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 $
33 #error IPFIREWALL3 requires INET.
37 * This module implements IP dummynet, a bandwidth limiter/delay emulator.
38 * Description of the data structures used is in ip_dummynet.h
39 * Here you mainly find the following blocks of code:
40 * + variable declarations;
41 * + heap management functions;
42 * + scheduler and dummynet functions;
43 * + configuration and initialization.
45 * Most important Changes:
48 * 010124: Fixed WF2Q behaviour
49 * 010122: Fixed spl protection.
50 * 000601: WF2Q support
51 * 000106: Large rewrite, use heaps to handle very many pipes.
52 * 980513: Initial release
55 #include <sys/param.h>
56 #include <sys/kernel.h>
57 #include <sys/malloc.h>
59 #include <sys/socketvar.h>
60 #include <sys/sysctl.h>
61 #include <sys/systimer.h>
62 #include <sys/thread2.h>
64 #include <net/ethernet.h>
65 #include <net/netmsg2.h>
66 #include <net/netisr2.h>
67 #include <net/route.h>
69 #include <netinet/in_var.h>
70 #include <netinet/ip_var.h>
72 #include <net/dummynet3/ip_dummynet3.h>
73 #include <net/ipfw3/ip_fw.h>
75 void check_pipe(int *cmd_ctl, int *cmd_val, struct ip_fw_args **args,
76 struct ip_fw **f, ipfw_insn *cmd, uint16_t ip_len);
77 void check_queue(int *cmd_ctl, int *cmd_val, struct ip_fw_args **args,
78 struct ip_fw **f, ipfw_insn *cmd, uint16_t ip_len);
81 check_pipe(int *cmd_ctl, int *cmd_val, struct ip_fw_args **args,
82 struct ip_fw **f, ipfw_insn *cmd, uint16_t ip_len)
85 (*args)->cookie = cmd->arg1;
86 *cmd_val = IP_FW_DUMMYNET;
87 *cmd_ctl = IP_FW_CTL_DONE;
91 check_queue(int *cmd_ctl, int *cmd_val, struct ip_fw_args **args,
92 struct ip_fw **f, ipfw_insn *cmd, uint16_t ip_len)
95 (*args)->cookie = cmd->arg1;
96 *cmd_val = IP_FW_DUMMYNET;
97 *cmd_ctl = IP_FW_CTL_DONE;
100 #ifdef DUMMYNET_DEBUG
101 #define DPRINTF(fmt, ...) kprintf(fmt, __VA_ARGS__)
103 #define DPRINTF(fmt, ...) ((void)0)
106 #ifndef DN_CALLOUT_FREQ_MAX
107 #define DN_CALLOUT_FREQ_MAX 10000
111 * The maximum/minimum hash table size for queues.
112 * These values must be a power of 2.
114 #define DN_MIN_HASH_SIZE 4
115 #define DN_MAX_HASH_SIZE 65536
118 * Some macros are used to compare key values and handle wraparounds.
119 * MAX64 returns the largest of two key values.
121 #define DN_KEY_LT(a, b) ((int64_t)((a) - (b)) < 0)
122 #define DN_KEY_LEQ(a, b) ((int64_t)((a) - (b)) <= 0)
123 #define DN_KEY_GT(a, b) ((int64_t)((a) - (b)) > 0)
124 #define DN_KEY_GEQ(a, b) ((int64_t)((a) - (b)) >= 0)
125 #define MAX64(x, y) ((((int64_t)((y) - (x))) > 0) ? (y) : (x))
127 #define DN_NR_HASH_MAX 16
128 #define DN_NR_HASH_MASK (DN_NR_HASH_MAX - 1)
129 #define DN_NR_HASH(nr) \
130 ((((nr) >> 12) ^ ((nr) >> 8) ^ ((nr) >> 4) ^ (nr)) & DN_NR_HASH_MASK)
132 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
134 extern int ip_dn_cpu;
136 static dn_key curr_time = 0; /* current simulation time */
137 static int dn_hash_size = 64; /* default hash size */
138 static int pipe_expire = 1; /* expire queue if empty */
139 static int dn_max_ratio = 16; /* max queues/buckets ratio */
142 * Statistics on number of queue searches and search steps
145 static int search_steps;
150 static int red_lookup_depth = 256; /* default lookup table depth */
151 static int red_avg_pkt_size = 512; /* default medium packet size */
152 static int red_max_pkt_size = 1500;/* default max packet size */
155 * Three heaps contain queues and pipes that the scheduler handles:
157 * + ready_heap contains all dn_flow_queue related to fixed-rate pipes.
158 * + wfq_ready_heap contains the pipes associated with WF2Q flows
159 * + extract_heap contains pipes associated with delay lines.
161 static struct dn_heap ready_heap;
162 static struct dn_heap extract_heap;
163 static struct dn_heap wfq_ready_heap;
165 static struct dn_pipe_head pipe_table[DN_NR_HASH_MAX];
166 static struct dn_flowset_head flowset_table[DN_NR_HASH_MAX];
169 * Variables for dummynet systimer
171 static struct netmsg_base dn_netmsg;
172 static struct systimer dn_clock;
173 static int dn_hz = 1000;
175 static int sysctl_dn_hz(SYSCTL_HANDLER_ARGS);
177 SYSCTL_DECL(_net_inet_ip_dummynet);
179 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size, CTLFLAG_RW,
180 &dn_hash_size, 0, "Default hash table size");
181 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time, CTLFLAG_RD,
182 &curr_time, 0, "Current tick");
183 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire, CTLFLAG_RW,
184 &pipe_expire, 0, "Expire queue if empty");
185 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len, CTLFLAG_RW,
186 &dn_max_ratio, 0, "Max ratio between dynamic queues and buckets");
188 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap, CTLFLAG_RD,
189 &ready_heap.size, 0, "Size of ready heap");
190 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap, CTLFLAG_RD,
191 &extract_heap.size, 0, "Size of extract heap");
193 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches, CTLFLAG_RD,
194 &searches, 0, "Number of queue searches");
195 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps, CTLFLAG_RD,
196 &search_steps, 0, "Number of queue search steps");
198 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth, CTLFLAG_RD,
199 &red_lookup_depth, 0, "Depth of RED lookup table");
200 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size, CTLFLAG_RD,
201 &red_avg_pkt_size, 0, "RED Medium packet size");
202 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size, CTLFLAG_RD,
203 &red_max_pkt_size, 0, "RED Max packet size");
205 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hz, CTLTYPE_INT | CTLFLAG_RW,
206 0, 0, sysctl_dn_hz, "I", "Dummynet callout frequency");
208 static int heap_init(struct dn_heap *, int);
209 static int heap_insert(struct dn_heap *, dn_key, void *);
210 static void heap_extract(struct dn_heap *, void *);
212 static void transmit_event(struct dn_pipe *);
213 static void ready_event(struct dn_flow_queue *);
214 static void ready_event_wfq(struct dn_pipe *);
216 static int config_pipe(struct dn_ioc_pipe *);
217 static void dummynet_flush(void);
219 static void dummynet_clock(systimer_t, int, struct intrframe *);
220 static void dummynet(netmsg_t);
222 static struct dn_pipe *dn_find_pipe(int);
223 static struct dn_flow_set *dn_locate_flowset(int, int);
225 typedef void (*dn_pipe_iter_t)(struct dn_pipe *, void *);
226 static void dn_iterate_pipe(dn_pipe_iter_t, void *);
228 typedef void (*dn_flowset_iter_t)(struct dn_flow_set *, void *);
229 static void dn_iterate_flowset(dn_flowset_iter_t, void *);
231 static ip_dn_io_t dummynet_io;
232 static ip_dn_ctl_t dummynet_ctl;
235 * Heap management functions.
237 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
238 * Some macros help finding parent/children so we can optimize them.
240 * heap_init() is called to expand the heap when needed.
241 * Increment size in blocks of 16 entries.
242 * XXX failure to allocate a new element is a pretty bad failure
243 * as we basically stall a whole queue forever!!
244 * Returns 1 on error, 0 on success
246 #define HEAP_FATHER(x) (((x) - 1) / 2)
247 #define HEAP_LEFT(x) (2*(x) + 1)
248 #define HEAP_IS_LEFT(x) ((x) & 1)
249 #define HEAP_RIGHT(x) (2*(x) + 2)
250 #define HEAP_SWAP(a, b, buffer) { buffer = a; a = b; b = buffer; }
251 #define HEAP_INCREMENT 15
254 heap_init(struct dn_heap *h, int new_size)
256 struct dn_heap_entry *p;
258 if (h->size >= new_size) {
259 kprintf("%s, Bogus call, have %d want %d\n", __func__,
264 new_size = (new_size + HEAP_INCREMENT) & ~HEAP_INCREMENT;
265 p = kmalloc(new_size * sizeof(*p), M_DUMMYNET, M_WAITOK | M_ZERO);
267 bcopy(h->p, p, h->size * sizeof(*p));
268 kfree(h->p, M_DUMMYNET);
276 * Insert element in heap. Normally, p != NULL, we insert p in
277 * a new position and bubble up. If p == NULL, then the element is
278 * already in place, and key is the position where to start the
280 * Returns 1 on failure (cannot allocate new heap entry)
282 * If offset > 0 the position (index, int) of the element in the heap is
283 * also stored in the element itself at the given offset in bytes.
285 #define SET_OFFSET(heap, node) \
286 if (heap->offset > 0) \
287 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node;
290 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
292 #define RESET_OFFSET(heap, node) \
293 if (heap->offset > 0) \
294 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1;
297 heap_insert(struct dn_heap *h, dn_key key1, void *p)
301 if (p == NULL) { /* Data already there, set starting point */
303 } else { /* Insert new element at the end, possibly resize */
305 if (son == h->size) { /* Need resize... */
306 if (heap_init(h, h->elements + 1))
307 return 1; /* Failure... */
309 h->p[son].object = p;
310 h->p[son].key = key1;
314 while (son > 0) { /* Bubble up */
315 int father = HEAP_FATHER(son);
316 struct dn_heap_entry tmp;
318 if (DN_KEY_LT(h->p[father].key, h->p[son].key))
319 break; /* Found right position */
321 /* 'son' smaller than 'father', swap and repeat */
322 HEAP_SWAP(h->p[son], h->p[father], tmp);
331 * Remove top element from heap, or obj if obj != NULL
334 heap_extract(struct dn_heap *h, void *obj)
336 int child, father, max = h->elements - 1;
339 kprintf("warning, extract from empty heap 0x%p\n", h);
343 father = 0; /* Default: move up smallest child */
344 if (obj != NULL) { /* Extract specific element, index is at offset */
346 panic("%s from middle not supported on this heap!!!", __func__);
348 father = *((int *)((char *)obj + h->offset));
349 if (father < 0 || father >= h->elements) {
350 panic("%s father %d out of bound 0..%d", __func__,
351 father, h->elements);
354 RESET_OFFSET(h, father);
356 child = HEAP_LEFT(father); /* Left child */
357 while (child <= max) { /* Valid entry */
358 if (child != max && DN_KEY_LT(h->p[child + 1].key, h->p[child].key))
359 child = child + 1; /* Take right child, otherwise left */
360 h->p[father] = h->p[child];
361 SET_OFFSET(h, father);
363 child = HEAP_LEFT(child); /* Left child for next loop */
368 * Fill hole with last entry and bubble up, reusing the insert code
370 h->p[father] = h->p[max];
371 heap_insert(h, father, NULL); /* This one cannot fail */
376 * heapify() will reorganize data inside an array to maintain the
377 * heap property. It is needed when we delete a bunch of entries.
380 heapify(struct dn_heap *h)
384 for (i = 0; i < h->elements; i++)
385 heap_insert(h, i , NULL);
389 * Cleanup the heap and free data structure
392 heap_free(struct dn_heap *h)
395 kfree(h->p, M_DUMMYNET);
396 bzero(h, sizeof(*h));
400 * --- End of heap management functions ---
404 * Scheduler functions:
406 * transmit_event() is called when the delay-line needs to enter
407 * the scheduler, either because of existing pkts getting ready,
408 * or new packets entering the queue. The event handled is the delivery
409 * time of the packet.
411 * ready_event() does something similar with fixed-rate queues, and the
412 * event handled is the finish time of the head pkt.
414 * ready_event_wfq() does something similar with WF2Q queues, and the
415 * event handled is the start time of the head pkt.
417 * In all cases, we make sure that the data structures are consistent
418 * before passing pkts out, because this might trigger recursive
419 * invocations of the procedures.
422 transmit_event(struct dn_pipe *pipe)
426 while ((pkt = TAILQ_FIRST(&pipe->p_queue)) &&
427 DN_KEY_LEQ(pkt->output_time, curr_time)) {
428 TAILQ_REMOVE(&pipe->p_queue, pkt, dn_next);
429 ip_dn_packet_redispatch(pkt);
433 * If there are leftover packets, put into the heap for next event
435 if ((pkt = TAILQ_FIRST(&pipe->p_queue)) != NULL) {
437 * XXX should check errors on heap_insert, by draining the
438 * whole pipe and hoping in the future we are more successful
440 heap_insert(&extract_heap, pkt->output_time, pipe);
445 * The following macro computes how many ticks we have to wait
446 * before being able to transmit a packet. The credit is taken from
447 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
449 #define SET_TICKS(pkt, q, p) \
450 (pkt->dn_m->m_pkthdr.len*8*dn_hz - (q)->numbytes + p->bandwidth - 1 ) / \
454 * Extract pkt from queue, compute output time (could be now)
455 * and put into delay line (p_queue)
458 move_pkt(struct dn_pkt *pkt, struct dn_flow_queue *q,
459 struct dn_pipe *p, int len)
461 TAILQ_REMOVE(&q->queue, pkt, dn_next);
465 pkt->output_time = curr_time + p->delay;
467 TAILQ_INSERT_TAIL(&p->p_queue, pkt, dn_next);
471 * ready_event() is invoked every time the queue must enter the
472 * scheduler, either because the first packet arrives, or because
473 * a previously scheduled event fired.
474 * On invokation, drain as many pkts as possible (could be 0) and then
475 * if there are leftover packets reinsert the pkt in the scheduler.
478 ready_event(struct dn_flow_queue *q)
481 struct dn_pipe *p = q->fs->pipe;
485 kprintf("ready_event- pipe is gone\n");
488 p_was_empty = TAILQ_EMPTY(&p->p_queue);
491 * Schedule fixed-rate queues linked to this pipe:
492 * Account for the bw accumulated since last scheduling, then
493 * drain as many pkts as allowed by q->numbytes and move to
494 * the delay line (in p) computing output time.
495 * bandwidth==0 (no limit) means we can drain the whole queue,
496 * setting len_scaled = 0 does the job.
498 q->numbytes += (curr_time - q->sched_time) * p->bandwidth;
499 while ((pkt = TAILQ_FIRST(&q->queue)) != NULL) {
500 int len = pkt->dn_m->m_pkthdr.len;
501 int len_scaled = p->bandwidth ? len*8*dn_hz : 0;
503 if (len_scaled > q->numbytes)
505 q->numbytes -= len_scaled;
506 move_pkt(pkt, q, p, len);
510 * If we have more packets queued, schedule next ready event
511 * (can only occur when bandwidth != 0, otherwise we would have
512 * flushed the whole queue in the previous loop).
513 * To this purpose we record the current time and compute how many
514 * ticks to go for the finish time of the packet.
516 if ((pkt = TAILQ_FIRST(&q->queue)) != NULL) {
517 /* This implies bandwidth != 0 */
518 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
520 q->sched_time = curr_time;
523 * XXX should check errors on heap_insert, and drain the whole
524 * queue on error hoping next time we are luckier.
526 heap_insert(&ready_heap, curr_time + t, q);
527 } else { /* RED needs to know when the queue becomes empty */
528 q->q_time = curr_time;
533 * If the delay line was empty call transmit_event(p) now.
534 * Otherwise, the scheduler will take care of it.
541 * Called when we can transmit packets on WF2Q queues. Take pkts out of
542 * the queues at their start time, and enqueue into the delay line.
543 * Packets are drained until p->numbytes < 0. As long as
544 * len_scaled >= p->numbytes, the packet goes into the delay line
545 * with a deadline p->delay. For the last packet, if p->numbytes < 0,
546 * there is an additional delay.
549 ready_event_wfq(struct dn_pipe *p)
551 int p_was_empty = TAILQ_EMPTY(&p->p_queue);
552 struct dn_heap *sch = &p->scheduler_heap;
553 struct dn_heap *neh = &p->not_eligible_heap;
555 p->numbytes += (curr_time - p->sched_time) * p->bandwidth;
558 * While we have backlogged traffic AND credit, we need to do
559 * something on the queue.
561 while (p->numbytes >= 0 && (sch->elements > 0 || neh->elements > 0)) {
562 if (sch->elements > 0) { /* Have some eligible pkts to send out */
563 struct dn_flow_queue *q = sch->p[0].object;
564 struct dn_pkt *pkt = TAILQ_FIRST(&q->queue);
565 struct dn_flow_set *fs = q->fs;
566 uint64_t len = pkt->dn_m->m_pkthdr.len;
567 int len_scaled = p->bandwidth ? len*8*dn_hz : 0;
569 heap_extract(sch, NULL); /* Remove queue from heap */
570 p->numbytes -= len_scaled;
571 move_pkt(pkt, q, p, len);
573 p->V += (len << MY_M) / p->sum; /* Update V */
574 q->S = q->F; /* Update start time */
576 if (q->len == 0) { /* Flow not backlogged any more */
578 heap_insert(&p->idle_heap, q->F, q);
579 } else { /* Still backlogged */
581 * Update F and position in backlogged queue, then
582 * put flow in not_eligible_heap (we will fix this later).
584 len = TAILQ_FIRST(&q->queue)->dn_m->m_pkthdr.len;
585 q->F += (len << MY_M) / (uint64_t)fs->weight;
586 if (DN_KEY_LEQ(q->S, p->V))
587 heap_insert(neh, q->S, q);
589 heap_insert(sch, q->F, q);
594 * Now compute V = max(V, min(S_i)). Remember that all elements in
595 * sch have by definition S_i <= V so if sch is not empty, V is surely
596 * the max and we must not update it. Conversely, if sch is empty
597 * we only need to look at neh.
599 if (sch->elements == 0 && neh->elements > 0)
600 p->V = MAX64(p->V, neh->p[0].key);
603 * Move from neh to sch any packets that have become eligible
605 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V)) {
606 struct dn_flow_queue *q = neh->p[0].object;
608 heap_extract(neh, NULL);
609 heap_insert(sch, q->F, q);
613 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0 &&
614 p->idle_heap.elements > 0) {
616 * No traffic and no events scheduled. We can get rid of idle-heap.
620 for (i = 0; i < p->idle_heap.elements; i++) {
621 struct dn_flow_queue *q = p->idle_heap.p[i].object;
628 p->idle_heap.elements = 0;
632 * If we are getting clocks from dummynet and if we are under credit,
633 * schedule the next ready event.
634 * Also fix the delivery time of the last packet.
636 if (p->numbytes < 0) { /* This implies bandwidth>0 */
637 dn_key t = 0; /* Number of ticks i have to wait */
639 if (p->bandwidth > 0)
640 t = (p->bandwidth - 1 - p->numbytes) / p->bandwidth;
641 TAILQ_LAST(&p->p_queue, dn_pkt_queue)->output_time += t;
642 p->sched_time = curr_time;
645 * XXX should check errors on heap_insert, and drain the whole
646 * queue on error hoping next time we are luckier.
648 heap_insert(&wfq_ready_heap, curr_time + t, p);
652 * If the delay line was empty call transmit_event(p) now.
653 * Otherwise, the scheduler will take care of it.
660 dn_expire_pipe_cb(struct dn_pipe *pipe, void *dummy __unused)
662 if (pipe->idle_heap.elements > 0 &&
663 DN_KEY_LT(pipe->idle_heap.p[0].key, pipe->V)) {
664 struct dn_flow_queue *q = pipe->idle_heap.p[0].object;
666 heap_extract(&pipe->idle_heap, NULL);
667 q->S = q->F + 1; /* Mark timestamp as invalid */
668 pipe->sum -= q->fs->weight;
673 * This is called once per tick, or dn_hz times per second. It is used to
674 * increment the current tick counter and schedule expired events.
677 dummynet(netmsg_t msg)
681 struct dn_heap *heaps[3];
684 heaps[0] = &ready_heap; /* Fixed-rate queues */
685 heaps[1] = &wfq_ready_heap; /* WF2Q queues */
686 heaps[2] = &extract_heap; /* Delay line */
690 lwkt_replymsg(&msg->lmsg, 0);
694 for (i = 0; i < 3; i++) {
696 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time)) {
697 if (h->p[0].key > curr_time) {
698 kprintf("-- dummynet: warning, heap %d is %d ticks late\n",
699 i, (int)(curr_time - h->p[0].key));
702 p = h->p[0].object; /* Store a copy before heap_extract */
703 heap_extract(h, NULL); /* Need to extract before processing */
714 /* Sweep pipes trying to expire idle flow_queues */
715 dn_iterate_pipe(dn_expire_pipe_cb, NULL);
719 * Unconditionally expire empty queues in case of shortage.
720 * Returns the number of queues freed.
723 expire_queues(struct dn_flow_set *fs)
725 int i, initial_elements = fs->rq_elements;
727 if (fs->last_expired == time_uptime)
730 fs->last_expired = time_uptime;
732 for (i = 0; i <= fs->rq_size; i++) { /* Last one is overflow */
733 struct dn_flow_queue *q, *qn;
735 LIST_FOREACH_MUTABLE(q, &fs->rq[i], q_link, qn) {
736 if (!TAILQ_EMPTY(&q->queue) || q->S != q->F + 1)
740 * Entry is idle, expire it
742 LIST_REMOVE(q, q_link);
743 kfree(q, M_DUMMYNET);
745 KASSERT(fs->rq_elements > 0,
746 ("invalid rq_elements %d", fs->rq_elements));
750 return initial_elements - fs->rq_elements;
754 * If room, create a new queue and put at head of slot i;
755 * otherwise, create or use the default queue.
757 static struct dn_flow_queue *
758 create_queue(struct dn_flow_set *fs, int i)
760 struct dn_flow_queue *q;
762 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
763 expire_queues(fs) == 0) {
765 * No way to get room, use or create overflow queue.
768 if (!LIST_EMPTY(&fs->rq[i]))
769 return LIST_FIRST(&fs->rq[i]);
772 q = kmalloc(sizeof(*q), M_DUMMYNET, M_INTWAIT | M_NULLOK | M_ZERO);
778 q->S = q->F + 1; /* hack - mark timestamp as invalid */
779 TAILQ_INIT(&q->queue);
781 LIST_INSERT_HEAD(&fs->rq[i], q, q_link);
788 * Given a flow_set and a pkt in last_pkt, find a matching queue
789 * after appropriate masking. The queue is moved to front
790 * so that further searches take less time.
792 static struct dn_flow_queue *
793 find_queue(struct dn_flow_set *fs, struct dn_flow_id *id)
795 struct dn_flow_queue *q;
798 if (!(fs->flags_fs & DN_HAVE_FLOW_MASK)) {
799 q = LIST_FIRST(&fs->rq[0]);
801 struct dn_flow_queue *qn;
803 /* First, do the masking */
804 id->fid_dst_ip &= fs->flow_mask.fid_dst_ip;
805 id->fid_src_ip &= fs->flow_mask.fid_src_ip;
806 id->fid_dst_port &= fs->flow_mask.fid_dst_port;
807 id->fid_src_port &= fs->flow_mask.fid_src_port;
808 id->fid_proto &= fs->flow_mask.fid_proto;
809 id->fid_flags = 0; /* we don't care about this one */
811 /* Then, hash function */
812 i = ((id->fid_dst_ip) & 0xffff) ^
813 ((id->fid_dst_ip >> 15) & 0xffff) ^
814 ((id->fid_src_ip << 1) & 0xffff) ^
815 ((id->fid_src_ip >> 16 ) & 0xffff) ^
816 (id->fid_dst_port << 1) ^ (id->fid_src_port) ^
821 * Finally, scan the current list for a match and
822 * expire idle flow queues
825 LIST_FOREACH_MUTABLE(q, &fs->rq[i], q_link, qn) {
827 if (id->fid_dst_ip == q->id.fid_dst_ip &&
828 id->fid_src_ip == q->id.fid_src_ip &&
829 id->fid_dst_port == q->id.fid_dst_port &&
830 id->fid_src_port == q->id.fid_src_port &&
831 id->fid_proto == q->id.fid_proto &&
832 id->fid_flags == q->id.fid_flags) {
834 } else if (pipe_expire && TAILQ_EMPTY(&q->queue) &&
837 * Entry is idle and not in any heap, expire it
839 LIST_REMOVE(q, q_link);
840 kfree(q, M_DUMMYNET);
842 KASSERT(fs->rq_elements > 0,
843 ("invalid rq_elements %d", fs->rq_elements));
847 if (q && LIST_FIRST(&fs->rq[i]) != q) { /* Found and not in front */
848 LIST_REMOVE(q, q_link);
849 LIST_INSERT_HEAD(&fs->rq[i], q, q_link);
852 if (q == NULL) { /* No match, need to allocate a new entry */
853 q = create_queue(fs, i);
861 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
866 * RED calculates the average queue size (avg) using a low-pass filter
867 * with an exponential weighted (w_q) moving average:
868 * avg <- (1-w_q) * avg + w_q * q_size
869 * where q_size is the queue length (measured in bytes or * packets).
871 * If q_size == 0, we compute the idle time for the link, and set
872 * avg = (1 - w_q)^(idle/s)
873 * where s is the time needed for transmitting a medium-sized packet.
875 * Now, if avg < min_th the packet is enqueued.
876 * If avg > max_th the packet is dropped. Otherwise, the packet is
877 * dropped with probability P function of avg.
881 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len;
883 IPFW3_DEBUG("\n%d q: %2u ", (int)curr_time, q_size);
885 /* Average queue size estimation */
888 * Queue is not empty, avg <- avg + (q_size - avg) * w_q
890 int diff = SCALE(q_size) - q->avg;
891 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);
896 * Queue is empty, find for how long the queue has been
897 * empty and use a lookup table for computing
898 * (1 - * w_q)^(idle_time/s) where s is the time to send a
903 u_int t = (curr_time - q->q_time) / fs->lookup_step;
905 q->avg = (t < fs->lookup_depth) ?
906 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
909 IPFW3_DEBUG("avg: %u ", SCALE_VAL(q->avg));
913 if (q->avg < fs->min_th) {
919 if (q->avg >= fs->max_th) { /* Average queue >= Max threshold */
920 if (fs->flags_fs & DN_IS_GENTLE_RED) {
922 * According to Gentle-RED, if avg is greater than max_th the
923 * packet is dropped with a probability
924 * p_b = c_3 * avg - c_4
925 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
927 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) - fs->c_4;
933 } else if (q->avg > fs->min_th) {
935 * We compute p_b using the linear dropping function p_b = c_1 *
936 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
937 * max_p * min_th / (max_th - min_th)
939 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
941 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
942 p_b = (p_b * len) / fs->max_pkt_size;
944 if (++q->count == 0) {
945 q->random = krandom() & 0xffff;
948 * q->count counts packets arrived since last drop, so a greater
949 * value of q->count means a greater packet drop probability.
951 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
953 IPFW3_DEBUG("%s", "- red drop");
954 /* After a drop we calculate a new random value */
955 q->random = krandom() & 0xffff;
959 /* End of RED algorithm */
960 return 0; /* Accept */
964 dn_iterate_pipe(dn_pipe_iter_t func, void *arg)
968 for (i = 0; i < DN_NR_HASH_MAX; ++i) {
969 struct dn_pipe_head *pipe_hdr = &pipe_table[i];
970 struct dn_pipe *pipe, *pipe_next;
972 LIST_FOREACH_MUTABLE(pipe, pipe_hdr, p_link, pipe_next)
978 dn_iterate_flowset(dn_flowset_iter_t func, void *arg)
982 for (i = 0; i < DN_NR_HASH_MAX; ++i) {
983 struct dn_flowset_head *fs_hdr = &flowset_table[i];
984 struct dn_flow_set *fs, *fs_next;
986 LIST_FOREACH_MUTABLE(fs, fs_hdr, fs_link, fs_next)
991 static struct dn_pipe *
992 dn_find_pipe(int pipe_nr)
994 struct dn_pipe_head *pipe_hdr;
997 pipe_hdr = &pipe_table[DN_NR_HASH(pipe_nr)];
998 LIST_FOREACH(p, pipe_hdr, p_link) {
999 if (p->pipe_nr == pipe_nr)
1005 static struct dn_flow_set *
1006 dn_find_flowset(int fs_nr)
1008 struct dn_flowset_head *fs_hdr;
1009 struct dn_flow_set *fs;
1011 fs_hdr = &flowset_table[DN_NR_HASH(fs_nr)];
1012 LIST_FOREACH(fs, fs_hdr, fs_link) {
1013 if (fs->fs_nr == fs_nr)
1019 static struct dn_flow_set *
1020 dn_locate_flowset(int pipe_nr, int is_pipe)
1022 struct dn_flow_set *fs = NULL;
1025 fs = dn_find_flowset(pipe_nr);
1029 p = dn_find_pipe(pipe_nr);
1037 * Dummynet hook for packets. Below 'pipe' is a pipe or a queue
1038 * depending on whether WF2Q or fixed bw is used.
1040 * pipe_nr pipe or queue the packet is destined for.
1041 * dir where shall we send the packet after dummynet.
1042 * m the mbuf with the packet
1043 * fwa->oif the 'ifp' parameter from the caller.
1044 * NULL in ip_input, destination interface in ip_output
1045 * fwa->ro route parameter (only used in ip_output, NULL otherwise)
1046 * fwa->dst destination address, only used by ip_output
1047 * fwa->rule matching rule, in case of multiple passes
1048 * fwa->flags flags from the caller, only used in ip_output
1051 dummynet_io(struct mbuf *m)
1055 struct dn_flow_set *fs;
1056 struct dn_pipe *pipe;
1057 uint64_t len = m->m_pkthdr.len;
1058 struct dn_flow_queue *q = NULL;
1059 int is_pipe, pipe_nr;
1061 tag = m_tag_find(m, PACKET_TAG_DUMMYNET, NULL);
1062 pkt = m_tag_data(tag);
1064 is_pipe = pkt->dn_flags & DN_FLAGS_IS_PIPE;
1065 pipe_nr = pkt->pipe_nr;
1068 * This is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule
1070 fs = dn_locate_flowset(pipe_nr, is_pipe);
1072 goto dropit; /* This queue/pipe does not exist! */
1075 if (pipe == NULL) { /* Must be a queue, try find a matching pipe */
1076 pipe = dn_find_pipe(fs->parent_nr);
1080 kprintf("No pipe %d for queue %d, drop pkt\n",
1081 fs->parent_nr, fs->fs_nr);
1086 q = find_queue(fs, &pkt->id);
1088 goto dropit; /* Cannot allocate queue */
1091 * Update statistics, then check reasons to drop pkt
1093 q->tot_bytes += len;
1096 if (fs->plr && krandom() < fs->plr)
1097 goto dropit; /* Random pkt drop */
1099 if (fs->flags_fs & DN_QSIZE_IS_BYTES) {
1100 if (q->len_bytes > fs->qsize)
1101 goto dropit; /* Queue size overflow */
1103 if (q->len >= fs->qsize)
1104 goto dropit; /* Queue count overflow */
1107 if ((fs->flags_fs & DN_IS_RED) && red_drops(fs, q, len))
1110 TAILQ_INSERT_TAIL(&q->queue, pkt, dn_next);
1112 q->len_bytes += len;
1114 if (TAILQ_FIRST(&q->queue) != pkt) /* Flow was not idle, we are done */
1118 * If we reach this point the flow was previously idle, so we need
1119 * to schedule it. This involves different actions for fixed-rate
1124 * Fixed-rate queue: just insert into the ready_heap.
1128 if (pipe->bandwidth)
1129 t = SET_TICKS(pkt, q, pipe);
1131 q->sched_time = curr_time;
1132 if (t == 0) /* Must process it now */
1135 heap_insert(&ready_heap, curr_time + t, q);
1139 * First, compute start time S: if the flow was idle (S=F+1)
1140 * set S to the virtual time V for the controlling pipe, and update
1141 * the sum of weights for the pipe; otherwise, remove flow from
1142 * idle_heap and set S to max(F, V).
1143 * Second, compute finish time F = S + len/weight.
1144 * Third, if pipe was idle, update V = max(S, V).
1145 * Fourth, count one more backlogged flow.
1147 if (DN_KEY_GT(q->S, q->F)) { /* Means timestamps are invalid */
1149 pipe->sum += fs->weight; /* Add weight of new queue */
1151 heap_extract(&pipe->idle_heap, q);
1152 q->S = MAX64(q->F, pipe->V);
1154 q->F = q->S + (len << MY_M) / (uint64_t)fs->weight;
1156 if (pipe->not_eligible_heap.elements == 0 &&
1157 pipe->scheduler_heap.elements == 0)
1158 pipe->V = MAX64(q->S, pipe->V);
1163 * Look at eligibility. A flow is not eligibile if S>V (when
1164 * this happens, it means that there is some other flow already
1165 * scheduled for the same pipe, so the scheduler_heap cannot be
1166 * empty). If the flow is not eligible we just store it in the
1167 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1168 * and possibly invoke ready_event_wfq() right now if there is
1170 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1171 * and for all flows in not_eligible_heap (NEH), S_i > V.
1172 * So when we need to compute max(V, min(S_i)) forall i in SCH+NEH,
1173 * we only need to look into NEH.
1175 if (DN_KEY_GT(q->S, pipe->V)) { /* Not eligible */
1176 if (pipe->scheduler_heap.elements == 0)
1177 kprintf("++ ouch! not eligible but empty scheduler!\n");
1178 heap_insert(&pipe->not_eligible_heap, q->S, q);
1180 heap_insert(&pipe->scheduler_heap, q->F, q);
1181 if (pipe->numbytes >= 0) { /* Pipe is idle */
1182 if (pipe->scheduler_heap.elements != 1)
1183 kprintf("*** OUCH! pipe should have been idle!\n");
1184 IPFW3_DEBUG("Waking up pipe %d at %d\n",
1185 pipe->pipe_nr, (int)(q->F >> MY_M));
1186 pipe->sched_time = curr_time;
1187 ready_event_wfq(pipe);
1201 * Dispose all packets and flow_queues on a flow_set.
1202 * If all=1, also remove red lookup table and other storage,
1203 * including the descriptor itself.
1204 * For the one in dn_pipe MUST also cleanup ready_heap...
1207 purge_flow_set(struct dn_flow_set *fs, int all)
1211 int rq_elements = 0;
1214 for (i = 0; i <= fs->rq_size; i++) {
1215 struct dn_flow_queue *q;
1217 while ((q = LIST_FIRST(&fs->rq[i])) != NULL) {
1220 while ((pkt = TAILQ_FIRST(&q->queue)) != NULL) {
1221 TAILQ_REMOVE(&q->queue, pkt, dn_next);
1222 ip_dn_packet_free(pkt);
1225 LIST_REMOVE(q, q_link);
1226 kfree(q, M_DUMMYNET);
1233 KASSERT(rq_elements == fs->rq_elements,
1234 ("# rq elements mismatch, freed %d, total %d",
1235 rq_elements, fs->rq_elements));
1236 fs->rq_elements = 0;
1239 /* RED - free lookup table */
1241 kfree(fs->w_q_lookup, M_DUMMYNET);
1244 kfree(fs->rq, M_DUMMYNET);
1247 * If this fs is not part of a pipe, free it
1249 * fs->pipe == NULL could happen, if 'fs' is a WF2Q and
1250 * - No packet belongs to that flow set is delivered by
1251 * dummynet_io(), i.e. parent pipe is not installed yet.
1252 * - Parent pipe is deleted.
1254 if (fs->pipe == NULL || (fs->pipe && fs != &fs->pipe->fs))
1255 kfree(fs, M_DUMMYNET);
1260 * Dispose all packets queued on a pipe (not a flow_set).
1261 * Also free all resources associated to a pipe, which is about
1265 purge_pipe(struct dn_pipe *pipe)
1269 purge_flow_set(&pipe->fs, 1);
1271 while ((pkt = TAILQ_FIRST(&pipe->p_queue)) != NULL) {
1272 TAILQ_REMOVE(&pipe->p_queue, pkt, dn_next);
1273 ip_dn_packet_free(pkt);
1276 heap_free(&pipe->scheduler_heap);
1277 heap_free(&pipe->not_eligible_heap);
1278 heap_free(&pipe->idle_heap);
1282 * Delete all pipes and heaps returning memory.
1285 dummynet_flush(void)
1287 struct dn_pipe_head pipe_list;
1288 struct dn_flowset_head fs_list;
1290 struct dn_flow_set *fs;
1294 * Prevent future matches...
1296 LIST_INIT(&pipe_list);
1297 for (i = 0; i < DN_NR_HASH_MAX; ++i) {
1298 struct dn_pipe_head *pipe_hdr = &pipe_table[i];
1300 while ((p = LIST_FIRST(pipe_hdr)) != NULL) {
1301 LIST_REMOVE(p, p_link);
1302 LIST_INSERT_HEAD(&pipe_list, p, p_link);
1306 LIST_INIT(&fs_list);
1307 for (i = 0; i < DN_NR_HASH_MAX; ++i) {
1308 struct dn_flowset_head *fs_hdr = &flowset_table[i];
1310 while ((fs = LIST_FIRST(fs_hdr)) != NULL) {
1311 LIST_REMOVE(fs, fs_link);
1312 LIST_INSERT_HEAD(&fs_list, fs, fs_link);
1316 /* Free heaps so we don't have unwanted events */
1317 heap_free(&ready_heap);
1318 heap_free(&wfq_ready_heap);
1319 heap_free(&extract_heap);
1322 * Now purge all queued pkts and delete all pipes
1324 /* Scan and purge all flow_sets. */
1325 while ((fs = LIST_FIRST(&fs_list)) != NULL) {
1326 LIST_REMOVE(fs, fs_link);
1327 purge_flow_set(fs, 1);
1330 while ((p = LIST_FIRST(&pipe_list)) != NULL) {
1331 LIST_REMOVE(p, p_link);
1333 kfree(p, M_DUMMYNET);
1338 * setup RED parameters
1341 config_red(const struct dn_ioc_flowset *ioc_fs, struct dn_flow_set *x)
1345 x->w_q = ioc_fs->w_q;
1346 x->min_th = SCALE(ioc_fs->min_th);
1347 x->max_th = SCALE(ioc_fs->max_th);
1348 x->max_p = ioc_fs->max_p;
1350 x->c_1 = ioc_fs->max_p / (ioc_fs->max_th - ioc_fs->min_th);
1351 x->c_2 = SCALE_MUL(x->c_1, SCALE(ioc_fs->min_th));
1352 if (x->flags_fs & DN_IS_GENTLE_RED) {
1353 x->c_3 = (SCALE(1) - ioc_fs->max_p) / ioc_fs->max_th;
1354 x->c_4 = (SCALE(1) - 2 * ioc_fs->max_p);
1357 /* If the lookup table already exist, free and create it again */
1358 if (x->w_q_lookup) {
1359 kfree(x->w_q_lookup, M_DUMMYNET);
1360 x->w_q_lookup = NULL ;
1363 if (red_lookup_depth == 0) {
1364 kprintf("net.inet.ip.dummynet.red_lookup_depth must be > 0\n");
1365 kfree(x, M_DUMMYNET);
1368 x->lookup_depth = red_lookup_depth;
1369 x->w_q_lookup = kmalloc(x->lookup_depth * sizeof(int),
1370 M_DUMMYNET, M_WAITOK);
1372 /* Fill the lookup table with (1 - w_q)^x */
1373 x->lookup_step = ioc_fs->lookup_step;
1374 x->lookup_weight = ioc_fs->lookup_weight;
1376 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1377 for (i = 1; i < x->lookup_depth; i++)
1378 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1380 if (red_avg_pkt_size < 1)
1381 red_avg_pkt_size = 512;
1382 x->avg_pkt_size = red_avg_pkt_size;
1384 if (red_max_pkt_size < 1)
1385 red_max_pkt_size = 1500;
1386 x->max_pkt_size = red_max_pkt_size;
1392 alloc_hash(struct dn_flow_set *x, const struct dn_ioc_flowset *ioc_fs)
1396 if (x->flags_fs & DN_HAVE_FLOW_MASK) {
1397 int l = ioc_fs->rq_size;
1399 /* Allocate some slots */
1403 if (l < DN_MIN_HASH_SIZE)
1404 l = DN_MIN_HASH_SIZE;
1405 else if (l > DN_MAX_HASH_SIZE)
1406 l = DN_MAX_HASH_SIZE;
1410 /* One is enough for null mask */
1413 alloc_size = x->rq_size + 1;
1415 x->rq = kmalloc(alloc_size * sizeof(struct dn_flowqueue_head),
1416 M_DUMMYNET, M_WAITOK | M_ZERO);
1419 for (i = 0; i < alloc_size; ++i)
1420 LIST_INIT(&x->rq[i]);
1424 set_flowid_parms(struct dn_flow_id *id, const struct dn_ioc_flowid *ioc_id)
1426 id->fid_dst_ip = ioc_id->u.ip.dst_ip;
1427 id->fid_src_ip = ioc_id->u.ip.src_ip;
1428 id->fid_dst_port = ioc_id->u.ip.dst_port;
1429 id->fid_src_port = ioc_id->u.ip.src_port;
1430 id->fid_proto = ioc_id->u.ip.proto;
1431 id->fid_flags = ioc_id->u.ip.flags;
1435 set_fs_parms(struct dn_flow_set *x, const struct dn_ioc_flowset *ioc_fs)
1437 x->flags_fs = ioc_fs->flags_fs;
1438 x->qsize = ioc_fs->qsize;
1439 x->plr = ioc_fs->plr;
1440 set_flowid_parms(&x->flow_mask, &ioc_fs->flow_mask);
1441 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1442 if (x->qsize > 1024 * 1024)
1443 x->qsize = 1024 * 1024;
1445 if (x->qsize == 0 || x->qsize > 100)
1449 /* Configuring RED */
1450 if (x->flags_fs & DN_IS_RED)
1451 config_red(ioc_fs, x); /* XXX should check errors */
1455 * setup pipe or queue parameters.
1459 config_pipe(struct dn_ioc_pipe *ioc_pipe)
1461 struct dn_ioc_flowset *ioc_fs = &ioc_pipe->fs;
1465 * The config program passes parameters as follows:
1466 * bw bits/second (0 means no limits)
1467 * delay ms (must be translated into ticks)
1468 * qsize slots or bytes
1470 ioc_pipe->delay = (ioc_pipe->delay * dn_hz) / 1000;
1473 * We need either a pipe number or a flow_set number
1475 if (ioc_pipe->pipe_nr == 0 && ioc_fs->fs_nr == 0)
1477 if (ioc_pipe->pipe_nr != 0 && ioc_fs->fs_nr != 0)
1481 * Validate pipe number
1483 if (ioc_pipe->pipe_nr > DN_PIPE_NR_MAX || ioc_pipe->pipe_nr < 0)
1487 if (ioc_pipe->pipe_nr != 0) { /* This is a pipe */
1488 struct dn_pipe *x, *p;
1491 p = dn_find_pipe(ioc_pipe->pipe_nr);
1493 if (p == NULL) { /* New pipe */
1494 x = kmalloc(sizeof(struct dn_pipe), M_DUMMYNET, M_WAITOK | M_ZERO);
1495 x->pipe_nr = ioc_pipe->pipe_nr;
1497 TAILQ_INIT(&x->p_queue);
1500 * idle_heap is the only one from which we extract from the middle.
1502 x->idle_heap.size = x->idle_heap.elements = 0;
1503 x->idle_heap.offset = __offsetof(struct dn_flow_queue, heap_pos);
1509 /* Flush accumulated credit for all queues */
1510 for (i = 0; i <= x->fs.rq_size; i++) {
1511 struct dn_flow_queue *q;
1513 LIST_FOREACH(q, &x->fs.rq[i], q_link)
1518 x->bandwidth = ioc_pipe->bandwidth;
1519 x->numbytes = 0; /* Just in case... */
1520 x->delay = ioc_pipe->delay;
1522 set_fs_parms(&x->fs, ioc_fs);
1524 if (x->fs.rq == NULL) { /* A new pipe */
1525 struct dn_pipe_head *pipe_hdr;
1527 alloc_hash(&x->fs, ioc_fs);
1529 pipe_hdr = &pipe_table[DN_NR_HASH(x->pipe_nr)];
1530 LIST_INSERT_HEAD(pipe_hdr, x, p_link);
1532 } else { /* Config flow_set */
1533 struct dn_flow_set *x, *fs;
1535 /* Locate flow_set */
1536 fs = dn_find_flowset(ioc_fs->fs_nr);
1538 if (fs == NULL) { /* New flow_set */
1539 if (ioc_fs->parent_nr == 0) /* Need link to a pipe */
1542 x = kmalloc(sizeof(struct dn_flow_set), M_DUMMYNET,
1544 x->fs_nr = ioc_fs->fs_nr;
1545 x->parent_nr = ioc_fs->parent_nr;
1546 x->weight = ioc_fs->weight;
1549 else if (x->weight > 100)
1552 /* Change parent pipe not allowed; must delete and recreate */
1553 if (ioc_fs->parent_nr != 0 && fs->parent_nr != ioc_fs->parent_nr)
1558 set_fs_parms(x, ioc_fs);
1560 if (x->rq == NULL) { /* A new flow_set */
1561 struct dn_flowset_head *fs_hdr;
1563 alloc_hash(x, ioc_fs);
1565 fs_hdr = &flowset_table[DN_NR_HASH(x->fs_nr)];
1566 LIST_INSERT_HEAD(fs_hdr, x, fs_link);
1576 * Helper function to remove from a heap queues which are linked to
1577 * a flow_set about to be deleted.
1580 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1582 int i = 0, found = 0;
1584 while (i < h->elements) {
1585 if (((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1587 h->p[i] = h->p[h->elements];
1598 * helper function to remove a pipe from a heap (can be there at most once)
1601 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1603 if (h->elements > 0) {
1606 for (i = 0; i < h->elements; i++) {
1607 if (h->p[i].object == p) { /* found it */
1609 h->p[i] = h->p[h->elements];
1618 dn_unref_pipe_cb(struct dn_flow_set *fs, void *pipe0)
1620 struct dn_pipe *pipe = pipe0;
1622 if (fs->pipe == pipe) {
1623 kprintf("++ ref to pipe %d from fs %d\n",
1624 pipe->pipe_nr, fs->fs_nr);
1626 purge_flow_set(fs, 0);
1631 * Fully delete a pipe or a queue, cleaning up associated info.
1634 delete_pipe(const struct dn_ioc_pipe *ioc_pipe)
1639 if (ioc_pipe->pipe_nr == 0 && ioc_pipe->fs.fs_nr == 0)
1641 if (ioc_pipe->pipe_nr != 0 && ioc_pipe->fs.fs_nr != 0)
1644 if (ioc_pipe->pipe_nr > DN_NR_HASH_MAX || ioc_pipe->pipe_nr < 0)
1648 if (ioc_pipe->pipe_nr != 0) { /* This is an old-style pipe */
1650 p = dn_find_pipe(ioc_pipe->pipe_nr);
1652 goto back; /* Not found */
1654 /* Unlink from pipe hash table */
1655 LIST_REMOVE(p, p_link);
1657 /* Remove all references to this pipe from flow_sets */
1658 dn_iterate_flowset(dn_unref_pipe_cb, p);
1660 fs_remove_from_heap(&ready_heap, &p->fs);
1661 purge_pipe(p); /* Remove all data associated to this pipe */
1663 /* Remove reference to here from extract_heap and wfq_ready_heap */
1664 pipe_remove_from_heap(&extract_heap, p);
1665 pipe_remove_from_heap(&wfq_ready_heap, p);
1667 kfree(p, M_DUMMYNET);
1668 } else { /* This is a WF2Q queue (dn_flow_set) */
1669 struct dn_flow_set *fs;
1671 /* Locate flow_set */
1672 fs = dn_find_flowset(ioc_pipe->fs.fs_nr);
1674 goto back; /* Not found */
1676 LIST_REMOVE(fs, fs_link);
1678 if ((p = fs->pipe) != NULL) {
1679 /* Update total weight on parent pipe and cleanup parent heaps */
1680 p->sum -= fs->weight * fs->backlogged;
1681 fs_remove_from_heap(&p->not_eligible_heap, fs);
1682 fs_remove_from_heap(&p->scheduler_heap, fs);
1683 #if 1 /* XXX should i remove from idle_heap as well ? */
1684 fs_remove_from_heap(&p->idle_heap, fs);
1687 purge_flow_set(fs, 1);
1696 * helper function used to copy data from kernel in DUMMYNET_GET
1699 dn_copy_flowid(const struct dn_flow_id *id, struct dn_ioc_flowid *ioc_id)
1701 ioc_id->type = ETHERTYPE_IP;
1702 ioc_id->u.ip.dst_ip = id->fid_dst_ip;
1703 ioc_id->u.ip.src_ip = id->fid_src_ip;
1704 ioc_id->u.ip.dst_port = id->fid_dst_port;
1705 ioc_id->u.ip.src_port = id->fid_src_port;
1706 ioc_id->u.ip.proto = id->fid_proto;
1707 ioc_id->u.ip.flags = id->fid_flags;
1711 dn_copy_flowqueues(const struct dn_flow_set *fs, void *bp)
1713 struct dn_ioc_flowqueue *ioc_fq = bp;
1716 for (i = 0; i <= fs->rq_size; i++) {
1717 const struct dn_flow_queue *q;
1719 LIST_FOREACH(q, &fs->rq[i], q_link) {
1720 if (q->hash_slot != i) { /* XXX ASSERT */
1721 kprintf("++ at %d: wrong slot (have %d, "
1723 copied, q->hash_slot, i);
1725 if (q->fs != fs) { /* XXX ASSERT */
1726 kprintf("++ at %d: wrong fs ptr (have %p, should be %p)\n",
1732 ioc_fq->len = q->len;
1733 ioc_fq->len_bytes = q->len_bytes;
1734 ioc_fq->tot_pkts = q->tot_pkts;
1735 ioc_fq->tot_bytes = q->tot_bytes;
1736 ioc_fq->drops = q->drops;
1737 ioc_fq->hash_slot = q->hash_slot;
1740 dn_copy_flowid(&q->id, &ioc_fq->id);
1746 if (copied != fs->rq_elements) { /* XXX ASSERT */
1747 kprintf("++ wrong count, have %d should be %d\n",
1748 copied, fs->rq_elements);
1754 dn_copy_flowset(const struct dn_flow_set *fs, struct dn_ioc_flowset *ioc_fs,
1757 ioc_fs->fs_type = fs_type;
1759 ioc_fs->fs_nr = fs->fs_nr;
1760 ioc_fs->flags_fs = fs->flags_fs;
1761 ioc_fs->parent_nr = fs->parent_nr;
1763 ioc_fs->weight = fs->weight;
1764 ioc_fs->qsize = fs->qsize;
1765 ioc_fs->plr = fs->plr;
1767 ioc_fs->rq_size = fs->rq_size;
1768 ioc_fs->rq_elements = fs->rq_elements;
1770 ioc_fs->w_q = fs->w_q;
1771 ioc_fs->max_th = fs->max_th;
1772 ioc_fs->min_th = fs->min_th;
1773 ioc_fs->max_p = fs->max_p;
1775 dn_copy_flowid(&fs->flow_mask, &ioc_fs->flow_mask);
1779 dn_calc_pipe_size_cb(struct dn_pipe *pipe, void *sz)
1783 *size += sizeof(struct dn_ioc_pipe) +
1784 pipe->fs.rq_elements * sizeof(struct dn_ioc_flowqueue);
1788 dn_calc_fs_size_cb(struct dn_flow_set *fs, void *sz)
1792 *size += sizeof(struct dn_ioc_flowset) +
1793 fs->rq_elements * sizeof(struct dn_ioc_flowqueue);
1797 dn_copyout_pipe_cb(struct dn_pipe *pipe, void *bp0)
1800 struct dn_ioc_pipe *ioc_pipe = (struct dn_ioc_pipe *)(*bp);
1803 * Copy flow set descriptor associated with this pipe
1805 dn_copy_flowset(&pipe->fs, &ioc_pipe->fs, DN_IS_PIPE);
1808 * Copy pipe descriptor
1810 ioc_pipe->bandwidth = pipe->bandwidth;
1811 ioc_pipe->pipe_nr = pipe->pipe_nr;
1812 ioc_pipe->V = pipe->V;
1813 /* Convert delay to milliseconds */
1814 ioc_pipe->delay = (pipe->delay * 1000) / dn_hz;
1817 * Copy flow queue descriptors
1819 *bp += sizeof(*ioc_pipe);
1820 *bp = dn_copy_flowqueues(&pipe->fs, *bp);
1824 dn_copyout_fs_cb(struct dn_flow_set *fs, void *bp0)
1827 struct dn_ioc_flowset *ioc_fs = (struct dn_ioc_flowset *)(*bp);
1830 * Copy flow set descriptor
1832 dn_copy_flowset(fs, ioc_fs, DN_IS_QUEUE);
1835 * Copy flow queue descriptors
1837 *bp += sizeof(*ioc_fs);
1838 *bp = dn_copy_flowqueues(fs, *bp);
1842 dummynet_get(struct dn_sopt *dn_sopt)
1848 * Compute size of data structures: list of pipes and flow_sets.
1850 dn_iterate_pipe(dn_calc_pipe_size_cb, &size);
1851 dn_iterate_flowset(dn_calc_fs_size_cb, &size);
1854 * Copyout pipe/flow_set/flow_queue
1856 bp = buf = kmalloc(size, M_TEMP, M_WAITOK | M_ZERO);
1857 dn_iterate_pipe(dn_copyout_pipe_cb, &bp);
1858 dn_iterate_flowset(dn_copyout_fs_cb, &bp);
1860 /* Temp memory will be freed by caller */
1861 dn_sopt->dn_sopt_arg = buf;
1862 dn_sopt->dn_sopt_arglen = size;
1867 * Handler for the various dummynet socket options (get, flush, config, del)
1870 dummynet_ctl(struct dn_sopt *dn_sopt)
1874 switch (dn_sopt->dn_sopt_name) {
1875 case IP_DUMMYNET_GET:
1876 error = dummynet_get(dn_sopt);
1879 case IP_DUMMYNET_FLUSH:
1883 case IP_DUMMYNET_CONFIGURE:
1884 KKASSERT(dn_sopt->dn_sopt_arglen == sizeof(struct dn_ioc_pipe));
1885 error = config_pipe(dn_sopt->dn_sopt_arg);
1888 case IP_DUMMYNET_DEL: /* Remove a pipe or flow_set */
1889 KKASSERT(dn_sopt->dn_sopt_arglen == sizeof(struct dn_ioc_pipe));
1890 error = delete_pipe(dn_sopt->dn_sopt_arg);
1894 kprintf("%s -- unknown option %d\n", __func__, dn_sopt->dn_sopt_name);
1902 dummynet_clock(systimer_t info __unused, int in_ipi __unused,
1903 struct intrframe *frame __unused)
1905 KASSERT(mycpuid == ip_dn_cpu,
1906 ("dummynet systimer comes on cpu%d, should be %d!",
1907 mycpuid, ip_dn_cpu));
1910 if (DUMMYNET_LOADED && (dn_netmsg.lmsg.ms_flags & MSGF_DONE))
1911 lwkt_sendmsg_oncpu(netisr_cpuport(mycpuid), &dn_netmsg.lmsg);
1916 sysctl_dn_hz(SYSCTL_HANDLER_ARGS)
1921 error = sysctl_handle_int(oidp, &val, 0, req);
1922 if (error || req->newptr == NULL)
1926 else if (val > DN_CALLOUT_FREQ_MAX)
1927 val = DN_CALLOUT_FREQ_MAX;
1931 systimer_adjust_periodic(&dn_clock, val);
1938 ip_dn_init_dispatch(netmsg_t msg)
1942 KASSERT(mycpuid == ip_dn_cpu,
1943 ("%s runs on cpu%d, instead of cpu%d", __func__,
1944 mycpuid, ip_dn_cpu));
1948 if (DUMMYNET_LOADED) {
1949 kprintf("DUMMYNET already loaded\n");
1954 kprintf("DUMMYNET initialized (011031)\n");
1956 for (i = 0; i < DN_NR_HASH_MAX; ++i)
1957 LIST_INIT(&pipe_table[i]);
1959 for (i = 0; i < DN_NR_HASH_MAX; ++i)
1960 LIST_INIT(&flowset_table[i]);
1962 ready_heap.size = ready_heap.elements = 0;
1963 ready_heap.offset = 0;
1965 wfq_ready_heap.size = wfq_ready_heap.elements = 0;
1966 wfq_ready_heap.offset = 0;
1968 extract_heap.size = extract_heap.elements = 0;
1969 extract_heap.offset = 0;
1971 ip_dn_ctl_ptr = dummynet_ctl;
1972 ip_dn_io_ptr = dummynet_io;
1974 netmsg_init(&dn_netmsg, NULL, &netisr_adone_rport,
1976 systimer_init_periodic_nq(&dn_clock, dummynet_clock, NULL, dn_hz);
1980 lwkt_replymsg(&msg->lmsg, error);
1986 struct netmsg_base smsg;
1988 if (ip_dn_cpu >= ncpus) {
1989 kprintf("%s: CPU%d does not exist, switch to CPU0\n",
1990 __func__, ip_dn_cpu);
1994 ip_fw3_register_module(MODULE_DUMMYNET_ID, MODULE_DUMMYNET_NAME);
1995 ip_fw3_register_filter_funcs(MODULE_DUMMYNET_ID, O_DUMMYNET_PIPE,
1996 (filter_func)check_pipe);
1997 ip_fw3_register_filter_funcs(MODULE_DUMMYNET_ID, O_DUMMYNET_QUEUE,
1998 (filter_func)check_pipe);
2000 netmsg_init(&smsg, NULL, &curthread->td_msgport,
2001 0, ip_dn_init_dispatch);
2002 lwkt_domsg(netisr_cpuport(ip_dn_cpu), &smsg.lmsg, 0);
2003 return smsg.lmsg.ms_error;
2009 ip_dn_stop_dispatch(netmsg_t msg)
2015 ip_dn_ctl_ptr = NULL;
2016 ip_dn_io_ptr = NULL;
2018 systimer_del(&dn_clock);
2021 lwkt_replymsg(&msg->lmsg, 0);
2027 struct netmsg_base smsg;
2029 netmsg_init(&smsg, NULL, &curthread->td_msgport,
2030 0, ip_dn_stop_dispatch);
2031 lwkt_domsg(netisr_cpuport(ip_dn_cpu), &smsg.lmsg, 0);
2033 netmsg_service_sync();
2036 #endif /* KLD_MODULE */
2039 dummynet_modevent(module_t mod, int type, void *data)
2043 return ip_dn_init();
2047 kprintf("dummynet statically compiled, cannot unload\n");
2060 static moduledata_t dummynet_mod = {
2065 DECLARE_MODULE(dummynet3, dummynet_mod, SI_SUB_PROTO_END, SI_ORDER_ANY);
2066 MODULE_DEPEND(dummynet3, ipfw3_basic, 1, 1, 1);
2067 MODULE_VERSION(dummynet3, 1);