2 * Copyright (c) 1998-2002 Luigi Rizzo, Universita` di Pisa
3 * Portions Copyright (c) 2000 Akamba Corp.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * $FreeBSD: src/sys/netinet/ip_dummynet.c,v 1.24.2.22 2003/05/13 09:31:06 maxim Exp $
28 * $DragonFly: src/sys/net/dummynet/ip_dummynet.c,v 1.38 2007/11/02 10:28:50 sephe Exp $
32 #include "opt_ipfw.h" /* for IPFW2 definition */
39 * This module implements IP dummynet, a bandwidth limiter/delay emulator
40 * used in conjunction with the ipfw package.
41 * Description of the data structures used is in ip_dummynet.h
42 * Here you mainly find the following blocks of code:
43 * + variable declarations;
44 * + heap management functions;
45 * + scheduler and dummynet functions;
46 * + configuration and initialization.
48 * 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
58 #include <sys/param.h>
59 #include <sys/kernel.h>
60 #include <sys/malloc.h>
62 #include <sys/socketvar.h>
63 #include <sys/sysctl.h>
64 #include <sys/systimer.h>
65 #include <sys/thread2.h>
67 #include <net/ethernet.h>
68 #include <net/route.h>
69 #include <net/netmsg2.h>
71 #include <netinet/in.h>
72 #include <netinet/in_var.h>
73 #include <netinet/ip.h>
74 #include <netinet/ip_var.h>
76 #include <net/ipfw/ip_fw.h>
77 #include <net/dummynet/ip_dummynet.h>
79 #ifndef DUMMYNET_CALLOUT_FREQ_MAX
80 #define DUMMYNET_CALLOUT_FREQ_MAX 30000
84 * We keep a private variable for the simulation time, but we could
85 * probably use an existing one ("softticks" in sys/kern/kern_timer.c)
87 static dn_key curr_time = 0; /* current simulation time */
89 static int dn_hash_size = 64; /* default hash size */
91 /* statistics on number of queue searches and search steps */
92 static int searches, search_steps;
93 static int pipe_expire = 1; /* expire queue if empty */
94 static int dn_max_ratio = 16; /* max queues/buckets ratio */
96 static int red_lookup_depth = 256; /* RED - default lookup table depth */
97 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
98 static int red_max_pkt_size = 1500; /* RED - default max packet size */
101 * Three heaps contain queues and pipes that the scheduler handles:
103 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
105 * wfq_ready_heap contains the pipes associated with WF2Q flows
107 * extract_heap contains pipes associated with delay lines.
111 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
113 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap;
115 static int heap_init(struct dn_heap *h, int size);
116 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
117 static void heap_extract(struct dn_heap *h, void *obj);
119 static void transmit_event(struct dn_pipe *pipe);
120 static void ready_event(struct dn_flow_queue *q);
122 static int sysctl_dn_hz(SYSCTL_HANDLER_ARGS);
124 static struct dn_pipe *all_pipes = NULL; /* list of all pipes */
125 static struct dn_flow_set *all_flow_sets = NULL;/* list of all flow_sets */
127 static struct netmsg dn_netmsg;
128 static struct systimer dn_clock;
129 static int dn_hz = 1000;
130 static int dn_cpu = 0; /* TODO tunable */
132 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
133 CTLFLAG_RW, 0, "Dummynet");
134 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
135 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
136 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time,
137 CTLFLAG_RD, &curr_time, 0, "Current tick");
138 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
139 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
140 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
141 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
142 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches,
143 CTLFLAG_RD, &searches, 0, "Number of queue searches");
144 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps,
145 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
146 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
147 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
148 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
149 CTLFLAG_RW, &dn_max_ratio, 0,
150 "Max ratio between dynamic queues and buckets");
151 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
152 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
153 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
154 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
155 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
156 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
157 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hz, CTLTYPE_INT | CTLFLAG_RW,
158 0, 0, sysctl_dn_hz, "I", "Dummynet callout frequency");
160 static int config_pipe(struct dn_pipe *p);
161 static int ip_dn_ctl(struct sockopt *sopt);
163 static void rt_unref(struct rtentry *);
164 static void dummynet_clock(systimer_t, struct intrframe *);
165 static void dummynet(struct netmsg *);
166 static void dummynet_flush(void);
167 static ip_dn_io_t dummynet_io;
168 static void dn_rule_delete(void *);
170 void dummynet_drain(void); /* XXX unused */
173 rt_unref(struct rtentry *rt)
177 if (rt->rt_refcnt <= 0)
178 kprintf("-- warning, refcnt now %ld, decreasing\n", rt->rt_refcnt);
183 * Heap management functions.
185 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
186 * Some macros help finding parent/children so we can optimize them.
188 * heap_init() is called to expand the heap when needed.
189 * Increment size in blocks of 16 entries.
190 * XXX failure to allocate a new element is a pretty bad failure
191 * as we basically stall a whole queue forever!!
192 * Returns 1 on error, 0 on success
194 #define HEAP_FATHER(x) (((x) - 1) / 2)
195 #define HEAP_LEFT(x) (2*(x) + 1)
196 #define HEAP_IS_LEFT(x) ((x) & 1)
197 #define HEAP_RIGHT(x) (2*(x) + 2)
198 #define HEAP_SWAP(a, b, buffer) { buffer = a; a = b; b = buffer; }
199 #define HEAP_INCREMENT 15
202 heap_init(struct dn_heap *h, int new_size)
204 struct dn_heap_entry *p;
206 if (h->size >= new_size) {
207 kprintf("%s, Bogus call, have %d want %d\n", __func__,
212 new_size = (new_size + HEAP_INCREMENT) & ~HEAP_INCREMENT;
213 p = kmalloc(new_size * sizeof(*p), M_DUMMYNET, M_WAITOK | M_ZERO);
215 bcopy(h->p, p, h->size * sizeof(*p));
216 kfree(h->p, M_DUMMYNET);
224 * Insert element in heap. Normally, p != NULL, we insert p in
225 * a new position and bubble up. If p == NULL, then the element is
226 * already in place, and key is the position where to start the
228 * Returns 1 on failure (cannot allocate new heap entry)
230 * If offset > 0 the position (index, int) of the element in the heap is
231 * also stored in the element itself at the given offset in bytes.
233 #define SET_OFFSET(heap, node) \
234 if (heap->offset > 0) \
235 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node;
238 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
240 #define RESET_OFFSET(heap, node) \
241 if (heap->offset > 0) \
242 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1;
245 heap_insert(struct dn_heap *h, dn_key key1, void *p)
247 int son = h->elements;
249 if (p == NULL) { /* Data already there, set starting point */
251 } else { /* Insert new element at the end, possibly resize */
253 if (son == h->size) { /* Need resize... */
254 if (heap_init(h, h->elements + 1))
255 return 1; /* Failure... */
257 h->p[son].object = p;
258 h->p[son].key = key1;
262 while (son > 0) { /* Bubble up */
263 int father = HEAP_FATHER(son);
264 struct dn_heap_entry tmp;
266 if (DN_KEY_LT(h->p[father].key, h->p[son].key))
267 break; /* Found right position */
269 /* 'son' smaller than 'father', swap and repeat */
270 HEAP_SWAP(h->p[son], h->p[father], tmp);
279 * Remove top element from heap, or obj if obj != NULL
282 heap_extract(struct dn_heap *h, void *obj)
284 int child, father, max = h->elements - 1;
287 kprintf("warning, extract from empty heap 0x%p\n", h);
291 father = 0; /* Default: move up smallest child */
292 if (obj != NULL) { /* Extract specific element, index is at offset */
294 panic("%s from middle not supported on this heap!!!\n", __func__);
296 father = *((int *)((char *)obj + h->offset));
297 if (father < 0 || father >= h->elements) {
298 panic("%s father %d out of bound 0..%d\n", __func__,
299 father, h->elements);
302 RESET_OFFSET(h, father);
304 child = HEAP_LEFT(father); /* Left child */
305 while (child <= max) { /* Valid entry */
306 if (child != max && DN_KEY_LT(h->p[child + 1].key, h->p[child].key))
307 child = child + 1; /* Take right child, otherwise left */
308 h->p[father] = h->p[child];
309 SET_OFFSET(h, father);
311 child = HEAP_LEFT(child); /* Left child for next loop */
316 * Fill hole with last entry and bubble up, reusing the insert code
318 h->p[father] = h->p[max];
319 heap_insert(h, father, NULL); /* This one cannot fail */
324 * heapify() will reorganize data inside an array to maintain the
325 * heap property. It is needed when we delete a bunch of entries.
328 heapify(struct dn_heap *h)
332 for (i = 0; i < h->elements; i++)
333 heap_insert(h, i , NULL);
337 * Cleanup the heap and free data structure
340 heap_free(struct dn_heap *h)
343 kfree(h->p, M_DUMMYNET);
344 bzero(h, sizeof(*h));
348 * --- End of heap management functions ---
352 * Scheduler functions:
354 * transmit_event() is called when the delay-line needs to enter
355 * the scheduler, either because of existing pkts getting ready,
356 * or new packets entering the queue. The event handled is the delivery
357 * time of the packet.
359 * ready_event() does something similar with fixed-rate queues, and the
360 * event handled is the finish time of the head pkt.
362 * wfq_ready_event() does something similar with WF2Q queues, and the
363 * event handled is the start time of the head pkt.
365 * In all cases, we make sure that the data structures are consistent
366 * before passing pkts out, because this might trigger recursive
367 * invocations of the procedures.
370 transmit_event(struct dn_pipe *pipe)
374 while ((pkt = pipe->head) && DN_KEY_LEQ(pkt->output_time, curr_time)) {
378 * First unlink, then call procedures, since ip_input() can invoke
379 * ip_output() and viceversa, thus causing nested calls
381 pipe->head = pkt->dn_next;
385 * 'pkt' should _not_ be touched after calling
386 * ip_output(), ip_input(), ether_demux() and ether_output_frame()
388 switch (pkt->dn_dir) {
391 * 'pkt' will be freed in ip_output, so we keep
392 * a reference of the 'rtentry' beforehand.
395 ip_output(pkt->dn_m, NULL, NULL, 0, NULL, NULL);
403 case DN_TO_ETH_DEMUX:
405 struct mbuf *m = pkt->dn_m;
406 struct ether_header *eh;
408 if (m->m_len < ETHER_HDR_LEN &&
409 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
410 kprintf("dummynet: pullup fail, dropping pkt\n");
414 * Same as ether_input, make eh be a pointer into the mbuf
416 eh = mtod(m, struct ether_header *);
417 m_adj(m, ETHER_HDR_LEN);
418 ether_demux(NULL, eh, m);
423 ether_output_frame(pkt->ifp, pkt->dn_m);
427 kprintf("dummynet: bad switch %d!\n", pkt->dn_dir);
434 * If there are leftover packets, put into the heap for next event
436 if ((pkt = pipe->head)) {
438 * XXX should check errors on heap_insert, by draining the
439 * whole pipe and hoping in the future we are more successful
441 heap_insert(&extract_heap, pkt->output_time, pipe);
446 * The following macro computes how many ticks we have to wait
447 * before being able to transmit a packet. The credit is taken from
448 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
450 #define SET_TICKS(pkt, q, p) \
451 (pkt->dn_m->m_pkthdr.len*8*dn_hz - (q)->numbytes + p->bandwidth - 1 ) / \
455 * Extract pkt from queue, compute output time (could be now)
456 * and put into delay line (p_queue)
459 move_pkt(struct dn_pkt *pkt, struct dn_flow_queue *q,
460 struct dn_pipe *p, int len)
462 q->head = pkt->dn_next;
466 pkt->output_time = curr_time + p->delay;
471 p->tail->dn_next = pkt;
473 p->tail->dn_next = NULL;
477 * ready_event() is invoked every time the queue must enter the
478 * scheduler, either because the first packet arrives, or because
479 * a previously scheduled event fired.
480 * On invokation, drain as many pkts as possible (could be 0) and then
481 * if there are leftover packets reinsert the pkt in the scheduler.
484 ready_event(struct dn_flow_queue *q)
487 struct dn_pipe *p = q->fs->pipe;
491 kprintf("ready_event- pipe is gone\n");
494 p_was_empty = (p->head == NULL);
497 * Schedule fixed-rate queues linked to this pipe:
498 * Account for the bw accumulated since last scheduling, then
499 * drain as many pkts as allowed by q->numbytes and move to
500 * the delay line (in p) computing output time.
501 * bandwidth==0 (no limit) means we can drain the whole queue,
502 * setting len_scaled = 0 does the job.
504 q->numbytes += (curr_time - q->sched_time) * p->bandwidth;
505 while ((pkt = q->head) != NULL) {
506 int len = pkt->dn_m->m_pkthdr.len;
507 int len_scaled = p->bandwidth ? len*8*dn_hz : 0;
509 if (len_scaled > q->numbytes)
511 q->numbytes -= len_scaled;
512 move_pkt(pkt, q, p, len);
516 * If we have more packets queued, schedule next ready event
517 * (can only occur when bandwidth != 0, otherwise we would have
518 * flushed the whole queue in the previous loop).
519 * To this purpose we record the current time and compute how many
520 * ticks to go for the finish time of the packet.
522 if ((pkt = q->head) != NULL) { /* this implies bandwidth != 0 */
523 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
525 q->sched_time = curr_time;
528 * XXX should check errors on heap_insert, and drain the whole
529 * queue on error hoping next time we are luckier.
531 heap_insert(&ready_heap, curr_time + t, q);
532 } else { /* RED needs to know when the queue becomes empty */
533 q->q_time = curr_time;
538 * If the delay line was empty call transmit_event(p) now.
539 * Otherwise, the scheduler will take care of it.
546 * Called when we can transmit packets on WF2Q queues. Take pkts out of
547 * the queues at their start time, and enqueue into the delay line.
548 * Packets are drained until p->numbytes < 0. As long as
549 * len_scaled >= p->numbytes, the packet goes into the delay line
550 * with a deadline p->delay. For the last packet, if p->numbytes < 0,
551 * there is an additional delay.
554 ready_event_wfq(struct dn_pipe *p)
556 int p_was_empty = (p->head == NULL);
557 struct dn_heap *sch = &p->scheduler_heap;
558 struct dn_heap *neh = &p->not_eligible_heap;
560 if (p->if_name[0] == 0) { /* tx clock is simulated */
561 p->numbytes += (curr_time - p->sched_time) * p->bandwidth;
562 } else { /* tx clock is for real, the ifq must be empty or this is a NOP */
563 if (p->ifp && p->ifp->if_snd.ifq_head != NULL) {
566 DEB(kprintf("pipe %d ready from %s --\n",
567 p->pipe_nr, p->if_name);)
572 * While we have backlogged traffic AND credit, we need to do
573 * something on the queue.
575 while (p->numbytes >= 0 && (sch->elements > 0 || neh->elements > 0)) {
576 if (sch->elements > 0) { /* Have some eligible pkts to send out */
577 struct dn_flow_queue *q = sch->p[0].object;
578 struct dn_pkt *pkt = q->head;
579 struct dn_flow_set *fs = q->fs;
580 uint64_t len = pkt->dn_m->m_pkthdr.len;
581 int len_scaled = p->bandwidth ? len*8*dn_hz : 0;
583 heap_extract(sch, NULL); /* Remove queue from heap */
584 p->numbytes -= len_scaled;
585 move_pkt(pkt, q, p, len);
587 p->V += (len << MY_M) / p->sum; /* Update V */
588 q->S = q->F; /* Update start time */
590 if (q->len == 0) { /* Flow not backlogged any more */
592 heap_insert(&p->idle_heap, q->F, q);
593 } else { /* Still backlogged */
595 * Update F and position in backlogged queue, then
596 * put flow in not_eligible_heap (we will fix this later).
598 len = q->head->dn_m->m_pkthdr.len;
599 q->F += (len << MY_M) / (uint64_t)fs->weight;
600 if (DN_KEY_LEQ(q->S, p->V))
601 heap_insert(neh, q->S, q);
603 heap_insert(sch, q->F, q);
608 * Now compute V = max(V, min(S_i)). Remember that all elements in
609 * sch have by definition S_i <= V so if sch is not empty, V is surely
610 * the max and we must not update it. Conversely, if sch is empty
611 * we only need to look at neh.
613 if (sch->elements == 0 && neh->elements > 0)
614 p->V = MAX64(p->V, neh->p[0].key);
617 * Move from neh to sch any packets that have become eligible
619 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V)) {
620 struct dn_flow_queue *q = neh->p[0].object;
622 heap_extract(neh, NULL);
623 heap_insert(sch, q->F, q);
626 if (p->if_name[0] != '\0') { /* tx clock is from a real thing */
627 p->numbytes = -1; /* mark not ready for I/O */
632 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0 &&
633 p->idle_heap.elements > 0) {
635 * No traffic and no events scheduled. We can get rid of idle-heap.
639 for (i = 0; i < p->idle_heap.elements; i++) {
640 struct dn_flow_queue *q = p->idle_heap.p[i].object;
647 p->idle_heap.elements = 0;
651 * If we are getting clocks from dummynet (not a real interface) and
652 * If we are under credit, schedule the next ready event.
653 * Also fix the delivery time of the last packet.
655 if (p->if_name[0] == 0 && p->numbytes < 0) { /* This implies bandwidth>0 */
656 dn_key t = 0; /* Number of ticks i have to wait */
658 if (p->bandwidth > 0)
659 t = (p->bandwidth - 1 - p->numbytes) / p->bandwidth;
660 p->tail->output_time += t;
661 p->sched_time = curr_time;
664 * XXX should check errors on heap_insert, and drain the whole
665 * queue on error hoping next time we are luckier.
667 heap_insert(&wfq_ready_heap, curr_time + t, p);
671 * If the delay line was empty call transmit_event(p) now.
672 * Otherwise, the scheduler will take care of it.
679 * This is called once per tick, or dn_hz times per second. It is used to
680 * increment the current tick counter and schedule expired events.
683 dummynet(struct netmsg *msg)
687 struct dn_heap *heaps[3];
691 heaps[0] = &ready_heap; /* Fixed-rate queues */
692 heaps[1] = &wfq_ready_heap; /* WF2Q queues */
693 heaps[2] = &extract_heap; /* Delay line */
698 lwkt_replymsg(&msg->nm_lmsg, 0);
701 for (i = 0; i < 3; i++) {
703 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time)) {
704 DDB(if (h->p[0].key > curr_time)
705 kprintf("-- dummynet: warning, heap %d is %d ticks late\n",
706 i, (int)(curr_time - h->p[0].key));)
708 p = h->p[0].object; /* Store a copy before heap_extract */
709 heap_extract(h, NULL); /* Need to extract before processing */
714 struct dn_pipe *pipe = p;
716 if (pipe->if_name[0] != '\0') {
717 kprintf("*** bad ready_event_wfq for pipe %s\n",
729 * Sweep pipes trying to expire idle flow_queues
731 for (pe = all_pipes; pe; pe = pe->next) {
732 if (pe->idle_heap.elements > 0 &&
733 DN_KEY_LT(pe->idle_heap.p[0].key, pe->V)) {
734 struct dn_flow_queue *q = pe->idle_heap.p[0].object;
736 heap_extract(&pe->idle_heap, NULL);
737 q->S = q->F + 1; /* Mark timestamp as invalid */
738 pe->sum -= q->fs->weight;
746 * Unconditionally expire empty queues in case of shortage.
747 * Returns the number of queues freed.
750 expire_queues(struct dn_flow_set *fs)
752 struct dn_flow_queue *q, *prev;
753 int i, initial_elements = fs->rq_elements;
755 if (fs->last_expired == time_second)
758 fs->last_expired = time_second;
760 for (i = 0; i <= fs->rq_size; i++) { /* Last one is overflow */
761 for (prev = NULL, q = fs->rq[i]; q != NULL;) {
762 if (q->head != NULL || q->S != q->F + 1) {
765 } else { /* Entry is idle, expire it */
766 struct dn_flow_queue *old_q = q;
769 prev->next = q = q->next;
771 fs->rq[i] = q = q->next;
773 kfree(old_q, M_DUMMYNET);
777 return initial_elements - fs->rq_elements;
781 * If room, create a new queue and put at head of slot i;
782 * otherwise, create or use the default queue.
784 static struct dn_flow_queue *
785 create_queue(struct dn_flow_set *fs, int i)
787 struct dn_flow_queue *q;
789 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
790 expire_queues(fs) == 0) {
792 * No way to get room, use or create overflow queue.
795 if (fs->rq[i] != NULL)
799 q = kmalloc(sizeof(*q), M_DUMMYNET, M_INTWAIT | M_NULLOK | M_ZERO);
806 q->S = q->F + 1; /* hack - mark timestamp as invalid */
814 * Given a flow_set and a pkt in last_pkt, find a matching queue
815 * after appropriate masking. The queue is moved to front
816 * so that further searches take less time.
818 static struct dn_flow_queue *
819 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
821 struct dn_flow_queue *q, *prev;
824 if (!(fs->flags_fs & DN_HAVE_FLOW_MASK)) {
827 /* First, do the masking */
828 id->dst_ip &= fs->flow_mask.dst_ip;
829 id->src_ip &= fs->flow_mask.src_ip;
830 id->dst_port &= fs->flow_mask.dst_port;
831 id->src_port &= fs->flow_mask.src_port;
832 id->proto &= fs->flow_mask.proto;
833 id->flags = 0; /* we don't care about this one */
835 /* Then, hash function */
836 i = ((id->dst_ip) & 0xffff) ^
837 ((id->dst_ip >> 15) & 0xffff) ^
838 ((id->src_ip << 1) & 0xffff) ^
839 ((id->src_ip >> 16 ) & 0xffff) ^
840 (id->dst_port << 1) ^ (id->src_port) ^
844 /* Finally, scan the current list for a match */
846 for (prev = NULL, q = fs->rq[i]; q;) {
848 if (id->dst_ip == q->id.dst_ip &&
849 id->src_ip == q->id.src_ip &&
850 id->dst_port == q->id.dst_port &&
851 id->src_port == q->id.src_port &&
852 id->proto == q->id.proto &&
853 id->flags == q->id.flags) {
855 } else if (pipe_expire && q->head == NULL && q->S == q->F + 1) {
856 /* Entry is idle and not in any heap, expire it */
857 struct dn_flow_queue *old_q = q;
860 prev->next = q = q->next;
862 fs->rq[i] = q = q->next;
864 kfree(old_q, M_DUMMYNET);
870 if (q && prev != NULL) { /* Found and not in front */
871 prev->next = q->next;
876 if (q == NULL) { /* No match, need to allocate a new entry */
877 q = create_queue(fs, i);
885 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
890 * RED calculates the average queue size (avg) using a low-pass filter
891 * with an exponential weighted (w_q) moving average:
892 * avg <- (1-w_q) * avg + w_q * q_size
893 * where q_size is the queue length (measured in bytes or * packets).
895 * If q_size == 0, we compute the idle time for the link, and set
896 * avg = (1 - w_q)^(idle/s)
897 * where s is the time needed for transmitting a medium-sized packet.
899 * Now, if avg < min_th the packet is enqueued.
900 * If avg > max_th the packet is dropped. Otherwise, the packet is
901 * dropped with probability P function of avg.
905 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len;
907 DEB(kprintf("\n%d q: %2u ", (int) curr_time, q_size);)
909 /* Average queue size estimation */
912 * Queue is not empty, avg <- avg + (q_size - avg) * w_q
914 int diff = SCALE(q_size) - q->avg;
915 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);
920 * Queue is empty, find for how long the queue has been
921 * empty and use a lookup table for computing
922 * (1 - * w_q)^(idle_time/s) where s is the time to send a
927 u_int t = (curr_time - q->q_time) / fs->lookup_step;
929 q->avg = (t < fs->lookup_depth) ?
930 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
933 DEB(kprintf("avg: %u ", SCALE_VAL(q->avg));)
937 if (q->avg < fs->min_th) {
943 if (q->avg >= fs->max_th) { /* Average queue >= Max threshold */
944 if (fs->flags_fs & DN_IS_GENTLE_RED) {
946 * According to Gentle-RED, if avg is greater than max_th the
947 * packet is dropped with a probability
948 * p_b = c_3 * avg - c_4
949 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
951 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) - fs->c_4;
957 } else if (q->avg > fs->min_th) {
959 * We compute p_b using the linear dropping function p_b = c_1 *
960 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
961 * max_p * min_th / (max_th - min_th)
963 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
965 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
966 p_b = (p_b * len) / fs->max_pkt_size;
968 if (++q->count == 0) {
969 q->random = krandom() & 0xffff;
972 * q->count counts packets arrived since last drop, so a greater
973 * value of q->count means a greater packet drop probability.
975 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
977 DEB(kprintf("- red drop");)
978 /* After a drop we calculate a new random value */
979 q->random = krandom() & 0xffff;
983 /* End of RED algorithm */
984 return 0; /* Accept */
987 static __inline struct dn_flow_set *
988 locate_flowset(int pipe_nr, struct ip_fw *rule)
990 ipfw_insn *cmd = rule->cmd + rule->act_ofs;
991 struct dn_flow_set *fs;
993 if (cmd->opcode == O_LOG)
996 fs = ((ipfw_insn_pipe *)cmd)->pipe_ptr;
1000 if (cmd->opcode == O_QUEUE) {
1001 for (fs = all_flow_sets; fs && fs->fs_nr != pipe_nr; fs = fs->next)
1006 for (p = all_pipes; p && p->pipe_nr != pipe_nr; p = p->next)
1012 /* record for the future */
1013 ((ipfw_insn_pipe *)cmd)->pipe_ptr = fs;
1018 * Dummynet hook for packets. Below 'pipe' is a pipe or a queue
1019 * depending on whether WF2Q or fixed bw is used.
1021 * pipe_nr pipe or queue the packet is destined for.
1022 * dir where shall we send the packet after dummynet.
1023 * m the mbuf with the packet
1024 * fwa->oif the 'ifp' parameter from the caller.
1025 * NULL in ip_input, destination interface in ip_output
1026 * fwa->ro route parameter (only used in ip_output, NULL otherwise)
1027 * fwa->dst destination address, only used by ip_output
1028 * fwa->rule matching rule, in case of multiple passes
1029 * fwa->flags flags from the caller, only used in ip_output
1032 dummynet_io(struct mbuf *m, int pipe_nr, int dir, struct ip_fw_args *fwa)
1036 struct dn_flow_set *fs;
1037 struct dn_pipe *pipe;
1038 uint64_t len = m->m_pkthdr.len;
1039 struct dn_flow_queue *q = NULL;
1045 cmd = fwa->rule->cmd + fwa->rule->act_ofs;
1046 if (cmd->opcode == O_LOG)
1049 KASSERT(cmd->opcode == O_PIPE || cmd->opcode == O_QUEUE,
1050 ("Rule is not PIPE or QUEUE, opcode %d\n", cmd->opcode));
1052 is_pipe = (cmd->opcode == O_PIPE);
1056 * This is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule
1058 fs = locate_flowset(pipe_nr, fwa->rule);
1060 goto dropit; /* This queue/pipe does not exist! */
1063 if (pipe == NULL) { /* Must be a queue, try find a matching pipe */
1064 for (pipe = all_pipes; pipe && pipe->pipe_nr != fs->parent_nr;
1070 kprintf("No pipe %d for queue %d, drop pkt\n",
1071 fs->parent_nr, fs->fs_nr);
1076 q = find_queue(fs, &fwa->f_id);
1078 goto dropit; /* Cannot allocate queue */
1081 * Update statistics, then check reasons to drop pkt
1083 q->tot_bytes += len;
1086 if (fs->plr && krandom() < fs->plr)
1087 goto dropit; /* Random pkt drop */
1089 if (fs->flags_fs & DN_QSIZE_IS_BYTES) {
1090 if (q->len_bytes > fs->qsize)
1091 goto dropit; /* Queue size overflow */
1093 if (q->len >= fs->qsize)
1094 goto dropit; /* Queue count overflow */
1097 if ((fs->flags_fs & DN_IS_RED) && red_drops(fs, q, len))
1101 * Build and enqueue packet + parameters
1103 tag = m_tag_get(PACKET_TAG_DUMMYNET, sizeof(*pkt), MB_DONTWAIT /* XXX */);
1106 m_tag_prepend(m, tag);
1108 pkt = m_tag_data(tag);
1109 bzero(pkt, sizeof(*pkt)); /* XXX expensive to zero */
1111 pkt->rule = fwa->rule;
1112 pkt->dn_next = NULL;
1116 pkt->ifp = fwa->oif;
1117 if (dir == DN_TO_IP_OUT) {
1119 * We need to copy *ro because for ICMP pkts (and maybe others)
1120 * the caller passed a pointer into the stack; dst might also be
1121 * a pointer into *ro so it needs to be updated.
1123 pkt->ro = *(fwa->ro);
1125 fwa->ro->ro_rt->rt_refcnt++;
1126 if (fwa->dst == (struct sockaddr_in *)&fwa->ro->ro_dst) {
1127 /* 'dst' points into 'ro' */
1128 fwa->dst = (struct sockaddr_in *)&(pkt->ro.ro_dst);
1131 pkt->dn_dst = fwa->dst;
1132 pkt->flags = fwa->flags;
1134 if (q->head == NULL)
1137 q->tail->dn_next = pkt;
1140 q->len_bytes += len;
1142 if (q->head != pkt) /* Flow was not idle, we are done */
1146 * If we reach this point the flow was previously idle, so we need
1147 * to schedule it. This involves different actions for fixed-rate
1152 * Fixed-rate queue: just insert into the ready_heap.
1156 if (pipe->bandwidth)
1157 t = SET_TICKS(pkt, q, pipe);
1159 q->sched_time = curr_time;
1160 if (t == 0) /* Must process it now */
1163 heap_insert(&ready_heap, curr_time + t, q);
1167 * First, compute start time S: if the flow was idle (S=F+1)
1168 * set S to the virtual time V for the controlling pipe, and update
1169 * the sum of weights for the pipe; otherwise, remove flow from
1170 * idle_heap and set S to max(F, V).
1171 * Second, compute finish time F = S + len/weight.
1172 * Third, if pipe was idle, update V = max(S, V).
1173 * Fourth, count one more backlogged flow.
1175 if (DN_KEY_GT(q->S, q->F)) { /* Means timestamps are invalid */
1177 pipe->sum += fs->weight; /* Add weight of new queue */
1179 heap_extract(&pipe->idle_heap, q);
1180 q->S = MAX64(q->F, pipe->V);
1182 q->F = q->S + (len << MY_M) / (uint64_t)fs->weight;
1184 if (pipe->not_eligible_heap.elements == 0 &&
1185 pipe->scheduler_heap.elements == 0)
1186 pipe->V = MAX64(q->S, pipe->V);
1191 * Look at eligibility. A flow is not eligibile if S>V (when
1192 * this happens, it means that there is some other flow already
1193 * scheduled for the same pipe, so the scheduler_heap cannot be
1194 * empty). If the flow is not eligible we just store it in the
1195 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1196 * and possibly invoke ready_event_wfq() right now if there is
1198 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1199 * and for all flows in not_eligible_heap (NEH), S_i > V.
1200 * So when we need to compute max(V, min(S_i)) forall i in SCH+NEH,
1201 * we only need to look into NEH.
1203 if (DN_KEY_GT(q->S, pipe->V)) { /* Not eligible */
1204 if (pipe->scheduler_heap.elements == 0)
1205 kprintf("++ ouch! not eligible but empty scheduler!\n");
1206 heap_insert(&pipe->not_eligible_heap, q->S, q);
1208 heap_insert(&pipe->scheduler_heap, q->F, q);
1209 if (pipe->numbytes >= 0) { /* Pipe is idle */
1210 if (pipe->scheduler_heap.elements != 1)
1211 kprintf("*** OUCH! pipe should have been idle!\n");
1212 DEB(kprintf("Waking up pipe %d at %d\n",
1213 pipe->pipe_nr, (int)(q->F >> MY_M)); )
1214 pipe->sched_time = curr_time;
1215 ready_event_wfq(pipe);
1228 return ((fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1232 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1233 * Doing this would probably save us the initial bzero of dn_pkt
1235 #define DN_FREE_PKT(pkt) { \
1236 struct dn_pkt *n = pkt; \
1238 rt_unref (n->ro.ro_rt); \
1242 * Dispose all packets and flow_queues on a flow_set.
1243 * If all=1, also remove red lookup table and other storage,
1244 * including the descriptor itself.
1245 * For the one in dn_pipe MUST also cleanup ready_heap...
1248 purge_flow_set(struct dn_flow_set *fs, int all)
1252 for (i = 0; i <= fs->rq_size; i++) {
1253 struct dn_flow_queue *q, *qn;
1255 for (q = fs->rq[i]; q; q = qn) {
1258 for (pkt = q->head; pkt;)
1262 kfree(q, M_DUMMYNET);
1266 fs->rq_elements = 0;
1269 /* RED - free lookup table */
1271 kfree(fs->w_q_lookup, M_DUMMYNET);
1274 kfree(fs->rq, M_DUMMYNET);
1276 /* If this fs is not part of a pipe, free it */
1277 if (fs->pipe && fs != &fs->pipe->fs)
1278 kfree(fs, M_DUMMYNET);
1283 * Dispose all packets queued on a pipe (not a flow_set).
1284 * Also free all resources associated to a pipe, which is about
1288 purge_pipe(struct dn_pipe *pipe)
1292 purge_flow_set(&pipe->fs, 1);
1294 for (pkt = pipe->head; pkt;)
1297 heap_free(&pipe->scheduler_heap);
1298 heap_free(&pipe->not_eligible_heap);
1299 heap_free(&pipe->idle_heap);
1303 * Delete all pipes and heaps returning memory. Must also
1304 * remove references from all ipfw rules to all pipes.
1307 dummynet_flush(void)
1310 struct dn_flow_set *fs;
1314 /* Remove all references to pipes ... */
1315 flush_pipe_ptrs(NULL);
1317 /* Prevent future matches... */
1321 all_flow_sets = NULL;
1323 /* Free heaps so we don't have unwanted events */
1324 heap_free(&ready_heap);
1325 heap_free(&wfq_ready_heap);
1326 heap_free(&extract_heap);
1331 * Now purge all queued pkts and delete all pipes
1333 /* Scan and purge all flow_sets. */
1334 while (fs != NULL) {
1335 struct dn_flow_set *curr_fs = fs;
1338 purge_flow_set(curr_fs, 1);
1341 struct dn_pipe *curr_p = p;
1345 kfree(curr_p, M_DUMMYNET);
1350 extern struct ip_fw *ip_fw_default_rule;
1353 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1357 for (i = 0; i <= fs->rq_size; i++) { /* Last one is ovflow */
1358 struct dn_flow_queue *q;
1360 for (q = fs->rq[i]; q; q = q->next) {
1363 for (pkt = q->head; pkt; pkt = pkt->dn_next) {
1365 pkt->rule = ip_fw_default_rule;
1372 * When a firewall rule is deleted, scan all queues and remove the flow-id
1373 * from packets matching this rule.
1376 dn_rule_delete(void *r)
1379 struct dn_flow_set *fs;
1382 * If the rule references a queue (dn_flow_set), then scan
1383 * the flow set, otherwise scan pipes. Should do either, but doing
1384 * both does not harm.
1387 for (fs = all_flow_sets; fs; fs = fs->next)
1388 dn_rule_delete_fs(fs, r);
1390 for (p = all_pipes; p; p = p->next) {
1394 dn_rule_delete_fs(fs, r);
1396 for (pkt = p->head; pkt; pkt = pkt->dn_next) {
1398 pkt->rule = ip_fw_default_rule;
1404 * setup RED parameters
1407 config_red(struct dn_flow_set *p, struct dn_flow_set *x)
1412 x->min_th = SCALE(p->min_th);
1413 x->max_th = SCALE(p->max_th);
1414 x->max_p = p->max_p;
1416 x->c_1 = p->max_p / (p->max_th - p->min_th);
1417 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1418 if (x->flags_fs & DN_IS_GENTLE_RED) {
1419 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1420 x->c_4 = (SCALE(1) - 2 * p->max_p);
1423 /* If the lookup table already exist, free and create it again */
1424 if (x->w_q_lookup) {
1425 kfree(x->w_q_lookup, M_DUMMYNET);
1426 x->w_q_lookup = NULL ;
1429 if (red_lookup_depth == 0) {
1430 kprintf("net.inet.ip.dummynet.red_lookup_depth must be > 0\n");
1431 kfree(x, M_DUMMYNET);
1434 x->lookup_depth = red_lookup_depth;
1435 x->w_q_lookup = kmalloc(x->lookup_depth * sizeof(int),
1436 M_DUMMYNET, M_WAITOK);
1438 /* Fill the lookup table with (1 - w_q)^x */
1439 x->lookup_step = p->lookup_step;
1440 x->lookup_weight = p->lookup_weight;
1442 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1443 for (i = 1; i < x->lookup_depth; i++)
1444 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1446 if (red_avg_pkt_size < 1)
1447 red_avg_pkt_size = 512;
1448 x->avg_pkt_size = red_avg_pkt_size;
1450 if (red_max_pkt_size < 1)
1451 red_max_pkt_size = 1500;
1452 x->max_pkt_size = red_max_pkt_size;
1458 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1460 if (x->flags_fs & DN_HAVE_FLOW_MASK) {
1461 int l = pfs->rq_size;
1463 /* Allocate some slots */
1467 if (l < DN_MIN_HASH_SIZE)
1468 l = DN_MIN_HASH_SIZE;
1469 else if (l > DN_MAX_HASH_SIZE)
1470 l = DN_MAX_HASH_SIZE;
1474 /* One is enough for null mask */
1477 x->rq = kmalloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1478 M_DUMMYNET, M_WAITOK | M_ZERO);
1483 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1485 x->flags_fs = src->flags_fs;
1486 x->qsize = src->qsize;
1488 x->flow_mask = src->flow_mask;
1489 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1490 if (x->qsize > 1024 * 1024)
1491 x->qsize = 1024 * 1024;
1493 if (x->qsize == 0 || x->qsize > 100)
1497 /* Configuring RED */
1498 if (x->flags_fs & DN_IS_RED)
1499 config_red(src, x); /* XXX should check errors */
1503 * setup pipe or queue parameters.
1507 config_pipe(struct dn_pipe *p)
1509 struct dn_flow_set *pfs = &p->fs;
1513 * The config program passes parameters as follows:
1514 * bw bits/second (0 means no limits)
1515 * delay ms (must be translated into ticks)
1516 * qsize slots or bytes
1518 p->delay = (p->delay * dn_hz) / 1000;
1521 * We need either a pipe number or a flow_set number
1523 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1525 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1531 if (p->pipe_nr != 0) { /* This is a pipe */
1532 struct dn_pipe *x, *a, *b;
1535 for (a = NULL, b = all_pipes; b && b->pipe_nr < p->pipe_nr;
1539 if (b == NULL || b->pipe_nr != p->pipe_nr) { /* New pipe */
1540 x = kmalloc(sizeof(struct dn_pipe), M_DUMMYNET, M_WAITOK | M_ZERO);
1541 x->pipe_nr = p->pipe_nr;
1545 * idle_heap is the only one from which we extract from the middle.
1547 x->idle_heap.size = x->idle_heap.elements = 0;
1548 x->idle_heap.offset = __offsetof(struct dn_flow_queue, heap_pos);
1554 /* Flush accumulated credit for all queues */
1555 for (i = 0; i <= x->fs.rq_size; i++) {
1556 struct dn_flow_queue *q;
1558 for (q = x->fs.rq[i]; q; q = q->next)
1563 x->bandwidth = p->bandwidth;
1564 x->numbytes = 0; /* Just in case... */
1565 bcopy(p->if_name, x->if_name, sizeof(x->if_name));
1566 x->ifp = NULL; /* Reset interface ptr */
1567 x->delay = p->delay;
1569 set_fs_parms(&x->fs, pfs);
1571 if (x->fs.rq == NULL) { /* A new pipe */
1572 alloc_hash(&x->fs, pfs);
1580 } else { /* Config flow_set */
1581 struct dn_flow_set *x, *a, *b;
1583 /* Locate flow_set */
1584 for (a = NULL, b = all_flow_sets; b && b->fs_nr < pfs->fs_nr;
1588 if (b == NULL || b->fs_nr != pfs->fs_nr) { /* New flow_set */
1589 if (pfs->parent_nr == 0) /* Need link to a pipe */
1592 x = kmalloc(sizeof(struct dn_flow_set), M_DUMMYNET,
1594 x->fs_nr = pfs->fs_nr;
1595 x->parent_nr = pfs->parent_nr;
1596 x->weight = pfs->weight;
1599 else if (x->weight > 100)
1602 /* Change parent pipe not allowed; must delete and recreate */
1603 if (pfs->parent_nr != 0 && b->parent_nr != pfs->parent_nr)
1608 set_fs_parms(x, pfs);
1610 if (x->rq == NULL) { /* A new flow_set */
1628 * Helper function to remove from a heap queues which are linked to
1629 * a flow_set about to be deleted.
1632 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1634 int i = 0, found = 0;
1636 while (i < h->elements) {
1637 if (((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1639 h->p[i] = h->p[h->elements];
1650 * helper function to remove a pipe from a heap (can be there at most once)
1653 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1655 if (h->elements > 0) {
1658 for (i = 0; i < h->elements; i++) {
1659 if (h->p[i].object == p) { /* found it */
1661 h->p[i] = h->p[h->elements];
1670 * drain all queues. Called in case of severe mbuf shortage.
1673 dummynet_drain(void)
1675 struct dn_flow_set *fs;
1679 heap_free(&ready_heap);
1680 heap_free(&wfq_ready_heap);
1681 heap_free(&extract_heap);
1683 /* remove all references to this pipe from flow_sets */
1684 for (fs = all_flow_sets; fs; fs= fs->next)
1685 purge_flow_set(fs, 0);
1687 for (p = all_pipes; p; p= p->next) {
1688 purge_flow_set(&p->fs, 0);
1689 for (pkt = p->head; pkt ;)
1691 p->head = p->tail = NULL;
1696 * Fully delete a pipe or a queue, cleaning up associated info.
1699 delete_pipe(struct dn_pipe *p)
1703 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1705 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1711 if (p->pipe_nr != 0) { /* This is an old-style pipe */
1712 struct dn_pipe *a, *b;
1713 struct dn_flow_set *fs;
1716 for (a = NULL, b = all_pipes; b && b->pipe_nr < p->pipe_nr;
1719 if (b == NULL || b->pipe_nr != p->pipe_nr)
1720 goto back; /* Not found */
1722 /* Unlink from list of pipes */
1724 all_pipes = b->next;
1728 /* Remove references to this pipe from the ip_fw rules. */
1729 flush_pipe_ptrs(&b->fs);
1731 /* Remove all references to this pipe from flow_sets */
1732 for (fs = all_flow_sets; fs; fs = fs->next) {
1733 if (fs->pipe == b) {
1734 kprintf("++ ref to pipe %d from fs %d\n",
1735 p->pipe_nr, fs->fs_nr);
1737 purge_flow_set(fs, 0);
1740 fs_remove_from_heap(&ready_heap, &b->fs);
1741 purge_pipe(b); /* Remove all data associated to this pipe */
1743 /* Remove reference to here from extract_heap and wfq_ready_heap */
1744 pipe_remove_from_heap(&extract_heap, b);
1745 pipe_remove_from_heap(&wfq_ready_heap, b);
1747 kfree(b, M_DUMMYNET);
1748 } else { /* This is a WF2Q queue (dn_flow_set) */
1749 struct dn_flow_set *a, *b;
1751 /* Locate flow_set */
1752 for (a = NULL, b = all_flow_sets; b && b->fs_nr < p->fs.fs_nr;
1755 if (b == NULL || b->fs_nr != p->fs.fs_nr)
1756 goto back; /* Not found */
1759 all_flow_sets = b->next;
1763 /* Remove references to this flow_set from the ip_fw rules. */
1766 if (b->pipe != NULL) {
1767 /* Update total weight on parent pipe and cleanup parent heaps */
1768 b->pipe->sum -= b->weight * b->backlogged;
1769 fs_remove_from_heap(&b->pipe->not_eligible_heap, b);
1770 fs_remove_from_heap(&b->pipe->scheduler_heap, b);
1771 #if 1 /* XXX should i remove from idle_heap as well ? */
1772 fs_remove_from_heap(&b->pipe->idle_heap, b);
1775 purge_flow_set(b, 1);
1785 * helper function used to copy data from kernel in DUMMYNET_GET
1788 dn_copy_set(struct dn_flow_set *set, char *bp)
1791 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1793 for (i = 0; i <= set->rq_size; i++) {
1794 for (q = set->rq[i]; q; q = q->next, qp++) {
1795 if (q->hash_slot != i) { /* XXX ASSERT */
1796 kprintf("++ at %d: wrong slot (have %d, "
1797 "should be %d)\n", copied, q->hash_slot, i);
1800 if (q->fs != set) { /* XXX ASSERT */
1801 kprintf("++ at %d: wrong fs ptr (have %p, should be %p)\n",
1806 bcopy(q, qp, sizeof(*q));
1808 /* cleanup pointers */
1810 qp->head = qp->tail = NULL;
1815 if (copied != set->rq_elements) { /* XXX ASSERT */
1816 kprintf("++ wrong count, have %d should be %d\n",
1817 copied, set->rq_elements);
1823 dummynet_get(struct sockopt *sopt)
1827 struct dn_flow_set *set;
1834 * Compute size of data structures: list of pipes and flow_sets.
1836 for (p = all_pipes, size = 0; p; p = p->next) {
1837 size += sizeof(*p) +
1838 p->fs.rq_elements * sizeof(struct dn_flow_queue);
1841 for (set = all_flow_sets; set; set = set->next) {
1842 size += sizeof(*set) +
1843 set->rq_elements * sizeof(struct dn_flow_queue);
1846 buf = kmalloc(size, M_TEMP, M_WAITOK);
1847 for (p = all_pipes, bp = buf; p; p = p->next) {
1848 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp;
1851 * Copy pipe descriptor into *bp, convert delay back to ms,
1852 * then copy the flow_set descriptor(s) one at a time.
1853 * After each flow_set, copy the queue descriptor it owns.
1855 bcopy(p, bp, sizeof(*p));
1856 pipe_bp->delay = (pipe_bp->delay * 1000) / dn_hz;
1859 * XXX the following is a hack based on ->next being the
1860 * first field in dn_pipe and dn_flow_set. The correct
1861 * solution would be to move the dn_flow_set to the beginning
1862 * of struct dn_pipe.
1864 pipe_bp->next = (struct dn_pipe *)DN_IS_PIPE;
1866 /* Clean pointers */
1867 pipe_bp->head = pipe_bp->tail = NULL;
1868 pipe_bp->fs.next = NULL;
1869 pipe_bp->fs.pipe = NULL;
1870 pipe_bp->fs.rq = NULL;
1873 bp = dn_copy_set(&p->fs, bp);
1876 for (set = all_flow_sets; set; set = set->next) {
1877 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp;
1879 bcopy(set, bp, sizeof(*set));
1881 /* XXX same hack as above */
1882 fs_bp->next = (struct dn_flow_set *)DN_IS_QUEUE;
1887 bp = dn_copy_set(set, bp);
1892 error = sooptcopyout(sopt, buf, size);
1898 * Handler for the various dummynet socket options (get, flush, config, del)
1901 ip_dn_ctl(struct sockopt *sopt)
1903 struct dn_pipe *p, tmp_pipe;
1906 /* Disallow sets in really-really secure mode. */
1907 if (sopt->sopt_dir == SOPT_SET) {
1908 if (securelevel >= 3)
1912 switch (sopt->sopt_name) {
1913 case IP_DUMMYNET_GET:
1914 error = dummynet_get(sopt);
1917 case IP_DUMMYNET_FLUSH:
1921 case IP_DUMMYNET_CONFIGURE:
1923 error = sooptcopyin(sopt, p, sizeof(*p), sizeof(*p));
1926 error = config_pipe(p);
1929 case IP_DUMMYNET_DEL: /* Remove a pipe or flow_set */
1931 error = sooptcopyin(sopt, p, sizeof(*p), sizeof(*p));
1934 error = delete_pipe(p);
1938 kprintf("%s -- unknown option %d\n", __func__, sopt->sopt_name);
1946 dummynet_clock(systimer_t info __unused, struct intrframe *frame __unused)
1948 KASSERT(mycpu->gd_cpuid == dn_cpu,
1949 ("systimer comes on a different cpu!\n"));
1952 if (dn_netmsg.nm_lmsg.ms_flags & MSGF_DONE)
1953 lwkt_sendmsg(cpu_portfn(mycpu->gd_cpuid), &dn_netmsg.nm_lmsg);
1958 sysctl_dn_hz(SYSCTL_HANDLER_ARGS)
1963 error = sysctl_handle_int(oidp, &val, 0, req);
1964 if (error || req->newptr == NULL)
1968 else if (val > DUMMYNET_CALLOUT_FREQ_MAX)
1969 val = DUMMYNET_CALLOUT_FREQ_MAX;
1973 systimer_adjust_periodic(&dn_clock, val);
1980 ip_dn_register_systimer(struct netmsg *msg)
1982 systimer_init_periodic_nq(&dn_clock, dummynet_clock, NULL, dn_hz);
1983 lwkt_replymsg(&msg->nm_lmsg, 0);
1987 ip_dn_deregister_systimer(struct netmsg *msg)
1989 systimer_del(&dn_clock);
1990 lwkt_replymsg(&msg->nm_lmsg, 0);
1999 kprintf("DUMMYNET initialized (011031)\n");
2002 all_flow_sets = NULL;
2004 ready_heap.size = ready_heap.elements = 0;
2005 ready_heap.offset = 0;
2007 wfq_ready_heap.size = wfq_ready_heap.elements = 0;
2008 wfq_ready_heap.offset = 0;
2010 extract_heap.size = extract_heap.elements = 0;
2011 extract_heap.offset = 0;
2013 ip_dn_ctl_ptr = ip_dn_ctl;
2014 ip_dn_io_ptr = dummynet_io;
2015 ip_dn_ruledel_ptr = dn_rule_delete;
2017 netmsg_init(&dn_netmsg, &netisr_adone_rport, 0, dummynet);
2019 netmsg_init(&smsg, &curthread->td_msgport, 0, ip_dn_register_systimer);
2020 port = cpu_portfn(dn_cpu);
2021 lwkt_domsg(port, &smsg.nm_lmsg, 0);
2030 netmsg_init(&smsg, &curthread->td_msgport, 0, ip_dn_deregister_systimer);
2031 port = cpu_portfn(dn_cpu);
2032 lwkt_domsg(port, &smsg.nm_lmsg, 0);
2036 ip_dn_ctl_ptr = NULL;
2037 ip_dn_io_ptr = NULL;
2038 ip_dn_ruledel_ptr = NULL;
2040 netmsg_service_sync();
2044 dummynet_modevent(module_t mod, int type, void *data)
2049 if (DUMMYNET_LOADED) {
2051 kprintf("DUMMYNET already loaded\n");
2060 kprintf("dummynet statically compiled, cannot unload\n");
2075 static moduledata_t dummynet_mod = {
2080 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PROTO_END, SI_ORDER_ANY);
2081 MODULE_DEPEND(dummynet, ipfw, 1, 1, 1);
2082 MODULE_VERSION(dummynet, 1);