static void rt_unref(struct rtentry *);
164 static void dummynet(void *);
165 static void dummynet_flush(void);
166 void dummynet_drain(void);
167 static ip_dn_io_t dummynet_io;
168 static void dn_rule_delete(void *);
170 int if_tx_rdy(struct ifnet *ifp);
173 rt_unref(struct rtentry *rt)
177 if (rt->rt_refcnt <= 0)
178 printf("-- warning, refcnt now %ld, decreasing\n", rt->rt_refcnt);
183 * Heap management functions.
185 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
186 * Some macros help finding parent/children so we can optimize them.
188 * heap_init() is called to expand the heap when needed.
189 * Increment size in blocks of 16 entries.
190 * XXX failure to allocate a new element is a pretty bad failure
191 * as we basically stall a whole queue forever!!
192 * Returns 1 on error, 0 on success
194 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
195 #define HEAP_LEFT(x) ( 2*(x) + 1 )
196 #define HEAP_IS_LEFT(x) ( (x) & 1 )
197 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
198 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
199 #define HEAP_INCREMENT 15
202 heap_init(struct dn_heap *h, int new_size)
204 struct dn_heap_entry *p;
206 if (h->size >= new_size ) {
207 printf("heap_init, Bogus call, have %d want %d\n",
211 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
212 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_NOWAIT);
214 printf(" heap_init, resize %d failed\n", new_size );
215 return 1 ; /* error */
218 bcopy(h->p, p, h->size * sizeof(*p) );
219 free(h->p, M_DUMMYNET);
227 * Insert element in heap. Normally, p != NULL, we insert p in
228 * a new position and bubble up. If p == NULL, then the element is
229 * already in place, and key is the position where to start the
231 * Returns 1 on failure (cannot allocate new heap entry)
233 * If offset > 0 the position (index, int) of the element in the heap is
234 * also stored in the element itself at the given offset in bytes.
236 #define SET_OFFSET(heap, node) \
237 if (heap->offset > 0) \
238 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
240 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
242 #define RESET_OFFSET(heap, node) \
243 if (heap->offset > 0) \
244 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
246 heap_insert(struct dn_heap *h, dn_key key1, void *p)
248 int son = h->elements ;
250 if (p == NULL) /* data already there, set starting point */
252 else { /* insert new element at the end, possibly resize */
254 if (son == h->size) /* need resize... */
255 if (heap_init(h, h->elements+1) )
256 return 1 ; /* failure... */
257 h->p[son].object = p ;
258 h->p[son].key = key1 ;
261 while (son > 0) { /* bubble up */
262 int father = HEAP_FATHER(son) ;
263 struct dn_heap_entry tmp ;
265 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
266 break ; /* found right position */
267 /* son smaller than father, swap and repeat */
268 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
277 * remove top element from heap, or obj if obj != NULL
280 heap_extract(struct dn_heap *h, void *obj)
282 int child, father, max = h->elements - 1 ;
285 printf("warning, extract from empty heap 0x%p\n", h);
288 father = 0 ; /* default: move up smallest child */
289 if (obj != NULL) { /* extract specific element, index is at offset */
291 panic("*** heap_extract from middle not supported on this heap!!!\n");
292 father = *((int *)((char *)obj + h->offset)) ;
293 if (father < 0 || father >= h->elements) {
294 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
295 father, h->elements);
296 panic("heap_extract");
299 RESET_OFFSET(h, father);
300 child = HEAP_LEFT(father) ; /* left child */
301 while (child <= max) { /* valid entry */
302 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
303 child = child+1 ; /* take right child, otherwise left */
304 h->p[father] = h->p[child] ;
305 SET_OFFSET(h, father);
307 child = HEAP_LEFT(child) ; /* left child for next loop */
312 * Fill hole with last entry and bubble up, reusing the insert code
314 h->p[father] = h->p[max] ;
315 heap_insert(h, father, NULL); /* this one cannot fail */
321 * change object position and update references
322 * XXX this one is never used!
325 heap_move(struct dn_heap *h, dn_key new_key, void *object)
329 int max = h->elements-1 ;
330 struct dn_heap_entry buf ;
333 panic("cannot move items on this heap");
335 i = *((int *)((char *)object + h->offset));
336 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
337 h->p[i].key = new_key ;
338 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
339 i = temp ) { /* bubble up */
340 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
343 } else { /* must move down */
344 h->p[i].key = new_key ;
345 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
346 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
347 temp++ ; /* select child with min key */
348 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
349 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
358 #endif /* heap_move, unused */
361 * heapify() will reorganize data inside an array to maintain the
362 * heap property. It is needed when we delete a bunch of entries.
365 heapify(struct dn_heap *h)
369 for (i = 0 ; i < h->elements ; i++ )
370 heap_insert(h, i , NULL) ;
374 * cleanup the heap and free data structure
377 heap_free(struct dn_heap *h)
380 free(h->p, M_DUMMYNET);
381 bzero(h, sizeof(*h) );
385 * --- end of heap management functions ---
389 * Scheduler functions:
391 * transmit_event() is called when the delay-line needs to enter
392 * the scheduler, either because of existing pkts getting ready,
393 * or new packets entering the queue. The event handled is the delivery
394 * time of the packet.
396 * ready_event() does something similar with fixed-rate queues, and the
397 * event handled is the finish time of the head pkt.
399 * wfq_ready_event() does something similar with WF2Q queues, and the
400 * event handled is the start time of the head pkt.
402 * In all cases, we make sure that the data structures are consistent
403 * before passing pkts out, because this might trigger recursive
404 * invocations of the procedures.
407 transmit_event(struct dn_pipe *pipe)
411 while ( (pkt = pipe->head) && DN_KEY_LEQ(pkt->output_time, curr_time) ) {
413 * first unlink, then call procedures, since ip_input() can invoke
414 * ip_output() and viceversa, thus causing nested calls
416 pipe->head = DN_NEXT(pkt) ;
419 * The actual mbuf is preceded by a struct dn_pkt, resembling an mbuf
420 * (NOT A REAL one, just a small block of malloc'ed memory) with
421 * m_type = MT_TAG, m_flags = PACKET_TAG_DUMMYNET
422 * dn_m (m_next) = actual mbuf to be processed by ip_input/output
423 * and some other fields.
424 * The block IS FREED HERE because it contains parameters passed
425 * to the called routine.
427 switch (pkt->dn_dir) {
429 (void)ip_output((struct mbuf *)pkt, NULL, NULL, 0, NULL, NULL);
430 rt_unref (pkt->ro.ro_rt) ;
434 ip_input((struct mbuf *)pkt) ;
439 /* somebody unloaded the bridge module. Drop pkt */
440 printf("-- dropping bridged packet trapped in pipe--\n");
444 case DN_TO_ETH_DEMUX:
446 struct mbuf *m = (struct mbuf *)pkt ;
447 struct ether_header *eh;
449 if (pkt->dn_m->m_len < ETHER_HDR_LEN &&
450 (pkt->dn_m = m_pullup(pkt->dn_m, ETHER_HDR_LEN)) == NULL) {
451 printf("dummynet/bridge: pullup fail, dropping pkt\n");
455 * same as ether_input, make eh be a pointer into the mbuf
457 eh = mtod(pkt->dn_m, struct ether_header *);
458 m_adj(pkt->dn_m, ETHER_HDR_LEN);
460 * bdg_forward() wants a pointer to the pseudo-mbuf-header, but
461 * on return it will supply the pointer to the actual packet
462 * (originally pkt->dn_m, but could be something else now) if
463 * it has not consumed it.
465 if (pkt->dn_dir == DN_TO_BDG_FWD) {
466 m = bdg_forward_ptr(m, eh, pkt->ifp);
470 ether_demux(NULL, eh, m); /* which consumes the mbuf */
474 ether_output_frame(pkt->ifp, (struct mbuf *)pkt);
478 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
482 free(pkt, M_DUMMYNET);
484 /* if there are leftover packets, put into the heap for next event */
485 if ( (pkt = pipe->head) )
486 heap_insert(&extract_heap, pkt->output_time, pipe ) ;
487 /* XXX should check errors on heap_insert, by draining the
488 * whole pipe p and hoping in the future we are more successful
493 * the following macro computes how many ticks we have to wait
494 * before being able to transmit a packet. The credit is taken from
495 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
497 #define SET_TICKS(pkt, q, p) \
498 (pkt->dn_m->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \
502 * extract pkt from queue, compute output time (could be now)
503 * and put into delay line (p_queue)
506 move_pkt(struct dn_pkt *pkt, struct dn_flow_queue *q,
507 struct dn_pipe *p, int len)
509 q->head = DN_NEXT(pkt) ;
511 q->len_bytes -= len ;
513 pkt->output_time = curr_time + p->delay ;
518 DN_NEXT(p->tail) = pkt;
520 DN_NEXT(p->tail) = NULL;
524 * ready_event() is invoked every time the queue must enter the
525 * scheduler, either because the first packet arrives, or because
526 * a previously scheduled event fired.
527 * On invokation, drain as many pkts as possible (could be 0) and then
528 * if there are leftover packets reinsert the pkt in the scheduler.
531 ready_event(struct dn_flow_queue *q)
534 struct dn_pipe *p = q->fs->pipe ;
538 printf("ready_event- pipe is gone\n");
541 p_was_empty = (p->head == NULL) ;
544 * schedule fixed-rate queues linked to this pipe:
545 * Account for the bw accumulated since last scheduling, then
546 * drain as many pkts as allowed by q->numbytes and move to
547 * the delay line (in p) computing output time.
548 * bandwidth==0 (no limit) means we can drain the whole queue,
549 * setting len_scaled = 0 does the job.
551 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth;
552 while ( (pkt = q->head) != NULL ) {
553 int len = pkt->dn_m->m_pkthdr.len;
554 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
555 if (len_scaled > q->numbytes )
557 q->numbytes -= len_scaled ;
558 move_pkt(pkt, q, p, len);
561 * If we have more packets queued, schedule next ready event
562 * (can only occur when bandwidth != 0, otherwise we would have
563 * flushed the whole queue in the previous loop).
564 * To this purpose we record the current time and compute how many
565 * ticks to go for the finish time of the packet.
567 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */
568 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
569 q->sched_time = curr_time ;
570 heap_insert(&ready_heap, curr_time + t, (void *)q );
571 /* XXX should check errors on heap_insert, and drain the whole
572 * queue on error hoping next time we are luckier.
574 } else { /* RED needs to know when the queue becomes empty */
575 q->q_time = curr_time;
579 * If the delay line was empty call transmit_event(p) now.
580 * Otherwise, the scheduler will take care of it.
587 * Called when we can transmit packets on WF2Q queues. Take pkts out of
588 * the queues at their start time, and enqueue into the delay line.
589 * Packets are drained until p->numbytes < 0. As long as
590 * len_scaled >= p->numbytes, the packet goes into the delay line
591 * with a deadline p->delay. For the last packet, if p->numbytes<0,
592 * there is an additional delay.
595 ready_event_wfq(struct dn_pipe *p)
597 int p_was_empty = (p->head == NULL) ;
598 struct dn_heap *sch = &(p->scheduler_heap);
599 struct dn_heap *neh = &(p->not_eligible_heap) ;
601 if (p->if_name[0] == 0) /* tx clock is simulated */
602 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth;
603 else { /* tx clock is for real, the ifq must be empty or this is a NOP */
604 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
607 DEB(printf("pipe %d ready from %s --\n",
608 p->pipe_nr, p->if_name);)
613 * While we have backlogged traffic AND credit, we need to do
614 * something on the queue.
616 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) {
617 if (sch->elements > 0) { /* have some eligible pkts to send out */
618 struct dn_flow_queue *q = sch->p[0].object ;
619 struct dn_pkt *pkt = q->head;
620 struct dn_flow_set *fs = q->fs;
621 u_int64_t len = pkt->dn_m->m_pkthdr.len;
622 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
624 heap_extract(sch, NULL); /* remove queue from heap */
625 p->numbytes -= len_scaled ;
626 move_pkt(pkt, q, p, len);
628 p->V += (len<<MY_M) / p->sum ; /* update V */
629 q->S = q->F ; /* update start time */
630 if (q->len == 0) { /* Flow not backlogged any more */
632 heap_insert(&(p->idle_heap), q->F, q);
633 } else { /* still backlogged */
635 * update F and position in backlogged queue, then
636 * put flow in not_eligible_heap (we will fix this later).
638 len = (q->head)->dn_m->m_pkthdr.len;
639 q->F += (len<<MY_M)/(u_int64_t) fs->weight ;
640 if (DN_KEY_LEQ(q->S, p->V))
641 heap_insert(neh, q->S, q);
643 heap_insert(sch, q->F, q);
647 * now compute V = max(V, min(S_i)). Remember that all elements in sch
648 * have by definition S_i <= V so if sch is not empty, V is surely
649 * the max and we must not update it. Conversely, if sch is empty
650 * we only need to look at neh.
652 if (sch->elements == 0 && neh->elements > 0)
653 p->V = MAX64 ( p->V, neh->p[0].key );
654 /* move from neh to sch any packets that have become eligible */
655 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) {
656 struct dn_flow_queue *q = neh->p[0].object ;
657 heap_extract(neh, NULL);
658 heap_insert(sch, q->F, q);
661 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */
662 p->numbytes = -1 ; /* mark not ready for I/O */
666 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0
667 && p->idle_heap.elements > 0) {
669 * no traffic and no events scheduled. We can get rid of idle-heap.
673 for (i = 0 ; i < p->idle_heap.elements ; i++) {
674 struct dn_flow_queue *q = p->idle_heap.p[i].object ;
681 p->idle_heap.elements = 0 ;
684 * If we are getting clocks from dummynet (not a real interface) and
685 * If we are under credit, schedule the next ready event.
686 * Also fix the delivery time of the last packet.
688 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */
689 dn_key t=0 ; /* number of ticks i have to wait */
691 if (p->bandwidth > 0)
692 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ;
693 p->tail->output_time += t ;
694 p->sched_time = curr_time ;
695 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
696 /* XXX should check errors on heap_insert, and drain the whole
697 * queue on error hoping next time we are luckier.
701 * If the delay line was empty call transmit_event(p) now.
702 * Otherwise, the scheduler will take care of it.
709 * This is called once per tick, or HZ times per second. It is used to
710 * increment the current tick counter and schedule expired events.
713 dummynet(void * __unused unused)
715 void *p ; /* generic parameter to handler */
718 struct dn_heap *heaps[3];
722 heaps[0] = &ready_heap ; /* fixed-rate queues */
723 heaps[1] = &wfq_ready_heap ; /* wfq queues */
724 heaps[2] = &extract_heap ; /* delay line */
725 s = splimp(); /* see note on top, splnet() is not enough */
727 for (i=0; i < 3 ; i++) {
729 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time) ) {
730 DDB(if (h->p[0].key > curr_time)
731 printf("-- dummynet: warning, heap %d is %d ticks late\n",
732 i, (int)(curr_time - h->p[0].key));)
733 p = h->p[0].object ; /* store a copy before heap_extract */
734 heap_extract(h, NULL); /* need to extract before processing */
738 struct dn_pipe *pipe = p;
739 if (pipe->if_name[0] != '\0')
740 printf("*** bad ready_event_wfq for pipe %s\n",
748 /* sweep pipes trying to expire idle flow_queues */
749 for (pe = all_pipes; pe ; pe = pe->next )
750 if (pe->idle_heap.elements > 0 &&
751 DN_KEY_LT(pe->idle_heap.p[0].key, pe->V) ) {
752 struct dn_flow_queue *q = pe->idle_heap.p[0].object ;
754 heap_extract(&(pe->idle_heap), NULL);
755 q->S = q->F + 1 ; /* mark timestamp as invalid */
756 pe->sum -= q->fs->weight ;
759 dn_timeout = timeout(dummynet, NULL, 1);
763 * called by an interface when tx_rdy occurs.
766 if_tx_rdy(struct ifnet *ifp)
770 for (p = all_pipes; p ; p = p->next )
775 sprintf(buf, "%s%d",ifp->if_name, ifp->if_unit);
776 for (p = all_pipes; p ; p = p->next )
777 if (!strcmp(p->if_name, buf) ) {
779 DEB(printf("++ tx rdy from %s (now found)\n", buf);)
784 DEB(printf("++ tx rdy from %s%d - qlen %d\n", ifp->if_name,
785 ifp->if_unit, ifp->if_snd.ifq_len);)
786 p->numbytes = 0 ; /* mark ready for I/O */
793 * Unconditionally expire empty queues in case of shortage.
794 * Returns the number of queues freed.
797 expire_queues(struct dn_flow_set *fs)
799 struct dn_flow_queue *q, *prev ;
800 int i, initial_elements = fs->rq_elements ;
802 if (fs->last_expired == time_second)
804 fs->last_expired = time_second ;
805 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
806 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
807 if (q->head != NULL || q->S != q->F+1) {
810 } else { /* entry is idle, expire it */
811 struct dn_flow_queue *old_q = q ;
814 prev->next = q = q->next ;
816 fs->rq[i] = q = q->next ;
818 free(old_q, M_DUMMYNET);
820 return initial_elements - fs->rq_elements ;
824 * If room, create a new queue and put at head of slot i;
825 * otherwise, create or use the default queue.
827 static struct dn_flow_queue *
828 create_queue(struct dn_flow_set *fs, int i)
830 struct dn_flow_queue *q ;
832 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
833 expire_queues(fs) == 0) {
835 * No way to get room, use or create overflow queue.
838 if ( fs->rq[i] != NULL )
841 q = malloc(sizeof(*q), M_DUMMYNET, M_NOWAIT | M_ZERO);
843 printf("sorry, cannot allocate queue for new flow\n");
848 q->next = fs->rq[i] ;
849 q->S = q->F + 1; /* hack - mark timestamp as invalid */
856 * Given a flow_set and a pkt in last_pkt, find a matching queue
857 * after appropriate masking. The queue is moved to front
858 * so that further searches take less time.
860 static struct dn_flow_queue *
861 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
863 int i = 0 ; /* we need i and q for new allocations */
864 struct dn_flow_queue *q, *prev;
866 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
869 /* first, do the masking */
870 id->dst_ip &= fs->flow_mask.dst_ip ;
871 id->src_ip &= fs->flow_mask.src_ip ;
872 id->dst_port &= fs->flow_mask.dst_port ;
873 id->src_port &= fs->flow_mask.src_port ;
874 id->proto &= fs->flow_mask.proto ;
875 id->flags = 0 ; /* we don't care about this one */
876 /* then, hash function */
877 i = ( (id->dst_ip) & 0xffff ) ^
878 ( (id->dst_ip >> 15) & 0xffff ) ^
879 ( (id->src_ip << 1) & 0xffff ) ^
880 ( (id->src_ip >> 16 ) & 0xffff ) ^
881 (id->dst_port << 1) ^ (id->src_port) ^
883 i = i % fs->rq_size ;
884 /* finally, scan the current list for a match */
886 for (prev=NULL, q = fs->rq[i] ; q ; ) {
888 if (id->dst_ip == q->id.dst_ip &&
889 id->src_ip == q->id.src_ip &&
890 id->dst_port == q->id.dst_port &&
891 id->src_port == q->id.src_port &&
892 id->proto == q->id.proto &&
893 id->flags == q->id.flags)
895 else if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
896 /* entry is idle and not in any heap, expire it */
897 struct dn_flow_queue *old_q = q ;
900 prev->next = q = q->next ;
902 fs->rq[i] = q = q->next ;
904 free(old_q, M_DUMMYNET);
910 if (q && prev != NULL) { /* found and not in front */
911 prev->next = q->next ;
912 q->next = fs->rq[i] ;
916 if (q == NULL) { /* no match, need to allocate a new entry */
917 q = create_queue(fs, i);
925 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
930 * RED calculates the average queue size (avg) using a low-pass filter
931 * with an exponential weighted (w_q) moving average:
932 * avg <- (1-w_q) * avg + w_q * q_size
933 * where q_size is the queue length (measured in bytes or * packets).
935 * If q_size == 0, we compute the idle time for the link, and set
936 * avg = (1 - w_q)^(idle/s)
937 * where s is the time needed for transmitting a medium-sized packet.
939 * Now, if avg < min_th the packet is enqueued.
940 * If avg > max_th the packet is dropped. Otherwise, the packet is
941 * dropped with probability P function of avg.
946 /* queue in bytes or packets ? */
947 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len;
949 DEB(printf("\n%d q: %2u ", (int) curr_time, q_size);)
951 /* average queue size estimation */
954 * queue is not empty, avg <- avg + (q_size - avg) * w_q
956 int diff = SCALE(q_size) - q->avg;
957 int64_t v = SCALE_MUL((int64_t) diff, (int64_t) fs->w_q);
962 * queue is empty, find for how long the queue has been
963 * empty and use a lookup table for computing
964 * (1 - * w_q)^(idle_time/s) where s is the time to send a
969 u_int t = (curr_time - q->q_time) / fs->lookup_step;
971 q->avg = (t < fs->lookup_depth) ?
972 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
975 DEB(printf("avg: %u ", SCALE_VAL(q->avg));)
977 /* should i drop ? */
979 if (q->avg < fs->min_th) {
981 return 0; /* accept packet ; */
983 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
984 if (fs->flags_fs & DN_IS_GENTLE_RED) {
986 * According to Gentle-RED, if avg is greater than max_th the
987 * packet is dropped with a probability
988 * p_b = c_3 * avg - c_4
989 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
991 p_b = SCALE_MUL((int64_t) fs->c_3, (int64_t) q->avg) - fs->c_4;
997 } else if (q->avg > fs->min_th) {
999 * we compute p_b using the linear dropping function p_b = c_1 *
1000 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
1001 * max_p * min_th / (max_th - min_th)
1003 p_b = SCALE_MUL((int64_t) fs->c_1, (int64_t) q->avg) - fs->c_2;
1005 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
1006 p_b = (p_b * len) / fs->max_pkt_size;
1007 if (++q->count == 0)
1008 q->random = random() & 0xffff;
1011 * q->count counts packets arrived since last drop, so a greater
1012 * value of q->count means a greater packet drop probability.
1014 if (SCALE_MUL(p_b, SCALE((int64_t) q->count)) > q->random) {
1016 DEB(printf("- red drop");)
1017 /* after a drop we calculate a new random value */
1018 q->random = random() & 0xffff;
1019 return 1; /* drop */
1022 /* end of RED algorithm */
1023 return 0 ; /* accept */
1027 struct dn_flow_set *
1028 locate_flowset(int pipe_nr, struct ip_fw *rule)
1031 struct dn_flow_set *fs;
1032 ipfw_insn *cmd = rule->cmd + rule->act_ofs;
1034 if (cmd->opcode == O_LOG)
1036 fs = ((ipfw_insn_pipe *)cmd)->pipe_ptr;
1041 if (cmd->opcode == O_QUEUE)
1043 struct dn_flow_set *fs = NULL ;
1045 if ( (rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_QUEUE )
1047 for (fs=all_flow_sets; fs && fs->fs_nr != pipe_nr; fs=fs->next)
1051 for (p1 = all_pipes; p1 && p1->pipe_nr != pipe_nr; p1 = p1->next)
1056 /* record for the future */
1058 ((ipfw_insn_pipe *)cmd)->pipe_ptr = fs;
1061 rule->pipe_ptr = fs;
1067 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1068 * depending on whether WF2Q or fixed bw is used.
1070 * pipe_nr pipe or queue the packet is destined for.
1071 * dir where shall we send the packet after dummynet.
1072 * m the mbuf with the packet
1073 * ifp the 'ifp' parameter from the caller.
1074 * NULL in ip_input, destination interface in ip_output,
1075 * real_dst in bdg_forward
1076 * ro route parameter (only used in ip_output, NULL otherwise)
1077 * dst destination address, only used by ip_output
1078 * rule matching rule, in case of multiple passes
1079 * flags flags from the caller, only used in ip_output
1083 dummynet_io(struct mbuf *m, int pipe_nr, int dir, struct ip_fw_args *fwa)
1086 struct dn_flow_set *fs;
1087 struct dn_pipe *pipe ;
1088 u_int64_t len = m->m_pkthdr.len ;
1089 struct dn_flow_queue *q = NULL ;
1093 ipfw_insn *cmd = fwa->rule->cmd + fwa->rule->act_ofs;
1095 if (cmd->opcode == O_LOG)
1097 is_pipe = (cmd->opcode == O_PIPE);
1099 is_pipe = (fwa->rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_PIPE;
1105 * this is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule
1107 fs = locate_flowset(pipe_nr, fwa->rule);
1109 goto dropit ; /* this queue/pipe does not exist! */
1111 if (pipe == NULL) { /* must be a queue, try find a matching pipe */
1112 for (pipe = all_pipes; pipe && pipe->pipe_nr != fs->parent_nr;
1118 printf("No pipe %d for queue %d, drop pkt\n",
1119 fs->parent_nr, fs->fs_nr);
1123 q = find_queue(fs, &(fwa->f_id));
1125 goto dropit ; /* cannot allocate queue */
1127 * update statistics, then check reasons to drop pkt
1129 q->tot_bytes += len ;
1131 if ( fs->plr && random() < fs->plr )
1132 goto dropit ; /* random pkt drop */
1133 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) {
1134 if (q->len_bytes > fs->qsize)
1135 goto dropit ; /* queue size overflow */
1137 if (q->len >= fs->qsize)
1138 goto dropit ; /* queue count overflow */
1140 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) )
1143 /* XXX expensive to zero, see if we can remove it*/
1144 pkt = (struct dn_pkt *)malloc(sizeof (*pkt), M_DUMMYNET, M_NOWAIT|M_ZERO);
1146 goto dropit ; /* cannot allocate packet header */
1147 /* ok, i can handle the pkt now... */
1148 /* build and enqueue packet + parameters */
1149 pkt->hdr.mh_type = MT_TAG;
1150 pkt->hdr.mh_flags = PACKET_TAG_DUMMYNET;
1151 pkt->rule = fwa->rule ;
1152 DN_NEXT(pkt) = NULL;
1156 pkt->ifp = fwa->oif;
1157 if (dir == DN_TO_IP_OUT) {
1159 * We need to copy *ro because for ICMP pkts (and maybe others)
1160 * the caller passed a pointer into the stack; dst might also be
1161 * a pointer into *ro so it needs to be updated.
1163 pkt->ro = *(fwa->ro);
1165 fwa->ro->ro_rt->rt_refcnt++ ;
1166 if (fwa->dst == (struct sockaddr_in *)&fwa->ro->ro_dst) /* dst points into ro */
1167 fwa->dst = (struct sockaddr_in *)&(pkt->ro.ro_dst) ;
1169 pkt->dn_dst = fwa->dst;
1170 pkt->flags = fwa->flags;
1172 if (q->head == NULL)
1175 DN_NEXT(q->tail) = pkt;
1178 q->len_bytes += len ;
1180 if ( q->head != pkt ) /* flow was not idle, we are done */
1183 * If we reach this point the flow was previously idle, so we need
1184 * to schedule it. This involves different actions for fixed-rate or
1189 * Fixed-rate queue: just insert into the ready_heap.
1192 if (pipe->bandwidth)
1193 t = SET_TICKS(pkt, q, pipe);
1194 q->sched_time = curr_time ;
1195 if (t == 0) /* must process it now */
1198 heap_insert(&ready_heap, curr_time + t , q );
1201 * WF2Q. First, compute start time S: if the flow was idle (S=F+1)
1202 * set S to the virtual time V for the controlling pipe, and update
1203 * the sum of weights for the pipe; otherwise, remove flow from
1204 * idle_heap and set S to max(F,V).
1205 * Second, compute finish time F = S + len/weight.
1206 * Third, if pipe was idle, update V=max(S, V).
1207 * Fourth, count one more backlogged flow.
1209 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */
1211 pipe->sum += fs->weight ; /* add weight of new queue */
1213 heap_extract(&(pipe->idle_heap), q);
1214 q->S = MAX64(q->F, pipe->V ) ;
1216 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight;
1218 if (pipe->not_eligible_heap.elements == 0 &&
1219 pipe->scheduler_heap.elements == 0)
1220 pipe->V = MAX64 ( q->S, pipe->V );
1223 * Look at eligibility. A flow is not eligibile if S>V (when
1224 * this happens, it means that there is some other flow already
1225 * scheduled for the same pipe, so the scheduler_heap cannot be
1226 * empty). If the flow is not eligible we just store it in the
1227 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1228 * and possibly invoke ready_event_wfq() right now if there is
1230 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1231 * and for all flows in not_eligible_heap (NEH), S_i > V .
1232 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH,
1233 * we only need to look into NEH.
1235 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */
1236 if (pipe->scheduler_heap.elements == 0)
1237 printf("++ ouch! not eligible but empty scheduler!\n");
1238 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1240 heap_insert(&(pipe->scheduler_heap), q->F, q);
1241 if (pipe->numbytes >= 0) { /* pipe is idle */
1242 if (pipe->scheduler_heap.elements != 1)
1243 printf("*** OUCH! pipe should have been idle!\n");
1244 DEB(printf("Waking up pipe %d at %d\n",
1245 pipe->pipe_nr, (int)(q->F >> MY_M)); )
1246 pipe->sched_time = curr_time ;
1247 ready_event_wfq(pipe);
1260 return ( (fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1264 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1265 * Doing this would probably save us the initial bzero of dn_pkt
1267 #define DN_FREE_PKT(pkt) { \
1268 struct dn_pkt *n = pkt ; \
1269 rt_unref ( n->ro.ro_rt ) ; \
1271 pkt = DN_NEXT(n) ; \
1272 free(n, M_DUMMYNET) ; }
1275 * Dispose all packets and flow_queues on a flow_set.
1276 * If all=1, also remove red lookup table and other storage,
1277 * including the descriptor itself.
1278 * For the one in dn_pipe MUST also cleanup ready_heap...
1281 purge_flow_set(struct dn_flow_set *fs, int all)
1283 struct dn_pkt *pkt ;
1284 struct dn_flow_queue *q, *qn ;
1287 for (i = 0 ; i <= fs->rq_size ; i++ ) {
1288 for (q = fs->rq[i] ; q ; q = qn ) {
1289 for (pkt = q->head ; pkt ; )
1292 free(q, M_DUMMYNET);
1296 fs->rq_elements = 0 ;
1298 /* RED - free lookup table */
1300 free(fs->w_q_lookup, M_DUMMYNET);
1302 free(fs->rq, M_DUMMYNET);
1303 /* if this fs is not part of a pipe, free it */
1304 if (fs->pipe && fs != &(fs->pipe->fs) )
1305 free(fs, M_DUMMYNET);
1310 * Dispose all packets queued on a pipe (not a flow_set).
1311 * Also free all resources associated to a pipe, which is about
1315 purge_pipe(struct dn_pipe *pipe)
1317 struct dn_pkt *pkt ;
1319 purge_flow_set( &(pipe->fs), 1 );
1321 for (pkt = pipe->head ; pkt ; )
1324 heap_free( &(pipe->scheduler_heap) );
1325 heap_free( &(pipe->not_eligible_heap) );
1326 heap_free( &(pipe->idle_heap) );
1330 * Delete all pipes and heaps returning memory. Must also
1331 * remove references from all ipfw rules to all pipes.
1336 struct dn_pipe *curr_p, *p ;
1337 struct dn_flow_set *fs, *curr_fs;
1342 /* remove all references to pipes ...*/
1343 flush_pipe_ptrs(NULL);
1344 /* prevent future matches... */
1347 fs = all_flow_sets ;
1348 all_flow_sets = NULL ;
1349 /* and free heaps so we don't have unwanted events */
1350 heap_free(&ready_heap);
1351 heap_free(&wfq_ready_heap);
1352 heap_free(&extract_heap);
1355 * Now purge all queued pkts and delete all pipes
1357 /* scan and purge all flow_sets. */
1361 purge_flow_set(curr_fs, 1);
1367 free(curr_p, M_DUMMYNET);
1372 extern struct ip_fw *ip_fw_default_rule ;
1374 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1377 struct dn_flow_queue *q ;
1378 struct dn_pkt *pkt ;
1380 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1381 for (q = fs->rq[i] ; q ; q = q->next )
1382 for (pkt = q->head ; pkt ; pkt = DN_NEXT(pkt) )
1384 pkt->rule = ip_fw_default_rule ;
1387 * when a firewall rule is deleted, scan all queues and remove the flow-id
1388 * from packets matching this rule.
1391 dn_rule_delete(void *r)
1394 struct dn_pkt *pkt ;
1395 struct dn_flow_set *fs ;
1398 * If the rule references a queue (dn_flow_set), then scan
1399 * the flow set, otherwise scan pipes. Should do either, but doing
1400 * both does not harm.
1402 for ( fs = all_flow_sets ; fs ; fs = fs->next )
1403 dn_rule_delete_fs(fs, r);
1404 for ( p = all_pipes ; p ; p = p->next ) {
1406 dn_rule_delete_fs(fs, r);
1407 for (pkt = p->head ; pkt ; pkt = DN_NEXT(pkt) )
1409 pkt->rule = ip_fw_default_rule ;
1414 * setup RED parameters
1417 config_red(struct dn_flow_set *p, struct dn_flow_set * x)
1422 x->min_th = SCALE(p->min_th);
1423 x->max_th = SCALE(p->max_th);
1424 x->max_p = p->max_p;
1426 x->c_1 = p->max_p / (p->max_th - p->min_th);
1427 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1428 if (x->flags_fs & DN_IS_GENTLE_RED) {
1429 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1430 x->c_4 = (SCALE(1) - 2 * p->max_p);
1433 /* if the lookup table already exist, free and create it again */
1434 if (x->w_q_lookup) {
1435 free(x->w_q_lookup, M_DUMMYNET);
1436 x->w_q_lookup = NULL ;
1438 if (red_lookup_depth == 0) {
1439 printf("\nnet.inet.ip.dummynet.red_lookup_depth must be > 0");
1440 free(x, M_DUMMYNET);
1443 x->lookup_depth = red_lookup_depth;
1444 x->w_q_lookup = (u_int *) malloc(x->lookup_depth * sizeof(int),
1445 M_DUMMYNET, M_NOWAIT);
1446 if (x->w_q_lookup == NULL) {
1447 printf("sorry, cannot allocate red lookup table\n");
1448 free(x, M_DUMMYNET);
1452 /* fill the lookup table with (1 - w_q)^x */
1453 x->lookup_step = p->lookup_step ;
1454 x->lookup_weight = p->lookup_weight ;
1455 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1456 for (i = 1; i < x->lookup_depth; i++)
1457 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1458 if (red_avg_pkt_size < 1)
1459 red_avg_pkt_size = 512 ;
1460 x->avg_pkt_size = red_avg_pkt_size ;
1461 if (red_max_pkt_size < 1)
1462 red_max_pkt_size = 1500 ;
1463 x->max_pkt_size = red_max_pkt_size ;
1468 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1470 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1471 int l = pfs->rq_size;
1477 else if (l > DN_MAX_HASH_SIZE)
1478 l = DN_MAX_HASH_SIZE;
1480 } else /* one is enough for null mask */
1482 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1483 M_DUMMYNET, M_NOWAIT | M_ZERO);
1484 if (x->rq == NULL) {
1485 printf("sorry, cannot allocate queue\n");
1493 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1495 x->flags_fs = src->flags_fs;
1496 x->qsize = src->qsize;
1498 x->flow_mask = src->flow_mask;
1499 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1500 if (x->qsize > 1024*1024)
1501 x->qsize = 1024*1024 ;
1508 /* configuring RED */
1509 if ( x->flags_fs & DN_IS_RED )
1510 config_red(src, x) ; /* XXX should check errors */
1514 * setup pipe or queue parameters.
1518 config_pipe(struct dn_pipe *p)
1521 struct dn_flow_set *pfs = &(p->fs);
1522 struct dn_flow_queue *q;
1525 * The config program passes parameters as follows:
1526 * bw = bits/second (0 means no limits),
1527 * delay = ms, must be translated into ticks.
1528 * qsize = slots/bytes
1530 p->delay = ( p->delay * hz ) / 1000 ;
1531 /* We need either a pipe number or a flow_set number */
1532 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1534 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1536 if (p->pipe_nr != 0) { /* this is a pipe */
1537 struct dn_pipe *x, *a, *b;
1539 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1540 a = b , b = b->next) ;
1542 if (b == NULL || b->pipe_nr != p->pipe_nr) { /* new pipe */
1543 x = malloc(sizeof(struct dn_pipe), M_DUMMYNET, M_NOWAIT | M_ZERO);
1545 printf("ip_dummynet.c: no memory for new pipe\n");
1548 x->pipe_nr = p->pipe_nr;
1550 /* idle_heap is the only one from which we extract from the middle.
1552 x->idle_heap.size = x->idle_heap.elements = 0 ;
1553 x->idle_heap.offset=OFFSET_OF(struct dn_flow_queue, heap_pos);
1557 /* Flush accumulated credit for all queues */
1558 for (i = 0; i <= x->fs.rq_size; i++)
1559 for (q = x->fs.rq[i]; q; q = q->next)
1565 x->bandwidth = p->bandwidth ;
1566 x->numbytes = 0; /* just in case... */
1567 bcopy(p->if_name, x->if_name, sizeof(p->if_name) );
1568 x->ifp = NULL ; /* reset interface ptr */
1569 x->delay = p->delay ;
1570 set_fs_parms(&(x->fs), pfs);
1573 if ( x->fs.rq == NULL ) { /* a new pipe */
1574 s = alloc_hash(&(x->fs), pfs) ;
1576 free(x, M_DUMMYNET);
1586 } else { /* config queue */
1587 struct dn_flow_set *x, *a, *b ;
1589 /* locate flow_set */
1590 for (a=NULL, b=all_flow_sets ; b && b->fs_nr < pfs->fs_nr ;
1591 a = b , b = b->next) ;
1593 if (b == NULL || b->fs_nr != pfs->fs_nr) { /* new */
1594 if (pfs->parent_nr == 0) /* need link to a pipe */
1596 x = malloc(sizeof(struct dn_flow_set), M_DUMMYNET, M_NOWAIT|M_ZERO);
1598 printf("ip_dummynet.c: no memory for new flow_set\n");
1601 x->fs_nr = pfs->fs_nr;
1602 x->parent_nr = pfs->parent_nr;
1603 x->weight = pfs->weight ;
1606 else if (x->weight > 100)
1609 /* Change parent pipe not allowed; must delete and recreate */
1610 if (pfs->parent_nr != 0 && b->parent_nr != pfs->parent_nr)
1615 set_fs_parms(x, pfs);
1617 if ( x->rq == NULL ) { /* a new flow_set */
1618 s = alloc_hash(x, pfs) ;
1620 free(x, M_DUMMYNET);
1635 * Helper function to remove from a heap queues which are linked to
1636 * a flow_set about to be deleted.
1639 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1641 int i = 0, found = 0 ;
1642 for (; i < h->elements ;)
1643 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1645 h->p[i] = h->p[h->elements] ;
1654 * helper function to remove a pipe from a heap (can be there at most once)
1657 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1659 if (h->elements > 0) {
1661 for (i=0; i < h->elements ; i++ ) {
1662 if (h->p[i].object == p) { /* found it */
1664 h->p[i] = h->p[h->elements] ;
1673 * drain all queues. Called in case of severe mbuf shortage.
1678 struct dn_flow_set *fs;
1682 heap_free(&ready_heap);
1683 heap_free(&wfq_ready_heap);
1684 heap_free(&extract_heap);
1685 /* remove all references to this pipe from flow_sets */
1686 for (fs = all_flow_sets; fs; fs= fs->next )
1687 purge_flow_set(fs, 0);
1689 for (p = all_pipes; p; p= p->next ) {
1690 purge_flow_set(&(p->fs), 0);
1691 for (pkt = p->head ; pkt ; )
1693 p->head = p->tail = NULL ;
1698 * Fully delete a pipe or a queue, cleaning up associated info.
1701 delete_pipe(struct dn_pipe *p)
1705 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1707 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1709 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1710 struct dn_pipe *a, *b;
1711 struct dn_flow_set *fs;
1714 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1715 a = b , b = b->next) ;
1716 if (b == NULL || (b->pipe_nr != p->pipe_nr) )
1717 return EINVAL ; /* not found */
1721 /* unlink from list of pipes */
1723 all_pipes = b->next ;
1726 /* remove references to this pipe from the ip_fw rules. */
1727 flush_pipe_ptrs(&(b->fs));
1729 /* remove all references to this pipe from flow_sets */
1730 for (fs = all_flow_sets; fs; fs= fs->next )
1731 if (fs->pipe == b) {
1732 printf("++ ref to pipe %d from fs %d\n",
1733 p->pipe_nr, fs->fs_nr);
1735 purge_flow_set(fs, 0);
1737 fs_remove_from_heap(&ready_heap, &(b->fs));
1738 purge_pipe(b); /* remove all data associated to this pipe */
1739 /* remove reference to here from extract_heap and wfq_ready_heap */
1740 pipe_remove_from_heap(&extract_heap, b);
1741 pipe_remove_from_heap(&wfq_ready_heap, b);
1743 free(b, M_DUMMYNET);
1744 } else { /* this is a WF2Q queue (dn_flow_set) */
1745 struct dn_flow_set *a, *b;
1748 for (a = NULL, b = all_flow_sets ; b && b->fs_nr < p->fs.fs_nr ;
1749 a = b , b = b->next) ;
1750 if (b == NULL || (b->fs_nr != p->fs.fs_nr) )
1751 return EINVAL ; /* not found */
1755 all_flow_sets = b->next ;
1758 /* remove references to this flow_set from the ip_fw rules. */
1761 if (b->pipe != NULL) {
1762 /* Update total weight on parent pipe and cleanup parent heaps */
1763 b->pipe->sum -= b->weight * b->backlogged ;
1764 fs_remove_from_heap(&(b->pipe->not_eligible_heap), b);
1765 fs_remove_from_heap(&(b->pipe->scheduler_heap), b);
1766 #if 1 /* XXX should i remove from idle_heap as well ? */
1767 fs_remove_from_heap(&(b->pipe->idle_heap), b);
1770 purge_flow_set(b, 1);
1777 * helper function used to copy data from kernel in DUMMYNET_GET
1780 dn_copy_set(struct dn_flow_set *set, char *bp)
1783 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1785 for (i = 0 ; i <= set->rq_size ; i++)
1786 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
1787 if (q->hash_slot != i)
1788 printf("++ at %d: wrong slot (have %d, "
1789 "should be %d)\n", copied, q->hash_slot, i);
1791 printf("++ at %d: wrong fs ptr (have %p, should be %p)\n",
1794 bcopy(q, qp, sizeof( *q ) );
1795 /* cleanup pointers */
1797 qp->head = qp->tail = NULL ;
1800 if (copied != set->rq_elements)
1801 printf("++ wrong count, have %d should be %d\n",
1802 copied, set->rq_elements);
1807 dummynet_get(struct sockopt *sopt)
1809 char *buf, *bp ; /* bp is the "copy-pointer" */
1811 struct dn_flow_set *set ;
1817 * compute size of data structures: list of pipes and flow_sets.
1819 for (p = all_pipes, size = 0 ; p ; p = p->next )
1820 size += sizeof( *p ) +
1821 p->fs.rq_elements * sizeof(struct dn_flow_queue);
1822 for (set = all_flow_sets ; set ; set = set->next )
1823 size += sizeof ( *set ) +
1824 set->rq_elements * sizeof(struct dn_flow_queue);
1825 buf = malloc(size, M_TEMP, M_NOWAIT);
1830 for (p = all_pipes, bp = buf ; p ; p = p->next ) {
1831 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp ;
1834 * copy pipe descriptor into *bp, convert delay back to ms,
1835 * then copy the flow_set descriptor(s) one at a time.
1836 * After each flow_set, copy the queue descriptor it owns.
1838 bcopy(p, bp, sizeof( *p ) );
1839 pipe_bp->delay = (pipe_bp->delay * 1000) / hz ;
1841 * XXX the following is a hack based on ->next being the
1842 * first field in dn_pipe and dn_flow_set. The correct
1843 * solution would be to move the dn_flow_set to the beginning
1844 * of struct dn_pipe.
1846 pipe_bp->next = (struct dn_pipe *)DN_IS_PIPE ;
1847 /* clean pointers */
1848 pipe_bp->head = pipe_bp->tail = NULL ;
1849 pipe_bp->fs.next = NULL ;
1850 pipe_bp->fs.pipe = NULL ;
1851 pipe_bp->fs.rq = NULL ;
1853 bp += sizeof( *p ) ;
1854 bp = dn_copy_set( &(p->fs), bp );
1856 for (set = all_flow_sets ; set ; set = set->next ) {
1857 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp ;
1858 bcopy(set, bp, sizeof( *set ) );
1859 /* XXX same hack as above */
1860 fs_bp->next = (struct dn_flow_set *)DN_IS_QUEUE ;
1861 fs_bp->pipe = NULL ;
1863 bp += sizeof( *set ) ;
1864 bp = dn_copy_set( set, bp );
1867 error = sooptcopyout(sopt, buf, size);
1873 * Handler for the various dummynet socket options (get, flush, config, del)
1876 ip_dn_ctl(struct sockopt *sopt)
1879 struct dn_pipe *p, tmp_pipe;
1881 /* Disallow sets in really-really secure mode. */
1882 if (sopt->sopt_dir == SOPT_SET) {
1883 #if __FreeBSD_version >= 500034
1884 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
1888 if (securelevel >= 3)
1893 switch (sopt->sopt_name) {
1895 printf("ip_dn_ctl -- unknown option %d", sopt->sopt_name);
1898 case IP_DUMMYNET_GET :
1899 error = dummynet_get(sopt);
1902 case IP_DUMMYNET_FLUSH :
1906 case IP_DUMMYNET_CONFIGURE :
1908 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
1911 error = config_pipe(p);
1914 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
1916 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
1920 error = delete_pipe(p);
1929 printf("DUMMYNET initialized (011031)\n");
1931 all_flow_sets = NULL ;
1932 ready_heap.size = ready_heap.elements = 0 ;
1933 ready_heap.offset = 0 ;
1935 wfq_ready_heap.size = wfq_ready_heap.elements = 0 ;
1936 wfq_ready_heap.offset = 0 ;
1938 extract_heap.size = extract_heap.elements = 0 ;
1939 extract_heap.offset = 0 ;
1940 ip_dn_ctl_ptr = ip_dn_ctl;
1941 ip_dn_io_ptr = dummynet_io;
1942 ip_dn_ruledel_ptr = dn_rule_delete;
1943 bzero(&dn_timeout, sizeof(struct callout_handle));
1944 dn_timeout = timeout(dummynet, NULL, 1);
1948 dummynet_modevent(module_t mod, int type, void *data)
1954 if (DUMMYNET_LOADED) {
1956 printf("DUMMYNET already loaded\n");
1964 #if !defined(KLD_MODULE)
1965 printf("dummynet statically compiled, cannot unload\n");
1969 untimeout(dummynet, NULL, dn_timeout);
1971 ip_dn_ctl_ptr = NULL;
1972 ip_dn_io_ptr = NULL;
1973 ip_dn_ruledel_ptr = NULL;
1983 static moduledata_t dummynet_mod = {
1988 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);
1989 MODULE_DEPEND(dummynet, ipfw, 1, 1, 1);
1990 MODULE_VERSION(dummynet, 1);