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 #include "opt_inet6.h"
35 * This module implements IP dummynet, a bandwidth limiter/delay emulator
36 * used in conjunction with the ipfw package.
37 * Description of the data structures used is in ip_dummynet.h
38 * Here you mainly find the following blocks of code:
39 * + variable declarations;
40 * + heap management functions;
41 * + scheduler and dummynet functions;
42 * + configuration and initialization.
44 * NOTA BENE: critical sections are protected by the "dummynet lock".
46 * Most important Changes:
49 * 010124: Fixed WF2Q behaviour
50 * 010122: Fixed spl protection.
51 * 000601: WF2Q support
52 * 000106: large rewrite, use heaps to handle very many pipes.
53 * 980513: initial release
55 * include files marked with XXX are probably not needed
58 #include <sys/param.h>
59 #include <sys/systm.h>
60 #include <sys/malloc.h>
62 #include <sys/kernel.h>
63 #include <sys/module.h>
65 #include <sys/socket.h>
66 #include <sys/socketvar.h>
68 #include <sys/sysctl.h>
69 #include <sys/taskqueue.h>
71 #include <net/netisr.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 long 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 */
104 static struct timeval prev_t, t;
105 static long tick_last; /* Last tick duration (usec). */
106 static long tick_delta; /* Last vs standard tick diff (usec). */
107 static long tick_delta_sum; /* Accumulated tick difference (usec).*/
108 static long tick_adjustment; /* Tick adjustments done. */
109 static long tick_lost; /* Lost(coalesced) ticks number. */
110 /* Adjusted vs non-adjusted curr_time difference (ticks). */
111 static long tick_diff;
114 * Three heaps contain queues and pipes that the scheduler handles:
116 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
118 * wfq_ready_heap contains the pipes associated with WF2Q flows
120 * extract_heap contains pipes associated with delay lines.
124 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
126 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
128 static int heap_init(struct dn_heap *h, int size);
129 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
130 static void heap_extract(struct dn_heap *h, void *obj);
131 static void transmit_event(struct dn_pipe *pipe, struct mbuf **head,
133 static void ready_event(struct dn_flow_queue *q, struct mbuf **head,
135 static void ready_event_wfq(struct dn_pipe *p, struct mbuf **head,
139 #define HASH(num) ((((num) >> 8) ^ ((num) >> 4) ^ (num)) & 0x0f)
140 static struct dn_pipe_head pipehash[HASHSIZE]; /* all pipes */
141 static struct dn_flow_set_head flowsethash[HASHSIZE]; /* all flowsets */
143 static struct callout dn_timeout;
145 extern void (*bridge_dn_p)(struct mbuf *, struct ifnet *);
148 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW, 0, "Dummynet");
149 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
150 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
151 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, curr_time,
152 CTLFLAG_RD, &curr_time, 0, "Current tick");
153 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
154 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
155 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
156 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
157 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, searches,
158 CTLFLAG_RD, &searches, 0, "Number of queue searches");
159 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, search_steps,
160 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
161 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
162 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
163 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
164 CTLFLAG_RW, &dn_max_ratio, 0,
165 "Max ratio between dynamic queues and buckets");
166 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
167 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
168 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
169 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
170 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
171 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
172 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta,
173 CTLFLAG_RD, &tick_delta, 0, "Last vs standard tick difference (usec).");
174 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum,
175 CTLFLAG_RD, &tick_delta_sum, 0, "Accumulated tick difference (usec).");
176 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment,
177 CTLFLAG_RD, &tick_adjustment, 0, "Tick adjustments done.");
178 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff,
179 CTLFLAG_RD, &tick_diff, 0,
180 "Adjusted vs non-adjusted curr_time difference (ticks).");
181 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost,
182 CTLFLAG_RD, &tick_lost, 0,
183 "Number of ticks coalesced by dummynet taskqueue.");
186 #ifdef DUMMYNET_DEBUG
187 int dummynet_debug = 0;
189 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, CTLFLAG_RW, &dummynet_debug,
190 0, "control debugging printfs");
192 #define DPRINTF(X) if (dummynet_debug) printf X
197 static struct task dn_task;
198 static struct taskqueue *dn_tq = NULL;
199 static void dummynet_task(void *, int);
201 static struct mtx dummynet_mtx;
202 #define DUMMYNET_LOCK_INIT() \
203 mtx_init(&dummynet_mtx, "dummynet", NULL, MTX_DEF)
204 #define DUMMYNET_LOCK_DESTROY() mtx_destroy(&dummynet_mtx)
205 #define DUMMYNET_LOCK() mtx_lock(&dummynet_mtx)
206 #define DUMMYNET_UNLOCK() mtx_unlock(&dummynet_mtx)
207 #define DUMMYNET_LOCK_ASSERT() do { \
208 mtx_assert(&dummynet_mtx, MA_OWNED); \
209 NET_ASSERT_GIANT(); \
212 static int config_pipe(struct dn_pipe *p);
213 static int ip_dn_ctl(struct sockopt *sopt);
215 static void dummynet(void *);
216 static void dummynet_flush(void);
217 static void dummynet_send(struct mbuf *);
218 void dummynet_drain(void);
219 static ip_dn_io_t dummynet_io;
220 static void dn_rule_delete(void *);
223 * Heap management functions.
225 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
226 * Some macros help finding parent/children so we can optimize them.
228 * heap_init() is called to expand the heap when needed.
229 * Increment size in blocks of 16 entries.
230 * XXX failure to allocate a new element is a pretty bad failure
231 * as we basically stall a whole queue forever!!
232 * Returns 1 on error, 0 on success
234 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
235 #define HEAP_LEFT(x) ( 2*(x) + 1 )
236 #define HEAP_IS_LEFT(x) ( (x) & 1 )
237 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
238 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
239 #define HEAP_INCREMENT 15
242 heap_init(struct dn_heap *h, int new_size)
244 struct dn_heap_entry *p;
246 if (h->size >= new_size ) {
247 printf("dummynet: %s, Bogus call, have %d want %d\n", __func__,
251 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
252 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_NOWAIT);
254 printf("dummynet: %s, resize %d failed\n", __func__, new_size );
255 return 1 ; /* error */
258 bcopy(h->p, p, h->size * sizeof(*p) );
259 free(h->p, M_DUMMYNET);
267 * Insert element in heap. Normally, p != NULL, we insert p in
268 * a new position and bubble up. If p == NULL, then the element is
269 * already in place, and key is the position where to start the
271 * Returns 1 on failure (cannot allocate new heap entry)
273 * If offset > 0 the position (index, int) of the element in the heap is
274 * also stored in the element itself at the given offset in bytes.
276 #define SET_OFFSET(heap, node) \
277 if (heap->offset > 0) \
278 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
280 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
282 #define RESET_OFFSET(heap, node) \
283 if (heap->offset > 0) \
284 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
286 heap_insert(struct dn_heap *h, dn_key key1, void *p)
288 int son = h->elements ;
290 if (p == NULL) /* data already there, set starting point */
292 else { /* insert new element at the end, possibly resize */
294 if (son == h->size) /* need resize... */
295 if (heap_init(h, h->elements+1) )
296 return 1 ; /* failure... */
297 h->p[son].object = p ;
298 h->p[son].key = key1 ;
301 while (son > 0) { /* bubble up */
302 int father = HEAP_FATHER(son) ;
303 struct dn_heap_entry tmp ;
305 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
306 break ; /* found right position */
307 /* son smaller than father, swap and repeat */
308 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
317 * remove top element from heap, or obj if obj != NULL
320 heap_extract(struct dn_heap *h, void *obj)
322 int child, father, max = h->elements - 1 ;
325 printf("dummynet: warning, extract from empty heap 0x%p\n", h);
328 father = 0 ; /* default: move up smallest child */
329 if (obj != NULL) { /* extract specific element, index is at offset */
331 panic("dummynet: heap_extract from middle not supported on this heap!!!\n");
332 father = *((int *)((char *)obj + h->offset)) ;
333 if (father < 0 || father >= h->elements) {
334 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
335 father, h->elements);
336 panic("dummynet: heap_extract");
339 RESET_OFFSET(h, father);
340 child = HEAP_LEFT(father) ; /* left child */
341 while (child <= max) { /* valid entry */
342 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
343 child = child+1 ; /* take right child, otherwise left */
344 h->p[father] = h->p[child] ;
345 SET_OFFSET(h, father);
347 child = HEAP_LEFT(child) ; /* left child for next loop */
352 * Fill hole with last entry and bubble up, reusing the insert code
354 h->p[father] = h->p[max] ;
355 heap_insert(h, father, NULL); /* this one cannot fail */
361 * change object position and update references
362 * XXX this one is never used!
365 heap_move(struct dn_heap *h, dn_key new_key, void *object)
369 int max = h->elements-1 ;
370 struct dn_heap_entry buf ;
373 panic("cannot move items on this heap");
375 i = *((int *)((char *)object + h->offset));
376 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
377 h->p[i].key = new_key ;
378 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
379 i = temp ) { /* bubble up */
380 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
383 } else { /* must move down */
384 h->p[i].key = new_key ;
385 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
386 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
387 temp++ ; /* select child with min key */
388 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
389 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
398 #endif /* heap_move, unused */
401 * heapify() will reorganize data inside an array to maintain the
402 * heap property. It is needed when we delete a bunch of entries.
405 heapify(struct dn_heap *h)
409 for (i = 0 ; i < h->elements ; i++ )
410 heap_insert(h, i , NULL) ;
414 * cleanup the heap and free data structure
417 heap_free(struct dn_heap *h)
420 free(h->p, M_DUMMYNET);
421 bzero(h, sizeof(*h) );
425 * --- end of heap management functions ---
429 * Return the mbuf tag holding the dummynet state. As an optimization
430 * this is assumed to be the first tag on the list. If this turns out
431 * wrong we'll need to search the list.
433 static struct dn_pkt_tag *
434 dn_tag_get(struct mbuf *m)
436 struct m_tag *mtag = m_tag_first(m);
437 KASSERT(mtag != NULL &&
438 mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
439 mtag->m_tag_id == PACKET_TAG_DUMMYNET,
440 ("packet on dummynet queue w/o dummynet tag!"));
441 return (struct dn_pkt_tag *)(mtag+1);
445 * Scheduler functions:
447 * transmit_event() is called when the delay-line needs to enter
448 * the scheduler, either because of existing pkts getting ready,
449 * or new packets entering the queue. The event handled is the delivery
450 * time of the packet.
452 * ready_event() does something similar with fixed-rate queues, and the
453 * event handled is the finish time of the head pkt.
455 * wfq_ready_event() does something similar with WF2Q queues, and the
456 * event handled is the start time of the head pkt.
458 * In all cases, we make sure that the data structures are consistent
459 * before passing pkts out, because this might trigger recursive
460 * invocations of the procedures.
463 transmit_event(struct dn_pipe *pipe, struct mbuf **head, struct mbuf **tail)
466 struct dn_pkt_tag *pkt;
468 DUMMYNET_LOCK_ASSERT();
470 while ((m = pipe->head) != NULL) {
472 if (!DN_KEY_LEQ(pkt->output_time, curr_time))
475 pipe->head = m->m_nextpkt;
477 (*tail)->m_nextpkt = m;
483 (*tail)->m_nextpkt = NULL;
485 /* If there are leftover packets, put into the heap for next event. */
486 if ((m = pipe->head) != NULL) {
489 * XXX: Should check errors on heap_insert, by draining the
490 * whole pipe p and hoping in the future we are more successful.
492 heap_insert(&extract_heap, pkt->output_time, pipe);
497 * the following macro computes how many ticks we have to wait
498 * before being able to transmit a packet. The credit is taken from
499 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
501 #define SET_TICKS(_m, q, p) \
502 ((_m)->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \
506 * extract pkt from queue, compute output time (could be now)
507 * and put into delay line (p_queue)
510 move_pkt(struct mbuf *pkt, struct dn_flow_queue *q,
511 struct dn_pipe *p, int len)
513 struct dn_pkt_tag *dt = dn_tag_get(pkt);
515 q->head = pkt->m_nextpkt ;
517 q->len_bytes -= len ;
519 dt->output_time = curr_time + p->delay ;
524 p->tail->m_nextpkt = pkt;
526 p->tail->m_nextpkt = NULL;
530 * ready_event() is invoked every time the queue must enter the
531 * scheduler, either because the first packet arrives, or because
532 * a previously scheduled event fired.
533 * On invokation, drain as many pkts as possible (could be 0) and then
534 * if there are leftover packets reinsert the pkt in the scheduler.
537 ready_event(struct dn_flow_queue *q, struct mbuf **head, struct mbuf **tail)
540 struct dn_pipe *p = q->fs->pipe ;
543 DUMMYNET_LOCK_ASSERT();
546 printf("dummynet: ready_event- pipe is gone\n");
549 p_was_empty = (p->head == NULL) ;
552 * schedule fixed-rate queues linked to this pipe:
553 * Account for the bw accumulated since last scheduling, then
554 * drain as many pkts as allowed by q->numbytes and move to
555 * the delay line (in p) computing output time.
556 * bandwidth==0 (no limit) means we can drain the whole queue,
557 * setting len_scaled = 0 does the job.
559 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth;
560 while ( (pkt = q->head) != NULL ) {
561 int len = pkt->m_pkthdr.len;
562 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
563 if (len_scaled > q->numbytes )
565 q->numbytes -= len_scaled ;
566 move_pkt(pkt, q, p, len);
569 * If we have more packets queued, schedule next ready event
570 * (can only occur when bandwidth != 0, otherwise we would have
571 * flushed the whole queue in the previous loop).
572 * To this purpose we record the current time and compute how many
573 * ticks to go for the finish time of the packet.
575 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */
576 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
577 q->sched_time = curr_time ;
578 heap_insert(&ready_heap, curr_time + t, (void *)q );
579 /* XXX should check errors on heap_insert, and drain the whole
580 * queue on error hoping next time we are luckier.
582 } else { /* RED needs to know when the queue becomes empty */
583 q->q_time = curr_time;
587 * If the delay line was empty call transmit_event() now.
588 * Otherwise, the scheduler will take care of it.
591 transmit_event(p, head, tail);
595 * Called when we can transmit packets on WF2Q queues. Take pkts out of
596 * the queues at their start time, and enqueue into the delay line.
597 * Packets are drained until p->numbytes < 0. As long as
598 * len_scaled >= p->numbytes, the packet goes into the delay line
599 * with a deadline p->delay. For the last packet, if p->numbytes<0,
600 * there is an additional delay.
603 ready_event_wfq(struct dn_pipe *p, struct mbuf **head, struct mbuf **tail)
605 int p_was_empty = (p->head == NULL) ;
606 struct dn_heap *sch = &(p->scheduler_heap);
607 struct dn_heap *neh = &(p->not_eligible_heap) ;
609 DUMMYNET_LOCK_ASSERT();
611 if (p->if_name[0] == 0) /* tx clock is simulated */
612 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth;
613 else { /* tx clock is for real, the ifq must be empty or this is a NOP */
614 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
617 DPRINTF(("dummynet: pipe %d ready from %s --\n",
618 p->pipe_nr, p->if_name));
623 * While we have backlogged traffic AND credit, we need to do
624 * something on the queue.
626 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) {
627 if (sch->elements > 0) { /* have some eligible pkts to send out */
628 struct dn_flow_queue *q = sch->p[0].object ;
629 struct mbuf *pkt = q->head;
630 struct dn_flow_set *fs = q->fs;
631 u_int64_t len = pkt->m_pkthdr.len;
632 int len_scaled = p->bandwidth ? len*8*hz : 0 ;
634 heap_extract(sch, NULL); /* remove queue from heap */
635 p->numbytes -= len_scaled ;
636 move_pkt(pkt, q, p, len);
638 p->V += (len<<MY_M) / p->sum ; /* update V */
639 q->S = q->F ; /* update start time */
640 if (q->len == 0) { /* Flow not backlogged any more */
642 heap_insert(&(p->idle_heap), q->F, q);
643 } else { /* still backlogged */
645 * update F and position in backlogged queue, then
646 * put flow in not_eligible_heap (we will fix this later).
648 len = (q->head)->m_pkthdr.len;
649 q->F += (len<<MY_M)/(u_int64_t) fs->weight ;
650 if (DN_KEY_LEQ(q->S, p->V))
651 heap_insert(neh, q->S, q);
653 heap_insert(sch, q->F, q);
657 * now compute V = max(V, min(S_i)). Remember that all elements in sch
658 * have by definition S_i <= V so if sch is not empty, V is surely
659 * the max and we must not update it. Conversely, if sch is empty
660 * we only need to look at neh.
662 if (sch->elements == 0 && neh->elements > 0)
663 p->V = MAX64 ( p->V, neh->p[0].key );
664 /* move from neh to sch any packets that have become eligible */
665 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) {
666 struct dn_flow_queue *q = neh->p[0].object ;
667 heap_extract(neh, NULL);
668 heap_insert(sch, q->F, q);
671 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */
672 p->numbytes = -1 ; /* mark not ready for I/O */
676 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0
677 && p->idle_heap.elements > 0) {
679 * no traffic and no events scheduled. We can get rid of idle-heap.
683 for (i = 0 ; i < p->idle_heap.elements ; i++) {
684 struct dn_flow_queue *q = p->idle_heap.p[i].object ;
691 p->idle_heap.elements = 0 ;
694 * If we are getting clocks from dummynet (not a real interface) and
695 * If we are under credit, schedule the next ready event.
696 * Also fix the delivery time of the last packet.
698 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */
699 dn_key t=0 ; /* number of ticks i have to wait */
701 if (p->bandwidth > 0)
702 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ;
703 dn_tag_get(p->tail)->output_time += t ;
704 p->sched_time = curr_time ;
705 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
706 /* XXX should check errors on heap_insert, and drain the whole
707 * queue on error hoping next time we are luckier.
711 * If the delay line was empty call transmit_event() now.
712 * Otherwise, the scheduler will take care of it.
715 transmit_event(p, head, tail);
719 * This is called one tick, after previous run. It is used to
723 dummynet(void * __unused unused)
725 taskqueue_enqueue(dn_tq, &dn_task);
729 * The main dummynet processing function.
732 dummynet_task(void *context, int pending)
735 struct mbuf *head = NULL, *tail = NULL;
736 struct dn_pipe *pipe;
737 struct dn_heap *heaps[3];
739 void *p; /* generic parameter to handler */
745 heaps[0] = &ready_heap; /* fixed-rate queues */
746 heaps[1] = &wfq_ready_heap; /* wfq queues */
747 heaps[2] = &extract_heap; /* delay line */
749 /* Update number of lost(coalesced) ticks. */
750 tick_lost += pending - 1;
753 /* Last tick duration (usec). */
754 tick_last = (t.tv_sec - prev_t.tv_sec) * 1000000 +
755 (t.tv_usec - prev_t.tv_usec);
756 /* Last tick vs standard tick difference (usec). */
757 tick_delta = (tick_last * hz - 1000000) / hz;
758 /* Accumulated tick difference (usec). */
759 tick_delta_sum += tick_delta;
764 * Adjust curr_time if accumulated tick difference greater than
765 * 'standard' tick. Since curr_time should be monotonically increasing,
766 * we do positive adjustment as required and throttle curr_time in
767 * case of negative adjustment.
770 if (tick_delta_sum - tick >= 0) {
771 int diff = tick_delta_sum / tick;
775 tick_delta_sum %= tick;
777 } else if (tick_delta_sum + tick <= 0) {
780 tick_delta_sum += tick;
784 for (i = 0; i < 3; i++) {
786 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time)) {
787 if (h->p[0].key > curr_time)
788 printf("dummynet: warning, "
789 "heap %d is %d ticks late\n",
790 i, (int)(curr_time - h->p[0].key));
791 /* store a copy before heap_extract */
793 /* need to extract before processing */
794 heap_extract(h, NULL);
796 ready_event(p, &head, &tail);
798 struct dn_pipe *pipe = p;
799 if (pipe->if_name[0] != '\0')
800 printf("dummynet: bad ready_event_wfq "
801 "for pipe %s\n", pipe->if_name);
803 ready_event_wfq(p, &head, &tail);
805 transmit_event(p, &head, &tail);
809 /* Sweep pipes trying to expire idle flow_queues. */
810 for (i = 0; i < HASHSIZE; i++)
811 SLIST_FOREACH(pipe, &pipehash[i], next)
812 if (pipe->idle_heap.elements > 0 &&
813 DN_KEY_LT(pipe->idle_heap.p[0].key, pipe->V)) {
814 struct dn_flow_queue *q =
815 pipe->idle_heap.p[0].object;
817 heap_extract(&(pipe->idle_heap), NULL);
818 /* Mark timestamp as invalid. */
820 pipe->sum -= q->fs->weight;
828 callout_reset(&dn_timeout, 1, dummynet, NULL);
834 dummynet_send(struct mbuf *m)
836 struct dn_pkt_tag *pkt;
840 for (; m != NULL; m = n) {
844 switch (pkt->dn_dir) {
846 ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL);
849 ip = mtod(m, struct ip *);
850 ip->ip_len = htons(ip->ip_len);
851 ip->ip_off = htons(ip->ip_off);
852 netisr_dispatch(NETISR_IP, m);
856 netisr_dispatch(NETISR_IPV6, m);
860 ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL);
864 if (bridge_dn_p != NULL)
865 ((*bridge_dn_p)(m, pkt->ifp));
867 printf("dummynet: if_bridge not loaded\n");
872 * The bridge requires/assumes the Ethernet header is
873 * contiguous in the first mbuf header. Ensure this
877 if (m->m_len < ETHER_HDR_LEN &&
878 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
879 printf("dummynet/bridge: pullup fail, "
883 m = bdg_forward_ptr(m, pkt->ifp);
886 * somebody unloaded the bridge module.
890 printf("dummynet: dropping bridged packet "
891 "trapped in pipe\n");
896 case DN_TO_ETH_DEMUX:
898 * The Ethernet code assumes the Ethernet header is
899 * contiguous in the first mbuf header.
900 * Insure this is true.
902 if (m->m_len < ETHER_HDR_LEN &&
903 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
904 printf("dummynet/ether: pullup failed, "
905 "dropping packet\n");
908 ether_demux(m->m_pkthdr.rcvif, m);
911 ether_output_frame(pkt->ifp, m);
914 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
922 * Unconditionally expire empty queues in case of shortage.
923 * Returns the number of queues freed.
926 expire_queues(struct dn_flow_set *fs)
928 struct dn_flow_queue *q, *prev ;
929 int i, initial_elements = fs->rq_elements ;
931 if (fs->last_expired == time_second)
933 fs->last_expired = time_second ;
934 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
935 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
936 if (q->head != NULL || q->S != q->F+1) {
939 } else { /* entry is idle, expire it */
940 struct dn_flow_queue *old_q = q ;
943 prev->next = q = q->next ;
945 fs->rq[i] = q = q->next ;
947 free(old_q, M_DUMMYNET);
949 return initial_elements - fs->rq_elements ;
953 * If room, create a new queue and put at head of slot i;
954 * otherwise, create or use the default queue.
956 static struct dn_flow_queue *
957 create_queue(struct dn_flow_set *fs, int i)
959 struct dn_flow_queue *q ;
961 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
962 expire_queues(fs) == 0) {
964 * No way to get room, use or create overflow queue.
967 if ( fs->rq[i] != NULL )
970 q = malloc(sizeof(*q), M_DUMMYNET, M_NOWAIT | M_ZERO);
972 printf("dummynet: sorry, cannot allocate queue for new flow\n");
977 q->next = fs->rq[i] ;
978 q->S = q->F + 1; /* hack - mark timestamp as invalid */
985 * Given a flow_set and a pkt in last_pkt, find a matching queue
986 * after appropriate masking. The queue is moved to front
987 * so that further searches take less time.
989 static struct dn_flow_queue *
990 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
992 int i = 0 ; /* we need i and q for new allocations */
993 struct dn_flow_queue *q, *prev;
994 int is_v6 = IS_IP6_FLOW_ID(id);
996 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
999 /* first, do the masking, then hash */
1000 id->dst_port &= fs->flow_mask.dst_port ;
1001 id->src_port &= fs->flow_mask.src_port ;
1002 id->proto &= fs->flow_mask.proto ;
1003 id->flags = 0 ; /* we don't care about this one */
1005 APPLY_MASK(&id->dst_ip6, &fs->flow_mask.dst_ip6);
1006 APPLY_MASK(&id->src_ip6, &fs->flow_mask.src_ip6);
1007 id->flow_id6 &= fs->flow_mask.flow_id6;
1009 i = ((id->dst_ip6.__u6_addr.__u6_addr32[0]) & 0xffff)^
1010 ((id->dst_ip6.__u6_addr.__u6_addr32[1]) & 0xffff)^
1011 ((id->dst_ip6.__u6_addr.__u6_addr32[2]) & 0xffff)^
1012 ((id->dst_ip6.__u6_addr.__u6_addr32[3]) & 0xffff)^
1014 ((id->dst_ip6.__u6_addr.__u6_addr32[0] >> 15) & 0xffff)^
1015 ((id->dst_ip6.__u6_addr.__u6_addr32[1] >> 15) & 0xffff)^
1016 ((id->dst_ip6.__u6_addr.__u6_addr32[2] >> 15) & 0xffff)^
1017 ((id->dst_ip6.__u6_addr.__u6_addr32[3] >> 15) & 0xffff)^
1019 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 1) & 0xfffff)^
1020 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 1) & 0xfffff)^
1021 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 1) & 0xfffff)^
1022 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 1) & 0xfffff)^
1024 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 16) & 0xffff)^
1025 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 16) & 0xffff)^
1026 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 16) & 0xffff)^
1027 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 16) & 0xffff)^
1029 (id->dst_port << 1) ^ (id->src_port) ^
1033 id->dst_ip &= fs->flow_mask.dst_ip ;
1034 id->src_ip &= fs->flow_mask.src_ip ;
1036 i = ( (id->dst_ip) & 0xffff ) ^
1037 ( (id->dst_ip >> 15) & 0xffff ) ^
1038 ( (id->src_ip << 1) & 0xffff ) ^
1039 ( (id->src_ip >> 16 ) & 0xffff ) ^
1040 (id->dst_port << 1) ^ (id->src_port) ^
1043 i = i % fs->rq_size ;
1044 /* finally, scan the current list for a match */
1046 for (prev=NULL, q = fs->rq[i] ; q ; ) {
1049 IN6_ARE_ADDR_EQUAL(&id->dst_ip6,&q->id.dst_ip6) &&
1050 IN6_ARE_ADDR_EQUAL(&id->src_ip6,&q->id.src_ip6) &&
1051 id->dst_port == q->id.dst_port &&
1052 id->src_port == q->id.src_port &&
1053 id->proto == q->id.proto &&
1054 id->flags == q->id.flags &&
1055 id->flow_id6 == q->id.flow_id6)
1058 if (!is_v6 && id->dst_ip == q->id.dst_ip &&
1059 id->src_ip == q->id.src_ip &&
1060 id->dst_port == q->id.dst_port &&
1061 id->src_port == q->id.src_port &&
1062 id->proto == q->id.proto &&
1063 id->flags == q->id.flags)
1066 /* No match. Check if we can expire the entry */
1067 if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
1068 /* entry is idle and not in any heap, expire it */
1069 struct dn_flow_queue *old_q = q ;
1072 prev->next = q = q->next ;
1074 fs->rq[i] = q = q->next ;
1076 free(old_q, M_DUMMYNET);
1082 if (q && prev != NULL) { /* found and not in front */
1083 prev->next = q->next ;
1084 q->next = fs->rq[i] ;
1088 if (q == NULL) { /* no match, need to allocate a new entry */
1089 q = create_queue(fs, i);
1097 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
1102 * RED calculates the average queue size (avg) using a low-pass filter
1103 * with an exponential weighted (w_q) moving average:
1104 * avg <- (1-w_q) * avg + w_q * q_size
1105 * where q_size is the queue length (measured in bytes or * packets).
1107 * If q_size == 0, we compute the idle time for the link, and set
1108 * avg = (1 - w_q)^(idle/s)
1109 * where s is the time needed for transmitting a medium-sized packet.
1111 * Now, if avg < min_th the packet is enqueued.
1112 * If avg > max_th the packet is dropped. Otherwise, the packet is
1113 * dropped with probability P function of avg.
1118 /* Queue in bytes or packets? */
1119 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ?
1120 q->len_bytes : q->len;
1122 DPRINTF(("\ndummynet: %d q: %2u ", (int)curr_time, q_size));
1124 /* Average queue size estimation. */
1126 /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */
1127 int diff = SCALE(q_size) - q->avg;
1128 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);
1133 * Queue is empty, find for how long the queue has been
1134 * empty and use a lookup table for computing
1135 * (1 - * w_q)^(idle_time/s) where s is the time to send a
1137 * XXX check wraps...
1140 u_int t = (curr_time - q->q_time) / fs->lookup_step;
1142 q->avg = (t < fs->lookup_depth) ?
1143 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
1146 DPRINTF(("dummynet: avg: %u ", SCALE_VAL(q->avg)));
1148 /* Should i drop? */
1149 if (q->avg < fs->min_th) {
1151 return (0); /* accept packet */
1153 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
1154 if (fs->flags_fs & DN_IS_GENTLE_RED) {
1156 * According to Gentle-RED, if avg is greater than
1157 * max_th the packet is dropped with a probability
1158 * p_b = c_3 * avg - c_4
1159 * where c_3 = (1 - max_p) / max_th
1160 * c_4 = 1 - 2 * max_p
1162 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) -
1166 DPRINTF(("dummynet: - drop"));
1169 } else if (q->avg > fs->min_th) {
1171 * We compute p_b using the linear dropping function
1172 * p_b = c_1 * avg - c_2
1173 * where c_1 = max_p / (max_th - min_th)
1174 * c_2 = max_p * min_th / (max_th - min_th)
1176 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
1179 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
1180 p_b = (p_b * len) / fs->max_pkt_size;
1181 if (++q->count == 0)
1182 q->random = random() & 0xffff;
1185 * q->count counts packets arrived since last drop, so a greater
1186 * value of q->count means a greater packet drop probability.
1188 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
1190 DPRINTF(("dummynet: - red drop"));
1191 /* After a drop we calculate a new random value. */
1192 q->random = random() & 0xffff;
1193 return (1); /* drop */
1196 /* End of RED algorithm. */
1198 return (0); /* accept */
1201 static __inline struct dn_flow_set *
1202 locate_flowset(int fs_nr)
1204 struct dn_flow_set *fs;
1206 SLIST_FOREACH(fs, &flowsethash[HASH(fs_nr)], next)
1207 if (fs->fs_nr == fs_nr)
1213 static __inline struct dn_pipe *
1214 locate_pipe(int pipe_nr)
1216 struct dn_pipe *pipe;
1218 SLIST_FOREACH(pipe, &pipehash[HASH(pipe_nr)], next)
1219 if (pipe->pipe_nr == pipe_nr)
1226 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1227 * depending on whether WF2Q or fixed bw is used.
1229 * pipe_nr pipe or queue the packet is destined for.
1230 * dir where shall we send the packet after dummynet.
1231 * m the mbuf with the packet
1232 * ifp the 'ifp' parameter from the caller.
1233 * NULL in ip_input, destination interface in ip_output,
1234 * real_dst in bdg_forward
1235 * rule matching rule, in case of multiple passes
1239 dummynet_io(struct mbuf *m, int dir, struct ip_fw_args *fwa)
1241 struct mbuf *head = NULL, *tail = NULL;
1242 struct dn_pkt_tag *pkt;
1244 struct dn_flow_set *fs = NULL;
1245 struct dn_pipe *pipe ;
1246 u_int64_t len = m->m_pkthdr.len ;
1247 struct dn_flow_queue *q = NULL ;
1249 ipfw_insn *cmd = ACTION_PTR(fwa->rule);
1251 KASSERT(m->m_nextpkt == NULL,
1252 ("dummynet_io: mbuf queue passed to dummynet"));
1254 if (cmd->opcode == O_LOG)
1256 if (cmd->opcode == O_ALTQ)
1258 if (cmd->opcode == O_TAG)
1260 is_pipe = (cmd->opcode == O_PIPE);
1264 * This is a dummynet rule, so we expect an O_PIPE or O_QUEUE rule.
1266 * XXXGL: probably the pipe->fs and fs->pipe logic here
1267 * below can be simplified.
1270 pipe = locate_pipe(fwa->cookie);
1274 fs = locate_flowset(fwa->cookie);
1277 goto dropit; /* This queue/pipe does not exist! */
1279 if (pipe == NULL) { /* Must be a queue, try find a matching pipe. */
1280 pipe = locate_pipe(fs->parent_nr);
1284 printf("dummynet: no pipe %d for queue %d, drop pkt\n",
1285 fs->parent_nr, fs->fs_nr);
1289 q = find_queue(fs, &(fwa->f_id));
1291 goto dropit ; /* cannot allocate queue */
1293 * update statistics, then check reasons to drop pkt
1295 q->tot_bytes += len ;
1297 if ( fs->plr && random() < fs->plr )
1298 goto dropit ; /* random pkt drop */
1299 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) {
1300 if (q->len_bytes > fs->qsize)
1301 goto dropit ; /* queue size overflow */
1303 if (q->len >= fs->qsize)
1304 goto dropit ; /* queue count overflow */
1306 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) )
1309 /* XXX expensive to zero, see if we can remove it*/
1310 mtag = m_tag_get(PACKET_TAG_DUMMYNET,
1311 sizeof(struct dn_pkt_tag), M_NOWAIT|M_ZERO);
1313 goto dropit ; /* cannot allocate packet header */
1314 m_tag_prepend(m, mtag); /* attach to mbuf chain */
1316 pkt = (struct dn_pkt_tag *)(mtag+1);
1317 /* ok, i can handle the pkt now... */
1318 /* build and enqueue packet + parameters */
1319 pkt->rule = fwa->rule ;
1322 pkt->ifp = fwa->oif;
1324 if (q->head == NULL)
1327 q->tail->m_nextpkt = m;
1330 q->len_bytes += len ;
1332 if ( q->head != m ) /* flow was not idle, we are done */
1335 * If we reach this point the flow was previously idle, so we need
1336 * to schedule it. This involves different actions for fixed-rate or
1341 * Fixed-rate queue: just insert into the ready_heap.
1344 if (pipe->bandwidth)
1345 t = SET_TICKS(m, q, pipe);
1346 q->sched_time = curr_time ;
1347 if (t == 0) /* must process it now */
1348 ready_event(q, &head, &tail);
1350 heap_insert(&ready_heap, curr_time + t , q );
1353 * WF2Q. First, compute start time S: if the flow was idle (S=F+1)
1354 * set S to the virtual time V for the controlling pipe, and update
1355 * the sum of weights for the pipe; otherwise, remove flow from
1356 * idle_heap and set S to max(F,V).
1357 * Second, compute finish time F = S + len/weight.
1358 * Third, if pipe was idle, update V=max(S, V).
1359 * Fourth, count one more backlogged flow.
1361 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */
1363 pipe->sum += fs->weight ; /* add weight of new queue */
1365 heap_extract(&(pipe->idle_heap), q);
1366 q->S = MAX64(q->F, pipe->V ) ;
1368 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight;
1370 if (pipe->not_eligible_heap.elements == 0 &&
1371 pipe->scheduler_heap.elements == 0)
1372 pipe->V = MAX64 ( q->S, pipe->V );
1375 * Look at eligibility. A flow is not eligibile if S>V (when
1376 * this happens, it means that there is some other flow already
1377 * scheduled for the same pipe, so the scheduler_heap cannot be
1378 * empty). If the flow is not eligible we just store it in the
1379 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1380 * and possibly invoke ready_event_wfq() right now if there is
1382 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1383 * and for all flows in not_eligible_heap (NEH), S_i > V .
1384 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH,
1385 * we only need to look into NEH.
1387 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */
1388 if (pipe->scheduler_heap.elements == 0)
1389 printf("dummynet: ++ ouch! not eligible but empty scheduler!\n");
1390 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1392 heap_insert(&(pipe->scheduler_heap), q->F, q);
1393 if (pipe->numbytes >= 0) { /* pipe is idle */
1394 if (pipe->scheduler_heap.elements != 1)
1395 printf("dummynet: OUCH! pipe should have been idle!\n");
1396 DPRINTF(("dummynet: waking up pipe %d at %d\n",
1397 pipe->pipe_nr, (int)(q->F >> MY_M)));
1398 pipe->sched_time = curr_time ;
1399 ready_event_wfq(pipe, &head, &tail);
1406 dummynet_send(head);
1414 return ( (fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1418 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1419 * Doing this would probably save us the initial bzero of dn_pkt
1421 #define DN_FREE_PKT(_m) do { \
1426 * Dispose all packets and flow_queues on a flow_set.
1427 * If all=1, also remove red lookup table and other storage,
1428 * including the descriptor itself.
1429 * For the one in dn_pipe MUST also cleanup ready_heap...
1432 purge_flow_set(struct dn_flow_set *fs, int all)
1434 struct dn_flow_queue *q, *qn;
1437 DUMMYNET_LOCK_ASSERT();
1439 for (i = 0; i <= fs->rq_size; i++) {
1440 for (q = fs->rq[i]; q != NULL; q = qn) {
1441 struct mbuf *m, *mnext;
1444 while ((m = mnext) != NULL) {
1445 mnext = m->m_nextpkt;
1449 free(q, M_DUMMYNET);
1454 fs->rq_elements = 0;
1456 /* RED - free lookup table. */
1457 if (fs->w_q_lookup != NULL)
1458 free(fs->w_q_lookup, M_DUMMYNET);
1460 free(fs->rq, M_DUMMYNET);
1461 /* If this fs is not part of a pipe, free it. */
1462 if (fs->pipe == NULL || fs != &(fs->pipe->fs))
1463 free(fs, M_DUMMYNET);
1468 * Dispose all packets queued on a pipe (not a flow_set).
1469 * Also free all resources associated to a pipe, which is about
1473 purge_pipe(struct dn_pipe *pipe)
1475 struct mbuf *m, *mnext;
1477 purge_flow_set( &(pipe->fs), 1 );
1480 while ((m = mnext) != NULL) {
1481 mnext = m->m_nextpkt;
1485 heap_free( &(pipe->scheduler_heap) );
1486 heap_free( &(pipe->not_eligible_heap) );
1487 heap_free( &(pipe->idle_heap) );
1491 * Delete all pipes and heaps returning memory. Must also
1492 * remove references from all ipfw rules to all pipes.
1495 dummynet_flush(void)
1497 struct dn_pipe *pipe, *pipe1;
1498 struct dn_flow_set *fs, *fs1;
1502 /* Free heaps so we don't have unwanted events. */
1503 heap_free(&ready_heap);
1504 heap_free(&wfq_ready_heap);
1505 heap_free(&extract_heap);
1508 * Now purge all queued pkts and delete all pipes.
1510 * XXXGL: can we merge the for(;;) cycles into one or not?
1512 for (i = 0; i < HASHSIZE; i++)
1513 SLIST_FOREACH_SAFE(fs, &flowsethash[i], next, fs1) {
1514 SLIST_REMOVE(&flowsethash[i], fs, dn_flow_set, next);
1515 purge_flow_set(fs, 1);
1517 for (i = 0; i < HASHSIZE; i++)
1518 SLIST_FOREACH_SAFE(pipe, &pipehash[i], next, pipe1) {
1519 SLIST_REMOVE(&pipehash[i], pipe, dn_pipe, next);
1521 free(pipe, M_DUMMYNET);
1526 extern struct ip_fw *ip_fw_default_rule ;
1528 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1531 struct dn_flow_queue *q ;
1534 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1535 for (q = fs->rq[i] ; q ; q = q->next )
1536 for (m = q->head ; m ; m = m->m_nextpkt ) {
1537 struct dn_pkt_tag *pkt = dn_tag_get(m) ;
1539 pkt->rule = ip_fw_default_rule ;
1543 * when a firewall rule is deleted, scan all queues and remove the flow-id
1544 * from packets matching this rule.
1547 dn_rule_delete(void *r)
1549 struct dn_pipe *pipe;
1550 struct dn_flow_set *fs;
1551 struct dn_pkt_tag *pkt;
1557 * If the rule references a queue (dn_flow_set), then scan
1558 * the flow set, otherwise scan pipes. Should do either, but doing
1559 * both does not harm.
1561 for (i = 0; i < HASHSIZE; i++)
1562 SLIST_FOREACH(fs, &flowsethash[i], next)
1563 dn_rule_delete_fs(fs, r);
1565 for (i = 0; i < HASHSIZE; i++)
1566 SLIST_FOREACH(pipe, &pipehash[i], next) {
1568 dn_rule_delete_fs(fs, r);
1569 for (m = pipe->head ; m ; m = m->m_nextpkt ) {
1570 pkt = dn_tag_get(m);
1572 pkt->rule = ip_fw_default_rule;
1579 * setup RED parameters
1582 config_red(struct dn_flow_set *p, struct dn_flow_set * x)
1587 x->min_th = SCALE(p->min_th);
1588 x->max_th = SCALE(p->max_th);
1589 x->max_p = p->max_p;
1591 x->c_1 = p->max_p / (p->max_th - p->min_th);
1592 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1594 if (x->flags_fs & DN_IS_GENTLE_RED) {
1595 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1596 x->c_4 = SCALE(1) - 2 * p->max_p;
1599 /* If the lookup table already exist, free and create it again. */
1600 if (x->w_q_lookup) {
1601 free(x->w_q_lookup, M_DUMMYNET);
1602 x->w_q_lookup = NULL;
1604 if (red_lookup_depth == 0) {
1605 printf("\ndummynet: net.inet.ip.dummynet.red_lookup_depth"
1607 free(x, M_DUMMYNET);
1610 x->lookup_depth = red_lookup_depth;
1611 x->w_q_lookup = (u_int *)malloc(x->lookup_depth * sizeof(int),
1612 M_DUMMYNET, M_NOWAIT);
1613 if (x->w_q_lookup == NULL) {
1614 printf("dummynet: sorry, cannot allocate red lookup table\n");
1615 free(x, M_DUMMYNET);
1619 /* Fill the lookup table with (1 - w_q)^x */
1620 x->lookup_step = p->lookup_step;
1621 x->lookup_weight = p->lookup_weight;
1622 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1624 for (i = 1; i < x->lookup_depth; i++)
1626 SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1628 if (red_avg_pkt_size < 1)
1629 red_avg_pkt_size = 512;
1630 x->avg_pkt_size = red_avg_pkt_size;
1631 if (red_max_pkt_size < 1)
1632 red_max_pkt_size = 1500;
1633 x->max_pkt_size = red_max_pkt_size;
1638 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1640 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1641 int l = pfs->rq_size;
1647 else if (l > DN_MAX_HASH_SIZE)
1648 l = DN_MAX_HASH_SIZE;
1650 } else /* one is enough for null mask */
1652 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1653 M_DUMMYNET, M_NOWAIT | M_ZERO);
1654 if (x->rq == NULL) {
1655 printf("dummynet: sorry, cannot allocate queue\n");
1663 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1665 x->flags_fs = src->flags_fs;
1666 x->qsize = src->qsize;
1668 x->flow_mask = src->flow_mask;
1669 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1670 if (x->qsize > 1024 * 1024)
1671 x->qsize = 1024 * 1024;
1678 /* Configuring RED. */
1679 if (x->flags_fs & DN_IS_RED)
1680 config_red(src, x); /* XXX should check errors */
1684 * Setup pipe or queue parameters.
1687 config_pipe(struct dn_pipe *p)
1689 struct dn_flow_set *pfs = &(p->fs);
1690 struct dn_flow_queue *q;
1694 * The config program passes parameters as follows:
1695 * bw = bits/second (0 means no limits),
1696 * delay = ms, must be translated into ticks.
1697 * qsize = slots/bytes
1699 p->delay = (p->delay * hz) / 1000;
1700 /* We need either a pipe number or a flow_set number. */
1701 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1703 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1705 if (p->pipe_nr != 0) { /* this is a pipe */
1706 struct dn_pipe *pipe;
1709 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1711 if (pipe == NULL) { /* new pipe */
1712 pipe = malloc(sizeof(struct dn_pipe), M_DUMMYNET,
1716 printf("dummynet: no memory for new pipe\n");
1719 pipe->pipe_nr = p->pipe_nr;
1720 pipe->fs.pipe = pipe;
1722 * idle_heap is the only one from which
1723 * we extract from the middle.
1725 pipe->idle_heap.size = pipe->idle_heap.elements = 0;
1726 pipe->idle_heap.offset =
1727 OFFSET_OF(struct dn_flow_queue, heap_pos);
1729 /* Flush accumulated credit for all queues. */
1730 for (i = 0; i <= pipe->fs.rq_size; i++)
1731 for (q = pipe->fs.rq[i]; q; q = q->next)
1734 pipe->bandwidth = p->bandwidth;
1735 pipe->numbytes = 0; /* just in case... */
1736 bcopy(p->if_name, pipe->if_name, sizeof(p->if_name));
1737 pipe->ifp = NULL; /* reset interface ptr */
1738 pipe->delay = p->delay;
1739 set_fs_parms(&(pipe->fs), pfs);
1741 if (pipe->fs.rq == NULL) { /* a new pipe */
1742 error = alloc_hash(&(pipe->fs), pfs);
1745 free(pipe, M_DUMMYNET);
1748 SLIST_INSERT_HEAD(&pipehash[HASH(pipe->pipe_nr)],
1752 } else { /* config queue */
1753 struct dn_flow_set *fs;
1756 fs = locate_flowset(pfs->fs_nr); /* locate flow_set */
1758 if (fs == NULL) { /* new */
1759 if (pfs->parent_nr == 0) { /* need link to a pipe */
1763 fs = malloc(sizeof(struct dn_flow_set), M_DUMMYNET,
1768 "dummynet: no memory for new flow_set\n");
1771 fs->fs_nr = pfs->fs_nr;
1772 fs->parent_nr = pfs->parent_nr;
1773 fs->weight = pfs->weight;
1774 if (fs->weight == 0)
1776 else if (fs->weight > 100)
1780 * Change parent pipe not allowed;
1781 * must delete and recreate.
1783 if (pfs->parent_nr != 0 &&
1784 fs->parent_nr != pfs->parent_nr) {
1790 set_fs_parms(fs, pfs);
1792 if (fs->rq == NULL) { /* a new flow_set */
1793 error = alloc_hash(fs, pfs);
1796 free(fs, M_DUMMYNET);
1799 SLIST_INSERT_HEAD(&flowsethash[HASH(fs->fs_nr)],
1808 * Helper function to remove from a heap queues which are linked to
1809 * a flow_set about to be deleted.
1812 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1814 int i = 0, found = 0 ;
1815 for (; i < h->elements ;)
1816 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1818 h->p[i] = h->p[h->elements] ;
1827 * helper function to remove a pipe from a heap (can be there at most once)
1830 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1832 if (h->elements > 0) {
1834 for (i=0; i < h->elements ; i++ ) {
1835 if (h->p[i].object == p) { /* found it */
1837 h->p[i] = h->p[h->elements] ;
1846 * drain all queues. Called in case of severe mbuf shortage.
1851 struct dn_flow_set *fs;
1852 struct dn_pipe *pipe;
1853 struct mbuf *m, *mnext;
1856 DUMMYNET_LOCK_ASSERT();
1858 heap_free(&ready_heap);
1859 heap_free(&wfq_ready_heap);
1860 heap_free(&extract_heap);
1861 /* remove all references to this pipe from flow_sets */
1862 for (i = 0; i < HASHSIZE; i++)
1863 SLIST_FOREACH(fs, &flowsethash[i], next)
1864 purge_flow_set(fs, 0);
1866 for (i = 0; i < HASHSIZE; i++) {
1867 SLIST_FOREACH(pipe, &pipehash[i], next) {
1868 purge_flow_set(&(pipe->fs), 0);
1871 while ((m = mnext) != NULL) {
1872 mnext = m->m_nextpkt;
1875 pipe->head = pipe->tail = NULL;
1881 * Fully delete a pipe or a queue, cleaning up associated info.
1884 delete_pipe(struct dn_pipe *p)
1886 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1888 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1890 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1891 struct dn_pipe *pipe;
1892 struct dn_flow_set *fs;
1896 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1900 return (ENOENT); /* not found */
1903 /* Unlink from list of pipes. */
1904 SLIST_REMOVE(&pipehash[HASH(pipe->pipe_nr)], pipe, dn_pipe, next);
1906 /* Remove all references to this pipe from flow_sets. */
1907 for (i = 0; i < HASHSIZE; i++)
1908 SLIST_FOREACH(fs, &flowsethash[i], next)
1909 if (fs->pipe == pipe) {
1910 printf("dummynet: ++ ref to pipe %d from fs %d\n",
1911 p->pipe_nr, fs->fs_nr);
1913 purge_flow_set(fs, 0);
1915 fs_remove_from_heap(&ready_heap, &(pipe->fs));
1916 purge_pipe(pipe); /* remove all data associated to this pipe */
1917 /* remove reference to here from extract_heap and wfq_ready_heap */
1918 pipe_remove_from_heap(&extract_heap, pipe);
1919 pipe_remove_from_heap(&wfq_ready_heap, pipe);
1922 free(pipe, M_DUMMYNET);
1923 } else { /* this is a WF2Q queue (dn_flow_set) */
1924 struct dn_flow_set *fs;
1927 fs = locate_flowset(p->fs.fs_nr); /* locate set */
1931 return (ENOENT); /* not found */
1934 /* Unlink from list of flowsets. */
1935 SLIST_REMOVE( &flowsethash[HASH(fs->fs_nr)], fs, dn_flow_set, next);
1937 if (fs->pipe != NULL) {
1938 /* Update total weight on parent pipe and cleanup parent heaps. */
1939 fs->pipe->sum -= fs->weight * fs->backlogged ;
1940 fs_remove_from_heap(&(fs->pipe->not_eligible_heap), fs);
1941 fs_remove_from_heap(&(fs->pipe->scheduler_heap), fs);
1942 #if 1 /* XXX should i remove from idle_heap as well ? */
1943 fs_remove_from_heap(&(fs->pipe->idle_heap), fs);
1946 purge_flow_set(fs, 1);
1953 * helper function used to copy data from kernel in DUMMYNET_GET
1956 dn_copy_set(struct dn_flow_set *set, char *bp)
1959 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1961 DUMMYNET_LOCK_ASSERT();
1963 for (i = 0 ; i <= set->rq_size ; i++)
1964 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
1965 if (q->hash_slot != i)
1966 printf("dummynet: ++ at %d: wrong slot (have %d, "
1967 "should be %d)\n", copied, q->hash_slot, i);
1969 printf("dummynet: ++ at %d: wrong fs ptr (have %p, should be %p)\n",
1972 bcopy(q, qp, sizeof( *q ) );
1973 /* cleanup pointers */
1975 qp->head = qp->tail = NULL ;
1978 if (copied != set->rq_elements)
1979 printf("dummynet: ++ wrong count, have %d should be %d\n",
1980 copied, set->rq_elements);
1987 struct dn_flow_set *fs;
1988 struct dn_pipe *pipe;
1992 DUMMYNET_LOCK_ASSERT();
1994 * Compute size of data structures: list of pipes and flow_sets.
1996 for (i = 0; i < HASHSIZE; i++) {
1997 SLIST_FOREACH(pipe, &pipehash[i], next)
1998 size += sizeof(*pipe) +
1999 pipe->fs.rq_elements * sizeof(struct dn_flow_queue);
2000 SLIST_FOREACH(fs, &flowsethash[i], next)
2001 size += sizeof (*fs) +
2002 fs->rq_elements * sizeof(struct dn_flow_queue);
2008 dummynet_get(struct sockopt *sopt)
2010 char *buf, *bp ; /* bp is the "copy-pointer" */
2012 struct dn_flow_set *fs;
2013 struct dn_pipe *pipe;
2016 /* XXX lock held too long */
2019 * XXX: Ugly, but we need to allocate memory with M_WAITOK flag and we
2020 * cannot use this flag while holding a mutex.
2022 for (i = 0; i < 10; i++) {
2023 size = dn_calc_size();
2025 buf = malloc(size, M_TEMP, M_WAITOK);
2027 if (size == dn_calc_size())
2037 for (i = 0; i < HASHSIZE; i++)
2038 SLIST_FOREACH(pipe, &pipehash[i], next) {
2039 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp;
2042 * Copy pipe descriptor into *bp, convert delay back to ms,
2043 * then copy the flow_set descriptor(s) one at a time.
2044 * After each flow_set, copy the queue descriptor it owns.
2046 bcopy(pipe, bp, sizeof(*pipe));
2047 pipe_bp->delay = (pipe_bp->delay * 1000) / hz;
2049 * XXX the following is a hack based on ->next being the
2050 * first field in dn_pipe and dn_flow_set. The correct
2051 * solution would be to move the dn_flow_set to the beginning
2052 * of struct dn_pipe.
2054 pipe_bp->next.sle_next = (struct dn_pipe *)DN_IS_PIPE;
2055 /* Clean pointers. */
2056 pipe_bp->head = pipe_bp->tail = NULL;
2057 pipe_bp->fs.next.sle_next = NULL;
2058 pipe_bp->fs.pipe = NULL;
2059 pipe_bp->fs.rq = NULL;
2061 bp += sizeof(*pipe) ;
2062 bp = dn_copy_set(&(pipe->fs), bp);
2065 for (i = 0; i < HASHSIZE; i++)
2066 SLIST_FOREACH(fs, &flowsethash[i], next) {
2067 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp;
2069 bcopy(fs, bp, sizeof(*fs));
2070 /* XXX same hack as above */
2071 fs_bp->next.sle_next = (struct dn_flow_set *)DN_IS_QUEUE;
2075 bp = dn_copy_set(fs, bp);
2080 error = sooptcopyout(sopt, buf, size);
2086 * Handler for the various dummynet socket options (get, flush, config, del)
2089 ip_dn_ctl(struct sockopt *sopt)
2092 struct dn_pipe *p, tmp_pipe;
2094 /* Disallow sets in really-really secure mode. */
2095 if (sopt->sopt_dir == SOPT_SET) {
2096 #if __FreeBSD_version >= 500034
2097 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
2101 if (securelevel >= 3)
2106 switch (sopt->sopt_name) {
2108 printf("dummynet: -- unknown option %d", sopt->sopt_name);
2111 case IP_DUMMYNET_GET :
2112 error = dummynet_get(sopt);
2115 case IP_DUMMYNET_FLUSH :
2119 case IP_DUMMYNET_CONFIGURE :
2121 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2124 error = config_pipe(p);
2127 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
2129 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2133 error = delete_pipe(p);
2145 printf("DUMMYNET with IPv6 initialized (040826)\n");
2147 DUMMYNET_LOCK_INIT();
2149 for (i = 0; i < HASHSIZE; i++) {
2150 SLIST_INIT(&pipehash[i]);
2151 SLIST_INIT(&flowsethash[i]);
2153 ready_heap.size = ready_heap.elements = 0;
2154 ready_heap.offset = 0;
2156 wfq_ready_heap.size = wfq_ready_heap.elements = 0;
2157 wfq_ready_heap.offset = 0;
2159 extract_heap.size = extract_heap.elements = 0;
2160 extract_heap.offset = 0;
2162 ip_dn_ctl_ptr = ip_dn_ctl;
2163 ip_dn_io_ptr = dummynet_io;
2164 ip_dn_ruledel_ptr = dn_rule_delete;
2166 TASK_INIT(&dn_task, 0, dummynet_task, NULL);
2167 dn_tq = taskqueue_create_fast("dummynet", M_NOWAIT,
2168 taskqueue_thread_enqueue, &dn_tq);
2169 taskqueue_start_threads(&dn_tq, 1, PI_NET, "dummynet");
2171 callout_init(&dn_timeout, NET_CALLOUT_MPSAFE);
2172 callout_reset(&dn_timeout, 1, dummynet, NULL);
2174 /* Initialize curr_time adjustment mechanics. */
2175 getmicrouptime(&prev_t);
2182 ip_dn_ctl_ptr = NULL;
2183 ip_dn_io_ptr = NULL;
2184 ip_dn_ruledel_ptr = NULL;
2187 callout_stop(&dn_timeout);
2189 taskqueue_drain(dn_tq, &dn_task);
2190 taskqueue_free(dn_tq);
2194 DUMMYNET_LOCK_DESTROY();
2196 #endif /* KLD_MODULE */
2199 dummynet_modevent(module_t mod, int type, void *data)
2203 if (DUMMYNET_LOADED) {
2204 printf("DUMMYNET already loaded\n");
2211 #if !defined(KLD_MODULE)
2212 printf("dummynet statically compiled, cannot unload\n");
2225 static moduledata_t dummynet_mod = {
2230 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY);
2231 MODULE_DEPEND(dummynet, ipfw, 2, 2, 2);
2232 MODULE_VERSION(dummynet, 1);