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 1.53 2007/12/08 04:33:58 sephe Exp $
32 #define DPRINTF(fmt, ...) kprintf(fmt, __VA_ARGS__)
34 #define DPRINTF(fmt, ...) ((void)0)
38 * This module implements IP dummynet, a bandwidth limiter/delay emulator.
39 * Description of the data structures used is in ip_dummynet.h
40 * Here you mainly find the following blocks of code:
41 * + variable declarations;
42 * + heap management functions;
43 * + scheduler and dummynet functions;
44 * + configuration and initialization.
46 * Most important Changes:
49 * 010124: Fixed WF2Q behaviour
50 * 010122: Fixed spl protection.
51 * 000601: WF2Q support
52 * 000106: Large rewrite, use heaps to handle very many pipes.
53 * 980513: Initial release
56 #include <sys/param.h>
57 #include <sys/kernel.h>
58 #include <sys/malloc.h>
60 #include <sys/socketvar.h>
61 #include <sys/sysctl.h>
62 #include <sys/systimer.h>
63 #include <sys/thread2.h>
65 #include <net/ethernet.h>
66 #include <net/netmsg2.h>
67 #include <net/route.h>
69 #include <netinet/in_var.h>
70 #include <netinet/ip_var.h>
72 #include <net/dummynet/ip_dummynet.h>
74 #ifndef DN_CALLOUT_FREQ_MAX
75 #define DN_CALLOUT_FREQ_MAX 10000
79 * The maximum/minimum hash table size for queues.
80 * These values must be a power of 2.
82 #define DN_MIN_HASH_SIZE 4
83 #define DN_MAX_HASH_SIZE 65536
86 * Some macros are used to compare key values and handle wraparounds.
87 * MAX64 returns the largest of two key values.
89 #define DN_KEY_LT(a, b) ((int64_t)((a) - (b)) < 0)
90 #define DN_KEY_LEQ(a, b) ((int64_t)((a) - (b)) <= 0)
91 #define DN_KEY_GT(a, b) ((int64_t)((a) - (b)) > 0)
92 #define DN_KEY_GEQ(a, b) ((int64_t)((a) - (b)) >= 0)
93 #define MAX64(x, y) ((((int64_t)((y) - (x))) > 0) ? (y) : (x))
95 #define DN_NR_HASH_MAX 16
96 #define DN_NR_HASH_MASK (DN_NR_HASH_MAX - 1)
97 #define DN_NR_HASH(nr) \
98 ((((nr) >> 12) ^ ((nr) >> 8) ^ ((nr) >> 4) ^ (nr)) & DN_NR_HASH_MASK)
100 MALLOC_DEFINE(M_DUMMYNET
, "dummynet", "dummynet heap");
102 extern int ip_dn_cpu
;
104 static dn_key curr_time
= 0; /* current simulation time */
105 static int dn_hash_size
= 64; /* default hash size */
106 static int pipe_expire
= 1; /* expire queue if empty */
107 static int dn_max_ratio
= 16; /* max queues/buckets ratio */
110 * Statistics on number of queue searches and search steps
113 static int search_steps
;
118 static int red_lookup_depth
= 256; /* default lookup table depth */
119 static int red_avg_pkt_size
= 512; /* default medium packet size */
120 static int red_max_pkt_size
= 1500;/* default max packet size */
123 * Three heaps contain queues and pipes that the scheduler handles:
125 * + ready_heap contains all dn_flow_queue related to fixed-rate pipes.
126 * + wfq_ready_heap contains the pipes associated with WF2Q flows
127 * + extract_heap contains pipes associated with delay lines.
129 static struct dn_heap ready_heap
;
130 static struct dn_heap extract_heap
;
131 static struct dn_heap wfq_ready_heap
;
133 static struct dn_pipe_head pipe_table
[DN_NR_HASH_MAX
];
134 static struct dn_flowset_head flowset_table
[DN_NR_HASH_MAX
];
137 * Variables for dummynet systimer
139 static struct netmsg dn_netmsg
;
140 static struct systimer dn_clock
;
141 static int dn_hz
= 1000;
143 static int sysctl_dn_hz(SYSCTL_HANDLER_ARGS
);
145 SYSCTL_DECL(_net_inet_ip_dummynet
);
147 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, hash_size
, CTLFLAG_RW
,
148 &dn_hash_size
, 0, "Default hash table size");
149 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, curr_time
, CTLFLAG_RD
,
150 &curr_time
, 0, "Current tick");
151 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, expire
, CTLFLAG_RW
,
152 &pipe_expire
, 0, "Expire queue if empty");
153 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, max_chain_len
, CTLFLAG_RW
,
154 &dn_max_ratio
, 0, "Max ratio between dynamic queues and buckets");
156 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, ready_heap
, CTLFLAG_RD
,
157 &ready_heap
.size
, 0, "Size of ready heap");
158 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, extract_heap
, CTLFLAG_RD
,
159 &extract_heap
.size
, 0, "Size of extract heap");
161 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, searches
, CTLFLAG_RD
,
162 &searches
, 0, "Number of queue searches");
163 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, search_steps
, CTLFLAG_RD
,
164 &search_steps
, 0, "Number of queue search steps");
166 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, red_lookup_depth
, CTLFLAG_RD
,
167 &red_lookup_depth
, 0, "Depth of RED lookup table");
168 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, red_avg_pkt_size
, CTLFLAG_RD
,
169 &red_avg_pkt_size
, 0, "RED Medium packet size");
170 SYSCTL_INT(_net_inet_ip_dummynet
, OID_AUTO
, red_max_pkt_size
, CTLFLAG_RD
,
171 &red_max_pkt_size
, 0, "RED Max packet size");
173 SYSCTL_PROC(_net_inet_ip_dummynet
, OID_AUTO
, hz
, CTLTYPE_INT
| CTLFLAG_RW
,
174 0, 0, sysctl_dn_hz
, "I", "Dummynet callout frequency");
176 static int heap_init(struct dn_heap
*, int);
177 static int heap_insert(struct dn_heap
*, dn_key
, void *);
178 static void heap_extract(struct dn_heap
*, void *);
180 static void transmit_event(struct dn_pipe
*);
181 static void ready_event(struct dn_flow_queue
*);
182 static void ready_event_wfq(struct dn_pipe
*);
184 static int config_pipe(struct dn_ioc_pipe
*);
185 static void dummynet_flush(void);
187 static void dummynet_clock(systimer_t
, struct intrframe
*);
188 static void dummynet(struct netmsg
*);
190 static struct dn_pipe
*dn_find_pipe(int);
191 static struct dn_flow_set
*dn_locate_flowset(int, int);
193 typedef void (*dn_pipe_iter_t
)(struct dn_pipe
*, void *);
194 static void dn_iterate_pipe(dn_pipe_iter_t
, void *);
196 typedef void (*dn_flowset_iter_t
)(struct dn_flow_set
*, void *);
197 static void dn_iterate_flowset(dn_flowset_iter_t
, void *);
199 static ip_dn_io_t dummynet_io
;
200 static ip_dn_ctl_t dummynet_ctl
;
203 * Heap management functions.
205 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
206 * Some macros help finding parent/children so we can optimize them.
208 * heap_init() is called to expand the heap when needed.
209 * Increment size in blocks of 16 entries.
210 * XXX failure to allocate a new element is a pretty bad failure
211 * as we basically stall a whole queue forever!!
212 * Returns 1 on error, 0 on success
214 #define HEAP_FATHER(x) (((x) - 1) / 2)
215 #define HEAP_LEFT(x) (2*(x) + 1)
216 #define HEAP_IS_LEFT(x) ((x) & 1)
217 #define HEAP_RIGHT(x) (2*(x) + 2)
218 #define HEAP_SWAP(a, b, buffer) { buffer = a; a = b; b = buffer; }
219 #define HEAP_INCREMENT 15
222 heap_init(struct dn_heap
*h
, int new_size
)
224 struct dn_heap_entry
*p
;
226 if (h
->size
>= new_size
) {
227 kprintf("%s, Bogus call, have %d want %d\n", __func__
,
232 new_size
= (new_size
+ HEAP_INCREMENT
) & ~HEAP_INCREMENT
;
233 p
= kmalloc(new_size
* sizeof(*p
), M_DUMMYNET
, M_WAITOK
| M_ZERO
);
235 bcopy(h
->p
, p
, h
->size
* sizeof(*p
));
236 kfree(h
->p
, M_DUMMYNET
);
244 * Insert element in heap. Normally, p != NULL, we insert p in
245 * a new position and bubble up. If p == NULL, then the element is
246 * already in place, and key is the position where to start the
248 * Returns 1 on failure (cannot allocate new heap entry)
250 * If offset > 0 the position (index, int) of the element in the heap is
251 * also stored in the element itself at the given offset in bytes.
253 #define SET_OFFSET(heap, node) \
254 if (heap->offset > 0) \
255 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node;
258 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
260 #define RESET_OFFSET(heap, node) \
261 if (heap->offset > 0) \
262 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1;
265 heap_insert(struct dn_heap
*h
, dn_key key1
, void *p
)
267 int son
= h
->elements
;
269 if (p
== NULL
) { /* Data already there, set starting point */
271 } else { /* Insert new element at the end, possibly resize */
273 if (son
== h
->size
) { /* Need resize... */
274 if (heap_init(h
, h
->elements
+ 1))
275 return 1; /* Failure... */
277 h
->p
[son
].object
= p
;
278 h
->p
[son
].key
= key1
;
282 while (son
> 0) { /* Bubble up */
283 int father
= HEAP_FATHER(son
);
284 struct dn_heap_entry tmp
;
286 if (DN_KEY_LT(h
->p
[father
].key
, h
->p
[son
].key
))
287 break; /* Found right position */
289 /* 'son' smaller than 'father', swap and repeat */
290 HEAP_SWAP(h
->p
[son
], h
->p
[father
], tmp
);
299 * Remove top element from heap, or obj if obj != NULL
302 heap_extract(struct dn_heap
*h
, void *obj
)
304 int child
, father
, max
= h
->elements
- 1;
307 kprintf("warning, extract from empty heap 0x%p\n", h
);
311 father
= 0; /* Default: move up smallest child */
312 if (obj
!= NULL
) { /* Extract specific element, index is at offset */
314 panic("%s from middle not supported on this heap!!!\n", __func__
);
316 father
= *((int *)((char *)obj
+ h
->offset
));
317 if (father
< 0 || father
>= h
->elements
) {
318 panic("%s father %d out of bound 0..%d\n", __func__
,
319 father
, h
->elements
);
322 RESET_OFFSET(h
, father
);
324 child
= HEAP_LEFT(father
); /* Left child */
325 while (child
<= max
) { /* Valid entry */
326 if (child
!= max
&& DN_KEY_LT(h
->p
[child
+ 1].key
, h
->p
[child
].key
))
327 child
= child
+ 1; /* Take right child, otherwise left */
328 h
->p
[father
] = h
->p
[child
];
329 SET_OFFSET(h
, father
);
331 child
= HEAP_LEFT(child
); /* Left child for next loop */
336 * Fill hole with last entry and bubble up, reusing the insert code
338 h
->p
[father
] = h
->p
[max
];
339 heap_insert(h
, father
, NULL
); /* This one cannot fail */
344 * heapify() will reorganize data inside an array to maintain the
345 * heap property. It is needed when we delete a bunch of entries.
348 heapify(struct dn_heap
*h
)
352 for (i
= 0; i
< h
->elements
; i
++)
353 heap_insert(h
, i
, NULL
);
357 * Cleanup the heap and free data structure
360 heap_free(struct dn_heap
*h
)
363 kfree(h
->p
, M_DUMMYNET
);
364 bzero(h
, sizeof(*h
));
368 * --- End of heap management functions ---
372 * Scheduler functions:
374 * transmit_event() is called when the delay-line needs to enter
375 * the scheduler, either because of existing pkts getting ready,
376 * or new packets entering the queue. The event handled is the delivery
377 * time of the packet.
379 * ready_event() does something similar with fixed-rate queues, and the
380 * event handled is the finish time of the head pkt.
382 * ready_event_wfq() does something similar with WF2Q queues, and the
383 * event handled is the start time of the head pkt.
385 * In all cases, we make sure that the data structures are consistent
386 * before passing pkts out, because this might trigger recursive
387 * invocations of the procedures.
390 transmit_event(struct dn_pipe
*pipe
)
394 while ((pkt
= TAILQ_FIRST(&pipe
->p_queue
)) &&
395 DN_KEY_LEQ(pkt
->output_time
, curr_time
)) {
396 TAILQ_REMOVE(&pipe
->p_queue
, pkt
, dn_next
);
397 ip_dn_packet_redispatch(pkt
);
401 * If there are leftover packets, put into the heap for next event
403 if ((pkt
= TAILQ_FIRST(&pipe
->p_queue
)) != NULL
) {
405 * XXX should check errors on heap_insert, by draining the
406 * whole pipe and hoping in the future we are more successful
408 heap_insert(&extract_heap
, pkt
->output_time
, pipe
);
413 * The following macro computes how many ticks we have to wait
414 * before being able to transmit a packet. The credit is taken from
415 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
417 #define SET_TICKS(pkt, q, p) \
418 (pkt->dn_m->m_pkthdr.len*8*dn_hz - (q)->numbytes + p->bandwidth - 1 ) / \
422 * Extract pkt from queue, compute output time (could be now)
423 * and put into delay line (p_queue)
426 move_pkt(struct dn_pkt
*pkt
, struct dn_flow_queue
*q
,
427 struct dn_pipe
*p
, int len
)
429 TAILQ_REMOVE(&q
->queue
, pkt
, dn_next
);
433 pkt
->output_time
= curr_time
+ p
->delay
;
435 TAILQ_INSERT_TAIL(&p
->p_queue
, pkt
, dn_next
);
439 * ready_event() is invoked every time the queue must enter the
440 * scheduler, either because the first packet arrives, or because
441 * a previously scheduled event fired.
442 * On invokation, drain as many pkts as possible (could be 0) and then
443 * if there are leftover packets reinsert the pkt in the scheduler.
446 ready_event(struct dn_flow_queue
*q
)
449 struct dn_pipe
*p
= q
->fs
->pipe
;
453 kprintf("ready_event- pipe is gone\n");
456 p_was_empty
= TAILQ_EMPTY(&p
->p_queue
);
459 * Schedule fixed-rate queues linked to this pipe:
460 * Account for the bw accumulated since last scheduling, then
461 * drain as many pkts as allowed by q->numbytes and move to
462 * the delay line (in p) computing output time.
463 * bandwidth==0 (no limit) means we can drain the whole queue,
464 * setting len_scaled = 0 does the job.
466 q
->numbytes
+= (curr_time
- q
->sched_time
) * p
->bandwidth
;
467 while ((pkt
= TAILQ_FIRST(&q
->queue
)) != NULL
) {
468 int len
= pkt
->dn_m
->m_pkthdr
.len
;
469 int len_scaled
= p
->bandwidth
? len
*8*dn_hz
: 0;
471 if (len_scaled
> q
->numbytes
)
473 q
->numbytes
-= len_scaled
;
474 move_pkt(pkt
, q
, p
, len
);
478 * If we have more packets queued, schedule next ready event
479 * (can only occur when bandwidth != 0, otherwise we would have
480 * flushed the whole queue in the previous loop).
481 * To this purpose we record the current time and compute how many
482 * ticks to go for the finish time of the packet.
484 if ((pkt
= TAILQ_FIRST(&q
->queue
)) != NULL
) {
485 /* This implies bandwidth != 0 */
486 dn_key t
= SET_TICKS(pkt
, q
, p
); /* ticks i have to wait */
488 q
->sched_time
= curr_time
;
491 * XXX should check errors on heap_insert, and drain the whole
492 * queue on error hoping next time we are luckier.
494 heap_insert(&ready_heap
, curr_time
+ t
, q
);
495 } else { /* RED needs to know when the queue becomes empty */
496 q
->q_time
= curr_time
;
501 * If the delay line was empty call transmit_event(p) now.
502 * Otherwise, the scheduler will take care of it.
509 * Called when we can transmit packets on WF2Q queues. Take pkts out of
510 * the queues at their start time, and enqueue into the delay line.
511 * Packets are drained until p->numbytes < 0. As long as
512 * len_scaled >= p->numbytes, the packet goes into the delay line
513 * with a deadline p->delay. For the last packet, if p->numbytes < 0,
514 * there is an additional delay.
517 ready_event_wfq(struct dn_pipe
*p
)
519 int p_was_empty
= TAILQ_EMPTY(&p
->p_queue
);
520 struct dn_heap
*sch
= &p
->scheduler_heap
;
521 struct dn_heap
*neh
= &p
->not_eligible_heap
;
523 p
->numbytes
+= (curr_time
- p
->sched_time
) * p
->bandwidth
;
526 * While we have backlogged traffic AND credit, we need to do
527 * something on the queue.
529 while (p
->numbytes
>= 0 && (sch
->elements
> 0 || neh
->elements
> 0)) {
530 if (sch
->elements
> 0) { /* Have some eligible pkts to send out */
531 struct dn_flow_queue
*q
= sch
->p
[0].object
;
532 struct dn_pkt
*pkt
= TAILQ_FIRST(&q
->queue
);
533 struct dn_flow_set
*fs
= q
->fs
;
534 uint64_t len
= pkt
->dn_m
->m_pkthdr
.len
;
535 int len_scaled
= p
->bandwidth
? len
*8*dn_hz
: 0;
537 heap_extract(sch
, NULL
); /* Remove queue from heap */
538 p
->numbytes
-= len_scaled
;
539 move_pkt(pkt
, q
, p
, len
);
541 p
->V
+= (len
<< MY_M
) / p
->sum
; /* Update V */
542 q
->S
= q
->F
; /* Update start time */
544 if (q
->len
== 0) { /* Flow not backlogged any more */
546 heap_insert(&p
->idle_heap
, q
->F
, q
);
547 } else { /* Still backlogged */
549 * Update F and position in backlogged queue, then
550 * put flow in not_eligible_heap (we will fix this later).
552 len
= TAILQ_FIRST(&q
->queue
)->dn_m
->m_pkthdr
.len
;
553 q
->F
+= (len
<< MY_M
) / (uint64_t)fs
->weight
;
554 if (DN_KEY_LEQ(q
->S
, p
->V
))
555 heap_insert(neh
, q
->S
, q
);
557 heap_insert(sch
, q
->F
, q
);
562 * Now compute V = max(V, min(S_i)). Remember that all elements in
563 * sch have by definition S_i <= V so if sch is not empty, V is surely
564 * the max and we must not update it. Conversely, if sch is empty
565 * we only need to look at neh.
567 if (sch
->elements
== 0 && neh
->elements
> 0)
568 p
->V
= MAX64(p
->V
, neh
->p
[0].key
);
571 * Move from neh to sch any packets that have become eligible
573 while (neh
->elements
> 0 && DN_KEY_LEQ(neh
->p
[0].key
, p
->V
)) {
574 struct dn_flow_queue
*q
= neh
->p
[0].object
;
576 heap_extract(neh
, NULL
);
577 heap_insert(sch
, q
->F
, q
);
581 if (sch
->elements
== 0 && neh
->elements
== 0 && p
->numbytes
>= 0 &&
582 p
->idle_heap
.elements
> 0) {
584 * No traffic and no events scheduled. We can get rid of idle-heap.
588 for (i
= 0; i
< p
->idle_heap
.elements
; i
++) {
589 struct dn_flow_queue
*q
= p
->idle_heap
.p
[i
].object
;
596 p
->idle_heap
.elements
= 0;
600 * If we are getting clocks from dummynet and if we are under credit,
601 * schedule the next ready event.
602 * Also fix the delivery time of the last packet.
604 if (p
->numbytes
< 0) { /* This implies bandwidth>0 */
605 dn_key t
= 0; /* Number of ticks i have to wait */
607 if (p
->bandwidth
> 0)
608 t
= (p
->bandwidth
- 1 - p
->numbytes
) / p
->bandwidth
;
609 TAILQ_LAST(&p
->p_queue
, dn_pkt_queue
)->output_time
+= t
;
610 p
->sched_time
= curr_time
;
613 * XXX should check errors on heap_insert, and drain the whole
614 * queue on error hoping next time we are luckier.
616 heap_insert(&wfq_ready_heap
, curr_time
+ t
, p
);
620 * If the delay line was empty call transmit_event(p) now.
621 * Otherwise, the scheduler will take care of it.
628 dn_expire_pipe_cb(struct dn_pipe
*pipe
, void *dummy __unused
)
630 if (pipe
->idle_heap
.elements
> 0 &&
631 DN_KEY_LT(pipe
->idle_heap
.p
[0].key
, pipe
->V
)) {
632 struct dn_flow_queue
*q
= pipe
->idle_heap
.p
[0].object
;
634 heap_extract(&pipe
->idle_heap
, NULL
);
635 q
->S
= q
->F
+ 1; /* Mark timestamp as invalid */
636 pipe
->sum
-= q
->fs
->weight
;
641 * This is called once per tick, or dn_hz times per second. It is used to
642 * increment the current tick counter and schedule expired events.
645 dummynet(struct netmsg
*msg
)
649 struct dn_heap
*heaps
[3];
652 heaps
[0] = &ready_heap
; /* Fixed-rate queues */
653 heaps
[1] = &wfq_ready_heap
; /* WF2Q queues */
654 heaps
[2] = &extract_heap
; /* Delay line */
658 lwkt_replymsg(&msg
->nm_lmsg
, 0);
662 for (i
= 0; i
< 3; i
++) {
664 while (h
->elements
> 0 && DN_KEY_LEQ(h
->p
[0].key
, curr_time
)) {
665 if (h
->p
[0].key
> curr_time
) {
666 kprintf("-- dummynet: warning, heap %d is %d ticks late\n",
667 i
, (int)(curr_time
- h
->p
[0].key
));
670 p
= h
->p
[0].object
; /* Store a copy before heap_extract */
671 heap_extract(h
, NULL
); /* Need to extract before processing */
682 /* Sweep pipes trying to expire idle flow_queues */
683 dn_iterate_pipe(dn_expire_pipe_cb
, NULL
);
687 * Unconditionally expire empty queues in case of shortage.
688 * Returns the number of queues freed.
691 expire_queues(struct dn_flow_set
*fs
)
693 int i
, initial_elements
= fs
->rq_elements
;
695 if (fs
->last_expired
== time_second
)
698 fs
->last_expired
= time_second
;
700 for (i
= 0; i
<= fs
->rq_size
; i
++) { /* Last one is overflow */
701 struct dn_flow_queue
*q
, *qn
;
703 LIST_FOREACH_MUTABLE(q
, &fs
->rq
[i
], q_link
, qn
) {
704 if (!TAILQ_EMPTY(&q
->queue
) || q
->S
!= q
->F
+ 1)
708 * Entry is idle, expire it
710 LIST_REMOVE(q
, q_link
);
711 kfree(q
, M_DUMMYNET
);
713 KASSERT(fs
->rq_elements
> 0,
714 ("invalid rq_elements %d\n", fs
->rq_elements
));
718 return initial_elements
- fs
->rq_elements
;
722 * If room, create a new queue and put at head of slot i;
723 * otherwise, create or use the default queue.
725 static struct dn_flow_queue
*
726 create_queue(struct dn_flow_set
*fs
, int i
)
728 struct dn_flow_queue
*q
;
730 if (fs
->rq_elements
> fs
->rq_size
* dn_max_ratio
&&
731 expire_queues(fs
) == 0) {
733 * No way to get room, use or create overflow queue.
736 if (!LIST_EMPTY(&fs
->rq
[i
]))
737 return LIST_FIRST(&fs
->rq
[i
]);
740 q
= kmalloc(sizeof(*q
), M_DUMMYNET
, M_INTWAIT
| M_NULLOK
| M_ZERO
);
746 q
->S
= q
->F
+ 1; /* hack - mark timestamp as invalid */
747 TAILQ_INIT(&q
->queue
);
749 LIST_INSERT_HEAD(&fs
->rq
[i
], q
, q_link
);
756 * Given a flow_set and a pkt in last_pkt, find a matching queue
757 * after appropriate masking. The queue is moved to front
758 * so that further searches take less time.
760 static struct dn_flow_queue
*
761 find_queue(struct dn_flow_set
*fs
, struct dn_flow_id
*id
)
763 struct dn_flow_queue
*q
;
766 if (!(fs
->flags_fs
& DN_HAVE_FLOW_MASK
)) {
767 q
= LIST_FIRST(&fs
->rq
[0]);
769 struct dn_flow_queue
*qn
;
771 /* First, do the masking */
772 id
->fid_dst_ip
&= fs
->flow_mask
.fid_dst_ip
;
773 id
->fid_src_ip
&= fs
->flow_mask
.fid_src_ip
;
774 id
->fid_dst_port
&= fs
->flow_mask
.fid_dst_port
;
775 id
->fid_src_port
&= fs
->flow_mask
.fid_src_port
;
776 id
->fid_proto
&= fs
->flow_mask
.fid_proto
;
777 id
->fid_flags
= 0; /* we don't care about this one */
779 /* Then, hash function */
780 i
= ((id
->fid_dst_ip
) & 0xffff) ^
781 ((id
->fid_dst_ip
>> 15) & 0xffff) ^
782 ((id
->fid_src_ip
<< 1) & 0xffff) ^
783 ((id
->fid_src_ip
>> 16 ) & 0xffff) ^
784 (id
->fid_dst_port
<< 1) ^ (id
->fid_src_port
) ^
789 * Finally, scan the current list for a match and
790 * expire idle flow queues
793 LIST_FOREACH_MUTABLE(q
, &fs
->rq
[i
], q_link
, qn
) {
795 if (id
->fid_dst_ip
== q
->id
.fid_dst_ip
&&
796 id
->fid_src_ip
== q
->id
.fid_src_ip
&&
797 id
->fid_dst_port
== q
->id
.fid_dst_port
&&
798 id
->fid_src_port
== q
->id
.fid_src_port
&&
799 id
->fid_proto
== q
->id
.fid_proto
&&
800 id
->fid_flags
== q
->id
.fid_flags
) {
802 } else if (pipe_expire
&& TAILQ_EMPTY(&q
->queue
) &&
805 * Entry is idle and not in any heap, expire it
807 LIST_REMOVE(q
, q_link
);
808 kfree(q
, M_DUMMYNET
);
810 KASSERT(fs
->rq_elements
> 0,
811 ("invalid rq_elements %d\n", fs
->rq_elements
));
815 if (q
&& LIST_FIRST(&fs
->rq
[i
]) != q
) { /* Found and not in front */
816 LIST_REMOVE(q
, q_link
);
817 LIST_INSERT_HEAD(&fs
->rq
[i
], q
, q_link
);
820 if (q
== NULL
) { /* No match, need to allocate a new entry */
821 q
= create_queue(fs
, i
);
829 red_drops(struct dn_flow_set
*fs
, struct dn_flow_queue
*q
, int len
)
834 * RED calculates the average queue size (avg) using a low-pass filter
835 * with an exponential weighted (w_q) moving average:
836 * avg <- (1-w_q) * avg + w_q * q_size
837 * where q_size is the queue length (measured in bytes or * packets).
839 * If q_size == 0, we compute the idle time for the link, and set
840 * avg = (1 - w_q)^(idle/s)
841 * where s is the time needed for transmitting a medium-sized packet.
843 * Now, if avg < min_th the packet is enqueued.
844 * If avg > max_th the packet is dropped. Otherwise, the packet is
845 * dropped with probability P function of avg.
849 u_int q_size
= (fs
->flags_fs
& DN_QSIZE_IS_BYTES
) ? q
->len_bytes
: q
->len
;
851 DPRINTF("\n%d q: %2u ", (int)curr_time
, q_size
);
853 /* Average queue size estimation */
856 * Queue is not empty, avg <- avg + (q_size - avg) * w_q
858 int diff
= SCALE(q_size
) - q
->avg
;
859 int64_t v
= SCALE_MUL((int64_t)diff
, (int64_t)fs
->w_q
);
864 * Queue is empty, find for how long the queue has been
865 * empty and use a lookup table for computing
866 * (1 - * w_q)^(idle_time/s) where s is the time to send a
871 u_int t
= (curr_time
- q
->q_time
) / fs
->lookup_step
;
873 q
->avg
= (t
< fs
->lookup_depth
) ?
874 SCALE_MUL(q
->avg
, fs
->w_q_lookup
[t
]) : 0;
877 DPRINTF("avg: %u ", SCALE_VAL(q
->avg
));
881 if (q
->avg
< fs
->min_th
) {
887 if (q
->avg
>= fs
->max_th
) { /* Average queue >= Max threshold */
888 if (fs
->flags_fs
& DN_IS_GENTLE_RED
) {
890 * According to Gentle-RED, if avg is greater than max_th the
891 * packet is dropped with a probability
892 * p_b = c_3 * avg - c_4
893 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p
895 p_b
= SCALE_MUL((int64_t)fs
->c_3
, (int64_t)q
->avg
) - fs
->c_4
;
901 } else if (q
->avg
> fs
->min_th
) {
903 * We compute p_b using the linear dropping function p_b = c_1 *
904 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 =
905 * max_p * min_th / (max_th - min_th)
907 p_b
= SCALE_MUL((int64_t)fs
->c_1
, (int64_t)q
->avg
) - fs
->c_2
;
909 if (fs
->flags_fs
& DN_QSIZE_IS_BYTES
)
910 p_b
= (p_b
* len
) / fs
->max_pkt_size
;
912 if (++q
->count
== 0) {
913 q
->random
= krandom() & 0xffff;
916 * q->count counts packets arrived since last drop, so a greater
917 * value of q->count means a greater packet drop probability.
919 if (SCALE_MUL(p_b
, SCALE((int64_t)q
->count
)) > q
->random
) {
921 DPRINTF("%s", "- red drop");
922 /* After a drop we calculate a new random value */
923 q
->random
= krandom() & 0xffff;
927 /* End of RED algorithm */
928 return 0; /* Accept */
932 dn_iterate_pipe(dn_pipe_iter_t func
, void *arg
)
936 for (i
= 0; i
< DN_NR_HASH_MAX
; ++i
) {
937 struct dn_pipe_head
*pipe_hdr
= &pipe_table
[i
];
938 struct dn_pipe
*pipe
, *pipe_next
;
940 LIST_FOREACH_MUTABLE(pipe
, pipe_hdr
, p_link
, pipe_next
)
946 dn_iterate_flowset(dn_flowset_iter_t func
, void *arg
)
950 for (i
= 0; i
< DN_NR_HASH_MAX
; ++i
) {
951 struct dn_flowset_head
*fs_hdr
= &flowset_table
[i
];
952 struct dn_flow_set
*fs
, *fs_next
;
954 LIST_FOREACH_MUTABLE(fs
, fs_hdr
, fs_link
, fs_next
)
959 static struct dn_pipe
*
960 dn_find_pipe(int pipe_nr
)
962 struct dn_pipe_head
*pipe_hdr
;
965 pipe_hdr
= &pipe_table
[DN_NR_HASH(pipe_nr
)];
966 LIST_FOREACH(p
, pipe_hdr
, p_link
) {
967 if (p
->pipe_nr
== pipe_nr
)
973 static struct dn_flow_set
*
974 dn_find_flowset(int fs_nr
)
976 struct dn_flowset_head
*fs_hdr
;
977 struct dn_flow_set
*fs
;
979 fs_hdr
= &flowset_table
[DN_NR_HASH(fs_nr
)];
980 LIST_FOREACH(fs
, fs_hdr
, fs_link
) {
981 if (fs
->fs_nr
== fs_nr
)
987 static struct dn_flow_set
*
988 dn_locate_flowset(int pipe_nr
, int is_pipe
)
990 struct dn_flow_set
*fs
= NULL
;
993 fs
= dn_find_flowset(pipe_nr
);
997 p
= dn_find_pipe(pipe_nr
);
1005 * Dummynet hook for packets. Below 'pipe' is a pipe or a queue
1006 * depending on whether WF2Q or fixed bw is used.
1008 * pipe_nr pipe or queue the packet is destined for.
1009 * dir where shall we send the packet after dummynet.
1010 * m the mbuf with the packet
1011 * fwa->oif the 'ifp' parameter from the caller.
1012 * NULL in ip_input, destination interface in ip_output
1013 * fwa->ro route parameter (only used in ip_output, NULL otherwise)
1014 * fwa->dst destination address, only used by ip_output
1015 * fwa->rule matching rule, in case of multiple passes
1016 * fwa->flags flags from the caller, only used in ip_output
1019 dummynet_io(struct mbuf
*m
)
1023 struct dn_flow_set
*fs
;
1024 struct dn_pipe
*pipe
;
1025 uint64_t len
= m
->m_pkthdr
.len
;
1026 struct dn_flow_queue
*q
= NULL
;
1027 int is_pipe
, pipe_nr
;
1029 tag
= m_tag_find(m
, PACKET_TAG_DUMMYNET
, NULL
);
1030 pkt
= m_tag_data(tag
);
1032 is_pipe
= pkt
->dn_flags
& DN_FLAGS_IS_PIPE
;
1033 pipe_nr
= pkt
->pipe_nr
;
1036 * This is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule
1038 fs
= dn_locate_flowset(pipe_nr
, is_pipe
);
1040 goto dropit
; /* This queue/pipe does not exist! */
1043 if (pipe
== NULL
) { /* Must be a queue, try find a matching pipe */
1044 pipe
= dn_find_pipe(fs
->parent_nr
);
1048 kprintf("No pipe %d for queue %d, drop pkt\n",
1049 fs
->parent_nr
, fs
->fs_nr
);
1054 q
= find_queue(fs
, &pkt
->id
);
1056 goto dropit
; /* Cannot allocate queue */
1059 * Update statistics, then check reasons to drop pkt
1061 q
->tot_bytes
+= len
;
1064 if (fs
->plr
&& krandom() < fs
->plr
)
1065 goto dropit
; /* Random pkt drop */
1067 if (fs
->flags_fs
& DN_QSIZE_IS_BYTES
) {
1068 if (q
->len_bytes
> fs
->qsize
)
1069 goto dropit
; /* Queue size overflow */
1071 if (q
->len
>= fs
->qsize
)
1072 goto dropit
; /* Queue count overflow */
1075 if ((fs
->flags_fs
& DN_IS_RED
) && red_drops(fs
, q
, len
))
1078 TAILQ_INSERT_TAIL(&q
->queue
, pkt
, dn_next
);
1080 q
->len_bytes
+= len
;
1082 if (TAILQ_FIRST(&q
->queue
) != pkt
) /* Flow was not idle, we are done */
1086 * If we reach this point the flow was previously idle, so we need
1087 * to schedule it. This involves different actions for fixed-rate
1092 * Fixed-rate queue: just insert into the ready_heap.
1096 if (pipe
->bandwidth
)
1097 t
= SET_TICKS(pkt
, q
, pipe
);
1099 q
->sched_time
= curr_time
;
1100 if (t
== 0) /* Must process it now */
1103 heap_insert(&ready_heap
, curr_time
+ t
, q
);
1107 * First, compute start time S: if the flow was idle (S=F+1)
1108 * set S to the virtual time V for the controlling pipe, and update
1109 * the sum of weights for the pipe; otherwise, remove flow from
1110 * idle_heap and set S to max(F, V).
1111 * Second, compute finish time F = S + len/weight.
1112 * Third, if pipe was idle, update V = max(S, V).
1113 * Fourth, count one more backlogged flow.
1115 if (DN_KEY_GT(q
->S
, q
->F
)) { /* Means timestamps are invalid */
1117 pipe
->sum
+= fs
->weight
; /* Add weight of new queue */
1119 heap_extract(&pipe
->idle_heap
, q
);
1120 q
->S
= MAX64(q
->F
, pipe
->V
);
1122 q
->F
= q
->S
+ (len
<< MY_M
) / (uint64_t)fs
->weight
;
1124 if (pipe
->not_eligible_heap
.elements
== 0 &&
1125 pipe
->scheduler_heap
.elements
== 0)
1126 pipe
->V
= MAX64(q
->S
, pipe
->V
);
1131 * Look at eligibility. A flow is not eligibile if S>V (when
1132 * this happens, it means that there is some other flow already
1133 * scheduled for the same pipe, so the scheduler_heap cannot be
1134 * empty). If the flow is not eligible we just store it in the
1135 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1136 * and possibly invoke ready_event_wfq() right now if there is
1138 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1139 * and for all flows in not_eligible_heap (NEH), S_i > V.
1140 * So when we need to compute max(V, min(S_i)) forall i in SCH+NEH,
1141 * we only need to look into NEH.
1143 if (DN_KEY_GT(q
->S
, pipe
->V
)) { /* Not eligible */
1144 if (pipe
->scheduler_heap
.elements
== 0)
1145 kprintf("++ ouch! not eligible but empty scheduler!\n");
1146 heap_insert(&pipe
->not_eligible_heap
, q
->S
, q
);
1148 heap_insert(&pipe
->scheduler_heap
, q
->F
, q
);
1149 if (pipe
->numbytes
>= 0) { /* Pipe is idle */
1150 if (pipe
->scheduler_heap
.elements
!= 1)
1151 kprintf("*** OUCH! pipe should have been idle!\n");
1152 DPRINTF("Waking up pipe %d at %d\n",
1153 pipe
->pipe_nr
, (int)(q
->F
>> MY_M
));
1154 pipe
->sched_time
= curr_time
;
1155 ready_event_wfq(pipe
);
1169 * Dispose all packets and flow_queues on a flow_set.
1170 * If all=1, also remove red lookup table and other storage,
1171 * including the descriptor itself.
1172 * For the one in dn_pipe MUST also cleanup ready_heap...
1175 purge_flow_set(struct dn_flow_set
*fs
, int all
)
1179 int rq_elements
= 0;
1182 for (i
= 0; i
<= fs
->rq_size
; i
++) {
1183 struct dn_flow_queue
*q
;
1185 while ((q
= LIST_FIRST(&fs
->rq
[i
])) != NULL
) {
1188 while ((pkt
= TAILQ_FIRST(&q
->queue
)) != NULL
) {
1189 TAILQ_REMOVE(&q
->queue
, pkt
, dn_next
);
1190 ip_dn_packet_free(pkt
);
1193 LIST_REMOVE(q
, q_link
);
1194 kfree(q
, M_DUMMYNET
);
1201 KASSERT(rq_elements
== fs
->rq_elements
,
1202 ("# rq elements mismatch, freed %d, total %d\n",
1203 rq_elements
, fs
->rq_elements
));
1204 fs
->rq_elements
= 0;
1207 /* RED - free lookup table */
1209 kfree(fs
->w_q_lookup
, M_DUMMYNET
);
1212 kfree(fs
->rq
, M_DUMMYNET
);
1215 * If this fs is not part of a pipe, free it
1217 * fs->pipe == NULL could happen, if 'fs' is a WF2Q and
1218 * - No packet belongs to that flow set is delivered by
1219 * dummynet_io(), i.e. parent pipe is not installed yet.
1220 * - Parent pipe is deleted.
1222 if (fs
->pipe
== NULL
|| (fs
->pipe
&& fs
!= &fs
->pipe
->fs
))
1223 kfree(fs
, M_DUMMYNET
);
1228 * Dispose all packets queued on a pipe (not a flow_set).
1229 * Also free all resources associated to a pipe, which is about
1233 purge_pipe(struct dn_pipe
*pipe
)
1237 purge_flow_set(&pipe
->fs
, 1);
1239 while ((pkt
= TAILQ_FIRST(&pipe
->p_queue
)) != NULL
) {
1240 TAILQ_REMOVE(&pipe
->p_queue
, pkt
, dn_next
);
1241 ip_dn_packet_free(pkt
);
1244 heap_free(&pipe
->scheduler_heap
);
1245 heap_free(&pipe
->not_eligible_heap
);
1246 heap_free(&pipe
->idle_heap
);
1250 * Delete all pipes and heaps returning memory.
1253 dummynet_flush(void)
1255 struct dn_pipe_head pipe_list
;
1256 struct dn_flowset_head fs_list
;
1258 struct dn_flow_set
*fs
;
1262 * Prevent future matches...
1264 LIST_INIT(&pipe_list
);
1265 for (i
= 0; i
< DN_NR_HASH_MAX
; ++i
) {
1266 struct dn_pipe_head
*pipe_hdr
= &pipe_table
[i
];
1268 while ((p
= LIST_FIRST(pipe_hdr
)) != NULL
) {
1269 LIST_REMOVE(p
, p_link
);
1270 LIST_INSERT_HEAD(&pipe_list
, p
, p_link
);
1274 LIST_INIT(&fs_list
);
1275 for (i
= 0; i
< DN_NR_HASH_MAX
; ++i
) {
1276 struct dn_flowset_head
*fs_hdr
= &flowset_table
[i
];
1278 while ((fs
= LIST_FIRST(fs_hdr
)) != NULL
) {
1279 LIST_REMOVE(fs
, fs_link
);
1280 LIST_INSERT_HEAD(&fs_list
, fs
, fs_link
);
1284 /* Free heaps so we don't have unwanted events */
1285 heap_free(&ready_heap
);
1286 heap_free(&wfq_ready_heap
);
1287 heap_free(&extract_heap
);
1290 * Now purge all queued pkts and delete all pipes
1292 /* Scan and purge all flow_sets. */
1293 while ((fs
= LIST_FIRST(&fs_list
)) != NULL
) {
1294 LIST_REMOVE(fs
, fs_link
);
1295 purge_flow_set(fs
, 1);
1298 while ((p
= LIST_FIRST(&pipe_list
)) != NULL
) {
1299 LIST_REMOVE(p
, p_link
);
1301 kfree(p
, M_DUMMYNET
);
1306 * setup RED parameters
1309 config_red(const struct dn_ioc_flowset
*ioc_fs
, struct dn_flow_set
*x
)
1313 x
->w_q
= ioc_fs
->w_q
;
1314 x
->min_th
= SCALE(ioc_fs
->min_th
);
1315 x
->max_th
= SCALE(ioc_fs
->max_th
);
1316 x
->max_p
= ioc_fs
->max_p
;
1318 x
->c_1
= ioc_fs
->max_p
/ (ioc_fs
->max_th
- ioc_fs
->min_th
);
1319 x
->c_2
= SCALE_MUL(x
->c_1
, SCALE(ioc_fs
->min_th
));
1320 if (x
->flags_fs
& DN_IS_GENTLE_RED
) {
1321 x
->c_3
= (SCALE(1) - ioc_fs
->max_p
) / ioc_fs
->max_th
;
1322 x
->c_4
= (SCALE(1) - 2 * ioc_fs
->max_p
);
1325 /* If the lookup table already exist, free and create it again */
1326 if (x
->w_q_lookup
) {
1327 kfree(x
->w_q_lookup
, M_DUMMYNET
);
1328 x
->w_q_lookup
= NULL
;
1331 if (red_lookup_depth
== 0) {
1332 kprintf("net.inet.ip.dummynet.red_lookup_depth must be > 0\n");
1333 kfree(x
, M_DUMMYNET
);
1336 x
->lookup_depth
= red_lookup_depth
;
1337 x
->w_q_lookup
= kmalloc(x
->lookup_depth
* sizeof(int),
1338 M_DUMMYNET
, M_WAITOK
);
1340 /* Fill the lookup table with (1 - w_q)^x */
1341 x
->lookup_step
= ioc_fs
->lookup_step
;
1342 x
->lookup_weight
= ioc_fs
->lookup_weight
;
1344 x
->w_q_lookup
[0] = SCALE(1) - x
->w_q
;
1345 for (i
= 1; i
< x
->lookup_depth
; i
++)
1346 x
->w_q_lookup
[i
] = SCALE_MUL(x
->w_q_lookup
[i
- 1], x
->lookup_weight
);
1348 if (red_avg_pkt_size
< 1)
1349 red_avg_pkt_size
= 512;
1350 x
->avg_pkt_size
= red_avg_pkt_size
;
1352 if (red_max_pkt_size
< 1)
1353 red_max_pkt_size
= 1500;
1354 x
->max_pkt_size
= red_max_pkt_size
;
1360 alloc_hash(struct dn_flow_set
*x
, const struct dn_ioc_flowset
*ioc_fs
)
1364 if (x
->flags_fs
& DN_HAVE_FLOW_MASK
) {
1365 int l
= ioc_fs
->rq_size
;
1367 /* Allocate some slots */
1371 if (l
< DN_MIN_HASH_SIZE
)
1372 l
= DN_MIN_HASH_SIZE
;
1373 else if (l
> DN_MAX_HASH_SIZE
)
1374 l
= DN_MAX_HASH_SIZE
;
1378 /* One is enough for null mask */
1381 alloc_size
= x
->rq_size
+ 1;
1383 x
->rq
= kmalloc(alloc_size
* sizeof(struct dn_flowqueue_head
),
1384 M_DUMMYNET
, M_WAITOK
| M_ZERO
);
1387 for (i
= 0; i
< alloc_size
; ++i
)
1388 LIST_INIT(&x
->rq
[i
]);
1392 set_flowid_parms(struct dn_flow_id
*id
, const struct dn_ioc_flowid
*ioc_id
)
1394 id
->fid_dst_ip
= ioc_id
->u
.ip
.dst_ip
;
1395 id
->fid_src_ip
= ioc_id
->u
.ip
.src_ip
;
1396 id
->fid_dst_port
= ioc_id
->u
.ip
.dst_port
;
1397 id
->fid_src_port
= ioc_id
->u
.ip
.src_port
;
1398 id
->fid_proto
= ioc_id
->u
.ip
.proto
;
1399 id
->fid_flags
= ioc_id
->u
.ip
.flags
;
1403 set_fs_parms(struct dn_flow_set
*x
, const struct dn_ioc_flowset
*ioc_fs
)
1405 x
->flags_fs
= ioc_fs
->flags_fs
;
1406 x
->qsize
= ioc_fs
->qsize
;
1407 x
->plr
= ioc_fs
->plr
;
1408 set_flowid_parms(&x
->flow_mask
, &ioc_fs
->flow_mask
);
1409 if (x
->flags_fs
& DN_QSIZE_IS_BYTES
) {
1410 if (x
->qsize
> 1024 * 1024)
1411 x
->qsize
= 1024 * 1024;
1413 if (x
->qsize
== 0 || x
->qsize
> 100)
1417 /* Configuring RED */
1418 if (x
->flags_fs
& DN_IS_RED
)
1419 config_red(ioc_fs
, x
); /* XXX should check errors */
1423 * setup pipe or queue parameters.
1427 config_pipe(struct dn_ioc_pipe
*ioc_pipe
)
1429 struct dn_ioc_flowset
*ioc_fs
= &ioc_pipe
->fs
;
1433 * The config program passes parameters as follows:
1434 * bw bits/second (0 means no limits)
1435 * delay ms (must be translated into ticks)
1436 * qsize slots or bytes
1438 ioc_pipe
->delay
= (ioc_pipe
->delay
* dn_hz
) / 1000;
1441 * We need either a pipe number or a flow_set number
1443 if (ioc_pipe
->pipe_nr
== 0 && ioc_fs
->fs_nr
== 0)
1445 if (ioc_pipe
->pipe_nr
!= 0 && ioc_fs
->fs_nr
!= 0)
1449 * Validate pipe number
1451 if (ioc_pipe
->pipe_nr
> DN_PIPE_NR_MAX
|| ioc_pipe
->pipe_nr
< 0)
1455 if (ioc_pipe
->pipe_nr
!= 0) { /* This is a pipe */
1456 struct dn_pipe
*x
, *p
;
1459 p
= dn_find_pipe(ioc_pipe
->pipe_nr
);
1461 if (p
== NULL
) { /* New pipe */
1462 x
= kmalloc(sizeof(struct dn_pipe
), M_DUMMYNET
, M_WAITOK
| M_ZERO
);
1463 x
->pipe_nr
= ioc_pipe
->pipe_nr
;
1465 TAILQ_INIT(&x
->p_queue
);
1468 * idle_heap is the only one from which we extract from the middle.
1470 x
->idle_heap
.size
= x
->idle_heap
.elements
= 0;
1471 x
->idle_heap
.offset
= __offsetof(struct dn_flow_queue
, heap_pos
);
1477 /* Flush accumulated credit for all queues */
1478 for (i
= 0; i
<= x
->fs
.rq_size
; i
++) {
1479 struct dn_flow_queue
*q
;
1481 LIST_FOREACH(q
, &x
->fs
.rq
[i
], q_link
)
1486 x
->bandwidth
= ioc_pipe
->bandwidth
;
1487 x
->numbytes
= 0; /* Just in case... */
1488 x
->delay
= ioc_pipe
->delay
;
1490 set_fs_parms(&x
->fs
, ioc_fs
);
1492 if (x
->fs
.rq
== NULL
) { /* A new pipe */
1493 struct dn_pipe_head
*pipe_hdr
;
1495 alloc_hash(&x
->fs
, ioc_fs
);
1497 pipe_hdr
= &pipe_table
[DN_NR_HASH(x
->pipe_nr
)];
1498 LIST_INSERT_HEAD(pipe_hdr
, x
, p_link
);
1500 } else { /* Config flow_set */
1501 struct dn_flow_set
*x
, *fs
;
1503 /* Locate flow_set */
1504 fs
= dn_find_flowset(ioc_fs
->fs_nr
);
1506 if (fs
== NULL
) { /* New flow_set */
1507 if (ioc_fs
->parent_nr
== 0) /* Need link to a pipe */
1510 x
= kmalloc(sizeof(struct dn_flow_set
), M_DUMMYNET
,
1512 x
->fs_nr
= ioc_fs
->fs_nr
;
1513 x
->parent_nr
= ioc_fs
->parent_nr
;
1514 x
->weight
= ioc_fs
->weight
;
1517 else if (x
->weight
> 100)
1520 /* Change parent pipe not allowed; must delete and recreate */
1521 if (ioc_fs
->parent_nr
!= 0 && fs
->parent_nr
!= ioc_fs
->parent_nr
)
1526 set_fs_parms(x
, ioc_fs
);
1528 if (x
->rq
== NULL
) { /* A new flow_set */
1529 struct dn_flowset_head
*fs_hdr
;
1531 alloc_hash(x
, ioc_fs
);
1533 fs_hdr
= &flowset_table
[DN_NR_HASH(x
->fs_nr
)];
1534 LIST_INSERT_HEAD(fs_hdr
, x
, fs_link
);
1544 * Helper function to remove from a heap queues which are linked to
1545 * a flow_set about to be deleted.
1548 fs_remove_from_heap(struct dn_heap
*h
, struct dn_flow_set
*fs
)
1550 int i
= 0, found
= 0;
1552 while (i
< h
->elements
) {
1553 if (((struct dn_flow_queue
*)h
->p
[i
].object
)->fs
== fs
) {
1555 h
->p
[i
] = h
->p
[h
->elements
];
1566 * helper function to remove a pipe from a heap (can be there at most once)
1569 pipe_remove_from_heap(struct dn_heap
*h
, struct dn_pipe
*p
)
1571 if (h
->elements
> 0) {
1574 for (i
= 0; i
< h
->elements
; i
++) {
1575 if (h
->p
[i
].object
== p
) { /* found it */
1577 h
->p
[i
] = h
->p
[h
->elements
];
1586 dn_unref_pipe_cb(struct dn_flow_set
*fs
, void *pipe0
)
1588 struct dn_pipe
*pipe
= pipe0
;
1590 if (fs
->pipe
== pipe
) {
1591 kprintf("++ ref to pipe %d from fs %d\n",
1592 pipe
->pipe_nr
, fs
->fs_nr
);
1594 purge_flow_set(fs
, 0);
1599 * Fully delete a pipe or a queue, cleaning up associated info.
1602 delete_pipe(const struct dn_ioc_pipe
*ioc_pipe
)
1607 if (ioc_pipe
->pipe_nr
== 0 && ioc_pipe
->fs
.fs_nr
== 0)
1609 if (ioc_pipe
->pipe_nr
!= 0 && ioc_pipe
->fs
.fs_nr
!= 0)
1612 if (ioc_pipe
->pipe_nr
> DN_NR_HASH_MAX
|| ioc_pipe
->pipe_nr
< 0)
1616 if (ioc_pipe
->pipe_nr
!= 0) { /* This is an old-style pipe */
1618 p
= dn_find_pipe(ioc_pipe
->pipe_nr
);
1620 goto back
; /* Not found */
1622 /* Unlink from pipe hash table */
1623 LIST_REMOVE(p
, p_link
);
1625 /* Remove all references to this pipe from flow_sets */
1626 dn_iterate_flowset(dn_unref_pipe_cb
, p
);
1628 fs_remove_from_heap(&ready_heap
, &p
->fs
);
1629 purge_pipe(p
); /* Remove all data associated to this pipe */
1631 /* Remove reference to here from extract_heap and wfq_ready_heap */
1632 pipe_remove_from_heap(&extract_heap
, p
);
1633 pipe_remove_from_heap(&wfq_ready_heap
, p
);
1635 kfree(p
, M_DUMMYNET
);
1636 } else { /* This is a WF2Q queue (dn_flow_set) */
1637 struct dn_flow_set
*fs
;
1639 /* Locate flow_set */
1640 fs
= dn_find_flowset(ioc_pipe
->fs
.fs_nr
);
1642 goto back
; /* Not found */
1644 LIST_REMOVE(fs
, fs_link
);
1646 if ((p
= fs
->pipe
) != NULL
) {
1647 /* Update total weight on parent pipe and cleanup parent heaps */
1648 p
->sum
-= fs
->weight
* fs
->backlogged
;
1649 fs_remove_from_heap(&p
->not_eligible_heap
, fs
);
1650 fs_remove_from_heap(&p
->scheduler_heap
, fs
);
1651 #if 1 /* XXX should i remove from idle_heap as well ? */
1652 fs_remove_from_heap(&p
->idle_heap
, fs
);
1655 purge_flow_set(fs
, 1);
1664 * helper function used to copy data from kernel in DUMMYNET_GET
1667 dn_copy_flowid(const struct dn_flow_id
*id
, struct dn_ioc_flowid
*ioc_id
)
1669 ioc_id
->type
= ETHERTYPE_IP
;
1670 ioc_id
->u
.ip
.dst_ip
= id
->fid_dst_ip
;
1671 ioc_id
->u
.ip
.src_ip
= id
->fid_src_ip
;
1672 ioc_id
->u
.ip
.dst_port
= id
->fid_dst_port
;
1673 ioc_id
->u
.ip
.src_port
= id
->fid_src_port
;
1674 ioc_id
->u
.ip
.proto
= id
->fid_proto
;
1675 ioc_id
->u
.ip
.flags
= id
->fid_flags
;
1679 dn_copy_flowqueues(const struct dn_flow_set
*fs
, void *bp
)
1681 struct dn_ioc_flowqueue
*ioc_fq
= bp
;
1684 for (i
= 0; i
<= fs
->rq_size
; i
++) {
1685 const struct dn_flow_queue
*q
;
1687 LIST_FOREACH(q
, &fs
->rq
[i
], q_link
) {
1688 if (q
->hash_slot
!= i
) { /* XXX ASSERT */
1689 kprintf("++ at %d: wrong slot (have %d, "
1690 "should be %d)\n", copied
, q
->hash_slot
, i
);
1692 if (q
->fs
!= fs
) { /* XXX ASSERT */
1693 kprintf("++ at %d: wrong fs ptr (have %p, should be %p)\n",
1699 ioc_fq
->len
= q
->len
;
1700 ioc_fq
->len_bytes
= q
->len_bytes
;
1701 ioc_fq
->tot_pkts
= q
->tot_pkts
;
1702 ioc_fq
->tot_bytes
= q
->tot_bytes
;
1703 ioc_fq
->drops
= q
->drops
;
1704 ioc_fq
->hash_slot
= q
->hash_slot
;
1707 dn_copy_flowid(&q
->id
, &ioc_fq
->id
);
1713 if (copied
!= fs
->rq_elements
) { /* XXX ASSERT */
1714 kprintf("++ wrong count, have %d should be %d\n",
1715 copied
, fs
->rq_elements
);
1721 dn_copy_flowset(const struct dn_flow_set
*fs
, struct dn_ioc_flowset
*ioc_fs
,
1724 ioc_fs
->fs_type
= fs_type
;
1726 ioc_fs
->fs_nr
= fs
->fs_nr
;
1727 ioc_fs
->flags_fs
= fs
->flags_fs
;
1728 ioc_fs
->parent_nr
= fs
->parent_nr
;
1730 ioc_fs
->weight
= fs
->weight
;
1731 ioc_fs
->qsize
= fs
->qsize
;
1732 ioc_fs
->plr
= fs
->plr
;
1734 ioc_fs
->rq_size
= fs
->rq_size
;
1735 ioc_fs
->rq_elements
= fs
->rq_elements
;
1737 ioc_fs
->w_q
= fs
->w_q
;
1738 ioc_fs
->max_th
= fs
->max_th
;
1739 ioc_fs
->min_th
= fs
->min_th
;
1740 ioc_fs
->max_p
= fs
->max_p
;
1742 dn_copy_flowid(&fs
->flow_mask
, &ioc_fs
->flow_mask
);
1746 dn_calc_pipe_size_cb(struct dn_pipe
*pipe
, void *sz
)
1750 *size
+= sizeof(struct dn_ioc_pipe
) +
1751 pipe
->fs
.rq_elements
* sizeof(struct dn_ioc_flowqueue
);
1755 dn_calc_fs_size_cb(struct dn_flow_set
*fs
, void *sz
)
1759 *size
+= sizeof(struct dn_ioc_flowset
) +
1760 fs
->rq_elements
* sizeof(struct dn_ioc_flowqueue
);
1764 dn_copyout_pipe_cb(struct dn_pipe
*pipe
, void *bp0
)
1767 struct dn_ioc_pipe
*ioc_pipe
= (struct dn_ioc_pipe
*)(*bp
);
1770 * Copy flow set descriptor associated with this pipe
1772 dn_copy_flowset(&pipe
->fs
, &ioc_pipe
->fs
, DN_IS_PIPE
);
1775 * Copy pipe descriptor
1777 ioc_pipe
->bandwidth
= pipe
->bandwidth
;
1778 ioc_pipe
->pipe_nr
= pipe
->pipe_nr
;
1779 ioc_pipe
->V
= pipe
->V
;
1780 /* Convert delay to milliseconds */
1781 ioc_pipe
->delay
= (pipe
->delay
* 1000) / dn_hz
;
1784 * Copy flow queue descriptors
1786 *bp
+= sizeof(*ioc_pipe
);
1787 *bp
= dn_copy_flowqueues(&pipe
->fs
, *bp
);
1791 dn_copyout_fs_cb(struct dn_flow_set
*fs
, void *bp0
)
1794 struct dn_ioc_flowset
*ioc_fs
= (struct dn_ioc_flowset
*)(*bp
);
1797 * Copy flow set descriptor
1799 dn_copy_flowset(fs
, ioc_fs
, DN_IS_QUEUE
);
1802 * Copy flow queue descriptors
1804 *bp
+= sizeof(*ioc_fs
);
1805 *bp
= dn_copy_flowqueues(fs
, *bp
);
1809 dummynet_get(struct dn_sopt
*dn_sopt
)
1815 * Compute size of data structures: list of pipes and flow_sets.
1817 dn_iterate_pipe(dn_calc_pipe_size_cb
, &size
);
1818 dn_iterate_flowset(dn_calc_fs_size_cb
, &size
);
1821 * Copyout pipe/flow_set/flow_queue
1823 bp
= buf
= kmalloc(size
, M_TEMP
, M_WAITOK
| M_ZERO
);
1824 dn_iterate_pipe(dn_copyout_pipe_cb
, &bp
);
1825 dn_iterate_flowset(dn_copyout_fs_cb
, &bp
);
1827 /* Temp memory will be freed by caller */
1828 dn_sopt
->dn_sopt_arg
= buf
;
1829 dn_sopt
->dn_sopt_arglen
= size
;
1834 * Handler for the various dummynet socket options (get, flush, config, del)
1837 dummynet_ctl(struct dn_sopt
*dn_sopt
)
1841 switch (dn_sopt
->dn_sopt_name
) {
1842 case IP_DUMMYNET_GET
:
1843 error
= dummynet_get(dn_sopt
);
1846 case IP_DUMMYNET_FLUSH
:
1850 case IP_DUMMYNET_CONFIGURE
:
1851 KKASSERT(dn_sopt
->dn_sopt_arglen
== sizeof(struct dn_ioc_pipe
));
1852 error
= config_pipe(dn_sopt
->dn_sopt_arg
);
1855 case IP_DUMMYNET_DEL
: /* Remove a pipe or flow_set */
1856 KKASSERT(dn_sopt
->dn_sopt_arglen
== sizeof(struct dn_ioc_pipe
));
1857 error
= delete_pipe(dn_sopt
->dn_sopt_arg
);
1861 kprintf("%s -- unknown option %d\n", __func__
, dn_sopt
->dn_sopt_name
);
1869 dummynet_clock(systimer_t info __unused
, struct intrframe
*frame __unused
)
1871 KASSERT(mycpuid
== ip_dn_cpu
,
1872 ("dummynet systimer comes on cpu%d, should be %d!\n",
1873 mycpuid
, ip_dn_cpu
));
1876 if (DUMMYNET_LOADED
&& (dn_netmsg
.nm_lmsg
.ms_flags
& MSGF_DONE
))
1877 lwkt_sendmsg(cpu_portfn(mycpuid
), &dn_netmsg
.nm_lmsg
);
1882 sysctl_dn_hz(SYSCTL_HANDLER_ARGS
)
1887 error
= sysctl_handle_int(oidp
, &val
, 0, req
);
1888 if (error
|| req
->newptr
== NULL
)
1892 else if (val
> DN_CALLOUT_FREQ_MAX
)
1893 val
= DN_CALLOUT_FREQ_MAX
;
1897 systimer_adjust_periodic(&dn_clock
, val
);
1904 ip_dn_init_dispatch(struct netmsg
*msg
)
1908 KASSERT(mycpuid
== ip_dn_cpu
,
1909 ("%s runs on cpu%d, instead of cpu%d", __func__
,
1910 mycpuid
, ip_dn_cpu
));
1914 if (DUMMYNET_LOADED
) {
1915 kprintf("DUMMYNET already loaded\n");
1920 kprintf("DUMMYNET initialized (011031)\n");
1922 for (i
= 0; i
< DN_NR_HASH_MAX
; ++i
)
1923 LIST_INIT(&pipe_table
[i
]);
1925 for (i
= 0; i
< DN_NR_HASH_MAX
; ++i
)
1926 LIST_INIT(&flowset_table
[i
]);
1928 ready_heap
.size
= ready_heap
.elements
= 0;
1929 ready_heap
.offset
= 0;
1931 wfq_ready_heap
.size
= wfq_ready_heap
.elements
= 0;
1932 wfq_ready_heap
.offset
= 0;
1934 extract_heap
.size
= extract_heap
.elements
= 0;
1935 extract_heap
.offset
= 0;
1937 ip_dn_ctl_ptr
= dummynet_ctl
;
1938 ip_dn_io_ptr
= dummynet_io
;
1940 netmsg_init(&dn_netmsg
, &netisr_adone_rport
, 0, dummynet
);
1941 systimer_init_periodic_nq(&dn_clock
, dummynet_clock
, NULL
, dn_hz
);
1945 lwkt_replymsg(&msg
->nm_lmsg
, error
);
1949 ip_dn_stop_dispatch(struct netmsg
*msg
)
1955 ip_dn_ctl_ptr
= NULL
;
1956 ip_dn_io_ptr
= NULL
;
1958 systimer_del(&dn_clock
);
1961 lwkt_replymsg(&msg
->nm_lmsg
, 0);
1969 if (ip_dn_cpu
>= ncpus
) {
1970 kprintf("%s: CPU%d does not exist, switch to CPU0\n",
1971 __func__
, ip_dn_cpu
);
1975 netmsg_init(&smsg
, &curthread
->td_msgport
, 0, ip_dn_init_dispatch
);
1976 lwkt_domsg(cpu_portfn(ip_dn_cpu
), &smsg
.nm_lmsg
, 0);
1977 return smsg
.nm_lmsg
.ms_error
;
1985 netmsg_init(&smsg
, &curthread
->td_msgport
, 0, ip_dn_stop_dispatch
);
1986 lwkt_domsg(cpu_portfn(ip_dn_cpu
), &smsg
.nm_lmsg
, 0);
1988 netmsg_service_sync();
1992 dummynet_modevent(module_t mod
, int type
, void *data
)
1996 return ip_dn_init();
2000 kprintf("dummynet statically compiled, cannot unload\n");
2013 static moduledata_t dummynet_mod
= {
2018 DECLARE_MODULE(dummynet
, dummynet_mod
, SI_SUB_PROTO_END
, SI_ORDER_ANY
);
2019 MODULE_VERSION(dummynet
, 1);