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 */
83 #include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */
84 #include <netinet6/ip6_var.h>
87 * We keep a private variable for the simulation time, but we could
88 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c)
90 static dn_key curr_time = 0 ; /* current simulation time */
92 static int dn_hash_size = 64 ; /* default hash size */
94 /* statistics on number of queue searches and search steps */
95 static int searches, search_steps ;
96 static int pipe_expire = 1 ; /* expire queue if empty */
97 static int dn_max_ratio = 16 ; /* max queues/buckets ratio */
99 static int red_lookup_depth = 256; /* RED - default lookup table depth */
100 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
101 static int red_max_pkt_size = 1500; /* RED - default max packet size */
104 * Three heaps contain queues and pipes that the scheduler handles:
106 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
108 * wfq_ready_heap contains the pipes associated with WF2Q flows
110 * extract_heap contains pipes associated with delay lines.
114 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
116 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
118 static int heap_init(struct dn_heap *h, int size) ;
119 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
120 static void heap_extract(struct dn_heap *h, void *obj);
122 static void transmit_event(struct dn_pipe *pipe);
123 static void ready_event(struct dn_flow_queue *q);
126 #define HASH(num) ((((num) >> 8) ^ ((num) >> 4) ^ (num)) & 0x0f)
127 static struct dn_pipe_head pipehash[HASHSIZE]; /* all pipes */
128 static struct dn_flow_set_head flowsethash[HASHSIZE]; /* all flowsets */
130 static struct callout dn_timeout;
132 extern void (*bridge_dn_p)(struct mbuf *, struct ifnet *);
135 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
136 CTLFLAG_RW, 0, "Dummynet");
137 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
138 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
139 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time,
140 CTLFLAG_RD, &curr_time, 0, "Current tick");
141 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
142 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
143 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
144 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
145 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches,
146 CTLFLAG_RD, &searches, 0, "Number of queue searches");
147 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps,
148 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
149 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
150 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
151 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
152 CTLFLAG_RW, &dn_max_ratio, 0,
153 "Max ratio between dynamic queues and buckets");
154 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
155 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
156 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
157 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
158 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
159 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
162 #ifdef DUMMYNET_DEBUG
163 int dummynet_debug = 0;
165 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, CTLFLAG_RW, &dummynet_debug,
166 0, "control debugging printfs");
168 #define DPRINTF(X) if (dummynet_debug) printf X
173 static struct mtx dummynet_mtx;
175 * NB: Recursion is needed to deal with re-entry via ICMP. That is,
176 * a packet may be dispatched via ip_input from dummynet_io and
177 * re-enter through ip_output. Yech.
179 #define DUMMYNET_LOCK_INIT() \
180 mtx_init(&dummynet_mtx, "dummynet", NULL, MTX_DEF | MTX_RECURSE)
181 #define DUMMYNET_LOCK_DESTROY() mtx_destroy(&dummynet_mtx)
182 #define DUMMYNET_LOCK() mtx_lock(&dummynet_mtx)
183 #define DUMMYNET_UNLOCK() mtx_unlock(&dummynet_mtx)
184 #define DUMMYNET_LOCK_ASSERT() do { \
185 mtx_assert(&dummynet_mtx, MA_OWNED); \
186 NET_ASSERT_GIANT(); \
189 static int config_pipe(struct dn_pipe *p);
190 static int ip_dn_ctl(struct sockopt *sopt);
192 static void dummynet(void *);
193 static void dummynet_flush(void);
194 void dummynet_drain(void);
195 static ip_dn_io_t dummynet_io;
196 static void dn_rule_delete(void *);
198 int if_tx_rdy(struct ifnet *ifp);
201 * Heap management functions.
203 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
204 * Some macros help finding parent/children so we can optimize them.
206 * heap_init() is called to expand the heap when needed.
207 * Increment size in blocks of 16 entries.
208 * XXX failure to allocate a new element is a pretty bad failure
209 * as we basically stall a whole queue forever!!
210 * Returns 1 on error, 0 on success
212 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
213 #define HEAP_LEFT(x) ( 2*(x) + 1 )
214 #define HEAP_IS_LEFT(x) ( (x) & 1 )
215 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
216 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
217 #define HEAP_INCREMENT 15
220 heap_init(struct dn_heap *h, int new_size)
222 struct dn_heap_entry *p;
224 if (h->size >= new_size ) {
225 printf("dummynet: %s, Bogus call, have %d want %d\n", __func__,
229 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
230 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_NOWAIT);
232 printf("dummynet: %s, resize %d failed\n", __func__, new_size );
233 return 1 ; /* error */
236 bcopy(h->p, p, h->size * sizeof(*p) );
237 free(h->p, M_DUMMYNET);
245 * Insert element in heap. Normally, p != NULL, we insert p in
246 * a new position and bubble up. If p == NULL, then the element is
247 * already in place, and key is the position where to start the
249 * Returns 1 on failure (cannot allocate new heap entry)
251 * If offset > 0 the position (index, int) of the element in the heap is
252 * also stored in the element itself at the given offset in bytes.
254 #define SET_OFFSET(heap, node) \
255 if (heap->offset > 0) \
256 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
258 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
260 #define RESET_OFFSET(heap, node) \
261 if (heap->offset > 0) \
262 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
264 heap_insert(struct dn_heap *h, dn_key key1, void *p)
266 int son = h->elements ;
268 if (p == NULL) /* data already there, set starting point */
270 else { /* insert new element at the end, possibly resize */
272 if (son == h->size) /* need resize... */
273 if (heap_init(h, h->elements+1) )
274 return 1 ; /* failure... */
275 h->p[son].object = p ;
276 h->p[son].key = key1 ;
279 while (son > 0) { /* bubble up */
280 int father = HEAP_FATHER(son) ;
281 struct dn_heap_entry tmp ;
283 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
284 break ; /* found right position */
285 /* son smaller than father, swap and repeat */
286 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
295 * remove top element from heap, or obj if obj != NULL
298 heap_extract(struct dn_heap *h, void *obj)
300 int child, father, max = h->elements - 1 ;
303 printf("dummynet: warning, extract from empty heap 0x%p\n", h);
306 father = 0 ; /* default: move up smallest child */
307 if (obj != NULL) { /* extract specific element, index is at offset */
309 panic("dummynet: heap_extract from middle not supported on this heap!!!\n");
310 father = *((int *)((char *)obj + h->offset)) ;
311 if (father < 0 || father >= h->elements) {
312 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
313 father, h->elements);
314 panic("dummynet: heap_extract");
317 RESET_OFFSET(h, father);
318 child = HEAP_LEFT(father) ; /* left child */
319 while (child <= max) { /* valid entry */
320 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
321 child = child+1 ; /* take right child, otherwise left */
322 h->p[father] = h->p[child] ;
323 SET_OFFSET(h, father);
325 child = HEAP_LEFT(child) ; /* left child for next loop */
330 * Fill hole with last entry and bubble up, reusing the insert code
332 h->p[father] = h->p[max] ;
333 heap_insert(h, father, NULL); /* this one cannot fail */
339 * change object position and update references
340 * XXX this one is never used!
343 heap_move(struct dn_heap *h, dn_key new_key, void *object)
347 int max = h->elements-1 ;
348 struct dn_heap_entry buf ;
351 panic("cannot move items on this heap");
353 i = *((int *)((char *)object + h->offset));
354 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
355 h->p[i].key = new_key ;
356 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
357 i = temp ) { /* bubble up */
358 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
361 } else { /* must move down */
362 h->p[i].key = new_key ;
363 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
364 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
365 temp++ ; /* select child with min key */
366 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
367 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
376 #endif /* heap_move, unused */
379 * heapify() will reorganize data inside an array to maintain the
380 * heap property. It is needed when we delete a bunch of entries.
383 heapify(struct dn_heap *h)
387 for (i = 0 ; i < h->elements ; i++ )
388 heap_insert(h, i , NULL) ;
392 * cleanup the heap and free data structure
395 heap_free(struct dn_heap *h)
398 free(h->p, M_DUMMYNET);
399 bzero(h, sizeof(*h) );
403 * --- end of heap management functions ---
407 * Return the mbuf tag holding the dummynet state. As an optimization
408 * this is assumed to be the first tag on the list. If this turns out
409 * wrong we'll need to search the list.
411 static struct dn_pkt_tag *
412 dn_tag_get(struct mbuf *m)
414 struct m_tag *mtag = m_tag_first(m);
415 KASSERT(mtag != NULL &&
416 mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
417 mtag->m_tag_id == PACKET_TAG_DUMMYNET,
418 ("packet on dummynet queue w/o dummynet tag!"));
419 return (struct dn_pkt_tag *)(mtag+1);
423 * Scheduler functions:
425 * transmit_event() is called when the delay-line needs to enter
426 * the scheduler, either because of existing pkts getting ready,
427 * or new packets entering the queue. The event handled is the delivery
428 * time of the packet.
430 * ready_event() does something similar with fixed-rate queues, and the
431 * event handled is the finish time of the head pkt.
433 * wfq_ready_event() does something similar with WF2Q queues, and the
434 * event handled is the start time of the head pkt.
436 * In all cases, we make sure that the data structures are consistent
437 * before passing pkts out, because this might trigger recursive
438 * invocations of the procedures.
441 transmit_event(struct dn_pipe *pipe)
444 struct dn_pkt_tag *pkt ;
447 DUMMYNET_LOCK_ASSERT();
449 while ( (m = pipe->head) ) {
451 if ( !DN_KEY_LEQ(pkt->output_time, curr_time) )
454 * first unlink, then call procedures, since ip_input() can invoke
455 * ip_output() and viceversa, thus causing nested calls
457 pipe->head = m->m_nextpkt ;
460 /* XXX: drop the lock for now to avoid LOR's */
462 switch (pkt->dn_dir) {
464 (void)ip_output(m, NULL, NULL, pkt->flags, NULL, NULL);
468 ip = mtod(m, struct ip *);
469 ip->ip_len = htons(ip->ip_len);
470 ip->ip_off = htons(ip->ip_off);
480 (void)ip6_output(m, NULL, NULL, pkt->flags, NULL, NULL, NULL);
485 if (bridge_dn_p != NULL)
486 ((*bridge_dn_p)(m, pkt->ifp));
488 printf("dummynet: if_bridge not loaded\n");
492 case DN_TO_ETH_DEMUX:
494 * The Ethernet code assumes the Ethernet header is
495 * contiguous in the first mbuf header. Insure this is true.
497 if (m->m_len < ETHER_HDR_LEN &&
498 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
499 printf("dummynet/ether: pullup fail, dropping pkt\n");
502 ether_demux(m->m_pkthdr.rcvif, m); /* which consumes the mbuf */
506 ether_output_frame(pkt->ifp, m);
510 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
516 /* if there are leftover packets, put into the heap for next event */
517 if ( (m = pipe->head) ) {
518 pkt = dn_tag_get(m) ;
519 /* XXX should check errors on heap_insert, by draining the
520 * whole pipe p and hoping in the future we are more successful
522 heap_insert(&extract_heap, pkt->output_time, pipe ) ;
527 * the following macro computes how many ticks we have to wait
528 * before being able to transmit a packet. The credit is taken from
529 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
531 #define SET_TICKS(_m, q, p) \
532 ((_m)->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \
536 * extract pkt from queue, compute output time (could be now)
537 * and put into delay line (p_queue)
540 move_pkt(struct mbuf *pkt, struct dn_flow_queue *q,
541 struct dn_pipe *p, int len)
543 struct dn_pkt_tag *dt = dn_tag_get(pkt);
545 q->head = pkt->m_nextpkt ;
547 q->len_bytes -= len ;
549 dt->output_time = curr_time + p->delay ;
554 p->tail->m_nextpkt = pkt;
556 p->tail->m_nextpkt = NULL;
560 * ready_event() is invoked every time the queue must enter the
561 * scheduler, either because the first packet arrives, or because
562 * a previously scheduled event fired.
563 * On invokation, drain as many pkts as possible (could be 0) and then
564 * if there are leftover packets reinsert the pkt in the scheduler.
567 ready_event(struct dn_flow_queue *q)
570 struct dn_pipe *p = q->fs->pipe ;
573 DUMMYNET_LOCK_ASSERT();
576 printf("dummynet: ready_event- pipe is gone\n");
579 p_was_empty = (p->head == NULL) ;
582 * schedule fixed-rate queues linked to this pipe:
583 * Account for the bw accumulated since last scheduling, then
584 * drain as many pkts as allowed by q->numbytes and move to
585 * the delay line (in p) computing output time.
586 * bandwidth==0 (no limit) means we can drain the whole queue,
587 * setting len_scaled = 0 does the job.
589 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth;
590 while ( (pkt = q->head) != NULL ) {
591 int len = pkt->m_pkthdr.len;
592 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
593 if (len_scaled > q->numbytes )
595 q->numbytes -= len_scaled ;
596 move_pkt(pkt, q, p, len);
599 * If we have more packets queued, schedule next ready event
600 * (can only occur when bandwidth != 0, otherwise we would have
601 * flushed the whole queue in the previous loop).
602 * To this purpose we record the current time and compute how many
603 * ticks to go for the finish time of the packet.
605 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */
606 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
607 q->sched_time = curr_time ;
608 heap_insert(&ready_heap, curr_time + t, (void *)q );
609 /* XXX should check errors on heap_insert, and drain the whole
610 * queue on error hoping next time we are luckier.
612 } else { /* RED needs to know when the queue becomes empty */
613 q->q_time = curr_time;
617 * If the delay line was empty call transmit_event(p) now.
618 * Otherwise, the scheduler will take care of it.
625 * Called when we can transmit packets on WF2Q queues. Take pkts out of
626 * the queues at their start time, and enqueue into the delay line.
627 * Packets are drained until p->numbytes < 0. As long as
628 * len_scaled >= p->numbytes, the packet goes into the delay line
629 * with a deadline p->delay. For the last packet, if p->numbytes<0,
630 * there is an additional delay.
633 ready_event_wfq(struct dn_pipe *p)
635 int p_was_empty = (p->head == NULL) ;
636 struct dn_heap *sch = &(p->scheduler_heap);
637 struct dn_heap *neh = &(p->not_eligible_heap) ;
639 DUMMYNET_LOCK_ASSERT();
641 if (p->if_name[0] == 0) /* tx clock is simulated */
642 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth;
643 else { /* tx clock is for real, the ifq must be empty or this is a NOP */
644 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
647 DPRINTF(("dummynet: pipe %d ready from %s --\n",
648 p->pipe_nr, p->if_name));
653 * While we have backlogged traffic AND credit, we need to do
654 * something on the queue.
656 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) {
657 if (sch->elements > 0) { /* have some eligible pkts to send out */
658 struct dn_flow_queue *q = sch->p[0].object ;
659 struct mbuf *pkt = q->head;
660 struct dn_flow_set *fs = q->fs;
661 u_int64_t len = pkt->m_pkthdr.len;
662 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
664 heap_extract(sch, NULL); /* remove queue from heap */
665 p->numbytes -= len_scaled ;
666 move_pkt(pkt, q, p, len);
668 p->V += (len<<MY_M) / p->sum ; /* update V */
669 q->S = q->F ; /* update start time */
670 if (q->len == 0) { /* Flow not backlogged any more */
672 heap_insert(&(p->idle_heap), q->F, q);
673 } else { /* still backlogged */
675 * update F and position in backlogged queue, then
676 * put flow in not_eligible_heap (we will fix this later).
678 len = (q->head)->m_pkthdr.len;
679 q->F += (len<<MY_M)/(u_int64_t) fs->weight ;
680 if (DN_KEY_LEQ(q->S, p->V))
681 heap_insert(neh, q->S, q);
683 heap_insert(sch, q->F, q);
687 * now compute V = max(V, min(S_i)). Remember that all elements in sch
688 * have by definition S_i <= V so if sch is not empty, V is surely
689 * the max and we must not update it. Conversely, if sch is empty
690 * we only need to look at neh.
692 if (sch->elements == 0 && neh->elements > 0)
693 p->V = MAX64 ( p->V, neh->p[0].key );
694 /* move from neh to sch any packets that have become eligible */
695 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) {
696 struct dn_flow_queue *q = neh->p[0].object ;
697 heap_extract(neh, NULL);
698 heap_insert(sch, q->F, q);
701 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */
702 p->numbytes = -1 ; /* mark not ready for I/O */
706 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0
707 && p->idle_heap.elements > 0) {
709 * no traffic and no events scheduled. We can get rid of idle-heap.
713 for (i = 0 ; i < p->idle_heap.elements ; i++) {
714 struct dn_flow_queue *q = p->idle_heap.p[i].object ;
721 p->idle_heap.elements = 0 ;
724 * If we are getting clocks from dummynet (not a real interface) and
725 * If we are under credit, schedule the next ready event.
726 * Also fix the delivery time of the last packet.
728 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */
729 dn_key t=0 ; /* number of ticks i have to wait */
731 if (p->bandwidth > 0)
732 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ;
733 dn_tag_get(p->tail)->output_time += t ;
734 p->sched_time = curr_time ;
735 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
736 /* XXX should check errors on heap_insert, and drain the whole
737 * queue on error hoping next time we are luckier.
741 * If the delay line was empty call transmit_event(p) now.
742 * Otherwise, the scheduler will take care of it.
749 * This is called once per tick, or HZ times per second. It is used to
750 * increment the current tick counter and schedule expired events.
753 dummynet(void * __unused unused)
755 struct dn_pipe *pipe;
756 struct dn_heap *heaps[3];
758 void *p; /* generic parameter to handler */
761 heaps[0] = &ready_heap ; /* fixed-rate queues */
762 heaps[1] = &wfq_ready_heap ; /* wfq queues */
763 heaps[2] = &extract_heap ; /* delay line */
767 for (i=0; i < 3 ; i++) {
769 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time) ) {
770 if (h->p[0].key > curr_time)
771 printf("dummynet: warning, heap %d is %d ticks late\n",
772 i, (int)(curr_time - h->p[0].key));
773 p = h->p[0].object ; /* store a copy before heap_extract */
774 heap_extract(h, NULL); /* need to extract before processing */
778 struct dn_pipe *pipe = p;
779 if (pipe->if_name[0] != '\0')
780 printf("dummynet: bad ready_event_wfq for pipe %s\n",
788 /* Sweep pipes trying to expire idle flow_queues. */
789 for (i = 0; i < HASHSIZE; i++)
790 SLIST_FOREACH(pipe, &pipehash[i], next)
791 if (pipe->idle_heap.elements > 0 &&
792 DN_KEY_LT(pipe->idle_heap.p[0].key, pipe->V) ) {
793 struct dn_flow_queue *q = pipe->idle_heap.p[0].object;
795 heap_extract(&(pipe->idle_heap), NULL);
796 q->S = q->F + 1; /* Mark timestamp as invalid. */
797 pipe->sum -= q->fs->weight;
802 callout_reset(&dn_timeout, 1, dummynet, NULL);
806 * called by an interface when tx_rdy occurs.
809 if_tx_rdy(struct ifnet *ifp)
811 struct dn_pipe *pipe;
815 for (i = 0; i < HASHSIZE; i++)
816 SLIST_FOREACH(pipe, &pipehash[i], next)
817 if (pipe->ifp == ifp)
820 for (i = 0; i < HASHSIZE; i++)
821 SLIST_FOREACH(pipe, &pipehash[i], next)
822 if (!strcmp(pipe->if_name, ifp->if_xname) ) {
824 DPRINTF(("dummynet: ++ tx rdy from %s (now found)\n",
831 DPRINTF(("dummynet: ++ tx rdy from %s - qlen %d\n", ifp->if_xname,
832 ifp->if_snd.ifq_len));
833 pipe->numbytes = 0; /* Mark ready for I/O. */
834 ready_event_wfq(pipe);
842 * Unconditionally expire empty queues in case of shortage.
843 * Returns the number of queues freed.
846 expire_queues(struct dn_flow_set *fs)
848 struct dn_flow_queue *q, *prev ;
849 int i, initial_elements = fs->rq_elements ;
851 if (fs->last_expired == time_uptime)
853 fs->last_expired = time_uptime ;
854 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
855 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
856 if (q->head != NULL || q->S != q->F+1) {
859 } else { /* entry is idle, expire it */
860 struct dn_flow_queue *old_q = q ;
863 prev->next = q = q->next ;
865 fs->rq[i] = q = q->next ;
867 free(old_q, M_DUMMYNET);
869 return initial_elements - fs->rq_elements ;
873 * If room, create a new queue and put at head of slot i;
874 * otherwise, create or use the default queue.
876 static struct dn_flow_queue *
877 create_queue(struct dn_flow_set *fs, int i)
879 struct dn_flow_queue *q ;
881 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
882 expire_queues(fs) == 0) {
884 * No way to get room, use or create overflow queue.
887 if ( fs->rq[i] != NULL )
890 q = malloc(sizeof(*q), M_DUMMYNET, M_NOWAIT | M_ZERO);
892 printf("dummynet: sorry, cannot allocate queue for new flow\n");
897 q->next = fs->rq[i] ;
898 q->S = q->F + 1; /* hack - mark timestamp as invalid */
905 * Given a flow_set and a pkt in last_pkt, find a matching queue
906 * after appropriate masking. The queue is moved to front
907 * so that further searches take less time.
909 static struct dn_flow_queue *
910 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
912 int i = 0 ; /* we need i and q for new allocations */
913 struct dn_flow_queue *q, *prev;
914 int is_v6 = IS_IP6_FLOW_ID(id);
916 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
919 /* first, do the masking, then hash */
920 id->dst_port &= fs->flow_mask.dst_port ;
921 id->src_port &= fs->flow_mask.src_port ;
922 id->proto &= fs->flow_mask.proto ;
923 id->flags = 0 ; /* we don't care about this one */
925 APPLY_MASK(&id->dst_ip6, &fs->flow_mask.dst_ip6);
926 APPLY_MASK(&id->src_ip6, &fs->flow_mask.src_ip6);
927 id->flow_id6 &= fs->flow_mask.flow_id6;
929 i = ((id->dst_ip6.__u6_addr.__u6_addr32[0]) & 0xffff)^
930 ((id->dst_ip6.__u6_addr.__u6_addr32[1]) & 0xffff)^
931 ((id->dst_ip6.__u6_addr.__u6_addr32[2]) & 0xffff)^
932 ((id->dst_ip6.__u6_addr.__u6_addr32[3]) & 0xffff)^
934 ((id->dst_ip6.__u6_addr.__u6_addr32[0] >> 15) & 0xffff)^
935 ((id->dst_ip6.__u6_addr.__u6_addr32[1] >> 15) & 0xffff)^
936 ((id->dst_ip6.__u6_addr.__u6_addr32[2] >> 15) & 0xffff)^
937 ((id->dst_ip6.__u6_addr.__u6_addr32[3] >> 15) & 0xffff)^
939 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 1) & 0xfffff)^
940 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 1) & 0xfffff)^
941 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 1) & 0xfffff)^
942 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 1) & 0xfffff)^
944 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 16) & 0xffff)^
945 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 16) & 0xffff)^
946 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 16) & 0xffff)^
947 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 16) & 0xffff)^
949 (id->dst_port << 1) ^ (id->src_port) ^
953 id->dst_ip &= fs->flow_mask.dst_ip ;
954 id->src_ip &= fs->flow_mask.src_ip ;
956 i = ( (id->dst_ip) & 0xffff ) ^
957 ( (id->dst_ip >> 15) & 0xffff ) ^
958 ( (id->src_ip << 1) & 0xffff ) ^
959 ( (id->src_ip >> 16 ) & 0xffff ) ^
960 (id->dst_port << 1) ^ (id->src_port) ^
963 i = i % fs->rq_size ;
964 /* finally, scan the current list for a match */
966 for (prev=NULL, q = fs->rq[i] ; q ; ) {
969 IN6_ARE_ADDR_EQUAL(&id->dst_ip6,&q->id.dst_ip6) &&
970 IN6_ARE_ADDR_EQUAL(&id->src_ip6,&q->id.src_ip6) &&
971 id->dst_port == q->id.dst_port &&
972 id->src_port == q->id.src_port &&
973 id->proto == q->id.proto &&
974 id->flags == q->id.flags &&
975 id->flow_id6 == q->id.flow_id6)
978 if (!is_v6 && id->dst_ip == q->id.dst_ip &&
979 id->src_ip == q->id.src_ip &&
980 id->dst_port == q->id.dst_port &&
981 id->src_port == q->id.src_port &&
982 id->proto == q->id.proto &&
983 id->flags == q->id.flags)
986 /* No match. Check if we can expire the entry */
987 if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
988 /* entry is idle and not in any heap, expire it */
989 struct dn_flow_queue *old_q = q ;
992 prev->next = q = q->next ;
994 fs->rq[i] = q = q->next ;
996 free(old_q, M_DUMMYNET);
1002 if (q && prev != NULL) { /* found and not in front */
1003 prev->next = q->next ;
1004 q->next = fs->rq[i] ;
1008 if (q == NULL) { /* no match, need to allocate a new entry */
1009 q = create_queue(fs, i);
1017 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
1022 * RED calculates the average queue size (avg) using a low-pass filter
1023 * with an exponential weighted (w_q) moving average:
1024 * avg <- (1-w_q) * avg + w_q * q_size
1025 * where q_size is the queue length (measured in bytes or * packets).
1027 * If q_size == 0, we compute the idle time for the link, and set
1028 * avg = (1 - w_q)^(idle/s)
1029 * where s is the time needed for transmitting a medium-sized packet.
1031 * Now, if avg < min_th the packet is enqueued.
1032 * If avg > max_th the packet is dropped. Otherwise, the packet is
1033 * dropped with probability P function of avg.
1038 /* queue in bytes or packets ? */
1039 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len;
1041 DPRINTF(("\ndummynet: %d q: %2u ", (int) curr_time, q_size));
1043 /* average queue size estimation */
1046 * queue is not empty, avg <- avg + (q_size - avg) * w_q
1048 int diff = SCALE(q_size) - q->avg;
1049 int64_t v = SCALE_MUL((int64_t) diff, (int64_t) fs->w_q);
1054 * queue is empty, find for how long the queue has been
1055 * empty and use a lookup table for computing
1056 * (1 - * w_q)^(idle_time/s) where s is the time to send a
1058 * XXX check wraps...
1061 u_int t = (curr_time - q->q_time) / fs->lookup_step;
1063 q->avg = (t < fs->lookup_depth) ?
1064 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
1067 DPRINTF(("dummynet: avg: %u ", SCALE_VAL(q->avg)));
1069 /* should i drop ? */
1071 if (q->avg < fs->min_th) {
1073 return 0; /* accept packet ; */
1075 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
1076 if (fs->flags_fs & DN_IS_GENTLE_RED) {
1078 * According to Gentle-RED, if avg is greater than max_th the
1079 * packet is dropped with a probability
1080 * p_b = c_3 * avg - c_4
1081 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
1083 p_b = SCALE_MUL((int64_t) fs->c_3, (int64_t) q->avg) - fs->c_4;
1086 DPRINTF(("dummynet: - drop"));
1089 } else if (q->avg > fs->min_th) {
1091 * we compute p_b using the linear dropping function p_b = c_1 *
1092 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
1093 * max_p * min_th / (max_th - min_th)
1095 p_b = SCALE_MUL((int64_t) fs->c_1, (int64_t) q->avg) - fs->c_2;
1097 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
1098 p_b = (p_b * len) / fs->max_pkt_size;
1099 if (++q->count == 0)
1100 q->random = random() & 0xffff;
1103 * q->count counts packets arrived since last drop, so a greater
1104 * value of q->count means a greater packet drop probability.
1106 if (SCALE_MUL(p_b, SCALE((int64_t) q->count)) > q->random) {
1108 DPRINTF(("dummynet: - red drop"));
1109 /* after a drop we calculate a new random value */
1110 q->random = random() & 0xffff;
1111 return 1; /* drop */
1114 /* end of RED algorithm */
1115 return 0 ; /* accept */
1118 static __inline struct dn_flow_set *
1119 locate_flowset(int fs_nr)
1121 struct dn_flow_set *fs;
1123 SLIST_FOREACH(fs, &flowsethash[HASH(fs_nr)], next)
1124 if (fs->fs_nr == fs_nr)
1130 static __inline struct dn_pipe *
1131 locate_pipe(int pipe_nr)
1133 struct dn_pipe *pipe;
1135 SLIST_FOREACH(pipe, &pipehash[HASH(pipe_nr)], next)
1136 if (pipe->pipe_nr == pipe_nr)
1143 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1144 * depending on whether WF2Q or fixed bw is used.
1146 * pipe_nr pipe or queue the packet is destined for.
1147 * dir where shall we send the packet after dummynet.
1148 * m the mbuf with the packet
1149 * ifp the 'ifp' parameter from the caller.
1150 * NULL in ip_input, destination interface in ip_output,
1151 * rule matching rule, in case of multiple passes
1152 * flags flags from the caller, only used in ip_output
1156 dummynet_io(struct mbuf *m, int dir, struct ip_fw_args *fwa)
1158 struct dn_pkt_tag *pkt;
1160 struct dn_flow_set *fs = NULL;
1161 struct dn_pipe *pipe ;
1162 u_int64_t len = m->m_pkthdr.len ;
1163 struct dn_flow_queue *q = NULL ;
1165 ipfw_insn *cmd = ACTION_PTR(fwa->rule);
1167 KASSERT(m->m_nextpkt == NULL,
1168 ("dummynet_io: mbuf queue passed to dummynet"));
1170 if (cmd->opcode == O_LOG)
1172 is_pipe = (cmd->opcode == O_PIPE);
1176 * This is a dummynet rule, so we expect an O_PIPE or O_QUEUE rule.
1178 * XXXGL: probably the pipe->fs and fs->pipe logic here
1179 * below can be simplified.
1182 pipe = locate_pipe(fwa->cookie);
1186 fs = locate_flowset(fwa->cookie);
1189 goto dropit; /* This queue/pipe does not exist! */
1191 if (pipe == NULL) { /* Must be a queue, try find a matching pipe. */
1192 pipe = locate_pipe(fs->parent_nr);
1196 printf("dummynet: no pipe %d for queue %d, drop pkt\n",
1197 fs->parent_nr, fs->fs_nr);
1201 q = find_queue(fs, &(fwa->f_id));
1203 goto dropit ; /* cannot allocate queue */
1205 * update statistics, then check reasons to drop pkt
1207 q->tot_bytes += len ;
1209 if ( fs->plr && random() < fs->plr )
1210 goto dropit ; /* random pkt drop */
1211 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) {
1212 if (q->len_bytes > fs->qsize)
1213 goto dropit ; /* queue size overflow */
1215 if (q->len >= fs->qsize)
1216 goto dropit ; /* queue count overflow */
1218 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) )
1221 /* XXX expensive to zero, see if we can remove it*/
1222 mtag = m_tag_get(PACKET_TAG_DUMMYNET,
1223 sizeof(struct dn_pkt_tag), M_NOWAIT|M_ZERO);
1225 goto dropit ; /* cannot allocate packet header */
1226 m_tag_prepend(m, mtag); /* attach to mbuf chain */
1228 pkt = (struct dn_pkt_tag *)(mtag+1);
1229 /* ok, i can handle the pkt now... */
1230 /* build and enqueue packet + parameters */
1231 pkt->rule = fwa->rule ;
1234 pkt->ifp = fwa->oif;
1235 if (dir == DN_TO_IP_OUT || dir == DN_TO_IP6_OUT)
1236 pkt->flags = fwa->flags;
1238 if (q->head == NULL)
1241 q->tail->m_nextpkt = m;
1244 q->len_bytes += len ;
1246 if ( q->head != m ) /* flow was not idle, we are done */
1249 * If we reach this point the flow was previously idle, so we need
1250 * to schedule it. This involves different actions for fixed-rate or
1255 * Fixed-rate queue: just insert into the ready_heap.
1258 if (pipe->bandwidth)
1259 t = SET_TICKS(m, q, pipe);
1260 q->sched_time = curr_time ;
1261 if (t == 0) /* must process it now */
1264 heap_insert(&ready_heap, curr_time + t , q );
1267 * WF2Q. First, compute start time S: if the flow was idle (S=F+1)
1268 * set S to the virtual time V for the controlling pipe, and update
1269 * the sum of weights for the pipe; otherwise, remove flow from
1270 * idle_heap and set S to max(F,V).
1271 * Second, compute finish time F = S + len/weight.
1272 * Third, if pipe was idle, update V=max(S, V).
1273 * Fourth, count one more backlogged flow.
1275 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */
1277 pipe->sum += fs->weight ; /* add weight of new queue */
1279 heap_extract(&(pipe->idle_heap), q);
1280 q->S = MAX64(q->F, pipe->V ) ;
1282 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight;
1284 if (pipe->not_eligible_heap.elements == 0 &&
1285 pipe->scheduler_heap.elements == 0)
1286 pipe->V = MAX64 ( q->S, pipe->V );
1289 * Look at eligibility. A flow is not eligibile if S>V (when
1290 * this happens, it means that there is some other flow already
1291 * scheduled for the same pipe, so the scheduler_heap cannot be
1292 * empty). If the flow is not eligible we just store it in the
1293 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1294 * and possibly invoke ready_event_wfq() right now if there is
1296 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1297 * and for all flows in not_eligible_heap (NEH), S_i > V .
1298 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH,
1299 * we only need to look into NEH.
1301 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */
1302 if (pipe->scheduler_heap.elements == 0)
1303 printf("dummynet: ++ ouch! not eligible but empty scheduler!\n");
1304 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1306 heap_insert(&(pipe->scheduler_heap), q->F, q);
1307 if (pipe->numbytes >= 0) { /* pipe is idle */
1308 if (pipe->scheduler_heap.elements != 1)
1309 printf("dummynet: OUCH! pipe should have been idle!\n");
1310 DPRINTF(("dummynet: waking up pipe %d at %d\n",
1311 pipe->pipe_nr, (int)(q->F >> MY_M)));
1312 pipe->sched_time = curr_time ;
1313 ready_event_wfq(pipe);
1326 return ( (fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1330 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1331 * Doing this would probably save us the initial bzero of dn_pkt
1333 #define DN_FREE_PKT(_m) do { \
1338 * Dispose all packets and flow_queues on a flow_set.
1339 * If all=1, also remove red lookup table and other storage,
1340 * including the descriptor itself.
1341 * For the one in dn_pipe MUST also cleanup ready_heap...
1344 purge_flow_set(struct dn_flow_set *fs, int all)
1346 struct dn_flow_queue *q, *qn ;
1349 DUMMYNET_LOCK_ASSERT();
1351 for (i = 0 ; i <= fs->rq_size ; i++ ) {
1352 for (q = fs->rq[i] ; q ; q = qn ) {
1353 struct mbuf *m, *mnext;
1356 while ((m = mnext) != NULL) {
1357 mnext = m->m_nextpkt;
1361 free(q, M_DUMMYNET);
1365 fs->rq_elements = 0 ;
1367 /* RED - free lookup table */
1369 free(fs->w_q_lookup, M_DUMMYNET);
1371 free(fs->rq, M_DUMMYNET);
1372 /* if this fs is not part of a pipe, free it */
1373 if (fs->pipe && fs != &(fs->pipe->fs) )
1374 free(fs, M_DUMMYNET);
1379 * Dispose all packets queued on a pipe (not a flow_set).
1380 * Also free all resources associated to a pipe, which is about
1384 purge_pipe(struct dn_pipe *pipe)
1386 struct mbuf *m, *mnext;
1388 purge_flow_set( &(pipe->fs), 1 );
1391 while ((m = mnext) != NULL) {
1392 mnext = m->m_nextpkt;
1396 heap_free( &(pipe->scheduler_heap) );
1397 heap_free( &(pipe->not_eligible_heap) );
1398 heap_free( &(pipe->idle_heap) );
1402 * Delete all pipes and heaps returning memory. Must also
1403 * remove references from all ipfw rules to all pipes.
1406 dummynet_flush(void)
1408 struct dn_pipe *pipe, *pipe1;
1409 struct dn_flow_set *fs, *fs1;
1413 /* Free heaps so we don't have unwanted events. */
1414 heap_free(&ready_heap);
1415 heap_free(&wfq_ready_heap);
1416 heap_free(&extract_heap);
1419 * Now purge all queued pkts and delete all pipes.
1421 * XXXGL: can we merge the for(;;) cycles into one or not?
1423 for (i = 0; i < HASHSIZE; i++)
1424 SLIST_FOREACH_SAFE(fs, &flowsethash[i], next, fs1) {
1425 SLIST_REMOVE(&flowsethash[i], fs, dn_flow_set, next);
1426 purge_flow_set(fs, 1);
1428 for (i = 0; i < HASHSIZE; i++)
1429 SLIST_FOREACH_SAFE(pipe, &pipehash[i], next, pipe1) {
1430 SLIST_REMOVE(&pipehash[i], pipe, dn_pipe, next);
1432 free(pipe, M_DUMMYNET);
1437 extern struct ip_fw *ip_fw_default_rule ;
1439 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1442 struct dn_flow_queue *q ;
1445 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1446 for (q = fs->rq[i] ; q ; q = q->next )
1447 for (m = q->head ; m ; m = m->m_nextpkt ) {
1448 struct dn_pkt_tag *pkt = dn_tag_get(m) ;
1450 pkt->rule = ip_fw_default_rule ;
1454 * when a firewall rule is deleted, scan all queues and remove the flow-id
1455 * from packets matching this rule.
1458 dn_rule_delete(void *r)
1460 struct dn_pipe *pipe;
1461 struct dn_flow_set *fs;
1462 struct dn_pkt_tag *pkt;
1468 * If the rule references a queue (dn_flow_set), then scan
1469 * the flow set, otherwise scan pipes. Should do either, but doing
1470 * both does not harm.
1472 for (i = 0; i < HASHSIZE; i++)
1473 SLIST_FOREACH(fs, &flowsethash[i], next)
1474 dn_rule_delete_fs(fs, r);
1476 for (i = 0; i < HASHSIZE; i++)
1477 SLIST_FOREACH(pipe, &pipehash[i], next) {
1479 dn_rule_delete_fs(fs, r);
1480 for (m = pipe->head ; m ; m = m->m_nextpkt ) {
1481 pkt = dn_tag_get(m);
1483 pkt->rule = ip_fw_default_rule;
1490 * setup RED parameters
1493 config_red(struct dn_flow_set *p, struct dn_flow_set * x)
1498 x->min_th = SCALE(p->min_th);
1499 x->max_th = SCALE(p->max_th);
1500 x->max_p = p->max_p;
1502 x->c_1 = p->max_p / (p->max_th - p->min_th);
1503 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1504 if (x->flags_fs & DN_IS_GENTLE_RED) {
1505 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1506 x->c_4 = (SCALE(1) - 2 * p->max_p);
1509 /* if the lookup table already exist, free and create it again */
1510 if (x->w_q_lookup) {
1511 free(x->w_q_lookup, M_DUMMYNET);
1512 x->w_q_lookup = NULL ;
1514 if (red_lookup_depth == 0) {
1515 printf("\ndummynet: net.inet.ip.dummynet.red_lookup_depth must be > 0\n");
1516 free(x, M_DUMMYNET);
1519 x->lookup_depth = red_lookup_depth;
1520 x->w_q_lookup = (u_int *) malloc(x->lookup_depth * sizeof(int),
1521 M_DUMMYNET, M_NOWAIT);
1522 if (x->w_q_lookup == NULL) {
1523 printf("dummynet: sorry, cannot allocate red lookup table\n");
1524 free(x, M_DUMMYNET);
1528 /* fill the lookup table with (1 - w_q)^x */
1529 x->lookup_step = p->lookup_step ;
1530 x->lookup_weight = p->lookup_weight ;
1531 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1532 for (i = 1; i < x->lookup_depth; i++)
1533 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1534 if (red_avg_pkt_size < 1)
1535 red_avg_pkt_size = 512 ;
1536 x->avg_pkt_size = red_avg_pkt_size ;
1537 if (red_max_pkt_size < 1)
1538 red_max_pkt_size = 1500 ;
1539 x->max_pkt_size = red_max_pkt_size ;
1544 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1546 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1547 int l = pfs->rq_size;
1553 else if (l > DN_MAX_HASH_SIZE)
1554 l = DN_MAX_HASH_SIZE;
1556 } else /* one is enough for null mask */
1558 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1559 M_DUMMYNET, M_NOWAIT | M_ZERO);
1560 if (x->rq == NULL) {
1561 printf("dummynet: sorry, cannot allocate queue\n");
1569 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1571 x->flags_fs = src->flags_fs;
1572 x->qsize = src->qsize;
1574 x->flow_mask = src->flow_mask;
1575 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1576 if (x->qsize > 1024*1024)
1577 x->qsize = 1024*1024 ;
1584 /* configuring RED */
1585 if ( x->flags_fs & DN_IS_RED )
1586 config_red(src, x) ; /* XXX should check errors */
1590 * setup pipe or queue parameters.
1594 config_pipe(struct dn_pipe *p)
1596 struct dn_flow_set *pfs = &(p->fs);
1597 struct dn_flow_queue *q;
1601 * The config program passes parameters as follows:
1602 * bw = bits/second (0 means no limits),
1603 * delay = ms, must be translated into ticks.
1604 * qsize = slots/bytes
1606 p->delay = ( p->delay * hz ) / 1000 ;
1607 /* We need either a pipe number or a flow_set number */
1608 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1610 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1612 if (p->pipe_nr != 0) { /* this is a pipe */
1613 struct dn_pipe *pipe;
1616 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1618 if (pipe == NULL) { /* new pipe */
1619 pipe = malloc(sizeof(struct dn_pipe), M_DUMMYNET,
1623 printf("dummynet: no memory for new pipe\n");
1626 pipe->pipe_nr = p->pipe_nr;
1627 pipe->fs.pipe = pipe ;
1628 /* idle_heap is the only one from which we extract from the middle.
1630 pipe->idle_heap.size = pipe->idle_heap.elements = 0 ;
1631 pipe->idle_heap.offset=OFFSET_OF(struct dn_flow_queue, heap_pos);
1633 /* Flush accumulated credit for all queues */
1634 for (i = 0; i <= pipe->fs.rq_size; i++)
1635 for (q = pipe->fs.rq[i]; q; q = q->next)
1638 pipe->bandwidth = p->bandwidth ;
1639 pipe->numbytes = 0; /* just in case... */
1640 bcopy(p->if_name, pipe->if_name, sizeof(p->if_name) );
1641 pipe->ifp = NULL ; /* reset interface ptr */
1642 pipe->delay = p->delay ;
1643 set_fs_parms(&(pipe->fs), pfs);
1646 if (pipe->fs.rq == NULL) { /* a new pipe */
1647 error = alloc_hash(&(pipe->fs), pfs);
1650 free(pipe, M_DUMMYNET);
1653 SLIST_INSERT_HEAD(&pipehash[HASH(pipe->pipe_nr)], pipe, next);
1656 } else { /* config queue */
1657 struct dn_flow_set *fs;
1660 fs = locate_flowset(pfs->fs_nr); /* locate flow_set */
1662 if (fs == NULL) { /* new */
1663 if (pfs->parent_nr == 0) { /* need link to a pipe */
1667 fs = malloc(sizeof(struct dn_flow_set), M_DUMMYNET,
1671 printf("dummynet: no memory for new flow_set\n");
1674 fs->fs_nr = pfs->fs_nr;
1675 fs->parent_nr = pfs->parent_nr;
1676 fs->weight = pfs->weight;
1677 if (fs->weight == 0)
1679 else if (fs->weight > 100)
1682 /* Change parent pipe not allowed; must delete and recreate */
1683 if (pfs->parent_nr != 0 && fs->parent_nr != pfs->parent_nr) {
1688 set_fs_parms(fs, pfs);
1690 if (fs->rq == NULL) { /* a new flow_set */
1691 error = alloc_hash(fs, pfs);
1694 free(fs, M_DUMMYNET);
1697 SLIST_INSERT_HEAD(&flowsethash[HASH(fs->fs_nr)], fs, next);
1705 * Helper function to remove from a heap queues which are linked to
1706 * a flow_set about to be deleted.
1709 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1711 int i = 0, found = 0 ;
1712 for (; i < h->elements ;)
1713 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1715 h->p[i] = h->p[h->elements] ;
1724 * helper function to remove a pipe from a heap (can be there at most once)
1727 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1729 if (h->elements > 0) {
1731 for (i=0; i < h->elements ; i++ ) {
1732 if (h->p[i].object == p) { /* found it */
1734 h->p[i] = h->p[h->elements] ;
1743 * drain all queues. Called in case of severe mbuf shortage.
1748 struct dn_flow_set *fs;
1749 struct dn_pipe *pipe;
1750 struct mbuf *m, *mnext;
1753 DUMMYNET_LOCK_ASSERT();
1755 heap_free(&ready_heap);
1756 heap_free(&wfq_ready_heap);
1757 heap_free(&extract_heap);
1758 /* remove all references to this pipe from flow_sets */
1759 for (i = 0; i < HASHSIZE; i++)
1760 SLIST_FOREACH(fs, &flowsethash[i], next)
1761 purge_flow_set(fs, 0);
1763 for (i = 0; i < HASHSIZE; i++) {
1764 SLIST_FOREACH(pipe, &pipehash[i], next) {
1765 purge_flow_set(&(pipe->fs), 0);
1768 while ((m = mnext) != NULL) {
1769 mnext = m->m_nextpkt;
1772 pipe->head = pipe->tail = NULL;
1778 * Fully delete a pipe or a queue, cleaning up associated info.
1781 delete_pipe(struct dn_pipe *p)
1783 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1785 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1787 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1788 struct dn_pipe *pipe;
1789 struct dn_flow_set *fs;
1793 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1797 return (ENOENT); /* not found */
1800 /* Unlink from list of pipes. */
1801 SLIST_REMOVE(&pipehash[HASH(pipe->pipe_nr)], pipe, dn_pipe, next);
1803 /* Remove all references to this pipe from flow_sets. */
1804 for (i = 0; i < HASHSIZE; i++)
1805 SLIST_FOREACH(fs, &flowsethash[i], next)
1806 if (fs->pipe == pipe) {
1807 printf("dummynet: ++ ref to pipe %d from fs %d\n",
1808 p->pipe_nr, fs->fs_nr);
1810 purge_flow_set(fs, 0);
1812 fs_remove_from_heap(&ready_heap, &(pipe->fs));
1813 purge_pipe(pipe); /* remove all data associated to this pipe */
1814 /* remove reference to here from extract_heap and wfq_ready_heap */
1815 pipe_remove_from_heap(&extract_heap, pipe);
1816 pipe_remove_from_heap(&wfq_ready_heap, pipe);
1819 free(pipe, M_DUMMYNET);
1820 } else { /* this is a WF2Q queue (dn_flow_set) */
1821 struct dn_flow_set *fs;
1824 fs = locate_flowset(p->fs.fs_nr); /* locate set */
1828 return (ENOENT); /* not found */
1831 /* Unlink from list of flowsets. */
1832 SLIST_REMOVE( &flowsethash[HASH(fs->fs_nr)], fs, dn_flow_set, next);
1834 if (fs->pipe != NULL) {
1835 /* Update total weight on parent pipe and cleanup parent heaps. */
1836 fs->pipe->sum -= fs->weight * fs->backlogged ;
1837 fs_remove_from_heap(&(fs->pipe->not_eligible_heap), fs);
1838 fs_remove_from_heap(&(fs->pipe->scheduler_heap), fs);
1839 #if 1 /* XXX should i remove from idle_heap as well ? */
1840 fs_remove_from_heap(&(fs->pipe->idle_heap), fs);
1843 purge_flow_set(fs, 1);
1850 * helper function used to copy data from kernel in DUMMYNET_GET
1853 dn_copy_set(struct dn_flow_set *set, char *bp)
1856 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1858 DUMMYNET_LOCK_ASSERT();
1860 for (i = 0 ; i <= set->rq_size ; i++)
1861 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
1862 if (q->hash_slot != i)
1863 printf("dummynet: ++ at %d: wrong slot (have %d, "
1864 "should be %d)\n", copied, q->hash_slot, i);
1866 printf("dummynet: ++ at %d: wrong fs ptr (have %p, should be %p)\n",
1869 bcopy(q, qp, sizeof( *q ) );
1870 /* cleanup pointers */
1872 qp->head = qp->tail = NULL ;
1875 if (copied != set->rq_elements)
1876 printf("dummynet: ++ wrong count, have %d should be %d\n",
1877 copied, set->rq_elements);
1884 struct dn_flow_set *fs;
1885 struct dn_pipe *pipe;
1889 DUMMYNET_LOCK_ASSERT();
1891 * Compute size of data structures: list of pipes and flow_sets.
1893 for (i = 0; i < HASHSIZE; i++) {
1894 SLIST_FOREACH(pipe, &pipehash[i], next)
1895 size += sizeof(*pipe) +
1896 pipe->fs.rq_elements * sizeof(struct dn_flow_queue);
1897 SLIST_FOREACH(fs, &flowsethash[i], next)
1898 size += sizeof (*fs) +
1899 fs->rq_elements * sizeof(struct dn_flow_queue);
1905 dummynet_get(struct sockopt *sopt)
1907 char *buf, *bp ; /* bp is the "copy-pointer" */
1909 struct dn_flow_set *fs;
1910 struct dn_pipe *pipe;
1913 /* XXX lock held too long */
1916 * XXX: Ugly, but we need to allocate memory with M_WAITOK flag and we
1917 * cannot use this flag while holding a mutex.
1919 for (i = 0; i < 10; i++) {
1920 size = dn_calc_size();
1922 buf = malloc(size, M_TEMP, M_WAITOK);
1924 if (size == dn_calc_size())
1934 for (i = 0; i < HASHSIZE; i++)
1935 SLIST_FOREACH(pipe, &pipehash[i], next) {
1936 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp;
1939 * Copy pipe descriptor into *bp, convert delay back to ms,
1940 * then copy the flow_set descriptor(s) one at a time.
1941 * After each flow_set, copy the queue descriptor it owns.
1943 bcopy(pipe, bp, sizeof(*pipe));
1944 pipe_bp->delay = (pipe_bp->delay * 1000) / hz;
1946 * XXX the following is a hack based on ->next being the
1947 * first field in dn_pipe and dn_flow_set. The correct
1948 * solution would be to move the dn_flow_set to the beginning
1949 * of struct dn_pipe.
1951 pipe_bp->next.sle_next = (struct dn_pipe *)DN_IS_PIPE;
1952 /* Clean pointers. */
1953 pipe_bp->head = pipe_bp->tail = NULL;
1954 pipe_bp->fs.next.sle_next = NULL;
1955 pipe_bp->fs.pipe = NULL;
1956 pipe_bp->fs.rq = NULL;
1958 bp += sizeof(*pipe) ;
1959 bp = dn_copy_set(&(pipe->fs), bp);
1962 for (i = 0; i < HASHSIZE; i++)
1963 SLIST_FOREACH(fs, &flowsethash[i], next) {
1964 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp;
1966 bcopy(fs, bp, sizeof(*fs));
1967 /* XXX same hack as above */
1968 fs_bp->next.sle_next = (struct dn_flow_set *)DN_IS_QUEUE;
1972 bp = dn_copy_set(fs, bp);
1977 error = sooptcopyout(sopt, buf, size);
1983 * Handler for the various dummynet socket options (get, flush, config, del)
1986 ip_dn_ctl(struct sockopt *sopt)
1989 struct dn_pipe *p, tmp_pipe;
1991 /* Disallow sets in really-really secure mode. */
1992 if (sopt->sopt_dir == SOPT_SET) {
1993 #if __FreeBSD_version >= 500034
1994 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
1998 if (securelevel >= 3)
2003 switch (sopt->sopt_name) {
2005 printf("dummynet: -- unknown option %d", sopt->sopt_name);
2008 case IP_DUMMYNET_GET :
2009 error = dummynet_get(sopt);
2012 case IP_DUMMYNET_FLUSH :
2016 case IP_DUMMYNET_CONFIGURE :
2018 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2021 error = config_pipe(p);
2024 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
2026 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2030 error = delete_pipe(p);
2042 printf("DUMMYNET with IPv6 initialized (040826)\n");
2044 DUMMYNET_LOCK_INIT();
2046 for (i = 0; i < HASHSIZE; i++) {
2047 SLIST_INIT(&pipehash[i]);
2048 SLIST_INIT(&flowsethash[i]);
2050 ready_heap.size = ready_heap.elements = 0 ;
2051 ready_heap.offset = 0 ;
2053 wfq_ready_heap.size = wfq_ready_heap.elements = 0 ;
2054 wfq_ready_heap.offset = 0 ;
2056 extract_heap.size = extract_heap.elements = 0 ;
2057 extract_heap.offset = 0 ;
2059 ip_dn_ctl_ptr = ip_dn_ctl;
2060 ip_dn_io_ptr = dummynet_io;
2061 ip_dn_ruledel_ptr = dn_rule_delete;
2063 callout_init(&dn_timeout, NET_CALLOUT_MPSAFE);
2064 callout_reset(&dn_timeout, 1, dummynet, NULL);
2071 ip_dn_ctl_ptr = NULL;
2072 ip_dn_io_ptr = NULL;
2073 ip_dn_ruledel_ptr = NULL;
2075 callout_stop(&dn_timeout);
2078 DUMMYNET_LOCK_DESTROY();
2080 #endif /* KLD_MODULE */
2083 dummynet_modevent(module_t mod, int type, void *data)
2087 if (DUMMYNET_LOADED) {
2088 printf("DUMMYNET already loaded\n");
2095 #if !defined(KLD_MODULE)
2096 printf("dummynet statically compiled, cannot unload\n");
2109 static moduledata_t dummynet_mod = {
2114 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY);
2115 MODULE_DEPEND(dummynet, ipfw, 2, 2, 2);
2116 MODULE_VERSION(dummynet, 1);