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
30 #define DUMMYNET_DEBUG
32 #if !defined(KLD_MODULE)
33 #include "opt_inet6.h"
37 * This module implements IP dummynet, a bandwidth limiter/delay emulator
38 * used in conjunction with the ipfw package.
39 * Description of the data structures used is in ip_dummynet.h
40 * Here you mainly find the following blocks of code:
41 * + variable declarations;
42 * + heap management functions;
43 * + scheduler and dummynet functions;
44 * + configuration and initialization.
46 * NOTA BENE: critical sections are protected by the "dummynet lock".
48 * Most important Changes:
51 * 010124: Fixed WF2Q behaviour
52 * 010122: Fixed spl protection.
53 * 000601: WF2Q support
54 * 000106: large rewrite, use heaps to handle very many pipes.
55 * 980513: initial release
57 * include files marked with XXX are probably not needed
60 #include <sys/param.h>
61 #include <sys/systm.h>
62 #include <sys/malloc.h>
64 #include <sys/kernel.h>
65 #include <sys/module.h>
67 #include <sys/socket.h>
68 #include <sys/socketvar.h>
70 #include <sys/sysctl.h>
72 #include <net/route.h>
73 #include <netinet/in.h>
74 #include <netinet/in_systm.h>
75 #include <netinet/in_var.h>
76 #include <netinet/ip.h>
77 #include <netinet/ip_fw.h>
78 #include <netinet/ip_dummynet.h>
79 #include <netinet/ip_var.h>
81 #include <netinet/if_ether.h> /* for struct arpcom */
82 #include <net/bridge.h>
84 #include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */
85 #include <netinet6/ip6_var.h>
88 * We keep a private variable for the simulation time, but we could
89 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c)
91 static dn_key curr_time = 0 ; /* current simulation time */
93 static int dn_hash_size = 64 ; /* default hash size */
95 /* statistics on number of queue searches and search steps */
96 static int searches, search_steps ;
97 static int pipe_expire = 1 ; /* expire queue if empty */
98 static int dn_max_ratio = 16 ; /* max queues/buckets ratio */
100 static int red_lookup_depth = 256; /* RED - default lookup table depth */
101 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
102 static int red_max_pkt_size = 1500; /* RED - default max packet size */
105 * Three heaps contain queues and pipes that the scheduler handles:
107 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
109 * wfq_ready_heap contains the pipes associated with WF2Q flows
111 * extract_heap contains pipes associated with delay lines.
115 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
117 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
119 static int heap_init(struct dn_heap *h, int size) ;
120 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
121 static void heap_extract(struct dn_heap *h, void *obj);
123 static void transmit_event(struct dn_pipe *pipe);
124 static void ready_event(struct dn_flow_queue *q);
126 static struct dn_pipe *all_pipes = NULL ; /* list of all pipes */
127 static struct dn_flow_set *all_flow_sets = NULL ;/* list of all flow_sets */
129 static struct callout dn_timeout;
132 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
133 CTLFLAG_RW, 0, "Dummynet");
134 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
135 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
136 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time,
137 CTLFLAG_RD, &curr_time, 0, "Current tick");
138 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
139 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
140 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
141 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
142 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches,
143 CTLFLAG_RD, &searches, 0, "Number of queue searches");
144 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps,
145 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
146 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
147 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
148 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
149 CTLFLAG_RW, &dn_max_ratio, 0,
150 "Max ratio between dynamic queues and buckets");
151 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
152 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
153 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
154 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
155 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
156 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
159 #ifdef DUMMYNET_DEBUG
160 int dummynet_debug = 0;
162 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, CTLFLAG_RW, &dummynet_debug,
163 0, "control debugging printfs");
165 #define DPRINTF(X) if (dummynet_debug) printf X
170 static struct mtx dummynet_mtx;
172 * NB: Recursion is needed to deal with re-entry via ICMP. That is,
173 * a packet may be dispatched via ip_input from dummynet_io and
174 * re-enter through ip_output. Yech.
176 #define DUMMYNET_LOCK_INIT() \
177 mtx_init(&dummynet_mtx, "dummynet", NULL, MTX_DEF | MTX_RECURSE)
178 #define DUMMYNET_LOCK_DESTROY() mtx_destroy(&dummynet_mtx)
179 #define DUMMYNET_LOCK() mtx_lock(&dummynet_mtx)
180 #define DUMMYNET_UNLOCK() mtx_unlock(&dummynet_mtx)
181 #define DUMMYNET_LOCK_ASSERT() do { \
182 mtx_assert(&dummynet_mtx, MA_OWNED); \
183 NET_ASSERT_GIANT(); \
186 static int config_pipe(struct dn_pipe *p);
187 static int ip_dn_ctl(struct sockopt *sopt);
189 static void dummynet(void *);
190 static void dummynet_flush(void);
191 void dummynet_drain(void);
192 static ip_dn_io_t dummynet_io;
193 static void dn_rule_delete(void *);
195 int if_tx_rdy(struct ifnet *ifp);
198 * Heap management functions.
200 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
201 * Some macros help finding parent/children so we can optimize them.
203 * heap_init() is called to expand the heap when needed.
204 * Increment size in blocks of 16 entries.
205 * XXX failure to allocate a new element is a pretty bad failure
206 * as we basically stall a whole queue forever!!
207 * Returns 1 on error, 0 on success
209 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
210 #define HEAP_LEFT(x) ( 2*(x) + 1 )
211 #define HEAP_IS_LEFT(x) ( (x) & 1 )
212 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
213 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
214 #define HEAP_INCREMENT 15
217 heap_init(struct dn_heap *h, int new_size)
219 struct dn_heap_entry *p;
221 if (h->size >= new_size ) {
222 printf("dummynet: %s, Bogus call, have %d want %d\n", __func__,
226 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
227 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_NOWAIT);
229 printf("dummynet: %s, resize %d failed\n", __func__, new_size );
230 return 1 ; /* error */
233 bcopy(h->p, p, h->size * sizeof(*p) );
234 free(h->p, M_DUMMYNET);
242 * Insert element in heap. Normally, p != NULL, we insert p in
243 * a new position and bubble up. If p == NULL, then the element is
244 * already in place, and key is the position where to start the
246 * Returns 1 on failure (cannot allocate new heap entry)
248 * If offset > 0 the position (index, int) of the element in the heap is
249 * also stored in the element itself at the given offset in bytes.
251 #define SET_OFFSET(heap, node) \
252 if (heap->offset > 0) \
253 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
255 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
257 #define RESET_OFFSET(heap, node) \
258 if (heap->offset > 0) \
259 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
261 heap_insert(struct dn_heap *h, dn_key key1, void *p)
263 int son = h->elements ;
265 if (p == NULL) /* data already there, set starting point */
267 else { /* insert new element at the end, possibly resize */
269 if (son == h->size) /* need resize... */
270 if (heap_init(h, h->elements+1) )
271 return 1 ; /* failure... */
272 h->p[son].object = p ;
273 h->p[son].key = key1 ;
276 while (son > 0) { /* bubble up */
277 int father = HEAP_FATHER(son) ;
278 struct dn_heap_entry tmp ;
280 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
281 break ; /* found right position */
282 /* son smaller than father, swap and repeat */
283 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
292 * remove top element from heap, or obj if obj != NULL
295 heap_extract(struct dn_heap *h, void *obj)
297 int child, father, max = h->elements - 1 ;
300 printf("dummynet: warning, extract from empty heap 0x%p\n", h);
303 father = 0 ; /* default: move up smallest child */
304 if (obj != NULL) { /* extract specific element, index is at offset */
306 panic("dummynet: heap_extract from middle not supported on this heap!!!\n");
307 father = *((int *)((char *)obj + h->offset)) ;
308 if (father < 0 || father >= h->elements) {
309 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
310 father, h->elements);
311 panic("dummynet: heap_extract");
314 RESET_OFFSET(h, father);
315 child = HEAP_LEFT(father) ; /* left child */
316 while (child <= max) { /* valid entry */
317 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
318 child = child+1 ; /* take right child, otherwise left */
319 h->p[father] = h->p[child] ;
320 SET_OFFSET(h, father);
322 child = HEAP_LEFT(child) ; /* left child for next loop */
327 * Fill hole with last entry and bubble up, reusing the insert code
329 h->p[father] = h->p[max] ;
330 heap_insert(h, father, NULL); /* this one cannot fail */
336 * change object position and update references
337 * XXX this one is never used!
340 heap_move(struct dn_heap *h, dn_key new_key, void *object)
344 int max = h->elements-1 ;
345 struct dn_heap_entry buf ;
348 panic("cannot move items on this heap");
350 i = *((int *)((char *)object + h->offset));
351 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
352 h->p[i].key = new_key ;
353 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
354 i = temp ) { /* bubble up */
355 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
358 } else { /* must move down */
359 h->p[i].key = new_key ;
360 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
361 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
362 temp++ ; /* select child with min key */
363 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
364 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
373 #endif /* heap_move, unused */
376 * heapify() will reorganize data inside an array to maintain the
377 * heap property. It is needed when we delete a bunch of entries.
380 heapify(struct dn_heap *h)
384 for (i = 0 ; i < h->elements ; i++ )
385 heap_insert(h, i , NULL) ;
389 * cleanup the heap and free data structure
392 heap_free(struct dn_heap *h)
395 free(h->p, M_DUMMYNET);
396 bzero(h, sizeof(*h) );
400 * --- end of heap management functions ---
404 * Return the mbuf tag holding the dummynet state. As an optimization
405 * this is assumed to be the first tag on the list. If this turns out
406 * wrong we'll need to search the list.
408 static struct dn_pkt_tag *
409 dn_tag_get(struct mbuf *m)
411 struct m_tag *mtag = m_tag_first(m);
412 KASSERT(mtag != NULL &&
413 mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
414 mtag->m_tag_id == PACKET_TAG_DUMMYNET,
415 ("packet on dummynet queue w/o dummynet tag!"));
416 return (struct dn_pkt_tag *)(mtag+1);
420 * Scheduler functions:
422 * transmit_event() is called when the delay-line needs to enter
423 * the scheduler, either because of existing pkts getting ready,
424 * or new packets entering the queue. The event handled is the delivery
425 * time of the packet.
427 * ready_event() does something similar with fixed-rate queues, and the
428 * event handled is the finish time of the head pkt.
430 * wfq_ready_event() does something similar with WF2Q queues, and the
431 * event handled is the start time of the head pkt.
433 * In all cases, we make sure that the data structures are consistent
434 * before passing pkts out, because this might trigger recursive
435 * invocations of the procedures.
438 transmit_event(struct dn_pipe *pipe)
441 struct dn_pkt_tag *pkt ;
444 DUMMYNET_LOCK_ASSERT();
446 while ( (m = pipe->head) ) {
448 if ( !DN_KEY_LEQ(pkt->output_time, curr_time) )
451 * first unlink, then call procedures, since ip_input() can invoke
452 * ip_output() and viceversa, thus causing nested calls
454 pipe->head = m->m_nextpkt ;
457 /* XXX: drop the lock for now to avoid LOR's */
459 switch (pkt->dn_dir) {
461 (void)ip_output(m, NULL, NULL, pkt->flags, NULL, NULL);
465 ip = mtod(m, struct ip *);
466 ip->ip_len = htons(ip->ip_len);
467 ip->ip_off = htons(ip->ip_off);
477 (void)ip6_output(m, NULL, NULL, pkt->flags, NULL, NULL, NULL);
483 * The bridge requires/assumes the Ethernet header is
484 * contiguous in the first mbuf header. Insure this is true.
487 if (m->m_len < ETHER_HDR_LEN &&
488 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
489 printf("dummynet/bridge: pullup fail, dropping pkt\n");
492 m = bdg_forward_ptr(m, pkt->ifp);
494 /* somebody unloaded the bridge module. Drop pkt */
496 printf("dummynet: dropping bridged packet trapped in pipe\n");
502 case DN_TO_ETH_DEMUX:
504 * The Ethernet code assumes the Ethernet header is
505 * contiguous in the first mbuf header. Insure this is true.
507 if (m->m_len < ETHER_HDR_LEN &&
508 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
509 printf("dummynet/ether: pullup fail, dropping pkt\n");
512 ether_demux(m->m_pkthdr.rcvif, m); /* which consumes the mbuf */
516 ether_output_frame(pkt->ifp, m);
520 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
526 /* if there are leftover packets, put into the heap for next event */
527 if ( (m = pipe->head) ) {
528 pkt = dn_tag_get(m) ;
529 /* XXX should check errors on heap_insert, by draining the
530 * whole pipe p and hoping in the future we are more successful
532 heap_insert(&extract_heap, pkt->output_time, pipe ) ;
537 * the following macro computes how many ticks we have to wait
538 * before being able to transmit a packet. The credit is taken from
539 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
541 #define SET_TICKS(_m, q, p) \
542 ((_m)->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \
546 * extract pkt from queue, compute output time (could be now)
547 * and put into delay line (p_queue)
550 move_pkt(struct mbuf *pkt, struct dn_flow_queue *q,
551 struct dn_pipe *p, int len)
553 struct dn_pkt_tag *dt = dn_tag_get(pkt);
555 q->head = pkt->m_nextpkt ;
557 q->len_bytes -= len ;
559 dt->output_time = curr_time + p->delay ;
564 p->tail->m_nextpkt = pkt;
566 p->tail->m_nextpkt = NULL;
570 * ready_event() is invoked every time the queue must enter the
571 * scheduler, either because the first packet arrives, or because
572 * a previously scheduled event fired.
573 * On invokation, drain as many pkts as possible (could be 0) and then
574 * if there are leftover packets reinsert the pkt in the scheduler.
577 ready_event(struct dn_flow_queue *q)
580 struct dn_pipe *p = q->fs->pipe ;
583 DUMMYNET_LOCK_ASSERT();
586 printf("dummynet: ready_event- pipe is gone\n");
589 p_was_empty = (p->head == NULL) ;
592 * schedule fixed-rate queues linked to this pipe:
593 * Account for the bw accumulated since last scheduling, then
594 * drain as many pkts as allowed by q->numbytes and move to
595 * the delay line (in p) computing output time.
596 * bandwidth==0 (no limit) means we can drain the whole queue,
597 * setting len_scaled = 0 does the job.
599 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth;
600 while ( (pkt = q->head) != NULL ) {
601 int len = pkt->m_pkthdr.len;
602 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
603 if (len_scaled > q->numbytes )
605 q->numbytes -= len_scaled ;
606 move_pkt(pkt, q, p, len);
609 * If we have more packets queued, schedule next ready event
610 * (can only occur when bandwidth != 0, otherwise we would have
611 * flushed the whole queue in the previous loop).
612 * To this purpose we record the current time and compute how many
613 * ticks to go for the finish time of the packet.
615 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */
616 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
617 q->sched_time = curr_time ;
618 heap_insert(&ready_heap, curr_time + t, (void *)q );
619 /* XXX should check errors on heap_insert, and drain the whole
620 * queue on error hoping next time we are luckier.
622 } else { /* RED needs to know when the queue becomes empty */
623 q->q_time = curr_time;
627 * If the delay line was empty call transmit_event(p) now.
628 * Otherwise, the scheduler will take care of it.
635 * Called when we can transmit packets on WF2Q queues. Take pkts out of
636 * the queues at their start time, and enqueue into the delay line.
637 * Packets are drained until p->numbytes < 0. As long as
638 * len_scaled >= p->numbytes, the packet goes into the delay line
639 * with a deadline p->delay. For the last packet, if p->numbytes<0,
640 * there is an additional delay.
643 ready_event_wfq(struct dn_pipe *p)
645 int p_was_empty = (p->head == NULL) ;
646 struct dn_heap *sch = &(p->scheduler_heap);
647 struct dn_heap *neh = &(p->not_eligible_heap) ;
649 DUMMYNET_LOCK_ASSERT();
651 if (p->if_name[0] == 0) /* tx clock is simulated */
652 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth;
653 else { /* tx clock is for real, the ifq must be empty or this is a NOP */
654 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
657 DPRINTF(("dummynet: pipe %d ready from %s --\n",
658 p->pipe_nr, p->if_name));
663 * While we have backlogged traffic AND credit, we need to do
664 * something on the queue.
666 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) {
667 if (sch->elements > 0) { /* have some eligible pkts to send out */
668 struct dn_flow_queue *q = sch->p[0].object ;
669 struct mbuf *pkt = q->head;
670 struct dn_flow_set *fs = q->fs;
671 u_int64_t len = pkt->m_pkthdr.len;
672 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
674 heap_extract(sch, NULL); /* remove queue from heap */
675 p->numbytes -= len_scaled ;
676 move_pkt(pkt, q, p, len);
678 p->V += (len<<MY_M) / p->sum ; /* update V */
679 q->S = q->F ; /* update start time */
680 if (q->len == 0) { /* Flow not backlogged any more */
682 heap_insert(&(p->idle_heap), q->F, q);
683 } else { /* still backlogged */
685 * update F and position in backlogged queue, then
686 * put flow in not_eligible_heap (we will fix this later).
688 len = (q->head)->m_pkthdr.len;
689 q->F += (len<<MY_M)/(u_int64_t) fs->weight ;
690 if (DN_KEY_LEQ(q->S, p->V))
691 heap_insert(neh, q->S, q);
693 heap_insert(sch, q->F, q);
697 * now compute V = max(V, min(S_i)). Remember that all elements in sch
698 * have by definition S_i <= V so if sch is not empty, V is surely
699 * the max and we must not update it. Conversely, if sch is empty
700 * we only need to look at neh.
702 if (sch->elements == 0 && neh->elements > 0)
703 p->V = MAX64 ( p->V, neh->p[0].key );
704 /* move from neh to sch any packets that have become eligible */
705 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) {
706 struct dn_flow_queue *q = neh->p[0].object ;
707 heap_extract(neh, NULL);
708 heap_insert(sch, q->F, q);
711 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */
712 p->numbytes = -1 ; /* mark not ready for I/O */
716 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0
717 && p->idle_heap.elements > 0) {
719 * no traffic and no events scheduled. We can get rid of idle-heap.
723 for (i = 0 ; i < p->idle_heap.elements ; i++) {
724 struct dn_flow_queue *q = p->idle_heap.p[i].object ;
731 p->idle_heap.elements = 0 ;
734 * If we are getting clocks from dummynet (not a real interface) and
735 * If we are under credit, schedule the next ready event.
736 * Also fix the delivery time of the last packet.
738 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */
739 dn_key t=0 ; /* number of ticks i have to wait */
741 if (p->bandwidth > 0)
742 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ;
743 dn_tag_get(p->tail)->output_time += t ;
744 p->sched_time = curr_time ;
745 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
746 /* XXX should check errors on heap_insert, and drain the whole
747 * queue on error hoping next time we are luckier.
751 * If the delay line was empty call transmit_event(p) now.
752 * Otherwise, the scheduler will take care of it.
759 * This is called once per tick, or HZ times per second. It is used to
760 * increment the current tick counter and schedule expired events.
763 dummynet(void * __unused unused)
765 void *p ; /* generic parameter to handler */
767 struct dn_heap *heaps[3];
771 heaps[0] = &ready_heap ; /* fixed-rate queues */
772 heaps[1] = &wfq_ready_heap ; /* wfq queues */
773 heaps[2] = &extract_heap ; /* delay line */
777 for (i=0; i < 3 ; i++) {
779 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time) ) {
780 if (h->p[0].key > curr_time)
781 printf("dummynet: warning, heap %d is %d ticks late\n",
782 i, (int)(curr_time - h->p[0].key));
783 p = h->p[0].object ; /* store a copy before heap_extract */
784 heap_extract(h, NULL); /* need to extract before processing */
788 struct dn_pipe *pipe = p;
789 if (pipe->if_name[0] != '\0')
790 printf("dummynet: bad ready_event_wfq for pipe %s\n",
798 /* sweep pipes trying to expire idle flow_queues */
799 for (pe = all_pipes; pe ; pe = pe->next )
800 if (pe->idle_heap.elements > 0 &&
801 DN_KEY_LT(pe->idle_heap.p[0].key, pe->V) ) {
802 struct dn_flow_queue *q = pe->idle_heap.p[0].object ;
804 heap_extract(&(pe->idle_heap), NULL);
805 q->S = q->F + 1 ; /* mark timestamp as invalid */
806 pe->sum -= q->fs->weight ;
810 callout_reset(&dn_timeout, 1, dummynet, NULL);
814 * called by an interface when tx_rdy occurs.
817 if_tx_rdy(struct ifnet *ifp)
822 for (p = all_pipes; p ; p = p->next )
826 for (p = all_pipes; p ; p = p->next )
827 if (!strcmp(p->if_name, ifp->if_xname) ) {
829 DPRINTF(("dummynet: ++ tx rdy from %s (now found)\n",
835 DPRINTF(("dummynet: ++ tx rdy from %s - qlen %d\n", ifp->if_xname,
836 ifp->if_snd.ifq_len));
837 p->numbytes = 0 ; /* mark ready for I/O */
846 * Unconditionally expire empty queues in case of shortage.
847 * Returns the number of queues freed.
850 expire_queues(struct dn_flow_set *fs)
852 struct dn_flow_queue *q, *prev ;
853 int i, initial_elements = fs->rq_elements ;
855 if (fs->last_expired == time_second)
857 fs->last_expired = time_second ;
858 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
859 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
860 if (q->head != NULL || q->S != q->F+1) {
863 } else { /* entry is idle, expire it */
864 struct dn_flow_queue *old_q = q ;
867 prev->next = q = q->next ;
869 fs->rq[i] = q = q->next ;
871 free(old_q, M_DUMMYNET);
873 return initial_elements - fs->rq_elements ;
877 * If room, create a new queue and put at head of slot i;
878 * otherwise, create or use the default queue.
880 static struct dn_flow_queue *
881 create_queue(struct dn_flow_set *fs, int i)
883 struct dn_flow_queue *q ;
885 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
886 expire_queues(fs) == 0) {
888 * No way to get room, use or create overflow queue.
891 if ( fs->rq[i] != NULL )
894 q = malloc(sizeof(*q), M_DUMMYNET, M_NOWAIT | M_ZERO);
896 printf("dummynet: sorry, cannot allocate queue for new flow\n");
901 q->next = fs->rq[i] ;
902 q->S = q->F + 1; /* hack - mark timestamp as invalid */
909 * Given a flow_set and a pkt in last_pkt, find a matching queue
910 * after appropriate masking. The queue is moved to front
911 * so that further searches take less time.
913 static struct dn_flow_queue *
914 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
916 int i = 0 ; /* we need i and q for new allocations */
917 struct dn_flow_queue *q, *prev;
918 int is_v6 = IS_IP6_FLOW_ID(id);
920 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
923 /* first, do the masking, then hash */
924 id->dst_port &= fs->flow_mask.dst_port ;
925 id->src_port &= fs->flow_mask.src_port ;
926 id->proto &= fs->flow_mask.proto ;
927 id->flags = 0 ; /* we don't care about this one */
929 APPLY_MASK(&id->dst_ip6, &fs->flow_mask.dst_ip6);
930 APPLY_MASK(&id->src_ip6, &fs->flow_mask.src_ip6);
931 id->flow_id6 &= fs->flow_mask.flow_id6;
933 i = ((id->dst_ip6.__u6_addr.__u6_addr32[0]) & 0xffff)^
934 ((id->dst_ip6.__u6_addr.__u6_addr32[1]) & 0xffff)^
935 ((id->dst_ip6.__u6_addr.__u6_addr32[2]) & 0xffff)^
936 ((id->dst_ip6.__u6_addr.__u6_addr32[3]) & 0xffff)^
938 ((id->dst_ip6.__u6_addr.__u6_addr32[0] >> 15) & 0xffff)^
939 ((id->dst_ip6.__u6_addr.__u6_addr32[1] >> 15) & 0xffff)^
940 ((id->dst_ip6.__u6_addr.__u6_addr32[2] >> 15) & 0xffff)^
941 ((id->dst_ip6.__u6_addr.__u6_addr32[3] >> 15) & 0xffff)^
943 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 1) & 0xfffff)^
944 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 1) & 0xfffff)^
945 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 1) & 0xfffff)^
946 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 1) & 0xfffff)^
948 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 16) & 0xffff)^
949 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 16) & 0xffff)^
950 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 16) & 0xffff)^
951 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 16) & 0xffff)^
953 (id->dst_port << 1) ^ (id->src_port) ^
957 id->dst_ip &= fs->flow_mask.dst_ip ;
958 id->src_ip &= fs->flow_mask.src_ip ;
960 i = ( (id->dst_ip) & 0xffff ) ^
961 ( (id->dst_ip >> 15) & 0xffff ) ^
962 ( (id->src_ip << 1) & 0xffff ) ^
963 ( (id->src_ip >> 16 ) & 0xffff ) ^
964 (id->dst_port << 1) ^ (id->src_port) ^
967 i = i % fs->rq_size ;
968 /* finally, scan the current list for a match */
970 for (prev=NULL, q = fs->rq[i] ; q ; ) {
973 IN6_ARE_ADDR_EQUAL(&id->dst_ip6,&q->id.dst_ip6) &&
974 IN6_ARE_ADDR_EQUAL(&id->src_ip6,&q->id.src_ip6) &&
975 id->dst_port == q->id.dst_port &&
976 id->src_port == q->id.src_port &&
977 id->proto == q->id.proto &&
978 id->flags == q->id.flags &&
979 id->flow_id6 == q->id.flow_id6)
982 if (!is_v6 && id->dst_ip == q->id.dst_ip &&
983 id->src_ip == q->id.src_ip &&
984 id->dst_port == q->id.dst_port &&
985 id->src_port == q->id.src_port &&
986 id->proto == q->id.proto &&
987 id->flags == q->id.flags)
990 /* No match. Check if we can expire the entry */
991 if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
992 /* entry is idle and not in any heap, expire it */
993 struct dn_flow_queue *old_q = q ;
996 prev->next = q = q->next ;
998 fs->rq[i] = q = q->next ;
1000 free(old_q, M_DUMMYNET);
1006 if (q && prev != NULL) { /* found and not in front */
1007 prev->next = q->next ;
1008 q->next = fs->rq[i] ;
1012 if (q == NULL) { /* no match, need to allocate a new entry */
1013 q = create_queue(fs, i);
1021 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
1026 * RED calculates the average queue size (avg) using a low-pass filter
1027 * with an exponential weighted (w_q) moving average:
1028 * avg <- (1-w_q) * avg + w_q * q_size
1029 * where q_size is the queue length (measured in bytes or * packets).
1031 * If q_size == 0, we compute the idle time for the link, and set
1032 * avg = (1 - w_q)^(idle/s)
1033 * where s is the time needed for transmitting a medium-sized packet.
1035 * Now, if avg < min_th the packet is enqueued.
1036 * If avg > max_th the packet is dropped. Otherwise, the packet is
1037 * dropped with probability P function of avg.
1042 /* queue in bytes or packets ? */
1043 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len;
1045 DPRINTF(("\ndummynet: %d q: %2u ", (int) curr_time, q_size));
1047 /* average queue size estimation */
1050 * queue is not empty, avg <- avg + (q_size - avg) * w_q
1052 int diff = SCALE(q_size) - q->avg;
1053 int64_t v = SCALE_MUL((int64_t) diff, (int64_t) fs->w_q);
1058 * queue is empty, find for how long the queue has been
1059 * empty and use a lookup table for computing
1060 * (1 - * w_q)^(idle_time/s) where s is the time to send a
1062 * XXX check wraps...
1065 u_int t = (curr_time - q->q_time) / fs->lookup_step;
1067 q->avg = (t < fs->lookup_depth) ?
1068 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
1071 DPRINTF(("dummynet: avg: %u ", SCALE_VAL(q->avg)));
1073 /* should i drop ? */
1075 if (q->avg < fs->min_th) {
1077 return 0; /* accept packet ; */
1079 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
1080 if (fs->flags_fs & DN_IS_GENTLE_RED) {
1082 * According to Gentle-RED, if avg is greater than max_th the
1083 * packet is dropped with a probability
1084 * p_b = c_3 * avg - c_4
1085 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
1087 p_b = SCALE_MUL((int64_t) fs->c_3, (int64_t) q->avg) - fs->c_4;
1090 DPRINTF(("dummynet: - drop"));
1093 } else if (q->avg > fs->min_th) {
1095 * we compute p_b using the linear dropping function p_b = c_1 *
1096 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
1097 * max_p * min_th / (max_th - min_th)
1099 p_b = SCALE_MUL((int64_t) fs->c_1, (int64_t) q->avg) - fs->c_2;
1101 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
1102 p_b = (p_b * len) / fs->max_pkt_size;
1103 if (++q->count == 0)
1104 q->random = random() & 0xffff;
1107 * q->count counts packets arrived since last drop, so a greater
1108 * value of q->count means a greater packet drop probability.
1110 if (SCALE_MUL(p_b, SCALE((int64_t) q->count)) > q->random) {
1112 DPRINTF(("dummynet: - red drop"));
1113 /* after a drop we calculate a new random value */
1114 q->random = random() & 0xffff;
1115 return 1; /* drop */
1118 /* end of RED algorithm */
1119 return 0 ; /* accept */
1123 struct dn_flow_set *
1124 locate_flowset(int pipe_nr, struct ip_fw *rule)
1126 struct dn_flow_set *fs;
1127 ipfw_insn *cmd = ACTION_PTR(rule);
1129 if (cmd->opcode == O_LOG)
1132 fs = ((ipfw_insn_pipe *)cmd)->pipe_ptr;
1134 bcopy(& ((ipfw_insn_pipe *)cmd)->pipe_ptr, &fs, sizeof(fs));
1140 if (cmd->opcode == O_QUEUE)
1141 for (fs=all_flow_sets; fs && fs->fs_nr != pipe_nr; fs=fs->next)
1145 for (p1 = all_pipes; p1 && p1->pipe_nr != pipe_nr; p1 = p1->next)
1150 /* record for the future */
1152 ((ipfw_insn_pipe *)cmd)->pipe_ptr = fs;
1154 bcopy(&fs, & ((ipfw_insn_pipe *)cmd)->pipe_ptr, sizeof(fs));
1160 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1161 * depending on whether WF2Q or fixed bw is used.
1163 * pipe_nr pipe or queue the packet is destined for.
1164 * dir where shall we send the packet after dummynet.
1165 * m the mbuf with the packet
1166 * ifp the 'ifp' parameter from the caller.
1167 * NULL in ip_input, destination interface in ip_output,
1168 * real_dst in bdg_forward
1169 * rule matching rule, in case of multiple passes
1170 * flags flags from the caller, only used in ip_output
1174 dummynet_io(struct mbuf *m, int dir, struct ip_fw_args *fwa)
1176 struct dn_pkt_tag *pkt;
1178 struct dn_flow_set *fs;
1179 struct dn_pipe *pipe ;
1180 u_int64_t len = m->m_pkthdr.len ;
1181 struct dn_flow_queue *q = NULL ;
1183 ipfw_insn *cmd = ACTION_PTR(fwa->rule);
1185 KASSERT(m->m_nextpkt == NULL,
1186 ("dummynet_io: mbuf queue passed to dummynet"));
1188 if (cmd->opcode == O_LOG)
1190 is_pipe = (cmd->opcode == O_PIPE);
1194 * This is a dummynet rule, so we expect an O_PIPE or O_QUEUE rule.
1196 fs = locate_flowset(fwa->cookie, fwa->rule);
1198 goto dropit ; /* this queue/pipe does not exist! */
1200 if (pipe == NULL) { /* must be a queue, try find a matching pipe */
1201 for (pipe = all_pipes; pipe && pipe->pipe_nr != fs->parent_nr;
1207 printf("dummynet: no pipe %d for queue %d, drop pkt\n",
1208 fs->parent_nr, fs->fs_nr);
1212 q = find_queue(fs, &(fwa->f_id));
1214 goto dropit ; /* cannot allocate queue */
1216 * update statistics, then check reasons to drop pkt
1218 q->tot_bytes += len ;
1220 if ( fs->plr && random() < fs->plr )
1221 goto dropit ; /* random pkt drop */
1222 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) {
1223 if (q->len_bytes > fs->qsize)
1224 goto dropit ; /* queue size overflow */
1226 if (q->len >= fs->qsize)
1227 goto dropit ; /* queue count overflow */
1229 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) )
1232 /* XXX expensive to zero, see if we can remove it*/
1233 mtag = m_tag_get(PACKET_TAG_DUMMYNET,
1234 sizeof(struct dn_pkt_tag), M_NOWAIT|M_ZERO);
1236 goto dropit ; /* cannot allocate packet header */
1237 m_tag_prepend(m, mtag); /* attach to mbuf chain */
1239 pkt = (struct dn_pkt_tag *)(mtag+1);
1240 /* ok, i can handle the pkt now... */
1241 /* build and enqueue packet + parameters */
1242 pkt->rule = fwa->rule ;
1245 pkt->ifp = fwa->oif;
1246 if (dir == DN_TO_IP_OUT || dir == DN_TO_IP6_OUT)
1247 pkt->flags = fwa->flags;
1249 if (q->head == NULL)
1252 q->tail->m_nextpkt = m;
1255 q->len_bytes += len ;
1257 if ( q->head != m ) /* flow was not idle, we are done */
1260 * If we reach this point the flow was previously idle, so we need
1261 * to schedule it. This involves different actions for fixed-rate or
1266 * Fixed-rate queue: just insert into the ready_heap.
1269 if (pipe->bandwidth)
1270 t = SET_TICKS(m, q, pipe);
1271 q->sched_time = curr_time ;
1272 if (t == 0) /* must process it now */
1275 heap_insert(&ready_heap, curr_time + t , q );
1278 * WF2Q. First, compute start time S: if the flow was idle (S=F+1)
1279 * set S to the virtual time V for the controlling pipe, and update
1280 * the sum of weights for the pipe; otherwise, remove flow from
1281 * idle_heap and set S to max(F,V).
1282 * Second, compute finish time F = S + len/weight.
1283 * Third, if pipe was idle, update V=max(S, V).
1284 * Fourth, count one more backlogged flow.
1286 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */
1288 pipe->sum += fs->weight ; /* add weight of new queue */
1290 heap_extract(&(pipe->idle_heap), q);
1291 q->S = MAX64(q->F, pipe->V ) ;
1293 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight;
1295 if (pipe->not_eligible_heap.elements == 0 &&
1296 pipe->scheduler_heap.elements == 0)
1297 pipe->V = MAX64 ( q->S, pipe->V );
1300 * Look at eligibility. A flow is not eligibile if S>V (when
1301 * this happens, it means that there is some other flow already
1302 * scheduled for the same pipe, so the scheduler_heap cannot be
1303 * empty). If the flow is not eligible we just store it in the
1304 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1305 * and possibly invoke ready_event_wfq() right now if there is
1307 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1308 * and for all flows in not_eligible_heap (NEH), S_i > V .
1309 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH,
1310 * we only need to look into NEH.
1312 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */
1313 if (pipe->scheduler_heap.elements == 0)
1314 printf("dummynet: ++ ouch! not eligible but empty scheduler!\n");
1315 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1317 heap_insert(&(pipe->scheduler_heap), q->F, q);
1318 if (pipe->numbytes >= 0) { /* pipe is idle */
1319 if (pipe->scheduler_heap.elements != 1)
1320 printf("dummynet: OUCH! pipe should have been idle!\n");
1321 DPRINTF(("dummynet: waking up pipe %d at %d\n",
1322 pipe->pipe_nr, (int)(q->F >> MY_M)));
1323 pipe->sched_time = curr_time ;
1324 ready_event_wfq(pipe);
1337 return ( (fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1341 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1342 * Doing this would probably save us the initial bzero of dn_pkt
1344 #define DN_FREE_PKT(_m) do { \
1349 * Dispose all packets and flow_queues on a flow_set.
1350 * If all=1, also remove red lookup table and other storage,
1351 * including the descriptor itself.
1352 * For the one in dn_pipe MUST also cleanup ready_heap...
1355 purge_flow_set(struct dn_flow_set *fs, int all)
1357 struct dn_flow_queue *q, *qn ;
1360 DUMMYNET_LOCK_ASSERT();
1362 for (i = 0 ; i <= fs->rq_size ; i++ ) {
1363 for (q = fs->rq[i] ; q ; q = qn ) {
1364 struct mbuf *m, *mnext;
1367 while ((m = mnext) != NULL) {
1368 mnext = m->m_nextpkt;
1372 free(q, M_DUMMYNET);
1376 fs->rq_elements = 0 ;
1378 /* RED - free lookup table */
1380 free(fs->w_q_lookup, M_DUMMYNET);
1382 free(fs->rq, M_DUMMYNET);
1383 /* if this fs is not part of a pipe, free it */
1384 if (fs->pipe && fs != &(fs->pipe->fs) )
1385 free(fs, M_DUMMYNET);
1390 * Dispose all packets queued on a pipe (not a flow_set).
1391 * Also free all resources associated to a pipe, which is about
1395 purge_pipe(struct dn_pipe *pipe)
1397 struct mbuf *m, *mnext;
1399 purge_flow_set( &(pipe->fs), 1 );
1402 while ((m = mnext) != NULL) {
1403 mnext = m->m_nextpkt;
1407 heap_free( &(pipe->scheduler_heap) );
1408 heap_free( &(pipe->not_eligible_heap) );
1409 heap_free( &(pipe->idle_heap) );
1413 * Delete all pipes and heaps returning memory. Must also
1414 * remove references from all ipfw rules to all pipes.
1417 dummynet_flush(void)
1419 struct dn_pipe *curr_p, *p ;
1420 struct dn_flow_set *fs, *curr_fs;
1423 /* remove all references to pipes ...*/
1424 flush_pipe_ptrs(NULL);
1425 /* prevent future matches... */
1428 fs = all_flow_sets ;
1429 all_flow_sets = NULL ;
1430 /* and free heaps so we don't have unwanted events */
1431 heap_free(&ready_heap);
1432 heap_free(&wfq_ready_heap);
1433 heap_free(&extract_heap);
1436 * Now purge all queued pkts and delete all pipes
1438 /* scan and purge all flow_sets. */
1442 purge_flow_set(curr_fs, 1);
1448 free(curr_p, M_DUMMYNET);
1454 extern struct ip_fw *ip_fw_default_rule ;
1456 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1459 struct dn_flow_queue *q ;
1462 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1463 for (q = fs->rq[i] ; q ; q = q->next )
1464 for (m = q->head ; m ; m = m->m_nextpkt ) {
1465 struct dn_pkt_tag *pkt = dn_tag_get(m) ;
1467 pkt->rule = ip_fw_default_rule ;
1471 * when a firewall rule is deleted, scan all queues and remove the flow-id
1472 * from packets matching this rule.
1475 dn_rule_delete(void *r)
1478 struct dn_flow_set *fs ;
1479 struct dn_pkt_tag *pkt ;
1484 * If the rule references a queue (dn_flow_set), then scan
1485 * the flow set, otherwise scan pipes. Should do either, but doing
1486 * both does not harm.
1488 for ( fs = all_flow_sets ; fs ; fs = fs->next )
1489 dn_rule_delete_fs(fs, r);
1490 for ( p = all_pipes ; p ; p = p->next ) {
1492 dn_rule_delete_fs(fs, r);
1493 for (m = p->head ; m ; m = m->m_nextpkt ) {
1494 pkt = dn_tag_get(m) ;
1496 pkt->rule = ip_fw_default_rule ;
1503 * setup RED parameters
1506 config_red(struct dn_flow_set *p, struct dn_flow_set * x)
1511 x->min_th = SCALE(p->min_th);
1512 x->max_th = SCALE(p->max_th);
1513 x->max_p = p->max_p;
1515 x->c_1 = p->max_p / (p->max_th - p->min_th);
1516 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1517 if (x->flags_fs & DN_IS_GENTLE_RED) {
1518 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1519 x->c_4 = (SCALE(1) - 2 * p->max_p);
1522 /* if the lookup table already exist, free and create it again */
1523 if (x->w_q_lookup) {
1524 free(x->w_q_lookup, M_DUMMYNET);
1525 x->w_q_lookup = NULL ;
1527 if (red_lookup_depth == 0) {
1528 printf("\ndummynet: net.inet.ip.dummynet.red_lookup_depth must be > 0\n");
1529 free(x, M_DUMMYNET);
1532 x->lookup_depth = red_lookup_depth;
1533 x->w_q_lookup = (u_int *) malloc(x->lookup_depth * sizeof(int),
1534 M_DUMMYNET, M_NOWAIT);
1535 if (x->w_q_lookup == NULL) {
1536 printf("dummynet: sorry, cannot allocate red lookup table\n");
1537 free(x, M_DUMMYNET);
1541 /* fill the lookup table with (1 - w_q)^x */
1542 x->lookup_step = p->lookup_step ;
1543 x->lookup_weight = p->lookup_weight ;
1544 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1545 for (i = 1; i < x->lookup_depth; i++)
1546 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1547 if (red_avg_pkt_size < 1)
1548 red_avg_pkt_size = 512 ;
1549 x->avg_pkt_size = red_avg_pkt_size ;
1550 if (red_max_pkt_size < 1)
1551 red_max_pkt_size = 1500 ;
1552 x->max_pkt_size = red_max_pkt_size ;
1557 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1559 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1560 int l = pfs->rq_size;
1566 else if (l > DN_MAX_HASH_SIZE)
1567 l = DN_MAX_HASH_SIZE;
1569 } else /* one is enough for null mask */
1571 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1572 M_DUMMYNET, M_NOWAIT | M_ZERO);
1573 if (x->rq == NULL) {
1574 printf("dummynet: sorry, cannot allocate queue\n");
1582 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1584 x->flags_fs = src->flags_fs;
1585 x->qsize = src->qsize;
1587 x->flow_mask = src->flow_mask;
1588 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1589 if (x->qsize > 1024*1024)
1590 x->qsize = 1024*1024 ;
1597 /* configuring RED */
1598 if ( x->flags_fs & DN_IS_RED )
1599 config_red(src, x) ; /* XXX should check errors */
1603 * setup pipe or queue parameters.
1607 config_pipe(struct dn_pipe *p)
1610 struct dn_flow_set *pfs = &(p->fs);
1611 struct dn_flow_queue *q;
1614 * The config program passes parameters as follows:
1615 * bw = bits/second (0 means no limits),
1616 * delay = ms, must be translated into ticks.
1617 * qsize = slots/bytes
1619 p->delay = ( p->delay * hz ) / 1000 ;
1620 /* We need either a pipe number or a flow_set number */
1621 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1623 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1625 if (p->pipe_nr != 0) { /* this is a pipe */
1626 struct dn_pipe *x, *a, *b;
1630 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1631 a = b , b = b->next) ;
1633 if (b == NULL || b->pipe_nr != p->pipe_nr) { /* new pipe */
1634 x = malloc(sizeof(struct dn_pipe), M_DUMMYNET, M_NOWAIT | M_ZERO);
1637 printf("dummynet: no memory for new pipe\n");
1640 x->pipe_nr = p->pipe_nr;
1642 /* idle_heap is the only one from which we extract from the middle.
1644 x->idle_heap.size = x->idle_heap.elements = 0 ;
1645 x->idle_heap.offset=OFFSET_OF(struct dn_flow_queue, heap_pos);
1648 /* Flush accumulated credit for all queues */
1649 for (i = 0; i <= x->fs.rq_size; i++)
1650 for (q = x->fs.rq[i]; q; q = q->next)
1654 x->bandwidth = p->bandwidth ;
1655 x->numbytes = 0; /* just in case... */
1656 bcopy(p->if_name, x->if_name, sizeof(p->if_name) );
1657 x->ifp = NULL ; /* reset interface ptr */
1658 x->delay = p->delay ;
1659 set_fs_parms(&(x->fs), pfs);
1662 if ( x->fs.rq == NULL ) { /* a new pipe */
1663 r = alloc_hash(&(x->fs), pfs) ;
1666 free(x, M_DUMMYNET);
1676 } else { /* config queue */
1677 struct dn_flow_set *x, *a, *b ;
1680 /* locate flow_set */
1681 for (a=NULL, b=all_flow_sets ; b && b->fs_nr < pfs->fs_nr ;
1682 a = b , b = b->next) ;
1684 if (b == NULL || b->fs_nr != pfs->fs_nr) { /* new */
1685 if (pfs->parent_nr == 0) { /* need link to a pipe */
1689 x = malloc(sizeof(struct dn_flow_set), M_DUMMYNET, M_NOWAIT|M_ZERO);
1692 printf("dummynet: no memory for new flow_set\n");
1695 x->fs_nr = pfs->fs_nr;
1696 x->parent_nr = pfs->parent_nr;
1697 x->weight = pfs->weight ;
1700 else if (x->weight > 100)
1703 /* Change parent pipe not allowed; must delete and recreate */
1704 if (pfs->parent_nr != 0 && b->parent_nr != pfs->parent_nr) {
1710 set_fs_parms(x, pfs);
1712 if ( x->rq == NULL ) { /* a new flow_set */
1713 r = alloc_hash(x, pfs) ;
1716 free(x, M_DUMMYNET);
1731 * Helper function to remove from a heap queues which are linked to
1732 * a flow_set about to be deleted.
1735 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1737 int i = 0, found = 0 ;
1738 for (; i < h->elements ;)
1739 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1741 h->p[i] = h->p[h->elements] ;
1750 * helper function to remove a pipe from a heap (can be there at most once)
1753 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1755 if (h->elements > 0) {
1757 for (i=0; i < h->elements ; i++ ) {
1758 if (h->p[i].object == p) { /* found it */
1760 h->p[i] = h->p[h->elements] ;
1769 * drain all queues. Called in case of severe mbuf shortage.
1774 struct dn_flow_set *fs;
1776 struct mbuf *m, *mnext;
1778 DUMMYNET_LOCK_ASSERT();
1780 heap_free(&ready_heap);
1781 heap_free(&wfq_ready_heap);
1782 heap_free(&extract_heap);
1783 /* remove all references to this pipe from flow_sets */
1784 for (fs = all_flow_sets; fs; fs= fs->next )
1785 purge_flow_set(fs, 0);
1787 for (p = all_pipes; p; p= p->next ) {
1788 purge_flow_set(&(p->fs), 0);
1791 while ((m = mnext) != NULL) {
1792 mnext = m->m_nextpkt;
1795 p->head = p->tail = NULL ;
1800 * Fully delete a pipe or a queue, cleaning up associated info.
1803 delete_pipe(struct dn_pipe *p)
1805 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1807 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1809 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1810 struct dn_pipe *a, *b;
1811 struct dn_flow_set *fs;
1815 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ;
1816 a = b , b = b->next) ;
1817 if (b == NULL || (b->pipe_nr != p->pipe_nr) ) {
1819 return EINVAL ; /* not found */
1822 /* unlink from list of pipes */
1824 all_pipes = b->next ;
1827 /* remove references to this pipe from the ip_fw rules. */
1828 flush_pipe_ptrs(&(b->fs));
1830 /* remove all references to this pipe from flow_sets */
1831 for (fs = all_flow_sets; fs; fs= fs->next )
1832 if (fs->pipe == b) {
1833 printf("dummynet: ++ ref to pipe %d from fs %d\n",
1834 p->pipe_nr, fs->fs_nr);
1836 purge_flow_set(fs, 0);
1838 fs_remove_from_heap(&ready_heap, &(b->fs));
1839 purge_pipe(b); /* remove all data associated to this pipe */
1840 /* remove reference to here from extract_heap and wfq_ready_heap */
1841 pipe_remove_from_heap(&extract_heap, b);
1842 pipe_remove_from_heap(&wfq_ready_heap, b);
1845 free(b, M_DUMMYNET);
1846 } else { /* this is a WF2Q queue (dn_flow_set) */
1847 struct dn_flow_set *a, *b;
1851 for (a = NULL, b = all_flow_sets ; b && b->fs_nr < p->fs.fs_nr ;
1852 a = b , b = b->next) ;
1853 if (b == NULL || (b->fs_nr != p->fs.fs_nr) ) {
1855 return EINVAL ; /* not found */
1859 all_flow_sets = b->next ;
1862 /* remove references to this flow_set from the ip_fw rules. */
1865 if (b->pipe != NULL) {
1866 /* Update total weight on parent pipe and cleanup parent heaps */
1867 b->pipe->sum -= b->weight * b->backlogged ;
1868 fs_remove_from_heap(&(b->pipe->not_eligible_heap), b);
1869 fs_remove_from_heap(&(b->pipe->scheduler_heap), b);
1870 #if 1 /* XXX should i remove from idle_heap as well ? */
1871 fs_remove_from_heap(&(b->pipe->idle_heap), b);
1874 purge_flow_set(b, 1);
1881 * helper function used to copy data from kernel in DUMMYNET_GET
1884 dn_copy_set(struct dn_flow_set *set, char *bp)
1887 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1889 DUMMYNET_LOCK_ASSERT();
1891 for (i = 0 ; i <= set->rq_size ; i++)
1892 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
1893 if (q->hash_slot != i)
1894 printf("dummynet: ++ at %d: wrong slot (have %d, "
1895 "should be %d)\n", copied, q->hash_slot, i);
1897 printf("dummynet: ++ at %d: wrong fs ptr (have %p, should be %p)\n",
1900 bcopy(q, qp, sizeof( *q ) );
1901 /* cleanup pointers */
1903 qp->head = qp->tail = NULL ;
1906 if (copied != set->rq_elements)
1907 printf("dummynet: ++ wrong count, have %d should be %d\n",
1908 copied, set->rq_elements);
1915 struct dn_flow_set *set ;
1919 DUMMYNET_LOCK_ASSERT();
1921 * compute size of data structures: list of pipes and flow_sets.
1923 for (p = all_pipes, size = 0 ; p ; p = p->next )
1924 size += sizeof( *p ) +
1925 p->fs.rq_elements * sizeof(struct dn_flow_queue);
1926 for (set = all_flow_sets ; set ; set = set->next )
1927 size += sizeof ( *set ) +
1928 set->rq_elements * sizeof(struct dn_flow_queue);
1933 dummynet_get(struct sockopt *sopt)
1935 char *buf, *bp ; /* bp is the "copy-pointer" */
1937 struct dn_flow_set *set ;
1941 /* XXX lock held too long */
1944 * XXX: Ugly, but we need to allocate memory with M_WAITOK flag and we
1945 * cannot use this flag while holding a mutex.
1947 for (i = 0; i < 10; i++) {
1948 size = dn_calc_size();
1950 buf = malloc(size, M_TEMP, M_WAITOK);
1952 if (size == dn_calc_size())
1961 for (p = all_pipes, bp = buf ; p ; p = p->next ) {
1962 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp ;
1965 * copy pipe descriptor into *bp, convert delay back to ms,
1966 * then copy the flow_set descriptor(s) one at a time.
1967 * After each flow_set, copy the queue descriptor it owns.
1969 bcopy(p, bp, sizeof( *p ) );
1970 pipe_bp->delay = (pipe_bp->delay * 1000) / hz ;
1972 * XXX the following is a hack based on ->next being the
1973 * first field in dn_pipe and dn_flow_set. The correct
1974 * solution would be to move the dn_flow_set to the beginning
1975 * of struct dn_pipe.
1977 pipe_bp->next = (struct dn_pipe *)DN_IS_PIPE ;
1978 /* clean pointers */
1979 pipe_bp->head = pipe_bp->tail = NULL ;
1980 pipe_bp->fs.next = NULL ;
1981 pipe_bp->fs.pipe = NULL ;
1982 pipe_bp->fs.rq = NULL ;
1984 bp += sizeof( *p ) ;
1985 bp = dn_copy_set( &(p->fs), bp );
1987 for (set = all_flow_sets ; set ; set = set->next ) {
1988 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp ;
1989 bcopy(set, bp, sizeof( *set ) );
1990 /* XXX same hack as above */
1991 fs_bp->next = (struct dn_flow_set *)DN_IS_QUEUE ;
1992 fs_bp->pipe = NULL ;
1994 bp += sizeof( *set ) ;
1995 bp = dn_copy_set( set, bp );
1999 error = sooptcopyout(sopt, buf, size);
2005 * Handler for the various dummynet socket options (get, flush, config, del)
2008 ip_dn_ctl(struct sockopt *sopt)
2011 struct dn_pipe *p, tmp_pipe;
2013 /* Disallow sets in really-really secure mode. */
2014 if (sopt->sopt_dir == SOPT_SET) {
2015 #if __FreeBSD_version >= 500034
2016 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
2020 if (securelevel >= 3)
2025 switch (sopt->sopt_name) {
2027 printf("dummynet: -- unknown option %d", sopt->sopt_name);
2030 case IP_DUMMYNET_GET :
2031 error = dummynet_get(sopt);
2034 case IP_DUMMYNET_FLUSH :
2038 case IP_DUMMYNET_CONFIGURE :
2040 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2043 error = config_pipe(p);
2046 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
2048 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2052 error = delete_pipe(p);
2062 printf("DUMMYNET with IPv6 initialized (040826)\n");
2064 DUMMYNET_LOCK_INIT();
2067 all_flow_sets = NULL ;
2068 ready_heap.size = ready_heap.elements = 0 ;
2069 ready_heap.offset = 0 ;
2071 wfq_ready_heap.size = wfq_ready_heap.elements = 0 ;
2072 wfq_ready_heap.offset = 0 ;
2074 extract_heap.size = extract_heap.elements = 0 ;
2075 extract_heap.offset = 0 ;
2077 ip_dn_ctl_ptr = ip_dn_ctl;
2078 ip_dn_io_ptr = dummynet_io;
2079 ip_dn_ruledel_ptr = dn_rule_delete;
2081 callout_init(&dn_timeout, NET_CALLOUT_MPSAFE);
2082 callout_reset(&dn_timeout, 1, dummynet, NULL);
2089 ip_dn_ctl_ptr = NULL;
2090 ip_dn_io_ptr = NULL;
2091 ip_dn_ruledel_ptr = NULL;
2093 callout_stop(&dn_timeout);
2096 DUMMYNET_LOCK_DESTROY();
2098 #endif /* KLD_MODULE */
2101 dummynet_modevent(module_t mod, int type, void *data)
2105 if (DUMMYNET_LOADED) {
2106 printf("DUMMYNET already loaded\n");
2113 #if !defined(KLD_MODULE)
2114 printf("dummynet statically compiled, cannot unload\n");
2127 static moduledata_t dummynet_mod = {
2132 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY);
2133 MODULE_DEPEND(dummynet, ipfw, 2, 2, 2);
2134 MODULE_VERSION(dummynet, 1);