1 /*-
2 * Copyright (c) 1998-2002 Luigi Rizzo, Universita` di Pisa
3 * Portions Copyright (c) 2000 Akamba Corp.
4 * All rights reserved
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
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.
14 *
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
25 * SUCH DAMAGE.
26 */
27
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30
31 #define DUMMYNET_DEBUG
32
33 #include "opt_inet6.h"
34
35 /*
36 * This module implements IP dummynet, a bandwidth limiter/delay emulator
37 * used in conjunction with the ipfw package.
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.
44 *
45 * NOTA BENE: critical sections are protected by the "dummynet lock".
46 *
47 * Most important Changes:
48 *
49 * 011004: KLDable
50 * 010124: Fixed WF2Q behaviour
51 * 010122: Fixed spl protection.
52 * 000601: WF2Q support
53 * 000106: large rewrite, use heaps to handle very many pipes.
54 * 980513: initial release
55 *
56 * include files marked with XXX are probably not needed
57 */
58
59 #include <sys/limits.h>
60 #include <sys/param.h>
61 #include <sys/systm.h>
62 #include <sys/malloc.h>
63 #include <sys/mbuf.h>
64 #include <sys/kernel.h>
65 #include <sys/lock.h>
66 #include <sys/module.h>
67 #include <sys/mutex.h>
68 #include <sys/priv.h>
69 #include <sys/proc.h>
70 #include <sys/socket.h>
71 #include <sys/socketvar.h>
72 #include <sys/time.h>
73 #include <sys/sysctl.h>
74 #include <sys/taskqueue.h>
75 #include <net/if.h> /* IFNAMSIZ, struct ifaddr, ifq head */
76 #include <net/netisr.h>
77 #include <netinet/in.h>
78 #include <netinet/ip.h> /* ip_len, ip_off */
79 #include <netinet/ip_fw.h>
80 #include <netinet/ip_dummynet.h>
81 #include <netinet/ip_var.h> /* ip_output(), IP_FORWARDING */
82
83 #include <netinet/if_ether.h> /* various ether_* routines */
84
85 #include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */
86 #include <netinet6/ip6_var.h>
87
88 /*
89 * We keep a private variable for the simulation time, but we could
90 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c)
91 */
92 static dn_key curr_time = 0 ; /* current simulation time */
93
94 static int dn_hash_size = 64 ; /* default hash size */
95
96 /* statistics on number of queue searches and search steps */
97 static long searches, search_steps ;
98 static int pipe_expire = 1 ; /* expire queue if empty */
99 static int dn_max_ratio = 16 ; /* max queues/buckets ratio */
100
101 static long pipe_slot_limit = 100; /* Foot shooting limit for pipe queues. */
102 static long pipe_byte_limit = 1024 * 1024;
103
104 static int red_lookup_depth = 256; /* RED - default lookup table depth */
105 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
106 static int red_max_pkt_size = 1500; /* RED - default max packet size */
107
108 static struct timeval prev_t, t;
109 static long tick_last; /* Last tick duration (usec). */
110 static long tick_delta; /* Last vs standard tick diff (usec). */
111 static long tick_delta_sum; /* Accumulated tick difference (usec).*/
112 static long tick_adjustment; /* Tick adjustments done. */
113 static long tick_lost; /* Lost(coalesced) ticks number. */
114 /* Adjusted vs non-adjusted curr_time difference (ticks). */
115 static long tick_diff;
116
117 static int io_fast;
118 static unsigned long io_pkt;
119 static unsigned long io_pkt_fast;
120 static unsigned long io_pkt_drop;
121
122 /*
123 * Three heaps contain queues and pipes that the scheduler handles:
124 *
125 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
126 *
127 * wfq_ready_heap contains the pipes associated with WF2Q flows
128 *
129 * extract_heap contains pipes associated with delay lines.
130 *
131 */
132
133 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
134
135 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
136
137 static int heap_init(struct dn_heap *h, int size);
138 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
139 static void heap_extract(struct dn_heap *h, void *obj);
140 static void transmit_event(struct dn_pipe *pipe, struct mbuf **head,
141 struct mbuf **tail);
142 static void ready_event(struct dn_flow_queue *q, struct mbuf **head,
143 struct mbuf **tail);
144 static void ready_event_wfq(struct dn_pipe *p, struct mbuf **head,
145 struct mbuf **tail);
146
147 #define HASHSIZE 16
148 #define HASH(num) ((((num) >> 8) ^ ((num) >> 4) ^ (num)) & 0x0f)
149 static struct dn_pipe_head pipehash[HASHSIZE]; /* all pipes */
150 static struct dn_flow_set_head flowsethash[HASHSIZE]; /* all flowsets */
151
152 static struct callout dn_timeout;
153
154 extern void (*bridge_dn_p)(struct mbuf *, struct ifnet *);
155
156 #ifdef SYSCTL_NODE
157 SYSCTL_DECL(_net_inet);
158 SYSCTL_DECL(_net_inet_ip);
159
160 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW, 0, "Dummynet");
161 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
162 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
163 #if 0 /* curr_time is 64 bit */
164 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, curr_time,
165 CTLFLAG_RD, &curr_time, 0, "Current tick");
166 #endif
167 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
168 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
169 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
170 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
171 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, searches,
172 CTLFLAG_RD, &searches, 0, "Number of queue searches");
173 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, search_steps,
174 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
175 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
176 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
177 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
178 CTLFLAG_RW, &dn_max_ratio, 0,
179 "Max ratio between dynamic queues and buckets");
180 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
181 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
182 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
183 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
184 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
185 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
186 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta,
187 CTLFLAG_RD, &tick_delta, 0, "Last vs standard tick difference (usec).");
188 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum,
189 CTLFLAG_RD, &tick_delta_sum, 0, "Accumulated tick difference (usec).");
190 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment,
191 CTLFLAG_RD, &tick_adjustment, 0, "Tick adjustments done.");
192 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff,
193 CTLFLAG_RD, &tick_diff, 0,
194 "Adjusted vs non-adjusted curr_time difference (ticks).");
195 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost,
196 CTLFLAG_RD, &tick_lost, 0,
197 "Number of ticks coalesced by dummynet taskqueue.");
198 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast,
199 CTLFLAG_RW, &io_fast, 0, "Enable fast dummynet io.");
200 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt,
201 CTLFLAG_RD, &io_pkt, 0,
202 "Number of packets passed to dummynet.");
203 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast,
204 CTLFLAG_RD, &io_pkt_fast, 0,
205 "Number of packets bypassed dummynet scheduler.");
206 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop,
207 CTLFLAG_RD, &io_pkt_drop, 0,
208 "Number of packets dropped by dummynet.");
209 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit,
210 CTLFLAG_RW, &pipe_slot_limit, 0, "Upper limit in slots for pipe queue.");
211 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit,
212 CTLFLAG_RW, &pipe_byte_limit, 0, "Upper limit in bytes for pipe queue.");
213 #endif
214
215 #ifdef DUMMYNET_DEBUG
216 int dummynet_debug = 0;
217 #ifdef SYSCTL_NODE
218 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, CTLFLAG_RW, &dummynet_debug,
219 0, "control debugging printfs");
220 #endif
221 #define DPRINTF(X) if (dummynet_debug) printf X
222 #else
223 #define DPRINTF(X)
224 #endif
225
226 static struct task dn_task;
227 static struct taskqueue *dn_tq = NULL;
228 static void dummynet_task(void *, int);
229
230 static struct mtx dummynet_mtx;
231 #define DUMMYNET_LOCK_INIT() \
232 mtx_init(&dummynet_mtx, "dummynet", NULL, MTX_DEF)
233 #define DUMMYNET_LOCK_DESTROY() mtx_destroy(&dummynet_mtx)
234 #define DUMMYNET_LOCK() mtx_lock(&dummynet_mtx)
235 #define DUMMYNET_UNLOCK() mtx_unlock(&dummynet_mtx)
236 #define DUMMYNET_LOCK_ASSERT() mtx_assert(&dummynet_mtx, MA_OWNED)
237
238 static int config_pipe(struct dn_pipe *p);
239 static int ip_dn_ctl(struct sockopt *sopt);
240
241 static void dummynet(void *);
242 static void dummynet_flush(void);
243 static void dummynet_send(struct mbuf *);
244 void dummynet_drain(void);
245 static ip_dn_io_t dummynet_io;
246 static void dn_rule_delete(void *);
247
248 /*
249 * Heap management functions.
250 *
251 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
252 * Some macros help finding parent/children so we can optimize them.
253 *
254 * heap_init() is called to expand the heap when needed.
255 * Increment size in blocks of 16 entries.
256 * XXX failure to allocate a new element is a pretty bad failure
257 * as we basically stall a whole queue forever!!
258 * Returns 1 on error, 0 on success
259 */
260 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
261 #define HEAP_LEFT(x) ( 2*(x) + 1 )
262 #define HEAP_IS_LEFT(x) ( (x) & 1 )
263 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
264 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
265 #define HEAP_INCREMENT 15
266
267 static int
268 heap_init(struct dn_heap *h, int new_size)
269 {
270 struct dn_heap_entry *p;
271
272 if (h->size >= new_size ) {
273 printf("dummynet: %s, Bogus call, have %d want %d\n", __func__,
274 h->size, new_size);
275 return 0 ;
276 }
277 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
278 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_NOWAIT);
279 if (p == NULL) {
280 printf("dummynet: %s, resize %d failed\n", __func__, new_size );
281 return 1 ; /* error */
282 }
283 if (h->size > 0) {
284 bcopy(h->p, p, h->size * sizeof(*p) );
285 free(h->p, M_DUMMYNET);
286 }
287 h->p = p ;
288 h->size = new_size ;
289 return 0 ;
290 }
291
292 /*
293 * Insert element in heap. Normally, p != NULL, we insert p in
294 * a new position and bubble up. If p == NULL, then the element is
295 * already in place, and key is the position where to start the
296 * bubble-up.
297 * Returns 1 on failure (cannot allocate new heap entry)
298 *
299 * If offset > 0 the position (index, int) of the element in the heap is
300 * also stored in the element itself at the given offset in bytes.
301 */
302 #define SET_OFFSET(heap, node) \
303 if (heap->offset > 0) \
304 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
305 /*
306 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
307 */
308 #define RESET_OFFSET(heap, node) \
309 if (heap->offset > 0) \
310 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
311 static int
312 heap_insert(struct dn_heap *h, dn_key key1, void *p)
313 {
314 int son = h->elements ;
315
316 if (p == NULL) /* data already there, set starting point */
317 son = key1 ;
318 else { /* insert new element at the end, possibly resize */
319 son = h->elements ;
320 if (son == h->size) /* need resize... */
321 if (heap_init(h, h->elements+1) )
322 return 1 ; /* failure... */
323 h->p[son].object = p ;
324 h->p[son].key = key1 ;
325 h->elements++ ;
326 }
327 while (son > 0) { /* bubble up */
328 int father = HEAP_FATHER(son) ;
329 struct dn_heap_entry tmp ;
330
331 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
332 break ; /* found right position */
333 /* son smaller than father, swap and repeat */
334 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
335 SET_OFFSET(h, son);
336 son = father ;
337 }
338 SET_OFFSET(h, son);
339 return 0 ;
340 }
341
342 /*
343 * remove top element from heap, or obj if obj != NULL
344 */
345 static void
346 heap_extract(struct dn_heap *h, void *obj)
347 {
348 int child, father, max = h->elements - 1 ;
349
350 if (max < 0) {
351 printf("dummynet: warning, extract from empty heap 0x%p\n", h);
352 return ;
353 }
354 father = 0 ; /* default: move up smallest child */
355 if (obj != NULL) { /* extract specific element, index is at offset */
356 if (h->offset <= 0)
357 panic("dummynet: heap_extract from middle not supported on this heap!!!\n");
358 father = *((int *)((char *)obj + h->offset)) ;
359 if (father < 0 || father >= h->elements) {
360 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
361 father, h->elements);
362 panic("dummynet: heap_extract");
363 }
364 }
365 RESET_OFFSET(h, father);
366 child = HEAP_LEFT(father) ; /* left child */
367 while (child <= max) { /* valid entry */
368 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
369 child = child+1 ; /* take right child, otherwise left */
370 h->p[father] = h->p[child] ;
371 SET_OFFSET(h, father);
372 father = child ;
373 child = HEAP_LEFT(child) ; /* left child for next loop */
374 }
375 h->elements-- ;
376 if (father != max) {
377 /*
378 * Fill hole with last entry and bubble up, reusing the insert code
379 */
380 h->p[father] = h->p[max] ;
381 heap_insert(h, father, NULL); /* this one cannot fail */
382 }
383 }
384
385 #if 0
386 /*
387 * change object position and update references
388 * XXX this one is never used!
389 */
390 static void
391 heap_move(struct dn_heap *h, dn_key new_key, void *object)
392 {
393 int temp;
394 int i ;
395 int max = h->elements-1 ;
396 struct dn_heap_entry buf ;
397
398 if (h->offset <= 0)
399 panic("cannot move items on this heap");
400
401 i = *((int *)((char *)object + h->offset));
402 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
403 h->p[i].key = new_key ;
404 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
405 i = temp ) { /* bubble up */
406 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
407 SET_OFFSET(h, i);
408 }
409 } else { /* must move down */
410 h->p[i].key = new_key ;
411 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
412 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
413 temp++ ; /* select child with min key */
414 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
415 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
416 SET_OFFSET(h, i);
417 } else
418 break ;
419 i = temp ;
420 }
421 }
422 SET_OFFSET(h, i);
423 }
424 #endif /* heap_move, unused */
425
426 /*
427 * heapify() will reorganize data inside an array to maintain the
428 * heap property. It is needed when we delete a bunch of entries.
429 */
430 static void
431 heapify(struct dn_heap *h)
432 {
433 int i ;
434
435 for (i = 0 ; i < h->elements ; i++ )
436 heap_insert(h, i , NULL) ;
437 }
438
439 /*
440 * cleanup the heap and free data structure
441 */
442 static void
443 heap_free(struct dn_heap *h)
444 {
445 if (h->size >0 )
446 free(h->p, M_DUMMYNET);
447 bzero(h, sizeof(*h) );
448 }
449
450 /*
451 * --- end of heap management functions ---
452 */
453
454 /*
455 * Return the mbuf tag holding the dummynet state. As an optimization
456 * this is assumed to be the first tag on the list. If this turns out
457 * wrong we'll need to search the list.
458 */
459 static struct dn_pkt_tag *
460 dn_tag_get(struct mbuf *m)
461 {
462 struct m_tag *mtag = m_tag_first(m);
463 KASSERT(mtag != NULL &&
464 mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
465 mtag->m_tag_id == PACKET_TAG_DUMMYNET,
466 ("packet on dummynet queue w/o dummynet tag!"));
467 return (struct dn_pkt_tag *)(mtag+1);
468 }
469
470 /*
471 * Scheduler functions:
472 *
473 * transmit_event() is called when the delay-line needs to enter
474 * the scheduler, either because of existing pkts getting ready,
475 * or new packets entering the queue. The event handled is the delivery
476 * time of the packet.
477 *
478 * ready_event() does something similar with fixed-rate queues, and the
479 * event handled is the finish time of the head pkt.
480 *
481 * wfq_ready_event() does something similar with WF2Q queues, and the
482 * event handled is the start time of the head pkt.
483 *
484 * In all cases, we make sure that the data structures are consistent
485 * before passing pkts out, because this might trigger recursive
486 * invocations of the procedures.
487 */
488 static void
489 transmit_event(struct dn_pipe *pipe, struct mbuf **head, struct mbuf **tail)
490 {
491 struct mbuf *m;
492 struct dn_pkt_tag *pkt;
493
494 DUMMYNET_LOCK_ASSERT();
495
496 while ((m = pipe->head) != NULL) {
497 pkt = dn_tag_get(m);
498 if (!DN_KEY_LEQ(pkt->output_time, curr_time))
499 break;
500
501 pipe->head = m->m_nextpkt;
502 if (*tail != NULL)
503 (*tail)->m_nextpkt = m;
504 else
505 *head = m;
506 *tail = m;
507 }
508 if (*tail != NULL)
509 (*tail)->m_nextpkt = NULL;
510
511 /* If there are leftover packets, put into the heap for next event. */
512 if ((m = pipe->head) != NULL) {
513 pkt = dn_tag_get(m);
514 /*
515 * XXX Should check errors on heap_insert, by draining the
516 * whole pipe p and hoping in the future we are more successful.
517 */
518 heap_insert(&extract_heap, pkt->output_time, pipe);
519 }
520 }
521
522 /*
523 * the following macro computes how many ticks we have to wait
524 * before being able to transmit a packet. The credit is taken from
525 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
526 */
527 #define SET_TICKS(_m, q, p) \
528 ((_m)->m_pkthdr.len * 8 * hz - (q)->numbytes + p->bandwidth - 1) / \
529 p->bandwidth;
530
531 /*
532 * extract pkt from queue, compute output time (could be now)
533 * and put into delay line (p_queue)
534 */
535 static void
536 move_pkt(struct mbuf *pkt, struct dn_flow_queue *q, struct dn_pipe *p,
537 int len)
538 {
539 struct dn_pkt_tag *dt = dn_tag_get(pkt);
540
541 q->head = pkt->m_nextpkt ;
542 q->len-- ;
543 q->len_bytes -= len ;
544
545 dt->output_time = curr_time + p->delay ;
546
547 if (p->head == NULL)
548 p->head = pkt;
549 else
550 p->tail->m_nextpkt = pkt;
551 p->tail = pkt;
552 p->tail->m_nextpkt = NULL;
553 }
554
555 /*
556 * ready_event() is invoked every time the queue must enter the
557 * scheduler, either because the first packet arrives, or because
558 * a previously scheduled event fired.
559 * On invokation, drain as many pkts as possible (could be 0) and then
560 * if there are leftover packets reinsert the pkt in the scheduler.
561 */
562 static void
563 ready_event(struct dn_flow_queue *q, struct mbuf **head, struct mbuf **tail)
564 {
565 struct mbuf *pkt;
566 struct dn_pipe *p = q->fs->pipe;
567 int p_was_empty;
568
569 DUMMYNET_LOCK_ASSERT();
570
571 if (p == NULL) {
572 printf("dummynet: ready_event- pipe is gone\n");
573 return;
574 }
575 p_was_empty = (p->head == NULL);
576
577 /*
578 * Schedule fixed-rate queues linked to this pipe:
579 * account for the bw accumulated since last scheduling, then
580 * drain as many pkts as allowed by q->numbytes and move to
581 * the delay line (in p) computing output time.
582 * bandwidth==0 (no limit) means we can drain the whole queue,
583 * setting len_scaled = 0 does the job.
584 */
585 q->numbytes += (curr_time - q->sched_time) * p->bandwidth;
586 while ((pkt = q->head) != NULL) {
587 int len = pkt->m_pkthdr.len;
588 int len_scaled = p->bandwidth ? len * 8 * hz : 0;
589
590 if (len_scaled > q->numbytes)
591 break;
592 q->numbytes -= len_scaled;
593 move_pkt(pkt, q, p, len);
594 }
595 /*
596 * If we have more packets queued, schedule next ready event
597 * (can only occur when bandwidth != 0, otherwise we would have
598 * flushed the whole queue in the previous loop).
599 * To this purpose we record the current time and compute how many
600 * ticks to go for the finish time of the packet.
601 */
602 if ((pkt = q->head) != NULL) { /* this implies bandwidth != 0 */
603 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
604
605 q->sched_time = curr_time;
606 heap_insert(&ready_heap, curr_time + t, (void *)q);
607 /*
608 * XXX Should check errors on heap_insert, and drain the whole
609 * queue on error hoping next time we are luckier.
610 */
611 } else /* RED needs to know when the queue becomes empty. */
612 q->q_time = curr_time;
613
614 /*
615 * If the delay line was empty call transmit_event() now.
616 * Otherwise, the scheduler will take care of it.
617 */
618 if (p_was_empty)
619 transmit_event(p, head, tail);
620 }
621
622 /*
623 * Called when we can transmit packets on WF2Q queues. Take pkts out of
624 * the queues at their start time, and enqueue into the delay line.
625 * Packets are drained until p->numbytes < 0. As long as
626 * len_scaled >= p->numbytes, the packet goes into the delay line
627 * with a deadline p->delay. For the last packet, if p->numbytes < 0,
628 * there is an additional delay.
629 */
630 static void
631 ready_event_wfq(struct dn_pipe *p, struct mbuf **head, struct mbuf **tail)
632 {
633 int p_was_empty = (p->head == NULL);
634 struct dn_heap *sch = &(p->scheduler_heap);
635 struct dn_heap *neh = &(p->not_eligible_heap);
636 int64_t p_numbytes = p->numbytes;
637
638 DUMMYNET_LOCK_ASSERT();
639
640 if (p->if_name[0] == 0) /* tx clock is simulated */
641 /*
642 * Since result may not fit into p->numbytes (32bit) we
643 * are using 64bit var here.
644 */
645 p_numbytes += (curr_time - p->sched_time) * p->bandwidth;
646 else { /*
647 * tx clock is for real,
648 * the ifq must be empty or this is a NOP.
649 */
650 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
651 return;
652 else {
653 DPRINTF(("dummynet: pipe %d ready from %s --\n",
654 p->pipe_nr, p->if_name));
655 }
656 }
657
658 /*
659 * While we have backlogged traffic AND credit, we need to do
660 * something on the queue.
661 */
662 while (p_numbytes >= 0 && (sch->elements > 0 || neh->elements > 0)) {
663 if (sch->elements > 0) {
664 /* Have some eligible pkts to send out. */
665 struct dn_flow_queue *q = sch->p[0].object;
666 struct mbuf *pkt = q->head;
667 struct dn_flow_set *fs = q->fs;
668 uint64_t len = pkt->m_pkthdr.len;
669 int len_scaled = p->bandwidth ? len * 8 * hz : 0;
670
671 heap_extract(sch, NULL); /* Remove queue from heap. */
672 p_numbytes -= len_scaled;
673 move_pkt(pkt, q, p, len);
674
675 p->V += (len << MY_M) / p->sum; /* Update V. */
676 q->S = q->F; /* Update start time. */
677 if (q->len == 0) {
678 /* Flow not backlogged any more. */
679 fs->backlogged--;
680 heap_insert(&(p->idle_heap), q->F, q);
681 } else {
682 /* Still backlogged. */
683
684 /*
685 * Update F and position in backlogged queue,
686 * then put flow in not_eligible_heap
687 * (we will fix this later).
688 */
689 len = (q->head)->m_pkthdr.len;
690 q->F += (len << MY_M) / (uint64_t)fs->weight;
691 if (DN_KEY_LEQ(q->S, p->V))
692 heap_insert(neh, q->S, q);
693 else
694 heap_insert(sch, q->F, q);
695 }
696 }
697 /*
698 * Now compute V = max(V, min(S_i)). Remember that all elements
699 * in sch have by definition S_i <= V so if sch is not empty,
700 * V is surely the max and we must not update it. Conversely,
701 * if sch is empty we only need to look at neh.
702 */
703 if (sch->elements == 0 && neh->elements > 0)
704 p->V = MAX64(p->V, neh->p[0].key);
705 /* Move from neh to sch any packets that have become eligible */
706 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V)) {
707 struct dn_flow_queue *q = neh->p[0].object;
708 heap_extract(neh, NULL);
709 heap_insert(sch, q->F, q);
710 }
711
712 if (p->if_name[0] != '\0') { /* Tx clock is from a real thing */
713 p_numbytes = -1; /* Mark not ready for I/O. */
714 break;
715 }
716 }
717 if (sch->elements == 0 && neh->elements == 0 && p_numbytes >= 0 &&
718 p->idle_heap.elements > 0) {
719 /*
720 * No traffic and no events scheduled.
721 * We can get rid of idle-heap.
722 */
723 int i;
724
725 for (i = 0; i < p->idle_heap.elements; i++) {
726 struct dn_flow_queue *q = p->idle_heap.p[i].object;
727
728 q->F = 0;
729 q->S = q->F + 1;
730 }
731 p->sum = 0;
732 p->V = 0;
733 p->idle_heap.elements = 0;
734 }
735 /*
736 * If we are getting clocks from dummynet (not a real interface) and
737 * If we are under credit, schedule the next ready event.
738 * Also fix the delivery time of the last packet.
739 */
740 if (p->if_name[0]==0 && p_numbytes < 0) { /* This implies bw > 0. */
741 dn_key t = 0; /* Number of ticks i have to wait. */
742
743 if (p->bandwidth > 0)
744 t = (p->bandwidth - 1 - p_numbytes) / p->bandwidth;
745 dn_tag_get(p->tail)->output_time += t;
746 p->sched_time = curr_time;
747 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
748 /*
749 * XXX Should check errors on heap_insert, and drain the whole
750 * queue on error hoping next time we are luckier.
751 */
752 }
753
754 /* Fit (adjust if necessary) 64bit result into 32bit variable. */
755 if (p_numbytes > INT_MAX)
756 p->numbytes = INT_MAX;
757 else if (p_numbytes < INT_MIN)
758 p->numbytes = INT_MIN;
759 else
760 p->numbytes = p_numbytes;
761
762 /*
763 * If the delay line was empty call transmit_event() now.
764 * Otherwise, the scheduler will take care of it.
765 */
766 if (p_was_empty)
767 transmit_event(p, head, tail);
768 }
769
770 /*
771 * This is called one tick, after previous run. It is used to
772 * schedule next run.
773 */
774 static void
775 dummynet(void * __unused unused)
776 {
777
778 taskqueue_enqueue(dn_tq, &dn_task);
779 }
780
781 /*
782 * The main dummynet processing function.
783 */
784 static void
785 dummynet_task(void *context, int pending)
786 {
787 struct mbuf *head = NULL, *tail = NULL;
788 struct dn_pipe *pipe;
789 struct dn_heap *heaps[3];
790 struct dn_heap *h;
791 void *p; /* generic parameter to handler */
792 int i;
793
794 DUMMYNET_LOCK();
795
796 heaps[0] = &ready_heap; /* fixed-rate queues */
797 heaps[1] = &wfq_ready_heap; /* wfq queues */
798 heaps[2] = &extract_heap; /* delay line */
799
800 /* Update number of lost(coalesced) ticks. */
801 tick_lost += pending - 1;
802
803 getmicrouptime(&t);
804 /* Last tick duration (usec). */
805 tick_last = (t.tv_sec - prev_t.tv_sec) * 1000000 +
806 (t.tv_usec - prev_t.tv_usec);
807 /* Last tick vs standard tick difference (usec). */
808 tick_delta = (tick_last * hz - 1000000) / hz;
809 /* Accumulated tick difference (usec). */
810 tick_delta_sum += tick_delta;
811
812 prev_t = t;
813
814 /*
815 * Adjust curr_time if accumulated tick difference greater than
816 * 'standard' tick. Since curr_time should be monotonically increasing,
817 * we do positive adjustment as required and throttle curr_time in
818 * case of negative adjustment.
819 */
820 curr_time++;
821 if (tick_delta_sum - tick >= 0) {
822 int diff = tick_delta_sum / tick;
823
824 curr_time += diff;
825 tick_diff += diff;
826 tick_delta_sum %= tick;
827 tick_adjustment++;
828 } else if (tick_delta_sum + tick <= 0) {
829 curr_time--;
830 tick_diff--;
831 tick_delta_sum += tick;
832 tick_adjustment++;
833 }
834
835 for (i = 0; i < 3; i++) {
836 h = heaps[i];
837 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time)) {
838 if (h->p[0].key > curr_time)
839 printf("dummynet: warning, "
840 "heap %d is %d ticks late\n",
841 i, (int)(curr_time - h->p[0].key));
842 /* store a copy before heap_extract */
843 p = h->p[0].object;
844 /* need to extract before processing */
845 heap_extract(h, NULL);
846 if (i == 0)
847 ready_event(p, &head, &tail);
848 else if (i == 1) {
849 struct dn_pipe *pipe = p;
850 if (pipe->if_name[0] != '\0')
851 printf("dummynet: bad ready_event_wfq "
852 "for pipe %s\n", pipe->if_name);
853 else
854 ready_event_wfq(p, &head, &tail);
855 } else
856 transmit_event(p, &head, &tail);
857 }
858 }
859
860 /* Sweep pipes trying to expire idle flow_queues. */
861 for (i = 0; i < HASHSIZE; i++)
862 SLIST_FOREACH(pipe, &pipehash[i], next)
863 if (pipe->idle_heap.elements > 0 &&
864 DN_KEY_LT(pipe->idle_heap.p[0].key, pipe->V)) {
865 struct dn_flow_queue *q =
866 pipe->idle_heap.p[0].object;
867
868 heap_extract(&(pipe->idle_heap), NULL);
869 /* Mark timestamp as invalid. */
870 q->S = q->F + 1;
871 pipe->sum -= q->fs->weight;
872 }
873
874 DUMMYNET_UNLOCK();
875
876 if (head != NULL)
877 dummynet_send(head);
878
879 callout_reset(&dn_timeout, 1, dummynet, NULL);
880 }
881
882 static void
883 dummynet_send(struct mbuf *m)
884 {
885 struct dn_pkt_tag *pkt;
886 struct mbuf *n;
887 struct ip *ip;
888
889 for (; m != NULL; m = n) {
890 n = m->m_nextpkt;
891 m->m_nextpkt = NULL;
892 pkt = dn_tag_get(m);
893 switch (pkt->dn_dir) {
894 case DN_TO_IP_OUT:
895 ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL);
896 break ;
897 case DN_TO_IP_IN :
898 ip = mtod(m, struct ip *);
899 ip->ip_len = htons(ip->ip_len);
900 ip->ip_off = htons(ip->ip_off);
901 netisr_dispatch(NETISR_IP, m);
902 break;
903 #ifdef INET6
904 case DN_TO_IP6_IN:
905 netisr_dispatch(NETISR_IPV6, m);
906 break;
907
908 case DN_TO_IP6_OUT:
909 ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL);
910 break;
911 #endif
912 case DN_TO_IFB_FWD:
913 if (bridge_dn_p != NULL)
914 ((*bridge_dn_p)(m, pkt->ifp));
915 else
916 printf("dummynet: if_bridge not loaded\n");
917
918 break;
919 case DN_TO_ETH_DEMUX:
920 /*
921 * The Ethernet code assumes the Ethernet header is
922 * contiguous in the first mbuf header.
923 * Insure this is true.
924 */
925 if (m->m_len < ETHER_HDR_LEN &&
926 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
927 printf("dummynet/ether: pullup failed, "
928 "dropping packet\n");
929 break;
930 }
931 ether_demux(m->m_pkthdr.rcvif, m);
932 break;
933 case DN_TO_ETH_OUT:
934 ether_output_frame(pkt->ifp, m);
935 break;
936 default:
937 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
938 m_freem(m);
939 break;
940 }
941 }
942 }
943
944 /*
945 * Unconditionally expire empty queues in case of shortage.
946 * Returns the number of queues freed.
947 */
948 static int
949 expire_queues(struct dn_flow_set *fs)
950 {
951 struct dn_flow_queue *q, *prev ;
952 int i, initial_elements = fs->rq_elements ;
953
954 if (fs->last_expired == time_uptime)
955 return 0 ;
956 fs->last_expired = time_uptime ;
957 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
958 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
959 if (q->head != NULL || q->S != q->F+1) {
960 prev = q ;
961 q = q->next ;
962 } else { /* entry is idle, expire it */
963 struct dn_flow_queue *old_q = q ;
964
965 if (prev != NULL)
966 prev->next = q = q->next ;
967 else
968 fs->rq[i] = q = q->next ;
969 fs->rq_elements-- ;
970 free(old_q, M_DUMMYNET);
971 }
972 return initial_elements - fs->rq_elements ;
973 }
974
975 /*
976 * If room, create a new queue and put at head of slot i;
977 * otherwise, create or use the default queue.
978 */
979 static struct dn_flow_queue *
980 create_queue(struct dn_flow_set *fs, int i)
981 {
982 struct dn_flow_queue *q;
983
984 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
985 expire_queues(fs) == 0) {
986 /* No way to get room, use or create overflow queue. */
987 i = fs->rq_size;
988 if (fs->rq[i] != NULL)
989 return fs->rq[i];
990 }
991 q = malloc(sizeof(*q), M_DUMMYNET, M_NOWAIT | M_ZERO);
992 if (q == NULL) {
993 printf("dummynet: sorry, cannot allocate queue for new flow\n");
994 return (NULL);
995 }
996 q->fs = fs;
997 q->hash_slot = i;
998 q->next = fs->rq[i];
999 q->S = q->F + 1; /* hack - mark timestamp as invalid. */
1000 q->numbytes = io_fast ? fs->pipe->bandwidth : 0;
1001 fs->rq[i] = q;
1002 fs->rq_elements++;
1003 return (q);
1004 }
1005
1006 /*
1007 * Given a flow_set and a pkt in last_pkt, find a matching queue
1008 * after appropriate masking. The queue is moved to front
1009 * so that further searches take less time.
1010 */
1011 static struct dn_flow_queue *
1012 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
1013 {
1014 int i = 0 ; /* we need i and q for new allocations */
1015 struct dn_flow_queue *q, *prev;
1016 int is_v6 = IS_IP6_FLOW_ID(id);
1017
1018 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
1019 q = fs->rq[0] ;
1020 else {
1021 /* first, do the masking, then hash */
1022 id->dst_port &= fs->flow_mask.dst_port ;
1023 id->src_port &= fs->flow_mask.src_port ;
1024 id->proto &= fs->flow_mask.proto ;
1025 id->flags = 0 ; /* we don't care about this one */
1026 if (is_v6) {
1027 APPLY_MASK(&id->dst_ip6, &fs->flow_mask.dst_ip6);
1028 APPLY_MASK(&id->src_ip6, &fs->flow_mask.src_ip6);
1029 id->flow_id6 &= fs->flow_mask.flow_id6;
1030
1031 i = ((id->dst_ip6.__u6_addr.__u6_addr32[0]) & 0xffff)^
1032 ((id->dst_ip6.__u6_addr.__u6_addr32[1]) & 0xffff)^
1033 ((id->dst_ip6.__u6_addr.__u6_addr32[2]) & 0xffff)^
1034 ((id->dst_ip6.__u6_addr.__u6_addr32[3]) & 0xffff)^
1035
1036 ((id->dst_ip6.__u6_addr.__u6_addr32[0] >> 15) & 0xffff)^
1037 ((id->dst_ip6.__u6_addr.__u6_addr32[1] >> 15) & 0xffff)^
1038 ((id->dst_ip6.__u6_addr.__u6_addr32[2] >> 15) & 0xffff)^
1039 ((id->dst_ip6.__u6_addr.__u6_addr32[3] >> 15) & 0xffff)^
1040
1041 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 1) & 0xfffff)^
1042 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 1) & 0xfffff)^
1043 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 1) & 0xfffff)^
1044 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 1) & 0xfffff)^
1045
1046 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 16) & 0xffff)^
1047 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 16) & 0xffff)^
1048 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 16) & 0xffff)^
1049 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 16) & 0xffff)^
1050
1051 (id->dst_port << 1) ^ (id->src_port) ^
1052 (id->proto ) ^
1053 (id->flow_id6);
1054 } else {
1055 id->dst_ip &= fs->flow_mask.dst_ip ;
1056 id->src_ip &= fs->flow_mask.src_ip ;
1057
1058 i = ( (id->dst_ip) & 0xffff ) ^
1059 ( (id->dst_ip >> 15) & 0xffff ) ^
1060 ( (id->src_ip << 1) & 0xffff ) ^
1061 ( (id->src_ip >> 16 ) & 0xffff ) ^
1062 (id->dst_port << 1) ^ (id->src_port) ^
1063 (id->proto );
1064 }
1065 i = i % fs->rq_size ;
1066 /* finally, scan the current list for a match */
1067 searches++ ;
1068 for (prev=NULL, q = fs->rq[i] ; q ; ) {
1069 search_steps++;
1070 if (is_v6 &&
1071 IN6_ARE_ADDR_EQUAL(&id->dst_ip6,&q->id.dst_ip6) &&
1072 IN6_ARE_ADDR_EQUAL(&id->src_ip6,&q->id.src_ip6) &&
1073 id->dst_port == q->id.dst_port &&
1074 id->src_port == q->id.src_port &&
1075 id->proto == q->id.proto &&
1076 id->flags == q->id.flags &&
1077 id->flow_id6 == q->id.flow_id6)
1078 break ; /* found */
1079
1080 if (!is_v6 && id->dst_ip == q->id.dst_ip &&
1081 id->src_ip == q->id.src_ip &&
1082 id->dst_port == q->id.dst_port &&
1083 id->src_port == q->id.src_port &&
1084 id->proto == q->id.proto &&
1085 id->flags == q->id.flags)
1086 break ; /* found */
1087
1088 /* No match. Check if we can expire the entry */
1089 if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
1090 /* entry is idle and not in any heap, expire it */
1091 struct dn_flow_queue *old_q = q ;
1092
1093 if (prev != NULL)
1094 prev->next = q = q->next ;
1095 else
1096 fs->rq[i] = q = q->next ;
1097 fs->rq_elements-- ;
1098 free(old_q, M_DUMMYNET);
1099 continue ;
1100 }
1101 prev = q ;
1102 q = q->next ;
1103 }
1104 if (q && prev != NULL) { /* found and not in front */
1105 prev->next = q->next ;
1106 q->next = fs->rq[i] ;
1107 fs->rq[i] = q ;
1108 }
1109 }
1110 if (q == NULL) { /* no match, need to allocate a new entry */
1111 q = create_queue(fs, i);
1112 if (q != NULL)
1113 q->id = *id ;
1114 }
1115 return q ;
1116 }
1117
1118 static int
1119 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
1120 {
1121 /*
1122 * RED algorithm
1123 *
1124 * RED calculates the average queue size (avg) using a low-pass filter
1125 * with an exponential weighted (w_q) moving average:
1126 * avg <- (1-w_q) * avg + w_q * q_size
1127 * where q_size is the queue length (measured in bytes or * packets).
1128 *
1129 * If q_size == 0, we compute the idle time for the link, and set
1130 * avg = (1 - w_q)^(idle/s)
1131 * where s is the time needed for transmitting a medium-sized packet.
1132 *
1133 * Now, if avg < min_th the packet is enqueued.
1134 * If avg > max_th the packet is dropped. Otherwise, the packet is
1135 * dropped with probability P function of avg.
1136 */
1137
1138 int64_t p_b = 0;
1139
1140 /* Queue in bytes or packets? */
1141 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ?
1142 q->len_bytes : q->len;
1143
1144 DPRINTF(("\ndummynet: %d q: %2u ", (int)curr_time, q_size));
1145
1146 /* Average queue size estimation. */
1147 if (q_size != 0) {
1148 /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */
1149 int diff = SCALE(q_size) - q->avg;
1150 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);
1151
1152 q->avg += (int)v;
1153 } else {
1154 /*
1155 * Queue is empty, find for how long the queue has been
1156 * empty and use a lookup table for computing
1157 * (1 - * w_q)^(idle_time/s) where s is the time to send a
1158 * (small) packet.
1159 * XXX check wraps...
1160 */
1161 if (q->avg) {
1162 u_int t = ((uint32_t)curr_time - q->q_time) /
1163 fs->lookup_step;
1164
1165 q->avg = (t < fs->lookup_depth) ?
1166 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
1167 }
1168 }
1169 DPRINTF(("dummynet: avg: %u ", SCALE_VAL(q->avg)));
1170
1171 /* Should i drop? */
1172 if (q->avg < fs->min_th) {
1173 q->count = -1;
1174 return (0); /* accept packet */
1175 }
1176 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
1177 if (fs->flags_fs & DN_IS_GENTLE_RED) {
1178 /*
1179 * According to Gentle-RED, if avg is greater than
1180 * max_th the packet is dropped with a probability
1181 * p_b = c_3 * avg - c_4
1182 * where c_3 = (1 - max_p) / max_th
1183 * c_4 = 1 - 2 * max_p
1184 */
1185 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) -
1186 fs->c_4;
1187 } else {
1188 q->count = -1;
1189 DPRINTF(("dummynet: - drop"));
1190 return (1);
1191 }
1192 } else if (q->avg > fs->min_th) {
1193 /*
1194 * We compute p_b using the linear dropping function
1195 * p_b = c_1 * avg - c_2
1196 * where c_1 = max_p / (max_th - min_th)
1197 * c_2 = max_p * min_th / (max_th - min_th)
1198 */
1199 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
1200 }
1201
1202 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
1203 p_b = (p_b * len) / fs->max_pkt_size;
1204 if (++q->count == 0)
1205 q->random = random() & 0xffff;
1206 else {
1207 /*
1208 * q->count counts packets arrived since last drop, so a greater
1209 * value of q->count means a greater packet drop probability.
1210 */
1211 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
1212 q->count = 0;
1213 DPRINTF(("dummynet: - red drop"));
1214 /* After a drop we calculate a new random value. */
1215 q->random = random() & 0xffff;
1216 return (1); /* drop */
1217 }
1218 }
1219 /* End of RED algorithm. */
1220
1221 return (0); /* accept */
1222 }
1223
1224 static __inline struct dn_flow_set *
1225 locate_flowset(int fs_nr)
1226 {
1227 struct dn_flow_set *fs;
1228
1229 SLIST_FOREACH(fs, &flowsethash[HASH(fs_nr)], next)
1230 if (fs->fs_nr == fs_nr)
1231 return (fs);
1232
1233 return (NULL);
1234 }
1235
1236 static __inline struct dn_pipe *
1237 locate_pipe(int pipe_nr)
1238 {
1239 struct dn_pipe *pipe;
1240
1241 SLIST_FOREACH(pipe, &pipehash[HASH(pipe_nr)], next)
1242 if (pipe->pipe_nr == pipe_nr)
1243 return (pipe);
1244
1245 return (NULL);
1246 }
1247
1248 /*
1249 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1250 * depending on whether WF2Q or fixed bw is used.
1251 *
1252 * pipe_nr pipe or queue the packet is destined for.
1253 * dir where shall we send the packet after dummynet.
1254 * m the mbuf with the packet
1255 * ifp the 'ifp' parameter from the caller.
1256 * NULL in ip_input, destination interface in ip_output,
1257 * rule matching rule, in case of multiple passes
1258 */
1259 static int
1260 dummynet_io(struct mbuf **m0, int dir, struct ip_fw_args *fwa)
1261 {
1262 struct mbuf *m = *m0, *head = NULL, *tail = NULL;
1263 struct dn_pkt_tag *pkt;
1264 struct m_tag *mtag;
1265 struct dn_flow_set *fs = NULL;
1266 struct dn_pipe *pipe;
1267 uint64_t len = m->m_pkthdr.len;
1268 struct dn_flow_queue *q = NULL;
1269 int is_pipe;
1270 ipfw_insn *cmd = ACTION_PTR(fwa->rule);
1271
1272 KASSERT(m->m_nextpkt == NULL,
1273 ("dummynet_io: mbuf queue passed to dummynet"));
1274
1275 if (cmd->opcode == O_LOG)
1276 cmd += F_LEN(cmd);
1277 if (cmd->opcode == O_ALTQ)
1278 cmd += F_LEN(cmd);
1279 if (cmd->opcode == O_TAG)
1280 cmd += F_LEN(cmd);
1281 is_pipe = (cmd->opcode == O_PIPE);
1282
1283 DUMMYNET_LOCK();
1284 io_pkt++;
1285 /*
1286 * This is a dummynet rule, so we expect an O_PIPE or O_QUEUE rule.
1287 *
1288 * XXXGL: probably the pipe->fs and fs->pipe logic here
1289 * below can be simplified.
1290 */
1291 if (is_pipe) {
1292 pipe = locate_pipe(fwa->cookie);
1293 if (pipe != NULL)
1294 fs = &(pipe->fs);
1295 } else
1296 fs = locate_flowset(fwa->cookie);
1297
1298 if (fs == NULL)
1299 goto dropit; /* This queue/pipe does not exist! */
1300 pipe = fs->pipe;
1301 if (pipe == NULL) { /* Must be a queue, try find a matching pipe. */
1302 pipe = locate_pipe(fs->parent_nr);
1303 if (pipe != NULL)
1304 fs->pipe = pipe;
1305 else {
1306 printf("dummynet: no pipe %d for queue %d, drop pkt\n",
1307 fs->parent_nr, fs->fs_nr);
1308 goto dropit;
1309 }
1310 }
1311 q = find_queue(fs, &(fwa->f_id));
1312 if (q == NULL)
1313 goto dropit; /* Cannot allocate queue. */
1314
1315 /* Update statistics, then check reasons to drop pkt. */
1316 q->tot_bytes += len;
1317 q->tot_pkts++;
1318 if (fs->plr && random() < fs->plr)
1319 goto dropit; /* Random pkt drop. */
1320 if (fs->flags_fs & DN_QSIZE_IS_BYTES) {
1321 if (q->len_bytes > fs->qsize)
1322 goto dropit; /* Queue size overflow. */
1323 } else {
1324 if (q->len >= fs->qsize)
1325 goto dropit; /* Queue count overflow. */
1326 }
1327 if (fs->flags_fs & DN_IS_RED && red_drops(fs, q, len))
1328 goto dropit;
1329
1330 /* XXX expensive to zero, see if we can remove it. */
1331 mtag = m_tag_get(PACKET_TAG_DUMMYNET,
1332 sizeof(struct dn_pkt_tag), M_NOWAIT | M_ZERO);
1333 if (mtag == NULL)
1334 goto dropit; /* Cannot allocate packet header. */
1335 m_tag_prepend(m, mtag); /* Attach to mbuf chain. */
1336
1337 pkt = (struct dn_pkt_tag *)(mtag + 1);
1338 /*
1339 * Ok, i can handle the pkt now...
1340 * Build and enqueue packet + parameters.
1341 */
1342 pkt->rule = fwa->rule;
1343 pkt->dn_dir = dir;
1344
1345 pkt->ifp = fwa->oif;
1346
1347 if (q->head == NULL)
1348 q->head = m;
1349 else
1350 q->tail->m_nextpkt = m;
1351 q->tail = m;
1352 q->len++;
1353 q->len_bytes += len;
1354
1355 if (q->head != m) /* Flow was not idle, we are done. */
1356 goto done;
1357
1358 if (q->q_time < (uint32_t)curr_time)
1359 q->numbytes = io_fast ? fs->pipe->bandwidth : 0;
1360 q->q_time = curr_time;
1361
1362 /*
1363 * If we reach this point the flow was previously idle, so we need
1364 * to schedule it. This involves different actions for fixed-rate or
1365 * WF2Q queues.
1366 */
1367 if (is_pipe) {
1368 /* Fixed-rate queue: just insert into the ready_heap. */
1369 dn_key t = 0;
1370
1371 if (pipe->bandwidth && m->m_pkthdr.len * 8 * hz > q->numbytes)
1372 t = SET_TICKS(m, q, pipe);
1373 q->sched_time = curr_time;
1374 if (t == 0) /* Must process it now. */
1375 ready_event(q, &head, &tail);
1376 else
1377 heap_insert(&ready_heap, curr_time + t , q);
1378 } else {
1379 /*
1380 * WF2Q. First, compute start time S: if the flow was
1381 * idle (S = F + 1) set S to the virtual time V for the
1382 * controlling pipe, and update the sum of weights for the pipe;
1383 * otherwise, remove flow from idle_heap and set S to max(F,V).
1384 * Second, compute finish time F = S + len / weight.
1385 * Third, if pipe was idle, update V = max(S, V).
1386 * Fourth, count one more backlogged flow.
1387 */
1388 if (DN_KEY_GT(q->S, q->F)) { /* Means timestamps are invalid. */
1389 q->S = pipe->V;
1390 pipe->sum += fs->weight; /* Add weight of new queue. */
1391 } else {
1392 heap_extract(&(pipe->idle_heap), q);
1393 q->S = MAX64(q->F, pipe->V);
1394 }
1395 q->F = q->S + (len << MY_M) / (uint64_t)fs->weight;
1396
1397 if (pipe->not_eligible_heap.elements == 0 &&
1398 pipe->scheduler_heap.elements == 0)
1399 pipe->V = MAX64(q->S, pipe->V);
1400 fs->backlogged++;
1401 /*
1402 * Look at eligibility. A flow is not eligibile if S>V (when
1403 * this happens, it means that there is some other flow already
1404 * scheduled for the same pipe, so the scheduler_heap cannot be
1405 * empty). If the flow is not eligible we just store it in the
1406 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1407 * and possibly invoke ready_event_wfq() right now if there is
1408 * leftover credit.
1409 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1410 * and for all flows in not_eligible_heap (NEH), S_i > V.
1411 * So when we need to compute max(V, min(S_i)) forall i in
1412 * SCH+NEH, we only need to look into NEH.
1413 */
1414 if (DN_KEY_GT(q->S, pipe->V)) { /* Not eligible. */
1415 if (pipe->scheduler_heap.elements == 0)
1416 printf("dummynet: ++ ouch! not eligible but empty scheduler!\n");
1417 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1418 } else {
1419 heap_insert(&(pipe->scheduler_heap), q->F, q);
1420 if (pipe->numbytes >= 0) { /* Pipe is idle. */
1421 if (pipe->scheduler_heap.elements != 1)
1422 printf("dummynet: OUCH! pipe should have been idle!\n");
1423 DPRINTF(("dummynet: waking up pipe %d at %d\n",
1424 pipe->pipe_nr, (int)(q->F >> MY_M)));
1425 pipe->sched_time = curr_time;
1426 ready_event_wfq(pipe, &head, &tail);
1427 }
1428 }
1429 }
1430 done:
1431 if (head == m && dir != DN_TO_IFB_FWD && dir != DN_TO_ETH_DEMUX &&
1432 dir != DN_TO_ETH_OUT) { /* Fast io. */
1433 io_pkt_fast++;
1434 if (m->m_nextpkt != NULL)
1435 printf("dummynet: fast io: pkt chain detected!\n");
1436 head = m->m_nextpkt = NULL;
1437 } else
1438 *m0 = NULL; /* Normal io. */
1439
1440 DUMMYNET_UNLOCK();
1441 if (head != NULL)
1442 dummynet_send(head);
1443 return (0);
1444
1445 dropit:
1446 io_pkt_drop++;
1447 if (q)
1448 q->drops++;
1449 DUMMYNET_UNLOCK();
1450 m_freem(m);
1451 *m0 = NULL;
1452 return ((fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1453 }
1454
1455 /*
1456 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1457 * Doing this would probably save us the initial bzero of dn_pkt
1458 */
1459 #define DN_FREE_PKT(_m) do { \
1460 m_freem(_m); \
1461 } while (0)
1462
1463 /*
1464 * Dispose all packets and flow_queues on a flow_set.
1465 * If all=1, also remove red lookup table and other storage,
1466 * including the descriptor itself.
1467 * For the one in dn_pipe MUST also cleanup ready_heap...
1468 */
1469 static void
1470 purge_flow_set(struct dn_flow_set *fs, int all)
1471 {
1472 struct dn_flow_queue *q, *qn;
1473 int i;
1474
1475 DUMMYNET_LOCK_ASSERT();
1476
1477 for (i = 0; i <= fs->rq_size; i++) {
1478 for (q = fs->rq[i]; q != NULL; q = qn) {
1479 struct mbuf *m, *mnext;
1480
1481 mnext = q->head;
1482 while ((m = mnext) != NULL) {
1483 mnext = m->m_nextpkt;
1484 DN_FREE_PKT(m);
1485 }
1486 qn = q->next;
1487 free(q, M_DUMMYNET);
1488 }
1489 fs->rq[i] = NULL;
1490 }
1491
1492 fs->rq_elements = 0;
1493 if (all) {
1494 /* RED - free lookup table. */
1495 if (fs->w_q_lookup != NULL)
1496 free(fs->w_q_lookup, M_DUMMYNET);
1497 if (fs->rq != NULL)
1498 free(fs->rq, M_DUMMYNET);
1499 /* If this fs is not part of a pipe, free it. */
1500 if (fs->pipe == NULL || fs != &(fs->pipe->fs))
1501 free(fs, M_DUMMYNET);
1502 }
1503 }
1504
1505 /*
1506 * Dispose all packets queued on a pipe (not a flow_set).
1507 * Also free all resources associated to a pipe, which is about
1508 * to be deleted.
1509 */
1510 static void
1511 purge_pipe(struct dn_pipe *pipe)
1512 {
1513 struct mbuf *m, *mnext;
1514
1515 purge_flow_set( &(pipe->fs), 1 );
1516
1517 mnext = pipe->head;
1518 while ((m = mnext) != NULL) {
1519 mnext = m->m_nextpkt;
1520 DN_FREE_PKT(m);
1521 }
1522
1523 heap_free( &(pipe->scheduler_heap) );
1524 heap_free( &(pipe->not_eligible_heap) );
1525 heap_free( &(pipe->idle_heap) );
1526 }
1527
1528 /*
1529 * Delete all pipes and heaps returning memory. Must also
1530 * remove references from all ipfw rules to all pipes.
1531 */
1532 static void
1533 dummynet_flush(void)
1534 {
1535 struct dn_pipe *pipe, *pipe1;
1536 struct dn_flow_set *fs, *fs1;
1537 int i;
1538
1539 DUMMYNET_LOCK();
1540 /* Free heaps so we don't have unwanted events. */
1541 heap_free(&ready_heap);
1542 heap_free(&wfq_ready_heap);
1543 heap_free(&extract_heap);
1544
1545 /*
1546 * Now purge all queued pkts and delete all pipes.
1547 *
1548 * XXXGL: can we merge the for(;;) cycles into one or not?
1549 */
1550 for (i = 0; i < HASHSIZE; i++)
1551 SLIST_FOREACH_SAFE(fs, &flowsethash[i], next, fs1) {
1552 SLIST_REMOVE(&flowsethash[i], fs, dn_flow_set, next);
1553 purge_flow_set(fs, 1);
1554 }
1555 for (i = 0; i < HASHSIZE; i++)
1556 SLIST_FOREACH_SAFE(pipe, &pipehash[i], next, pipe1) {
1557 SLIST_REMOVE(&pipehash[i], pipe, dn_pipe, next);
1558 purge_pipe(pipe);
1559 free(pipe, M_DUMMYNET);
1560 }
1561 DUMMYNET_UNLOCK();
1562 }
1563
1564 extern struct ip_fw *ip_fw_default_rule ;
1565 static void
1566 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1567 {
1568 int i ;
1569 struct dn_flow_queue *q ;
1570 struct mbuf *m ;
1571
1572 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1573 for (q = fs->rq[i] ; q ; q = q->next )
1574 for (m = q->head ; m ; m = m->m_nextpkt ) {
1575 struct dn_pkt_tag *pkt = dn_tag_get(m) ;
1576 if (pkt->rule == r)
1577 pkt->rule = ip_fw_default_rule ;
1578 }
1579 }
1580 /*
1581 * when a firewall rule is deleted, scan all queues and remove the flow-id
1582 * from packets matching this rule.
1583 */
1584 void
1585 dn_rule_delete(void *r)
1586 {
1587 struct dn_pipe *pipe;
1588 struct dn_flow_set *fs;
1589 struct dn_pkt_tag *pkt;
1590 struct mbuf *m;
1591 int i;
1592
1593 DUMMYNET_LOCK();
1594 /*
1595 * If the rule references a queue (dn_flow_set), then scan
1596 * the flow set, otherwise scan pipes. Should do either, but doing
1597 * both does not harm.
1598 */
1599 for (i = 0; i < HASHSIZE; i++)
1600 SLIST_FOREACH(fs, &flowsethash[i], next)
1601 dn_rule_delete_fs(fs, r);
1602
1603 for (i = 0; i < HASHSIZE; i++)
1604 SLIST_FOREACH(pipe, &pipehash[i], next) {
1605 fs = &(pipe->fs);
1606 dn_rule_delete_fs(fs, r);
1607 for (m = pipe->head ; m ; m = m->m_nextpkt ) {
1608 pkt = dn_tag_get(m);
1609 if (pkt->rule == r)
1610 pkt->rule = ip_fw_default_rule;
1611 }
1612 }
1613 DUMMYNET_UNLOCK();
1614 }
1615
1616 /*
1617 * setup RED parameters
1618 */
1619 static int
1620 config_red(struct dn_flow_set *p, struct dn_flow_set *x)
1621 {
1622 int i;
1623
1624 x->w_q = p->w_q;
1625 x->min_th = SCALE(p->min_th);
1626 x->max_th = SCALE(p->max_th);
1627 x->max_p = p->max_p;
1628
1629 x->c_1 = p->max_p / (p->max_th - p->min_th);
1630 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1631
1632 if (x->flags_fs & DN_IS_GENTLE_RED) {
1633 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1634 x->c_4 = SCALE(1) - 2 * p->max_p;
1635 }
1636
1637 /* If the lookup table already exist, free and create it again. */
1638 if (x->w_q_lookup) {
1639 free(x->w_q_lookup, M_DUMMYNET);
1640 x->w_q_lookup = NULL;
1641 }
1642 if (red_lookup_depth == 0) {
1643 printf("\ndummynet: net.inet.ip.dummynet.red_lookup_depth"
1644 "must be > 0\n");
1645 free(x, M_DUMMYNET);
1646 return (EINVAL);
1647 }
1648 x->lookup_depth = red_lookup_depth;
1649 x->w_q_lookup = (u_int *)malloc(x->lookup_depth * sizeof(int),
1650 M_DUMMYNET, M_NOWAIT);
1651 if (x->w_q_lookup == NULL) {
1652 printf("dummynet: sorry, cannot allocate red lookup table\n");
1653 free(x, M_DUMMYNET);
1654 return(ENOSPC);
1655 }
1656
1657 /* Fill the lookup table with (1 - w_q)^x */
1658 x->lookup_step = p->lookup_step;
1659 x->lookup_weight = p->lookup_weight;
1660 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1661
1662 for (i = 1; i < x->lookup_depth; i++)
1663 x->w_q_lookup[i] =
1664 SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1665
1666 if (red_avg_pkt_size < 1)
1667 red_avg_pkt_size = 512;
1668 x->avg_pkt_size = red_avg_pkt_size;
1669 if (red_max_pkt_size < 1)
1670 red_max_pkt_size = 1500;
1671 x->max_pkt_size = red_max_pkt_size;
1672 return (0);
1673 }
1674
1675 static int
1676 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1677 {
1678 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1679 int l = pfs->rq_size;
1680
1681 if (l == 0)
1682 l = dn_hash_size;
1683 if (l < 4)
1684 l = 4;
1685 else if (l > DN_MAX_HASH_SIZE)
1686 l = DN_MAX_HASH_SIZE;
1687 x->rq_size = l;
1688 } else /* one is enough for null mask */
1689 x->rq_size = 1;
1690 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1691 M_DUMMYNET, M_NOWAIT | M_ZERO);
1692 if (x->rq == NULL) {
1693 printf("dummynet: sorry, cannot allocate queue\n");
1694 return (ENOMEM);
1695 }
1696 x->rq_elements = 0;
1697 return 0 ;
1698 }
1699
1700 static void
1701 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1702 {
1703 x->flags_fs = src->flags_fs;
1704 x->qsize = src->qsize;
1705 x->plr = src->plr;
1706 x->flow_mask = src->flow_mask;
1707 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1708 if (x->qsize > pipe_byte_limit)
1709 x->qsize = 1024 * 1024;
1710 } else {
1711 if (x->qsize == 0)
1712 x->qsize = 50;
1713 if (x->qsize > pipe_slot_limit)
1714 x->qsize = 50;
1715 }
1716 /* Configuring RED. */
1717 if (x->flags_fs & DN_IS_RED)
1718 config_red(src, x); /* XXX should check errors */
1719 }
1720
1721 /*
1722 * Setup pipe or queue parameters.
1723 */
1724 static int
1725 config_pipe(struct dn_pipe *p)
1726 {
1727 struct dn_flow_set *pfs = &(p->fs);
1728 struct dn_flow_queue *q;
1729 int i, error;
1730
1731 /*
1732 * The config program passes parameters as follows:
1733 * bw = bits/second (0 means no limits),
1734 * delay = ms, must be translated into ticks.
1735 * qsize = slots/bytes
1736 */
1737 p->delay = (p->delay * hz) / 1000;
1738 /* We need either a pipe number or a flow_set number. */
1739 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1740 return (EINVAL);
1741 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1742 return (EINVAL);
1743 if (p->pipe_nr != 0) { /* this is a pipe */
1744 struct dn_pipe *pipe;
1745
1746 DUMMYNET_LOCK();
1747 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1748
1749 if (pipe == NULL) { /* new pipe */
1750 pipe = malloc(sizeof(struct dn_pipe), M_DUMMYNET,
1751 M_NOWAIT | M_ZERO);
1752 if (pipe == NULL) {
1753 DUMMYNET_UNLOCK();
1754 printf("dummynet: no memory for new pipe\n");
1755 return (ENOMEM);
1756 }
1757 pipe->pipe_nr = p->pipe_nr;
1758 pipe->fs.pipe = pipe;
1759 /*
1760 * idle_heap is the only one from which
1761 * we extract from the middle.
1762 */
1763 pipe->idle_heap.size = pipe->idle_heap.elements = 0;
1764 pipe->idle_heap.offset =
1765 offsetof(struct dn_flow_queue, heap_pos);
1766 } else
1767 /* Flush accumulated credit for all queues. */
1768 for (i = 0; i <= pipe->fs.rq_size; i++)
1769 for (q = pipe->fs.rq[i]; q; q = q->next)
1770 q->numbytes = io_fast ? p->bandwidth : 0;
1771
1772 pipe->bandwidth = p->bandwidth;
1773 pipe->numbytes = 0; /* just in case... */
1774 bcopy(p->if_name, pipe->if_name, sizeof(p->if_name));
1775 pipe->ifp = NULL; /* reset interface ptr */
1776 pipe->delay = p->delay;
1777 set_fs_parms(&(pipe->fs), pfs);
1778
1779 if (pipe->fs.rq == NULL) { /* a new pipe */
1780 error = alloc_hash(&(pipe->fs), pfs);
1781 if (error) {
1782 DUMMYNET_UNLOCK();
1783 free(pipe, M_DUMMYNET);
1784 return (error);
1785 }
1786 SLIST_INSERT_HEAD(&pipehash[HASH(pipe->pipe_nr)],
1787 pipe, next);
1788 }
1789 DUMMYNET_UNLOCK();
1790 } else { /* config queue */
1791 struct dn_flow_set *fs;
1792
1793 DUMMYNET_LOCK();
1794 fs = locate_flowset(pfs->fs_nr); /* locate flow_set */
1795
1796 if (fs == NULL) { /* new */
1797 if (pfs->parent_nr == 0) { /* need link to a pipe */
1798 DUMMYNET_UNLOCK();
1799 return (EINVAL);
1800 }
1801 fs = malloc(sizeof(struct dn_flow_set), M_DUMMYNET,
1802 M_NOWAIT | M_ZERO);
1803 if (fs == NULL) {
1804 DUMMYNET_UNLOCK();
1805 printf(
1806 "dummynet: no memory for new flow_set\n");
1807 return (ENOMEM);
1808 }
1809 fs->fs_nr = pfs->fs_nr;
1810 fs->parent_nr = pfs->parent_nr;
1811 fs->weight = pfs->weight;
1812 if (fs->weight == 0)
1813 fs->weight = 1;
1814 else if (fs->weight > 100)
1815 fs->weight = 100;
1816 } else {
1817 /*
1818 * Change parent pipe not allowed;
1819 * must delete and recreate.
1820 */
1821 if (pfs->parent_nr != 0 &&
1822 fs->parent_nr != pfs->parent_nr) {
1823 DUMMYNET_UNLOCK();
1824 return (EINVAL);
1825 }
1826 }
1827
1828 set_fs_parms(fs, pfs);
1829
1830 if (fs->rq == NULL) { /* a new flow_set */
1831 error = alloc_hash(fs, pfs);
1832 if (error) {
1833 DUMMYNET_UNLOCK();
1834 free(fs, M_DUMMYNET);
1835 return (error);
1836 }
1837 SLIST_INSERT_HEAD(&flowsethash[HASH(fs->fs_nr)],
1838 fs, next);
1839 }
1840 DUMMYNET_UNLOCK();
1841 }
1842 return (0);
1843 }
1844
1845 /*
1846 * Helper function to remove from a heap queues which are linked to
1847 * a flow_set about to be deleted.
1848 */
1849 static void
1850 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1851 {
1852 int i = 0, found = 0 ;
1853 for (; i < h->elements ;)
1854 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1855 h->elements-- ;
1856 h->p[i] = h->p[h->elements] ;
1857 found++ ;
1858 } else
1859 i++ ;
1860 if (found)
1861 heapify(h);
1862 }
1863
1864 /*
1865 * helper function to remove a pipe from a heap (can be there at most once)
1866 */
1867 static void
1868 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1869 {
1870 if (h->elements > 0) {
1871 int i = 0 ;
1872 for (i=0; i < h->elements ; i++ ) {
1873 if (h->p[i].object == p) { /* found it */
1874 h->elements-- ;
1875 h->p[i] = h->p[h->elements] ;
1876 heapify(h);
1877 break ;
1878 }
1879 }
1880 }
1881 }
1882
1883 /*
1884 * drain all queues. Called in case of severe mbuf shortage.
1885 */
1886 void
1887 dummynet_drain(void)
1888 {
1889 struct dn_flow_set *fs;
1890 struct dn_pipe *pipe;
1891 struct mbuf *m, *mnext;
1892 int i;
1893
1894 DUMMYNET_LOCK_ASSERT();
1895
1896 heap_free(&ready_heap);
1897 heap_free(&wfq_ready_heap);
1898 heap_free(&extract_heap);
1899 /* remove all references to this pipe from flow_sets */
1900 for (i = 0; i < HASHSIZE; i++)
1901 SLIST_FOREACH(fs, &flowsethash[i], next)
1902 purge_flow_set(fs, 0);
1903
1904 for (i = 0; i < HASHSIZE; i++) {
1905 SLIST_FOREACH(pipe, &pipehash[i], next) {
1906 purge_flow_set(&(pipe->fs), 0);
1907
1908 mnext = pipe->head;
1909 while ((m = mnext) != NULL) {
1910 mnext = m->m_nextpkt;
1911 DN_FREE_PKT(m);
1912 }
1913 pipe->head = pipe->tail = NULL;
1914 }
1915 }
1916 }
1917
1918 /*
1919 * Fully delete a pipe or a queue, cleaning up associated info.
1920 */
1921 static int
1922 delete_pipe(struct dn_pipe *p)
1923 {
1924
1925 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1926 return EINVAL ;
1927 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1928 return EINVAL ;
1929 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1930 struct dn_pipe *pipe;
1931 struct dn_flow_set *fs;
1932 int i;
1933
1934 DUMMYNET_LOCK();
1935 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1936
1937 if (pipe == NULL) {
1938 DUMMYNET_UNLOCK();
1939 return (ENOENT); /* not found */
1940 }
1941
1942 /* Unlink from list of pipes. */
1943 SLIST_REMOVE(&pipehash[HASH(pipe->pipe_nr)], pipe, dn_pipe, next);
1944
1945 /* Remove all references to this pipe from flow_sets. */
1946 for (i = 0; i < HASHSIZE; i++)
1947 SLIST_FOREACH(fs, &flowsethash[i], next)
1948 if (fs->pipe == pipe) {
1949 printf("dummynet: ++ ref to pipe %d from fs %d\n",
1950 p->pipe_nr, fs->fs_nr);
1951 fs->pipe = NULL ;
1952 purge_flow_set(fs, 0);
1953 }
1954 fs_remove_from_heap(&ready_heap, &(pipe->fs));
1955 purge_pipe(pipe); /* remove all data associated to this pipe */
1956 /* remove reference to here from extract_heap and wfq_ready_heap */
1957 pipe_remove_from_heap(&extract_heap, pipe);
1958 pipe_remove_from_heap(&wfq_ready_heap, pipe);
1959 DUMMYNET_UNLOCK();
1960
1961 free(pipe, M_DUMMYNET);
1962 } else { /* this is a WF2Q queue (dn_flow_set) */
1963 struct dn_flow_set *fs;
1964
1965 DUMMYNET_LOCK();
1966 fs = locate_flowset(p->fs.fs_nr); /* locate set */
1967
1968 if (fs == NULL) {
1969 DUMMYNET_UNLOCK();
1970 return (ENOENT); /* not found */
1971 }
1972
1973 /* Unlink from list of flowsets. */
1974 SLIST_REMOVE( &flowsethash[HASH(fs->fs_nr)], fs, dn_flow_set, next);
1975
1976 if (fs->pipe != NULL) {
1977 /* Update total weight on parent pipe and cleanup parent heaps. */
1978 fs->pipe->sum -= fs->weight * fs->backlogged ;
1979 fs_remove_from_heap(&(fs->pipe->not_eligible_heap), fs);
1980 fs_remove_from_heap(&(fs->pipe->scheduler_heap), fs);
1981 #if 1 /* XXX should i remove from idle_heap as well ? */
1982 fs_remove_from_heap(&(fs->pipe->idle_heap), fs);
1983 #endif
1984 }
1985 purge_flow_set(fs, 1);
1986 DUMMYNET_UNLOCK();
1987 }
1988 return 0 ;
1989 }
1990
1991 /*
1992 * helper function used to copy data from kernel in DUMMYNET_GET
1993 */
1994 static char *
1995 dn_copy_set(struct dn_flow_set *set, char *bp)
1996 {
1997 int i, copied = 0 ;
1998 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1999
2000 DUMMYNET_LOCK_ASSERT();
2001
2002 for (i = 0 ; i <= set->rq_size ; i++)
2003 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
2004 if (q->hash_slot != i)
2005 printf("dummynet: ++ at %d: wrong slot (have %d, "
2006 "should be %d)\n", copied, q->hash_slot, i);
2007 if (q->fs != set)
2008 printf("dummynet: ++ at %d: wrong fs ptr (have %p, should be %p)\n",
2009 i, q->fs, set);
2010 copied++ ;
2011 bcopy(q, qp, sizeof( *q ) );
2012 /* cleanup pointers */
2013 qp->next = NULL ;
2014 qp->head = qp->tail = NULL ;
2015 qp->fs = NULL ;
2016 }
2017 if (copied != set->rq_elements)
2018 printf("dummynet: ++ wrong count, have %d should be %d\n",
2019 copied, set->rq_elements);
2020 return (char *)qp ;
2021 }
2022
2023 static size_t
2024 dn_calc_size(void)
2025 {
2026 struct dn_flow_set *fs;
2027 struct dn_pipe *pipe;
2028 size_t size = 0;
2029 int i;
2030
2031 DUMMYNET_LOCK_ASSERT();
2032 /*
2033 * Compute size of data structures: list of pipes and flow_sets.
2034 */
2035 for (i = 0; i < HASHSIZE; i++) {
2036 SLIST_FOREACH(pipe, &pipehash[i], next)
2037 size += sizeof(*pipe) +
2038 pipe->fs.rq_elements * sizeof(struct dn_flow_queue);
2039 SLIST_FOREACH(fs, &flowsethash[i], next)
2040 size += sizeof (*fs) +
2041 fs->rq_elements * sizeof(struct dn_flow_queue);
2042 }
2043 return size;
2044 }
2045
2046 static int
2047 dummynet_get(struct sockopt *sopt)
2048 {
2049 char *buf, *bp ; /* bp is the "copy-pointer" */
2050 size_t size ;
2051 struct dn_flow_set *fs;
2052 struct dn_pipe *pipe;
2053 int error=0, i ;
2054
2055 /* XXX lock held too long */
2056 DUMMYNET_LOCK();
2057 /*
2058 * XXX: Ugly, but we need to allocate memory with M_WAITOK flag and we
2059 * cannot use this flag while holding a mutex.
2060 */
2061 for (i = 0; i < 10; i++) {
2062 size = dn_calc_size();
2063 DUMMYNET_UNLOCK();
2064 buf = malloc(size, M_TEMP, M_WAITOK);
2065 DUMMYNET_LOCK();
2066 if (size == dn_calc_size())
2067 break;
2068 free(buf, M_TEMP);
2069 buf = NULL;
2070 }
2071 if (buf == NULL) {
2072 DUMMYNET_UNLOCK();
2073 return ENOBUFS ;
2074 }
2075 bp = buf;
2076 for (i = 0; i < HASHSIZE; i++)
2077 SLIST_FOREACH(pipe, &pipehash[i], next) {
2078 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp;
2079
2080 /*
2081 * Copy pipe descriptor into *bp, convert delay back to ms,
2082 * then copy the flow_set descriptor(s) one at a time.
2083 * After each flow_set, copy the queue descriptor it owns.
2084 */
2085 bcopy(pipe, bp, sizeof(*pipe));
2086 pipe_bp->delay = (pipe_bp->delay * 1000) / hz;
2087 /*
2088 * XXX the following is a hack based on ->next being the
2089 * first field in dn_pipe and dn_flow_set. The correct
2090 * solution would be to move the dn_flow_set to the beginning
2091 * of struct dn_pipe.
2092 */
2093 pipe_bp->next.sle_next = (struct dn_pipe *)DN_IS_PIPE;
2094 /* Clean pointers. */
2095 pipe_bp->head = pipe_bp->tail = NULL;
2096 pipe_bp->fs.next.sle_next = NULL;
2097 pipe_bp->fs.pipe = NULL;
2098 pipe_bp->fs.rq = NULL;
2099
2100 bp += sizeof(*pipe) ;
2101 bp = dn_copy_set(&(pipe->fs), bp);
2102 }
2103
2104 for (i = 0; i < HASHSIZE; i++)
2105 SLIST_FOREACH(fs, &flowsethash[i], next) {
2106 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp;
2107
2108 bcopy(fs, bp, sizeof(*fs));
2109 /* XXX same hack as above */
2110 fs_bp->next.sle_next = (struct dn_flow_set *)DN_IS_QUEUE;
2111 fs_bp->pipe = NULL;
2112 fs_bp->rq = NULL;
2113 bp += sizeof(*fs);
2114 bp = dn_copy_set(fs, bp);
2115 }
2116
2117 DUMMYNET_UNLOCK();
2118
2119 error = sooptcopyout(sopt, buf, size);
2120 free(buf, M_TEMP);
2121 return error ;
2122 }
2123
2124 /*
2125 * Handler for the various dummynet socket options (get, flush, config, del)
2126 */
2127 static int
2128 ip_dn_ctl(struct sockopt *sopt)
2129 {
2130 int error = 0 ;
2131 struct dn_pipe *p, tmp_pipe;
2132
2133 error = priv_check(sopt->sopt_td, PRIV_NETINET_DUMMYNET);
2134 if (error)
2135 return (error);
2136
2137 /* Disallow sets in really-really secure mode. */
2138 if (sopt->sopt_dir == SOPT_SET) {
2139 #if __FreeBSD_version >= 500034
2140 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
2141 if (error)
2142 return (error);
2143 #else
2144 if (securelevel >= 3)
2145 return (EPERM);
2146 #endif
2147 }
2148
2149 switch (sopt->sopt_name) {
2150 default :
2151 printf("dummynet: -- unknown option %d", sopt->sopt_name);
2152 return EINVAL ;
2153
2154 case IP_DUMMYNET_GET :
2155 error = dummynet_get(sopt);
2156 break ;
2157
2158 case IP_DUMMYNET_FLUSH :
2159 dummynet_flush() ;
2160 break ;
2161
2162 case IP_DUMMYNET_CONFIGURE :
2163 p = &tmp_pipe ;
2164 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2165 if (error)
2166 break ;
2167 error = config_pipe(p);
2168 break ;
2169
2170 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
2171 p = &tmp_pipe ;
2172 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2173 if (error)
2174 break ;
2175
2176 error = delete_pipe(p);
2177 break ;
2178 }
2179 return error ;
2180 }
2181
2182 static void
2183 ip_dn_init(void)
2184 {
2185 int i;
2186
2187 if (bootverbose)
2188 printf("DUMMYNET with IPv6 initialized (040826)\n");
2189
2190 DUMMYNET_LOCK_INIT();
2191
2192 for (i = 0; i < HASHSIZE; i++) {
2193 SLIST_INIT(&pipehash[i]);
2194 SLIST_INIT(&flowsethash[i]);
2195 }
2196 ready_heap.size = ready_heap.elements = 0;
2197 ready_heap.offset = 0;
2198
2199 wfq_ready_heap.size = wfq_ready_heap.elements = 0;
2200 wfq_ready_heap.offset = 0;
2201
2202 extract_heap.size = extract_heap.elements = 0;
2203 extract_heap.offset = 0;
2204
2205 ip_dn_ctl_ptr = ip_dn_ctl;
2206 ip_dn_io_ptr = dummynet_io;
2207 ip_dn_ruledel_ptr = dn_rule_delete;
2208
2209 TASK_INIT(&dn_task, 0, dummynet_task, NULL);
2210 dn_tq = taskqueue_create_fast("dummynet", M_NOWAIT,
2211 taskqueue_thread_enqueue, &dn_tq);
2212 taskqueue_start_threads(&dn_tq, 1, PI_NET, "dummynet");
2213
2214 callout_init(&dn_timeout, CALLOUT_MPSAFE);
2215 callout_reset(&dn_timeout, 1, dummynet, NULL);
2216
2217 /* Initialize curr_time adjustment mechanics. */
2218 getmicrouptime(&prev_t);
2219 }
2220
2221 #ifdef KLD_MODULE
2222 static void
2223 ip_dn_destroy(void)
2224 {
2225 ip_dn_ctl_ptr = NULL;
2226 ip_dn_io_ptr = NULL;
2227 ip_dn_ruledel_ptr = NULL;
2228
2229 DUMMYNET_LOCK();
2230 callout_stop(&dn_timeout);
2231 DUMMYNET_UNLOCK();
2232 taskqueue_drain(dn_tq, &dn_task);
2233 taskqueue_free(dn_tq);
2234
2235 dummynet_flush();
2236
2237 DUMMYNET_LOCK_DESTROY();
2238 }
2239 #endif /* KLD_MODULE */
2240
2241 static int
2242 dummynet_modevent(module_t mod, int type, void *data)
2243 {
2244
2245 switch (type) {
2246 case MOD_LOAD:
2247 if (DUMMYNET_LOADED) {
2248 printf("DUMMYNET already loaded\n");
2249 return EEXIST ;
2250 }
2251 ip_dn_init();
2252 break;
2253
2254 case MOD_UNLOAD:
2255 #if !defined(KLD_MODULE)
2256 printf("dummynet statically compiled, cannot unload\n");
2257 return EINVAL ;
2258 #else
2259 ip_dn_destroy();
2260 #endif
2261 break ;
2262 default:
2263 return EOPNOTSUPP;
2264 break ;
2265 }
2266 return 0 ;
2267 }
2268
2269 static moduledata_t dummynet_mod = {
2270 "dummynet",
2271 dummynet_modevent,
2272 NULL
2273 };
2274 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY);
2275 MODULE_DEPEND(dummynet, ipfw, 2, 2, 2);
2276 MODULE_VERSION(dummynet, 1);
Cache object: 74841703d2552a7cc03bdc57a873da7e
|