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/module.h>
66 #include <sys/priv.h>
67 #include <sys/proc.h>
68 #include <sys/socket.h>
69 #include <sys/socketvar.h>
70 #include <sys/time.h>
71 #include <sys/sysctl.h>
72 #include <sys/taskqueue.h>
73 #include <net/if.h>
74 #include <net/netisr.h>
75 #include <net/route.h>
76 #include <netinet/in.h>
77 #include <netinet/in_systm.h>
78 #include <netinet/in_var.h>
79 #include <netinet/ip.h>
80 #include <netinet/ip_fw.h>
81 #include <netinet/ip_dummynet.h>
82 #include <netinet/ip_var.h>
83
84 #include <netinet/if_ether.h> /* for struct arpcom */
85
86 #include <netinet/ip6.h> /* for ip6_input, ip6_output prototypes */
87 #include <netinet6/ip6_var.h>
88
89 /*
90 * We keep a private variable for the simulation time, but we could
91 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c)
92 */
93 static dn_key curr_time = 0 ; /* current simulation time */
94
95 static int dn_hash_size = 64 ; /* default hash size */
96
97 /* statistics on number of queue searches and search steps */
98 static long searches, search_steps ;
99 static int pipe_expire = 1 ; /* expire queue if empty */
100 static int dn_max_ratio = 16 ; /* max queues/buckets ratio */
101
102 static long pipe_slot_limit = 100; /* Foot shooting limit for pipe queues. */
103 static long pipe_byte_limit = 1024 * 1024;
104
105 static int red_lookup_depth = 256; /* RED - default lookup table depth */
106 static int red_avg_pkt_size = 512; /* RED - default medium packet size */
107 static int red_max_pkt_size = 1500; /* RED - default max packet size */
108
109 static struct timeval prev_t, t;
110 static long tick_last; /* Last tick duration (usec). */
111 static long tick_delta; /* Last vs standard tick diff (usec). */
112 static long tick_delta_sum; /* Accumulated tick difference (usec).*/
113 static long tick_adjustment; /* Tick adjustments done. */
114 static long tick_lost; /* Lost(coalesced) ticks number. */
115 /* Adjusted vs non-adjusted curr_time difference (ticks). */
116 static long tick_diff;
117
118 static int io_fast;
119 static unsigned long io_pkt;
120 static unsigned long io_pkt_fast;
121 static unsigned long io_pkt_drop;
122
123 /*
124 * Three heaps contain queues and pipes that the scheduler handles:
125 *
126 * ready_heap contains all dn_flow_queue related to fixed-rate pipes.
127 *
128 * wfq_ready_heap contains the pipes associated with WF2Q flows
129 *
130 * extract_heap contains pipes associated with delay lines.
131 *
132 */
133
134 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
135
136 static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ;
137
138 static int heap_init(struct dn_heap *h, int size);
139 static int heap_insert (struct dn_heap *h, dn_key key1, void *p);
140 static void heap_extract(struct dn_heap *h, void *obj);
141 static void transmit_event(struct dn_pipe *pipe, struct mbuf **head,
142 struct mbuf **tail);
143 static void ready_event(struct dn_flow_queue *q, struct mbuf **head,
144 struct mbuf **tail);
145 static void ready_event_wfq(struct dn_pipe *p, struct mbuf **head,
146 struct mbuf **tail);
147
148 #define HASHSIZE 16
149 #define HASH(num) ((((num) >> 8) ^ ((num) >> 4) ^ (num)) & 0x0f)
150 static struct dn_pipe_head pipehash[HASHSIZE]; /* all pipes */
151 static struct dn_flow_set_head flowsethash[HASHSIZE]; /* all flowsets */
152
153 static struct callout dn_timeout;
154
155 extern void (*bridge_dn_p)(struct mbuf *, struct ifnet *);
156
157 #ifdef SYSCTL_NODE
158 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW, 0, "Dummynet");
159 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
160 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
161 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, curr_time,
162 CTLFLAG_RD, &curr_time, 0, "Current tick");
163 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
164 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
165 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
166 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
167 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, searches,
168 CTLFLAG_RD, &searches, 0, "Number of queue searches");
169 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, search_steps,
170 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
171 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
172 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
173 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
174 CTLFLAG_RW, &dn_max_ratio, 0,
175 "Max ratio between dynamic queues and buckets");
176 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
177 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
178 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
179 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
180 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
181 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
182 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta,
183 CTLFLAG_RD, &tick_delta, 0, "Last vs standard tick difference (usec).");
184 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum,
185 CTLFLAG_RD, &tick_delta_sum, 0, "Accumulated tick difference (usec).");
186 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment,
187 CTLFLAG_RD, &tick_adjustment, 0, "Tick adjustments done.");
188 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff,
189 CTLFLAG_RD, &tick_diff, 0,
190 "Adjusted vs non-adjusted curr_time difference (ticks).");
191 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost,
192 CTLFLAG_RD, &tick_lost, 0,
193 "Number of ticks coalesced by dummynet taskqueue.");
194 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast,
195 CTLFLAG_RW, &io_fast, 0, "Enable fast dummynet io.");
196 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt,
197 CTLFLAG_RD, &io_pkt, 0,
198 "Number of packets passed to dummynet.");
199 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast,
200 CTLFLAG_RD, &io_pkt_fast, 0,
201 "Number of packets bypassed dummynet scheduler.");
202 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop,
203 CTLFLAG_RD, &io_pkt_drop, 0,
204 "Number of packets dropped by dummynet.");
205 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit,
206 CTLFLAG_RW, &pipe_slot_limit, 0, "Upper limit in slots for pipe queue.");
207 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit,
208 CTLFLAG_RW, &pipe_byte_limit, 0, "Upper limit in bytes for pipe queue.");
209 #endif
210
211 #ifdef DUMMYNET_DEBUG
212 int dummynet_debug = 0;
213 #ifdef SYSCTL_NODE
214 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, CTLFLAG_RW, &dummynet_debug,
215 0, "control debugging printfs");
216 #endif
217 #define DPRINTF(X) if (dummynet_debug) printf X
218 #else
219 #define DPRINTF(X)
220 #endif
221
222 static struct task dn_task;
223 static struct taskqueue *dn_tq = NULL;
224 static void dummynet_task(void *, int);
225
226 static struct mtx dummynet_mtx;
227 #define DUMMYNET_LOCK_INIT() \
228 mtx_init(&dummynet_mtx, "dummynet", NULL, MTX_DEF)
229 #define DUMMYNET_LOCK_DESTROY() mtx_destroy(&dummynet_mtx)
230 #define DUMMYNET_LOCK() mtx_lock(&dummynet_mtx)
231 #define DUMMYNET_UNLOCK() mtx_unlock(&dummynet_mtx)
232 #define DUMMYNET_LOCK_ASSERT() mtx_assert(&dummynet_mtx, MA_OWNED)
233
234 static int config_pipe(struct dn_pipe *p);
235 static int ip_dn_ctl(struct sockopt *sopt);
236
237 static void dummynet(void *);
238 static void dummynet_flush(void);
239 static void dummynet_send(struct mbuf *);
240 void dummynet_drain(void);
241 static ip_dn_io_t dummynet_io;
242 static void dn_rule_delete(void *);
243
244 /*
245 * Heap management functions.
246 *
247 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
248 * Some macros help finding parent/children so we can optimize them.
249 *
250 * heap_init() is called to expand the heap when needed.
251 * Increment size in blocks of 16 entries.
252 * XXX failure to allocate a new element is a pretty bad failure
253 * as we basically stall a whole queue forever!!
254 * Returns 1 on error, 0 on success
255 */
256 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
257 #define HEAP_LEFT(x) ( 2*(x) + 1 )
258 #define HEAP_IS_LEFT(x) ( (x) & 1 )
259 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
260 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
261 #define HEAP_INCREMENT 15
262
263 static int
264 heap_init(struct dn_heap *h, int new_size)
265 {
266 struct dn_heap_entry *p;
267
268 if (h->size >= new_size ) {
269 printf("dummynet: %s, Bogus call, have %d want %d\n", __func__,
270 h->size, new_size);
271 return 0 ;
272 }
273 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
274 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_NOWAIT);
275 if (p == NULL) {
276 printf("dummynet: %s, resize %d failed\n", __func__, new_size );
277 return 1 ; /* error */
278 }
279 if (h->size > 0) {
280 bcopy(h->p, p, h->size * sizeof(*p) );
281 free(h->p, M_DUMMYNET);
282 }
283 h->p = p ;
284 h->size = new_size ;
285 return 0 ;
286 }
287
288 /*
289 * Insert element in heap. Normally, p != NULL, we insert p in
290 * a new position and bubble up. If p == NULL, then the element is
291 * already in place, and key is the position where to start the
292 * bubble-up.
293 * Returns 1 on failure (cannot allocate new heap entry)
294 *
295 * If offset > 0 the position (index, int) of the element in the heap is
296 * also stored in the element itself at the given offset in bytes.
297 */
298 #define SET_OFFSET(heap, node) \
299 if (heap->offset > 0) \
300 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
301 /*
302 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
303 */
304 #define RESET_OFFSET(heap, node) \
305 if (heap->offset > 0) \
306 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
307 static int
308 heap_insert(struct dn_heap *h, dn_key key1, void *p)
309 {
310 int son = h->elements ;
311
312 if (p == NULL) /* data already there, set starting point */
313 son = key1 ;
314 else { /* insert new element at the end, possibly resize */
315 son = h->elements ;
316 if (son == h->size) /* need resize... */
317 if (heap_init(h, h->elements+1) )
318 return 1 ; /* failure... */
319 h->p[son].object = p ;
320 h->p[son].key = key1 ;
321 h->elements++ ;
322 }
323 while (son > 0) { /* bubble up */
324 int father = HEAP_FATHER(son) ;
325 struct dn_heap_entry tmp ;
326
327 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
328 break ; /* found right position */
329 /* son smaller than father, swap and repeat */
330 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
331 SET_OFFSET(h, son);
332 son = father ;
333 }
334 SET_OFFSET(h, son);
335 return 0 ;
336 }
337
338 /*
339 * remove top element from heap, or obj if obj != NULL
340 */
341 static void
342 heap_extract(struct dn_heap *h, void *obj)
343 {
344 int child, father, max = h->elements - 1 ;
345
346 if (max < 0) {
347 printf("dummynet: warning, extract from empty heap 0x%p\n", h);
348 return ;
349 }
350 father = 0 ; /* default: move up smallest child */
351 if (obj != NULL) { /* extract specific element, index is at offset */
352 if (h->offset <= 0)
353 panic("dummynet: heap_extract from middle not supported on this heap!!!\n");
354 father = *((int *)((char *)obj + h->offset)) ;
355 if (father < 0 || father >= h->elements) {
356 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
357 father, h->elements);
358 panic("dummynet: heap_extract");
359 }
360 }
361 RESET_OFFSET(h, father);
362 child = HEAP_LEFT(father) ; /* left child */
363 while (child <= max) { /* valid entry */
364 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
365 child = child+1 ; /* take right child, otherwise left */
366 h->p[father] = h->p[child] ;
367 SET_OFFSET(h, father);
368 father = child ;
369 child = HEAP_LEFT(child) ; /* left child for next loop */
370 }
371 h->elements-- ;
372 if (father != max) {
373 /*
374 * Fill hole with last entry and bubble up, reusing the insert code
375 */
376 h->p[father] = h->p[max] ;
377 heap_insert(h, father, NULL); /* this one cannot fail */
378 }
379 }
380
381 #if 0
382 /*
383 * change object position and update references
384 * XXX this one is never used!
385 */
386 static void
387 heap_move(struct dn_heap *h, dn_key new_key, void *object)
388 {
389 int temp;
390 int i ;
391 int max = h->elements-1 ;
392 struct dn_heap_entry buf ;
393
394 if (h->offset <= 0)
395 panic("cannot move items on this heap");
396
397 i = *((int *)((char *)object + h->offset));
398 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
399 h->p[i].key = new_key ;
400 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
401 i = temp ) { /* bubble up */
402 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
403 SET_OFFSET(h, i);
404 }
405 } else { /* must move down */
406 h->p[i].key = new_key ;
407 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
408 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
409 temp++ ; /* select child with min key */
410 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
411 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
412 SET_OFFSET(h, i);
413 } else
414 break ;
415 i = temp ;
416 }
417 }
418 SET_OFFSET(h, i);
419 }
420 #endif /* heap_move, unused */
421
422 /*
423 * heapify() will reorganize data inside an array to maintain the
424 * heap property. It is needed when we delete a bunch of entries.
425 */
426 static void
427 heapify(struct dn_heap *h)
428 {
429 int i ;
430
431 for (i = 0 ; i < h->elements ; i++ )
432 heap_insert(h, i , NULL) ;
433 }
434
435 /*
436 * cleanup the heap and free data structure
437 */
438 static void
439 heap_free(struct dn_heap *h)
440 {
441 if (h->size >0 )
442 free(h->p, M_DUMMYNET);
443 bzero(h, sizeof(*h) );
444 }
445
446 /*
447 * --- end of heap management functions ---
448 */
449
450 /*
451 * Return the mbuf tag holding the dummynet state. As an optimization
452 * this is assumed to be the first tag on the list. If this turns out
453 * wrong we'll need to search the list.
454 */
455 static struct dn_pkt_tag *
456 dn_tag_get(struct mbuf *m)
457 {
458 struct m_tag *mtag = m_tag_first(m);
459 KASSERT(mtag != NULL &&
460 mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
461 mtag->m_tag_id == PACKET_TAG_DUMMYNET,
462 ("packet on dummynet queue w/o dummynet tag!"));
463 return (struct dn_pkt_tag *)(mtag+1);
464 }
465
466 /*
467 * Scheduler functions:
468 *
469 * transmit_event() is called when the delay-line needs to enter
470 * the scheduler, either because of existing pkts getting ready,
471 * or new packets entering the queue. The event handled is the delivery
472 * time of the packet.
473 *
474 * ready_event() does something similar with fixed-rate queues, and the
475 * event handled is the finish time of the head pkt.
476 *
477 * wfq_ready_event() does something similar with WF2Q queues, and the
478 * event handled is the start time of the head pkt.
479 *
480 * In all cases, we make sure that the data structures are consistent
481 * before passing pkts out, because this might trigger recursive
482 * invocations of the procedures.
483 */
484 static void
485 transmit_event(struct dn_pipe *pipe, struct mbuf **head, struct mbuf **tail)
486 {
487 struct mbuf *m;
488 struct dn_pkt_tag *pkt;
489
490 DUMMYNET_LOCK_ASSERT();
491
492 while ((m = pipe->head) != NULL) {
493 pkt = dn_tag_get(m);
494 if (!DN_KEY_LEQ(pkt->output_time, curr_time))
495 break;
496
497 pipe->head = m->m_nextpkt;
498 if (*tail != NULL)
499 (*tail)->m_nextpkt = m;
500 else
501 *head = m;
502 *tail = m;
503 }
504 if (*tail != NULL)
505 (*tail)->m_nextpkt = NULL;
506
507 /* If there are leftover packets, put into the heap for next event. */
508 if ((m = pipe->head) != NULL) {
509 pkt = dn_tag_get(m);
510 /*
511 * XXX Should check errors on heap_insert, by draining the
512 * whole pipe p and hoping in the future we are more successful.
513 */
514 heap_insert(&extract_heap, pkt->output_time, pipe);
515 }
516 }
517
518 /*
519 * the following macro computes how many ticks we have to wait
520 * before being able to transmit a packet. The credit is taken from
521 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
522 */
523 #define SET_TICKS(_m, q, p) \
524 ((_m)->m_pkthdr.len * 8 * hz - (q)->numbytes + p->bandwidth - 1) / \
525 p->bandwidth;
526
527 /*
528 * extract pkt from queue, compute output time (could be now)
529 * and put into delay line (p_queue)
530 */
531 static void
532 move_pkt(struct mbuf *pkt, struct dn_flow_queue *q, struct dn_pipe *p,
533 int len)
534 {
535 struct dn_pkt_tag *dt = dn_tag_get(pkt);
536
537 q->head = pkt->m_nextpkt ;
538 q->len-- ;
539 q->len_bytes -= len ;
540
541 dt->output_time = curr_time + p->delay ;
542
543 if (p->head == NULL)
544 p->head = pkt;
545 else
546 p->tail->m_nextpkt = pkt;
547 p->tail = pkt;
548 p->tail->m_nextpkt = NULL;
549 }
550
551 /*
552 * ready_event() is invoked every time the queue must enter the
553 * scheduler, either because the first packet arrives, or because
554 * a previously scheduled event fired.
555 * On invokation, drain as many pkts as possible (could be 0) and then
556 * if there are leftover packets reinsert the pkt in the scheduler.
557 */
558 static void
559 ready_event(struct dn_flow_queue *q, struct mbuf **head, struct mbuf **tail)
560 {
561 struct mbuf *pkt;
562 struct dn_pipe *p = q->fs->pipe;
563 int p_was_empty;
564
565 DUMMYNET_LOCK_ASSERT();
566
567 if (p == NULL) {
568 printf("dummynet: ready_event- pipe is gone\n");
569 return;
570 }
571 p_was_empty = (p->head == NULL);
572
573 /*
574 * Schedule fixed-rate queues linked to this pipe:
575 * account for the bw accumulated since last scheduling, then
576 * drain as many pkts as allowed by q->numbytes and move to
577 * the delay line (in p) computing output time.
578 * bandwidth==0 (no limit) means we can drain the whole queue,
579 * setting len_scaled = 0 does the job.
580 */
581 q->numbytes += (curr_time - q->sched_time) * p->bandwidth;
582 while ((pkt = q->head) != NULL) {
583 int len = pkt->m_pkthdr.len;
584 int len_scaled = p->bandwidth ? len * 8 * hz : 0;
585
586 if (len_scaled > q->numbytes)
587 break;
588 q->numbytes -= len_scaled;
589 move_pkt(pkt, q, p, len);
590 }
591 /*
592 * If we have more packets queued, schedule next ready event
593 * (can only occur when bandwidth != 0, otherwise we would have
594 * flushed the whole queue in the previous loop).
595 * To this purpose we record the current time and compute how many
596 * ticks to go for the finish time of the packet.
597 */
598 if ((pkt = q->head) != NULL) { /* this implies bandwidth != 0 */
599 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
600
601 q->sched_time = curr_time;
602 heap_insert(&ready_heap, curr_time + t, (void *)q);
603 /*
604 * XXX Should check errors on heap_insert, and drain the whole
605 * queue on error hoping next time we are luckier.
606 */
607 } else /* RED needs to know when the queue becomes empty. */
608 q->q_time = curr_time;
609
610 /*
611 * If the delay line was empty call transmit_event() now.
612 * Otherwise, the scheduler will take care of it.
613 */
614 if (p_was_empty)
615 transmit_event(p, head, tail);
616 }
617
618 /*
619 * Called when we can transmit packets on WF2Q queues. Take pkts out of
620 * the queues at their start time, and enqueue into the delay line.
621 * Packets are drained until p->numbytes < 0. As long as
622 * len_scaled >= p->numbytes, the packet goes into the delay line
623 * with a deadline p->delay. For the last packet, if p->numbytes < 0,
624 * there is an additional delay.
625 */
626 static void
627 ready_event_wfq(struct dn_pipe *p, struct mbuf **head, struct mbuf **tail)
628 {
629 int p_was_empty = (p->head == NULL);
630 struct dn_heap *sch = &(p->scheduler_heap);
631 struct dn_heap *neh = &(p->not_eligible_heap);
632 int64_t p_numbytes = p->numbytes;
633
634 DUMMYNET_LOCK_ASSERT();
635
636 if (p->if_name[0] == 0) /* tx clock is simulated */
637 /*
638 * Since result may not fit into p->numbytes (32bit) we
639 * are using 64bit var here.
640 */
641 p_numbytes += (curr_time - p->sched_time) * p->bandwidth;
642 else { /*
643 * tx clock is for real,
644 * the ifq must be empty or this is a NOP.
645 */
646 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
647 return;
648 else {
649 DPRINTF(("dummynet: pipe %d ready from %s --\n",
650 p->pipe_nr, p->if_name));
651 }
652 }
653
654 /*
655 * While we have backlogged traffic AND credit, we need to do
656 * something on the queue.
657 */
658 while (p_numbytes >= 0 && (sch->elements > 0 || neh->elements > 0)) {
659 if (sch->elements > 0) {
660 /* Have some eligible pkts to send out. */
661 struct dn_flow_queue *q = sch->p[0].object;
662 struct mbuf *pkt = q->head;
663 struct dn_flow_set *fs = q->fs;
664 uint64_t len = pkt->m_pkthdr.len;
665 int len_scaled = p->bandwidth ? len * 8 * hz : 0;
666
667 heap_extract(sch, NULL); /* Remove queue from heap. */
668 p_numbytes -= len_scaled;
669 move_pkt(pkt, q, p, len);
670
671 p->V += (len << MY_M) / p->sum; /* Update V. */
672 q->S = q->F; /* Update start time. */
673 if (q->len == 0) {
674 /* Flow not backlogged any more. */
675 fs->backlogged--;
676 heap_insert(&(p->idle_heap), q->F, q);
677 } else {
678 /* Still backlogged. */
679
680 /*
681 * Update F and position in backlogged queue,
682 * then put flow in not_eligible_heap
683 * (we will fix this later).
684 */
685 len = (q->head)->m_pkthdr.len;
686 q->F += (len << MY_M) / (uint64_t)fs->weight;
687 if (DN_KEY_LEQ(q->S, p->V))
688 heap_insert(neh, q->S, q);
689 else
690 heap_insert(sch, q->F, q);
691 }
692 }
693 /*
694 * Now compute V = max(V, min(S_i)). Remember that all elements
695 * in sch have by definition S_i <= V so if sch is not empty,
696 * V is surely the max and we must not update it. Conversely,
697 * if sch is empty we only need to look at neh.
698 */
699 if (sch->elements == 0 && neh->elements > 0)
700 p->V = MAX64(p->V, neh->p[0].key);
701 /* Move from neh to sch any packets that have become eligible */
702 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V)) {
703 struct dn_flow_queue *q = neh->p[0].object;
704 heap_extract(neh, NULL);
705 heap_insert(sch, q->F, q);
706 }
707
708 if (p->if_name[0] != '\0') { /* Tx clock is from a real thing */
709 p_numbytes = -1; /* Mark not ready for I/O. */
710 break;
711 }
712 }
713 if (sch->elements == 0 && neh->elements == 0 && p_numbytes >= 0 &&
714 p->idle_heap.elements > 0) {
715 /*
716 * No traffic and no events scheduled.
717 * We can get rid of idle-heap.
718 */
719 int i;
720
721 for (i = 0; i < p->idle_heap.elements; i++) {
722 struct dn_flow_queue *q = p->idle_heap.p[i].object;
723
724 q->F = 0;
725 q->S = q->F + 1;
726 }
727 p->sum = 0;
728 p->V = 0;
729 p->idle_heap.elements = 0;
730 }
731 /*
732 * If we are getting clocks from dummynet (not a real interface) and
733 * If we are under credit, schedule the next ready event.
734 * Also fix the delivery time of the last packet.
735 */
736 if (p->if_name[0]==0 && p_numbytes < 0) { /* This implies bw > 0. */
737 dn_key t = 0; /* Number of ticks i have to wait. */
738
739 if (p->bandwidth > 0)
740 t = (p->bandwidth - 1 - p_numbytes) / p->bandwidth;
741 dn_tag_get(p->tail)->output_time += t;
742 p->sched_time = curr_time;
743 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
744 /*
745 * XXX Should check errors on heap_insert, and drain the whole
746 * queue on error hoping next time we are luckier.
747 */
748 }
749
750 /* Fit (adjust if necessary) 64bit result into 32bit variable. */
751 if (p_numbytes > INT_MAX)
752 p->numbytes = INT_MAX;
753 else if (p_numbytes < INT_MIN)
754 p->numbytes = INT_MIN;
755 else
756 p->numbytes = p_numbytes;
757
758 /*
759 * If the delay line was empty call transmit_event() now.
760 * Otherwise, the scheduler will take care of it.
761 */
762 if (p_was_empty)
763 transmit_event(p, head, tail);
764 }
765
766 /*
767 * This is called one tick, after previous run. It is used to
768 * schedule next run.
769 */
770 static void
771 dummynet(void * __unused unused)
772 {
773
774 taskqueue_enqueue(dn_tq, &dn_task);
775 }
776
777 /*
778 * The main dummynet processing function.
779 */
780 static void
781 dummynet_task(void *context, int pending)
782 {
783 struct mbuf *head = NULL, *tail = NULL;
784 struct dn_pipe *pipe;
785 struct dn_heap *heaps[3];
786 struct dn_heap *h;
787 void *p; /* generic parameter to handler */
788 int i;
789
790 DUMMYNET_LOCK();
791
792 heaps[0] = &ready_heap; /* fixed-rate queues */
793 heaps[1] = &wfq_ready_heap; /* wfq queues */
794 heaps[2] = &extract_heap; /* delay line */
795
796 /* Update number of lost(coalesced) ticks. */
797 tick_lost += pending - 1;
798
799 getmicrouptime(&t);
800 /* Last tick duration (usec). */
801 tick_last = (t.tv_sec - prev_t.tv_sec) * 1000000 +
802 (t.tv_usec - prev_t.tv_usec);
803 /* Last tick vs standard tick difference (usec). */
804 tick_delta = (tick_last * hz - 1000000) / hz;
805 /* Accumulated tick difference (usec). */
806 tick_delta_sum += tick_delta;
807
808 prev_t = t;
809
810 /*
811 * Adjust curr_time if accumulated tick difference greater than
812 * 'standard' tick. Since curr_time should be monotonically increasing,
813 * we do positive adjustment as required and throttle curr_time in
814 * case of negative adjustment.
815 */
816 curr_time++;
817 if (tick_delta_sum - tick >= 0) {
818 int diff = tick_delta_sum / tick;
819
820 curr_time += diff;
821 tick_diff += diff;
822 tick_delta_sum %= tick;
823 tick_adjustment++;
824 } else if (tick_delta_sum + tick <= 0) {
825 curr_time--;
826 tick_diff--;
827 tick_delta_sum += tick;
828 tick_adjustment++;
829 }
830
831 for (i = 0; i < 3; i++) {
832 h = heaps[i];
833 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time)) {
834 if (h->p[0].key > curr_time)
835 printf("dummynet: warning, "
836 "heap %d is %d ticks late\n",
837 i, (int)(curr_time - h->p[0].key));
838 /* store a copy before heap_extract */
839 p = h->p[0].object;
840 /* need to extract before processing */
841 heap_extract(h, NULL);
842 if (i == 0)
843 ready_event(p, &head, &tail);
844 else if (i == 1) {
845 struct dn_pipe *pipe = p;
846 if (pipe->if_name[0] != '\0')
847 printf("dummynet: bad ready_event_wfq "
848 "for pipe %s\n", pipe->if_name);
849 else
850 ready_event_wfq(p, &head, &tail);
851 } else
852 transmit_event(p, &head, &tail);
853 }
854 }
855
856 /* Sweep pipes trying to expire idle flow_queues. */
857 for (i = 0; i < HASHSIZE; i++)
858 SLIST_FOREACH(pipe, &pipehash[i], next)
859 if (pipe->idle_heap.elements > 0 &&
860 DN_KEY_LT(pipe->idle_heap.p[0].key, pipe->V)) {
861 struct dn_flow_queue *q =
862 pipe->idle_heap.p[0].object;
863
864 heap_extract(&(pipe->idle_heap), NULL);
865 /* Mark timestamp as invalid. */
866 q->S = q->F + 1;
867 pipe->sum -= q->fs->weight;
868 }
869
870 DUMMYNET_UNLOCK();
871
872 if (head != NULL)
873 dummynet_send(head);
874
875 callout_reset(&dn_timeout, 1, dummynet, NULL);
876 }
877
878 static void
879 dummynet_send(struct mbuf *m)
880 {
881 struct dn_pkt_tag *pkt;
882 struct mbuf *n;
883 struct ip *ip;
884
885 for (; m != NULL; m = n) {
886 n = m->m_nextpkt;
887 m->m_nextpkt = NULL;
888 pkt = dn_tag_get(m);
889 switch (pkt->dn_dir) {
890 case DN_TO_IP_OUT:
891 ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL);
892 break ;
893 case DN_TO_IP_IN :
894 ip = mtod(m, struct ip *);
895 ip->ip_len = htons(ip->ip_len);
896 ip->ip_off = htons(ip->ip_off);
897 netisr_dispatch(NETISR_IP, m);
898 break;
899 #ifdef INET6
900 case DN_TO_IP6_IN:
901 netisr_dispatch(NETISR_IPV6, m);
902 break;
903
904 case DN_TO_IP6_OUT:
905 ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL);
906 break;
907 #endif
908 case DN_TO_IFB_FWD:
909 if (bridge_dn_p != NULL)
910 ((*bridge_dn_p)(m, pkt->ifp));
911 else
912 printf("dummynet: if_bridge not loaded\n");
913
914 break;
915 case DN_TO_ETH_DEMUX:
916 /*
917 * The Ethernet code assumes the Ethernet header is
918 * contiguous in the first mbuf header.
919 * Insure this is true.
920 */
921 if (m->m_len < ETHER_HDR_LEN &&
922 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
923 printf("dummynet/ether: pullup failed, "
924 "dropping packet\n");
925 break;
926 }
927 ether_demux(m->m_pkthdr.rcvif, m);
928 break;
929 case DN_TO_ETH_OUT:
930 ether_output_frame(pkt->ifp, m);
931 break;
932 default:
933 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
934 m_freem(m);
935 break;
936 }
937 }
938 }
939
940 /*
941 * Unconditionally expire empty queues in case of shortage.
942 * Returns the number of queues freed.
943 */
944 static int
945 expire_queues(struct dn_flow_set *fs)
946 {
947 struct dn_flow_queue *q, *prev ;
948 int i, initial_elements = fs->rq_elements ;
949
950 if (fs->last_expired == time_uptime)
951 return 0 ;
952 fs->last_expired = time_uptime ;
953 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
954 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
955 if (q->head != NULL || q->S != q->F+1) {
956 prev = q ;
957 q = q->next ;
958 } else { /* entry is idle, expire it */
959 struct dn_flow_queue *old_q = q ;
960
961 if (prev != NULL)
962 prev->next = q = q->next ;
963 else
964 fs->rq[i] = q = q->next ;
965 fs->rq_elements-- ;
966 free(old_q, M_DUMMYNET);
967 }
968 return initial_elements - fs->rq_elements ;
969 }
970
971 /*
972 * If room, create a new queue and put at head of slot i;
973 * otherwise, create or use the default queue.
974 */
975 static struct dn_flow_queue *
976 create_queue(struct dn_flow_set *fs, int i)
977 {
978 struct dn_flow_queue *q;
979
980 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
981 expire_queues(fs) == 0) {
982 /* No way to get room, use or create overflow queue. */
983 i = fs->rq_size;
984 if (fs->rq[i] != NULL)
985 return fs->rq[i];
986 }
987 q = malloc(sizeof(*q), M_DUMMYNET, M_NOWAIT | M_ZERO);
988 if (q == NULL) {
989 printf("dummynet: sorry, cannot allocate queue for new flow\n");
990 return (NULL);
991 }
992 q->fs = fs;
993 q->hash_slot = i;
994 q->next = fs->rq[i];
995 q->S = q->F + 1; /* hack - mark timestamp as invalid. */
996 q->numbytes = io_fast ? fs->pipe->bandwidth : 0;
997 fs->rq[i] = q;
998 fs->rq_elements++;
999 return (q);
1000 }
1001
1002 /*
1003 * Given a flow_set and a pkt in last_pkt, find a matching queue
1004 * after appropriate masking. The queue is moved to front
1005 * so that further searches take less time.
1006 */
1007 static struct dn_flow_queue *
1008 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
1009 {
1010 int i = 0 ; /* we need i and q for new allocations */
1011 struct dn_flow_queue *q, *prev;
1012 int is_v6 = IS_IP6_FLOW_ID(id);
1013
1014 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
1015 q = fs->rq[0] ;
1016 else {
1017 /* first, do the masking, then hash */
1018 id->dst_port &= fs->flow_mask.dst_port ;
1019 id->src_port &= fs->flow_mask.src_port ;
1020 id->proto &= fs->flow_mask.proto ;
1021 id->flags = 0 ; /* we don't care about this one */
1022 if (is_v6) {
1023 APPLY_MASK(&id->dst_ip6, &fs->flow_mask.dst_ip6);
1024 APPLY_MASK(&id->src_ip6, &fs->flow_mask.src_ip6);
1025 id->flow_id6 &= fs->flow_mask.flow_id6;
1026
1027 i = ((id->dst_ip6.__u6_addr.__u6_addr32[0]) & 0xffff)^
1028 ((id->dst_ip6.__u6_addr.__u6_addr32[1]) & 0xffff)^
1029 ((id->dst_ip6.__u6_addr.__u6_addr32[2]) & 0xffff)^
1030 ((id->dst_ip6.__u6_addr.__u6_addr32[3]) & 0xffff)^
1031
1032 ((id->dst_ip6.__u6_addr.__u6_addr32[0] >> 15) & 0xffff)^
1033 ((id->dst_ip6.__u6_addr.__u6_addr32[1] >> 15) & 0xffff)^
1034 ((id->dst_ip6.__u6_addr.__u6_addr32[2] >> 15) & 0xffff)^
1035 ((id->dst_ip6.__u6_addr.__u6_addr32[3] >> 15) & 0xffff)^
1036
1037 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 1) & 0xfffff)^
1038 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 1) & 0xfffff)^
1039 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 1) & 0xfffff)^
1040 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 1) & 0xfffff)^
1041
1042 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 16) & 0xffff)^
1043 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 16) & 0xffff)^
1044 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 16) & 0xffff)^
1045 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 16) & 0xffff)^
1046
1047 (id->dst_port << 1) ^ (id->src_port) ^
1048 (id->proto ) ^
1049 (id->flow_id6);
1050 } else {
1051 id->dst_ip &= fs->flow_mask.dst_ip ;
1052 id->src_ip &= fs->flow_mask.src_ip ;
1053
1054 i = ( (id->dst_ip) & 0xffff ) ^
1055 ( (id->dst_ip >> 15) & 0xffff ) ^
1056 ( (id->src_ip << 1) & 0xffff ) ^
1057 ( (id->src_ip >> 16 ) & 0xffff ) ^
1058 (id->dst_port << 1) ^ (id->src_port) ^
1059 (id->proto );
1060 }
1061 i = i % fs->rq_size ;
1062 /* finally, scan the current list for a match */
1063 searches++ ;
1064 for (prev=NULL, q = fs->rq[i] ; q ; ) {
1065 search_steps++;
1066 if (is_v6 &&
1067 IN6_ARE_ADDR_EQUAL(&id->dst_ip6,&q->id.dst_ip6) &&
1068 IN6_ARE_ADDR_EQUAL(&id->src_ip6,&q->id.src_ip6) &&
1069 id->dst_port == q->id.dst_port &&
1070 id->src_port == q->id.src_port &&
1071 id->proto == q->id.proto &&
1072 id->flags == q->id.flags &&
1073 id->flow_id6 == q->id.flow_id6)
1074 break ; /* found */
1075
1076 if (!is_v6 && id->dst_ip == q->id.dst_ip &&
1077 id->src_ip == q->id.src_ip &&
1078 id->dst_port == q->id.dst_port &&
1079 id->src_port == q->id.src_port &&
1080 id->proto == q->id.proto &&
1081 id->flags == q->id.flags)
1082 break ; /* found */
1083
1084 /* No match. Check if we can expire the entry */
1085 if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
1086 /* entry is idle and not in any heap, expire it */
1087 struct dn_flow_queue *old_q = q ;
1088
1089 if (prev != NULL)
1090 prev->next = q = q->next ;
1091 else
1092 fs->rq[i] = q = q->next ;
1093 fs->rq_elements-- ;
1094 free(old_q, M_DUMMYNET);
1095 continue ;
1096 }
1097 prev = q ;
1098 q = q->next ;
1099 }
1100 if (q && prev != NULL) { /* found and not in front */
1101 prev->next = q->next ;
1102 q->next = fs->rq[i] ;
1103 fs->rq[i] = q ;
1104 }
1105 }
1106 if (q == NULL) { /* no match, need to allocate a new entry */
1107 q = create_queue(fs, i);
1108 if (q != NULL)
1109 q->id = *id ;
1110 }
1111 return q ;
1112 }
1113
1114 static int
1115 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
1116 {
1117 /*
1118 * RED algorithm
1119 *
1120 * RED calculates the average queue size (avg) using a low-pass filter
1121 * with an exponential weighted (w_q) moving average:
1122 * avg <- (1-w_q) * avg + w_q * q_size
1123 * where q_size is the queue length (measured in bytes or * packets).
1124 *
1125 * If q_size == 0, we compute the idle time for the link, and set
1126 * avg = (1 - w_q)^(idle/s)
1127 * where s is the time needed for transmitting a medium-sized packet.
1128 *
1129 * Now, if avg < min_th the packet is enqueued.
1130 * If avg > max_th the packet is dropped. Otherwise, the packet is
1131 * dropped with probability P function of avg.
1132 */
1133
1134 int64_t p_b = 0;
1135
1136 /* Queue in bytes or packets? */
1137 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ?
1138 q->len_bytes : q->len;
1139
1140 DPRINTF(("\ndummynet: %d q: %2u ", (int)curr_time, q_size));
1141
1142 /* Average queue size estimation. */
1143 if (q_size != 0) {
1144 /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */
1145 int diff = SCALE(q_size) - q->avg;
1146 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);
1147
1148 q->avg += (int)v;
1149 } else {
1150 /*
1151 * Queue is empty, find for how long the queue has been
1152 * empty and use a lookup table for computing
1153 * (1 - * w_q)^(idle_time/s) where s is the time to send a
1154 * (small) packet.
1155 * XXX check wraps...
1156 */
1157 if (q->avg) {
1158 u_int t = ((uint32_t)curr_time - q->q_time) /
1159 fs->lookup_step;
1160
1161 q->avg = (t < fs->lookup_depth) ?
1162 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
1163 }
1164 }
1165 DPRINTF(("dummynet: avg: %u ", SCALE_VAL(q->avg)));
1166
1167 /* Should i drop? */
1168 if (q->avg < fs->min_th) {
1169 q->count = -1;
1170 return (0); /* accept packet */
1171 }
1172 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
1173 if (fs->flags_fs & DN_IS_GENTLE_RED) {
1174 /*
1175 * According to Gentle-RED, if avg is greater than
1176 * max_th the packet is dropped with a probability
1177 * p_b = c_3 * avg - c_4
1178 * where c_3 = (1 - max_p) / max_th
1179 * c_4 = 1 - 2 * max_p
1180 */
1181 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) -
1182 fs->c_4;
1183 } else {
1184 q->count = -1;
1185 DPRINTF(("dummynet: - drop"));
1186 return (1);
1187 }
1188 } else if (q->avg > fs->min_th) {
1189 /*
1190 * We compute p_b using the linear dropping function
1191 * p_b = c_1 * avg - c_2
1192 * where c_1 = max_p / (max_th - min_th)
1193 * c_2 = max_p * min_th / (max_th - min_th)
1194 */
1195 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
1196 }
1197
1198 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
1199 p_b = (p_b * len) / fs->max_pkt_size;
1200 if (++q->count == 0)
1201 q->random = random() & 0xffff;
1202 else {
1203 /*
1204 * q->count counts packets arrived since last drop, so a greater
1205 * value of q->count means a greater packet drop probability.
1206 */
1207 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
1208 q->count = 0;
1209 DPRINTF(("dummynet: - red drop"));
1210 /* After a drop we calculate a new random value. */
1211 q->random = random() & 0xffff;
1212 return (1); /* drop */
1213 }
1214 }
1215 /* End of RED algorithm. */
1216
1217 return (0); /* accept */
1218 }
1219
1220 static __inline struct dn_flow_set *
1221 locate_flowset(int fs_nr)
1222 {
1223 struct dn_flow_set *fs;
1224
1225 SLIST_FOREACH(fs, &flowsethash[HASH(fs_nr)], next)
1226 if (fs->fs_nr == fs_nr)
1227 return (fs);
1228
1229 return (NULL);
1230 }
1231
1232 static __inline struct dn_pipe *
1233 locate_pipe(int pipe_nr)
1234 {
1235 struct dn_pipe *pipe;
1236
1237 SLIST_FOREACH(pipe, &pipehash[HASH(pipe_nr)], next)
1238 if (pipe->pipe_nr == pipe_nr)
1239 return (pipe);
1240
1241 return (NULL);
1242 }
1243
1244 /*
1245 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1246 * depending on whether WF2Q or fixed bw is used.
1247 *
1248 * pipe_nr pipe or queue the packet is destined for.
1249 * dir where shall we send the packet after dummynet.
1250 * m the mbuf with the packet
1251 * ifp the 'ifp' parameter from the caller.
1252 * NULL in ip_input, destination interface in ip_output,
1253 * rule matching rule, in case of multiple passes
1254 */
1255 static int
1256 dummynet_io(struct mbuf **m0, int dir, struct ip_fw_args *fwa)
1257 {
1258 struct mbuf *m = *m0, *head = NULL, *tail = NULL;
1259 struct dn_pkt_tag *pkt;
1260 struct m_tag *mtag;
1261 struct dn_flow_set *fs = NULL;
1262 struct dn_pipe *pipe;
1263 uint64_t len = m->m_pkthdr.len;
1264 struct dn_flow_queue *q = NULL;
1265 int is_pipe;
1266 ipfw_insn *cmd = ACTION_PTR(fwa->rule);
1267
1268 KASSERT(m->m_nextpkt == NULL,
1269 ("dummynet_io: mbuf queue passed to dummynet"));
1270
1271 if (cmd->opcode == O_LOG)
1272 cmd += F_LEN(cmd);
1273 if (cmd->opcode == O_ALTQ)
1274 cmd += F_LEN(cmd);
1275 if (cmd->opcode == O_TAG)
1276 cmd += F_LEN(cmd);
1277 is_pipe = (cmd->opcode == O_PIPE);
1278
1279 DUMMYNET_LOCK();
1280 io_pkt++;
1281 /*
1282 * This is a dummynet rule, so we expect an O_PIPE or O_QUEUE rule.
1283 *
1284 * XXXGL: probably the pipe->fs and fs->pipe logic here
1285 * below can be simplified.
1286 */
1287 if (is_pipe) {
1288 pipe = locate_pipe(fwa->cookie);
1289 if (pipe != NULL)
1290 fs = &(pipe->fs);
1291 } else
1292 fs = locate_flowset(fwa->cookie);
1293
1294 if (fs == NULL)
1295 goto dropit; /* This queue/pipe does not exist! */
1296 pipe = fs->pipe;
1297 if (pipe == NULL) { /* Must be a queue, try find a matching pipe. */
1298 pipe = locate_pipe(fs->parent_nr);
1299 if (pipe != NULL)
1300 fs->pipe = pipe;
1301 else {
1302 printf("dummynet: no pipe %d for queue %d, drop pkt\n",
1303 fs->parent_nr, fs->fs_nr);
1304 goto dropit;
1305 }
1306 }
1307 q = find_queue(fs, &(fwa->f_id));
1308 if (q == NULL)
1309 goto dropit; /* Cannot allocate queue. */
1310
1311 /* Update statistics, then check reasons to drop pkt. */
1312 q->tot_bytes += len;
1313 q->tot_pkts++;
1314 if (fs->plr && random() < fs->plr)
1315 goto dropit; /* Random pkt drop. */
1316 if (fs->flags_fs & DN_QSIZE_IS_BYTES) {
1317 if (q->len_bytes > fs->qsize)
1318 goto dropit; /* Queue size overflow. */
1319 } else {
1320 if (q->len >= fs->qsize)
1321 goto dropit; /* Queue count overflow. */
1322 }
1323 if (fs->flags_fs & DN_IS_RED && red_drops(fs, q, len))
1324 goto dropit;
1325
1326 /* XXX expensive to zero, see if we can remove it. */
1327 mtag = m_tag_get(PACKET_TAG_DUMMYNET,
1328 sizeof(struct dn_pkt_tag), M_NOWAIT | M_ZERO);
1329 if (mtag == NULL)
1330 goto dropit; /* Cannot allocate packet header. */
1331 m_tag_prepend(m, mtag); /* Attach to mbuf chain. */
1332
1333 pkt = (struct dn_pkt_tag *)(mtag + 1);
1334 /*
1335 * Ok, i can handle the pkt now...
1336 * Build and enqueue packet + parameters.
1337 */
1338 pkt->rule = fwa->rule;
1339 pkt->dn_dir = dir;
1340
1341 pkt->ifp = fwa->oif;
1342
1343 if (q->head == NULL)
1344 q->head = m;
1345 else
1346 q->tail->m_nextpkt = m;
1347 q->tail = m;
1348 q->len++;
1349 q->len_bytes += len;
1350
1351 if (q->head != m) /* Flow was not idle, we are done. */
1352 goto done;
1353
1354 if (q->q_time < (uint32_t)curr_time)
1355 q->numbytes = io_fast ? fs->pipe->bandwidth : 0;
1356 q->q_time = curr_time;
1357
1358 /*
1359 * If we reach this point the flow was previously idle, so we need
1360 * to schedule it. This involves different actions for fixed-rate or
1361 * WF2Q queues.
1362 */
1363 if (is_pipe) {
1364 /* Fixed-rate queue: just insert into the ready_heap. */
1365 dn_key t = 0;
1366
1367 if (pipe->bandwidth && m->m_pkthdr.len * 8 * hz > q->numbytes)
1368 t = SET_TICKS(m, q, pipe);
1369 q->sched_time = curr_time;
1370 if (t == 0) /* Must process it now. */
1371 ready_event(q, &head, &tail);
1372 else
1373 heap_insert(&ready_heap, curr_time + t , q);
1374 } else {
1375 /*
1376 * WF2Q. First, compute start time S: if the flow was
1377 * idle (S = F + 1) set S to the virtual time V for the
1378 * controlling pipe, and update the sum of weights for the pipe;
1379 * otherwise, remove flow from idle_heap and set S to max(F,V).
1380 * Second, compute finish time F = S + len / weight.
1381 * Third, if pipe was idle, update V = max(S, V).
1382 * Fourth, count one more backlogged flow.
1383 */
1384 if (DN_KEY_GT(q->S, q->F)) { /* Means timestamps are invalid. */
1385 q->S = pipe->V;
1386 pipe->sum += fs->weight; /* Add weight of new queue. */
1387 } else {
1388 heap_extract(&(pipe->idle_heap), q);
1389 q->S = MAX64(q->F, pipe->V);
1390 }
1391 q->F = q->S + (len << MY_M) / (uint64_t)fs->weight;
1392
1393 if (pipe->not_eligible_heap.elements == 0 &&
1394 pipe->scheduler_heap.elements == 0)
1395 pipe->V = MAX64(q->S, pipe->V);
1396 fs->backlogged++;
1397 /*
1398 * Look at eligibility. A flow is not eligibile if S>V (when
1399 * this happens, it means that there is some other flow already
1400 * scheduled for the same pipe, so the scheduler_heap cannot be
1401 * empty). If the flow is not eligible we just store it in the
1402 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1403 * and possibly invoke ready_event_wfq() right now if there is
1404 * leftover credit.
1405 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1406 * and for all flows in not_eligible_heap (NEH), S_i > V.
1407 * So when we need to compute max(V, min(S_i)) forall i in
1408 * SCH+NEH, we only need to look into NEH.
1409 */
1410 if (DN_KEY_GT(q->S, pipe->V)) { /* Not eligible. */
1411 if (pipe->scheduler_heap.elements == 0)
1412 printf("dummynet: ++ ouch! not eligible but empty scheduler!\n");
1413 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1414 } else {
1415 heap_insert(&(pipe->scheduler_heap), q->F, q);
1416 if (pipe->numbytes >= 0) { /* Pipe is idle. */
1417 if (pipe->scheduler_heap.elements != 1)
1418 printf("dummynet: OUCH! pipe should have been idle!\n");
1419 DPRINTF(("dummynet: waking up pipe %d at %d\n",
1420 pipe->pipe_nr, (int)(q->F >> MY_M)));
1421 pipe->sched_time = curr_time;
1422 ready_event_wfq(pipe, &head, &tail);
1423 }
1424 }
1425 }
1426 done:
1427 if (head == m && dir != DN_TO_IFB_FWD && dir != DN_TO_ETH_DEMUX &&
1428 dir != DN_TO_ETH_OUT) { /* Fast io. */
1429 io_pkt_fast++;
1430 if (m->m_nextpkt != NULL)
1431 printf("dummynet: fast io: pkt chain detected!\n");
1432 head = m->m_nextpkt = NULL;
1433 } else
1434 *m0 = NULL; /* Normal io. */
1435
1436 DUMMYNET_UNLOCK();
1437 if (head != NULL)
1438 dummynet_send(head);
1439 return (0);
1440
1441 dropit:
1442 io_pkt_drop++;
1443 if (q)
1444 q->drops++;
1445 DUMMYNET_UNLOCK();
1446 m_freem(m);
1447 *m0 = NULL;
1448 return ((fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1449 }
1450
1451 /*
1452 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1453 * Doing this would probably save us the initial bzero of dn_pkt
1454 */
1455 #define DN_FREE_PKT(_m) do { \
1456 m_freem(_m); \
1457 } while (0)
1458
1459 /*
1460 * Dispose all packets and flow_queues on a flow_set.
1461 * If all=1, also remove red lookup table and other storage,
1462 * including the descriptor itself.
1463 * For the one in dn_pipe MUST also cleanup ready_heap...
1464 */
1465 static void
1466 purge_flow_set(struct dn_flow_set *fs, int all)
1467 {
1468 struct dn_flow_queue *q, *qn;
1469 int i;
1470
1471 DUMMYNET_LOCK_ASSERT();
1472
1473 for (i = 0; i <= fs->rq_size; i++) {
1474 for (q = fs->rq[i]; q != NULL; q = qn) {
1475 struct mbuf *m, *mnext;
1476
1477 mnext = q->head;
1478 while ((m = mnext) != NULL) {
1479 mnext = m->m_nextpkt;
1480 DN_FREE_PKT(m);
1481 }
1482 qn = q->next;
1483 free(q, M_DUMMYNET);
1484 }
1485 fs->rq[i] = NULL;
1486 }
1487
1488 fs->rq_elements = 0;
1489 if (all) {
1490 /* RED - free lookup table. */
1491 if (fs->w_q_lookup != NULL)
1492 free(fs->w_q_lookup, M_DUMMYNET);
1493 if (fs->rq != NULL)
1494 free(fs->rq, M_DUMMYNET);
1495 /* If this fs is not part of a pipe, free it. */
1496 if (fs->pipe == NULL || fs != &(fs->pipe->fs))
1497 free(fs, M_DUMMYNET);
1498 }
1499 }
1500
1501 /*
1502 * Dispose all packets queued on a pipe (not a flow_set).
1503 * Also free all resources associated to a pipe, which is about
1504 * to be deleted.
1505 */
1506 static void
1507 purge_pipe(struct dn_pipe *pipe)
1508 {
1509 struct mbuf *m, *mnext;
1510
1511 purge_flow_set( &(pipe->fs), 1 );
1512
1513 mnext = pipe->head;
1514 while ((m = mnext) != NULL) {
1515 mnext = m->m_nextpkt;
1516 DN_FREE_PKT(m);
1517 }
1518
1519 heap_free( &(pipe->scheduler_heap) );
1520 heap_free( &(pipe->not_eligible_heap) );
1521 heap_free( &(pipe->idle_heap) );
1522 }
1523
1524 /*
1525 * Delete all pipes and heaps returning memory. Must also
1526 * remove references from all ipfw rules to all pipes.
1527 */
1528 static void
1529 dummynet_flush(void)
1530 {
1531 struct dn_pipe *pipe, *pipe1;
1532 struct dn_flow_set *fs, *fs1;
1533 int i;
1534
1535 DUMMYNET_LOCK();
1536 /* Free heaps so we don't have unwanted events. */
1537 heap_free(&ready_heap);
1538 heap_free(&wfq_ready_heap);
1539 heap_free(&extract_heap);
1540
1541 /*
1542 * Now purge all queued pkts and delete all pipes.
1543 *
1544 * XXXGL: can we merge the for(;;) cycles into one or not?
1545 */
1546 for (i = 0; i < HASHSIZE; i++)
1547 SLIST_FOREACH_SAFE(fs, &flowsethash[i], next, fs1) {
1548 SLIST_REMOVE(&flowsethash[i], fs, dn_flow_set, next);
1549 purge_flow_set(fs, 1);
1550 }
1551 for (i = 0; i < HASHSIZE; i++)
1552 SLIST_FOREACH_SAFE(pipe, &pipehash[i], next, pipe1) {
1553 SLIST_REMOVE(&pipehash[i], pipe, dn_pipe, next);
1554 purge_pipe(pipe);
1555 free(pipe, M_DUMMYNET);
1556 }
1557 DUMMYNET_UNLOCK();
1558 }
1559
1560 extern struct ip_fw *ip_fw_default_rule ;
1561 static void
1562 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1563 {
1564 int i ;
1565 struct dn_flow_queue *q ;
1566 struct mbuf *m ;
1567
1568 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1569 for (q = fs->rq[i] ; q ; q = q->next )
1570 for (m = q->head ; m ; m = m->m_nextpkt ) {
1571 struct dn_pkt_tag *pkt = dn_tag_get(m) ;
1572 if (pkt->rule == r)
1573 pkt->rule = ip_fw_default_rule ;
1574 }
1575 }
1576 /*
1577 * when a firewall rule is deleted, scan all queues and remove the flow-id
1578 * from packets matching this rule.
1579 */
1580 void
1581 dn_rule_delete(void *r)
1582 {
1583 struct dn_pipe *pipe;
1584 struct dn_flow_set *fs;
1585 struct dn_pkt_tag *pkt;
1586 struct mbuf *m;
1587 int i;
1588
1589 DUMMYNET_LOCK();
1590 /*
1591 * If the rule references a queue (dn_flow_set), then scan
1592 * the flow set, otherwise scan pipes. Should do either, but doing
1593 * both does not harm.
1594 */
1595 for (i = 0; i < HASHSIZE; i++)
1596 SLIST_FOREACH(fs, &flowsethash[i], next)
1597 dn_rule_delete_fs(fs, r);
1598
1599 for (i = 0; i < HASHSIZE; i++)
1600 SLIST_FOREACH(pipe, &pipehash[i], next) {
1601 fs = &(pipe->fs);
1602 dn_rule_delete_fs(fs, r);
1603 for (m = pipe->head ; m ; m = m->m_nextpkt ) {
1604 pkt = dn_tag_get(m);
1605 if (pkt->rule == r)
1606 pkt->rule = ip_fw_default_rule;
1607 }
1608 }
1609 DUMMYNET_UNLOCK();
1610 }
1611
1612 /*
1613 * setup RED parameters
1614 */
1615 static int
1616 config_red(struct dn_flow_set *p, struct dn_flow_set *x)
1617 {
1618 int i;
1619
1620 x->w_q = p->w_q;
1621 x->min_th = SCALE(p->min_th);
1622 x->max_th = SCALE(p->max_th);
1623 x->max_p = p->max_p;
1624
1625 x->c_1 = p->max_p / (p->max_th - p->min_th);
1626 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1627
1628 if (x->flags_fs & DN_IS_GENTLE_RED) {
1629 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1630 x->c_4 = SCALE(1) - 2 * p->max_p;
1631 }
1632
1633 /* If the lookup table already exist, free and create it again. */
1634 if (x->w_q_lookup) {
1635 free(x->w_q_lookup, M_DUMMYNET);
1636 x->w_q_lookup = NULL;
1637 }
1638 if (red_lookup_depth == 0) {
1639 printf("\ndummynet: net.inet.ip.dummynet.red_lookup_depth"
1640 "must be > 0\n");
1641 free(x, M_DUMMYNET);
1642 return (EINVAL);
1643 }
1644 x->lookup_depth = red_lookup_depth;
1645 x->w_q_lookup = (u_int *)malloc(x->lookup_depth * sizeof(int),
1646 M_DUMMYNET, M_NOWAIT);
1647 if (x->w_q_lookup == NULL) {
1648 printf("dummynet: sorry, cannot allocate red lookup table\n");
1649 free(x, M_DUMMYNET);
1650 return(ENOSPC);
1651 }
1652
1653 /* Fill the lookup table with (1 - w_q)^x */
1654 x->lookup_step = p->lookup_step;
1655 x->lookup_weight = p->lookup_weight;
1656 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1657
1658 for (i = 1; i < x->lookup_depth; i++)
1659 x->w_q_lookup[i] =
1660 SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1661
1662 if (red_avg_pkt_size < 1)
1663 red_avg_pkt_size = 512;
1664 x->avg_pkt_size = red_avg_pkt_size;
1665 if (red_max_pkt_size < 1)
1666 red_max_pkt_size = 1500;
1667 x->max_pkt_size = red_max_pkt_size;
1668 return (0);
1669 }
1670
1671 static int
1672 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1673 {
1674 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1675 int l = pfs->rq_size;
1676
1677 if (l == 0)
1678 l = dn_hash_size;
1679 if (l < 4)
1680 l = 4;
1681 else if (l > DN_MAX_HASH_SIZE)
1682 l = DN_MAX_HASH_SIZE;
1683 x->rq_size = l;
1684 } else /* one is enough for null mask */
1685 x->rq_size = 1;
1686 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1687 M_DUMMYNET, M_NOWAIT | M_ZERO);
1688 if (x->rq == NULL) {
1689 printf("dummynet: sorry, cannot allocate queue\n");
1690 return (ENOMEM);
1691 }
1692 x->rq_elements = 0;
1693 return 0 ;
1694 }
1695
1696 static void
1697 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1698 {
1699 x->flags_fs = src->flags_fs;
1700 x->qsize = src->qsize;
1701 x->plr = src->plr;
1702 x->flow_mask = src->flow_mask;
1703 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1704 if (x->qsize > pipe_byte_limit)
1705 x->qsize = 1024 * 1024;
1706 } else {
1707 if (x->qsize == 0)
1708 x->qsize = 50;
1709 if (x->qsize > pipe_slot_limit)
1710 x->qsize = 50;
1711 }
1712 /* Configuring RED. */
1713 if (x->flags_fs & DN_IS_RED)
1714 config_red(src, x); /* XXX should check errors */
1715 }
1716
1717 /*
1718 * Setup pipe or queue parameters.
1719 */
1720 static int
1721 config_pipe(struct dn_pipe *p)
1722 {
1723 struct dn_flow_set *pfs = &(p->fs);
1724 struct dn_flow_queue *q;
1725 int i, error;
1726
1727 /*
1728 * The config program passes parameters as follows:
1729 * bw = bits/second (0 means no limits),
1730 * delay = ms, must be translated into ticks.
1731 * qsize = slots/bytes
1732 */
1733 p->delay = (p->delay * hz) / 1000;
1734 /* We need either a pipe number or a flow_set number. */
1735 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1736 return (EINVAL);
1737 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1738 return (EINVAL);
1739 if (p->pipe_nr != 0) { /* this is a pipe */
1740 struct dn_pipe *pipe;
1741
1742 DUMMYNET_LOCK();
1743 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1744
1745 if (pipe == NULL) { /* new pipe */
1746 pipe = malloc(sizeof(struct dn_pipe), M_DUMMYNET,
1747 M_NOWAIT | M_ZERO);
1748 if (pipe == NULL) {
1749 DUMMYNET_UNLOCK();
1750 printf("dummynet: no memory for new pipe\n");
1751 return (ENOMEM);
1752 }
1753 pipe->pipe_nr = p->pipe_nr;
1754 pipe->fs.pipe = pipe;
1755 /*
1756 * idle_heap is the only one from which
1757 * we extract from the middle.
1758 */
1759 pipe->idle_heap.size = pipe->idle_heap.elements = 0;
1760 pipe->idle_heap.offset =
1761 offsetof(struct dn_flow_queue, heap_pos);
1762 } else
1763 /* Flush accumulated credit for all queues. */
1764 for (i = 0; i <= pipe->fs.rq_size; i++)
1765 for (q = pipe->fs.rq[i]; q; q = q->next)
1766 q->numbytes = io_fast ? p->bandwidth : 0;
1767
1768 pipe->bandwidth = p->bandwidth;
1769 pipe->numbytes = 0; /* just in case... */
1770 bcopy(p->if_name, pipe->if_name, sizeof(p->if_name));
1771 pipe->ifp = NULL; /* reset interface ptr */
1772 pipe->delay = p->delay;
1773 set_fs_parms(&(pipe->fs), pfs);
1774
1775 if (pipe->fs.rq == NULL) { /* a new pipe */
1776 error = alloc_hash(&(pipe->fs), pfs);
1777 if (error) {
1778 DUMMYNET_UNLOCK();
1779 free(pipe, M_DUMMYNET);
1780 return (error);
1781 }
1782 SLIST_INSERT_HEAD(&pipehash[HASH(pipe->pipe_nr)],
1783 pipe, next);
1784 }
1785 DUMMYNET_UNLOCK();
1786 } else { /* config queue */
1787 struct dn_flow_set *fs;
1788
1789 DUMMYNET_LOCK();
1790 fs = locate_flowset(pfs->fs_nr); /* locate flow_set */
1791
1792 if (fs == NULL) { /* new */
1793 if (pfs->parent_nr == 0) { /* need link to a pipe */
1794 DUMMYNET_UNLOCK();
1795 return (EINVAL);
1796 }
1797 fs = malloc(sizeof(struct dn_flow_set), M_DUMMYNET,
1798 M_NOWAIT | M_ZERO);
1799 if (fs == NULL) {
1800 DUMMYNET_UNLOCK();
1801 printf(
1802 "dummynet: no memory for new flow_set\n");
1803 return (ENOMEM);
1804 }
1805 fs->fs_nr = pfs->fs_nr;
1806 fs->parent_nr = pfs->parent_nr;
1807 fs->weight = pfs->weight;
1808 if (fs->weight == 0)
1809 fs->weight = 1;
1810 else if (fs->weight > 100)
1811 fs->weight = 100;
1812 } else {
1813 /*
1814 * Change parent pipe not allowed;
1815 * must delete and recreate.
1816 */
1817 if (pfs->parent_nr != 0 &&
1818 fs->parent_nr != pfs->parent_nr) {
1819 DUMMYNET_UNLOCK();
1820 return (EINVAL);
1821 }
1822 }
1823
1824 set_fs_parms(fs, pfs);
1825
1826 if (fs->rq == NULL) { /* a new flow_set */
1827 error = alloc_hash(fs, pfs);
1828 if (error) {
1829 DUMMYNET_UNLOCK();
1830 free(fs, M_DUMMYNET);
1831 return (error);
1832 }
1833 SLIST_INSERT_HEAD(&flowsethash[HASH(fs->fs_nr)],
1834 fs, next);
1835 }
1836 DUMMYNET_UNLOCK();
1837 }
1838 return (0);
1839 }
1840
1841 /*
1842 * Helper function to remove from a heap queues which are linked to
1843 * a flow_set about to be deleted.
1844 */
1845 static void
1846 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1847 {
1848 int i = 0, found = 0 ;
1849 for (; i < h->elements ;)
1850 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1851 h->elements-- ;
1852 h->p[i] = h->p[h->elements] ;
1853 found++ ;
1854 } else
1855 i++ ;
1856 if (found)
1857 heapify(h);
1858 }
1859
1860 /*
1861 * helper function to remove a pipe from a heap (can be there at most once)
1862 */
1863 static void
1864 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1865 {
1866 if (h->elements > 0) {
1867 int i = 0 ;
1868 for (i=0; i < h->elements ; i++ ) {
1869 if (h->p[i].object == p) { /* found it */
1870 h->elements-- ;
1871 h->p[i] = h->p[h->elements] ;
1872 heapify(h);
1873 break ;
1874 }
1875 }
1876 }
1877 }
1878
1879 /*
1880 * drain all queues. Called in case of severe mbuf shortage.
1881 */
1882 void
1883 dummynet_drain(void)
1884 {
1885 struct dn_flow_set *fs;
1886 struct dn_pipe *pipe;
1887 struct mbuf *m, *mnext;
1888 int i;
1889
1890 DUMMYNET_LOCK_ASSERT();
1891
1892 heap_free(&ready_heap);
1893 heap_free(&wfq_ready_heap);
1894 heap_free(&extract_heap);
1895 /* remove all references to this pipe from flow_sets */
1896 for (i = 0; i < HASHSIZE; i++)
1897 SLIST_FOREACH(fs, &flowsethash[i], next)
1898 purge_flow_set(fs, 0);
1899
1900 for (i = 0; i < HASHSIZE; i++) {
1901 SLIST_FOREACH(pipe, &pipehash[i], next) {
1902 purge_flow_set(&(pipe->fs), 0);
1903
1904 mnext = pipe->head;
1905 while ((m = mnext) != NULL) {
1906 mnext = m->m_nextpkt;
1907 DN_FREE_PKT(m);
1908 }
1909 pipe->head = pipe->tail = NULL;
1910 }
1911 }
1912 }
1913
1914 /*
1915 * Fully delete a pipe or a queue, cleaning up associated info.
1916 */
1917 static int
1918 delete_pipe(struct dn_pipe *p)
1919 {
1920
1921 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1922 return EINVAL ;
1923 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1924 return EINVAL ;
1925 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1926 struct dn_pipe *pipe;
1927 struct dn_flow_set *fs;
1928 int i;
1929
1930 DUMMYNET_LOCK();
1931 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1932
1933 if (pipe == NULL) {
1934 DUMMYNET_UNLOCK();
1935 return (ENOENT); /* not found */
1936 }
1937
1938 /* Unlink from list of pipes. */
1939 SLIST_REMOVE(&pipehash[HASH(pipe->pipe_nr)], pipe, dn_pipe, next);
1940
1941 /* Remove all references to this pipe from flow_sets. */
1942 for (i = 0; i < HASHSIZE; i++)
1943 SLIST_FOREACH(fs, &flowsethash[i], next)
1944 if (fs->pipe == pipe) {
1945 printf("dummynet: ++ ref to pipe %d from fs %d\n",
1946 p->pipe_nr, fs->fs_nr);
1947 fs->pipe = NULL ;
1948 purge_flow_set(fs, 0);
1949 }
1950 fs_remove_from_heap(&ready_heap, &(pipe->fs));
1951 purge_pipe(pipe); /* remove all data associated to this pipe */
1952 /* remove reference to here from extract_heap and wfq_ready_heap */
1953 pipe_remove_from_heap(&extract_heap, pipe);
1954 pipe_remove_from_heap(&wfq_ready_heap, pipe);
1955 DUMMYNET_UNLOCK();
1956
1957 free(pipe, M_DUMMYNET);
1958 } else { /* this is a WF2Q queue (dn_flow_set) */
1959 struct dn_flow_set *fs;
1960
1961 DUMMYNET_LOCK();
1962 fs = locate_flowset(p->fs.fs_nr); /* locate set */
1963
1964 if (fs == NULL) {
1965 DUMMYNET_UNLOCK();
1966 return (ENOENT); /* not found */
1967 }
1968
1969 /* Unlink from list of flowsets. */
1970 SLIST_REMOVE( &flowsethash[HASH(fs->fs_nr)], fs, dn_flow_set, next);
1971
1972 if (fs->pipe != NULL) {
1973 /* Update total weight on parent pipe and cleanup parent heaps. */
1974 fs->pipe->sum -= fs->weight * fs->backlogged ;
1975 fs_remove_from_heap(&(fs->pipe->not_eligible_heap), fs);
1976 fs_remove_from_heap(&(fs->pipe->scheduler_heap), fs);
1977 #if 1 /* XXX should i remove from idle_heap as well ? */
1978 fs_remove_from_heap(&(fs->pipe->idle_heap), fs);
1979 #endif
1980 }
1981 purge_flow_set(fs, 1);
1982 DUMMYNET_UNLOCK();
1983 }
1984 return 0 ;
1985 }
1986
1987 /*
1988 * helper function used to copy data from kernel in DUMMYNET_GET
1989 */
1990 static char *
1991 dn_copy_set(struct dn_flow_set *set, char *bp)
1992 {
1993 int i, copied = 0 ;
1994 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
1995
1996 DUMMYNET_LOCK_ASSERT();
1997
1998 for (i = 0 ; i <= set->rq_size ; i++)
1999 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
2000 if (q->hash_slot != i)
2001 printf("dummynet: ++ at %d: wrong slot (have %d, "
2002 "should be %d)\n", copied, q->hash_slot, i);
2003 if (q->fs != set)
2004 printf("dummynet: ++ at %d: wrong fs ptr (have %p, should be %p)\n",
2005 i, q->fs, set);
2006 copied++ ;
2007 bcopy(q, qp, sizeof( *q ) );
2008 /* cleanup pointers */
2009 qp->next = NULL ;
2010 qp->head = qp->tail = NULL ;
2011 qp->fs = NULL ;
2012 }
2013 if (copied != set->rq_elements)
2014 printf("dummynet: ++ wrong count, have %d should be %d\n",
2015 copied, set->rq_elements);
2016 return (char *)qp ;
2017 }
2018
2019 static size_t
2020 dn_calc_size(void)
2021 {
2022 struct dn_flow_set *fs;
2023 struct dn_pipe *pipe;
2024 size_t size = 0;
2025 int i;
2026
2027 DUMMYNET_LOCK_ASSERT();
2028 /*
2029 * Compute size of data structures: list of pipes and flow_sets.
2030 */
2031 for (i = 0; i < HASHSIZE; i++) {
2032 SLIST_FOREACH(pipe, &pipehash[i], next)
2033 size += sizeof(*pipe) +
2034 pipe->fs.rq_elements * sizeof(struct dn_flow_queue);
2035 SLIST_FOREACH(fs, &flowsethash[i], next)
2036 size += sizeof (*fs) +
2037 fs->rq_elements * sizeof(struct dn_flow_queue);
2038 }
2039 return size;
2040 }
2041
2042 static int
2043 dummynet_get(struct sockopt *sopt)
2044 {
2045 char *buf, *bp ; /* bp is the "copy-pointer" */
2046 size_t size ;
2047 struct dn_flow_set *fs;
2048 struct dn_pipe *pipe;
2049 int error=0, i ;
2050
2051 /* XXX lock held too long */
2052 DUMMYNET_LOCK();
2053 /*
2054 * XXX: Ugly, but we need to allocate memory with M_WAITOK flag and we
2055 * cannot use this flag while holding a mutex.
2056 */
2057 for (i = 0; i < 10; i++) {
2058 size = dn_calc_size();
2059 DUMMYNET_UNLOCK();
2060 buf = malloc(size, M_TEMP, M_WAITOK);
2061 DUMMYNET_LOCK();
2062 if (size == dn_calc_size())
2063 break;
2064 free(buf, M_TEMP);
2065 buf = NULL;
2066 }
2067 if (buf == NULL) {
2068 DUMMYNET_UNLOCK();
2069 return ENOBUFS ;
2070 }
2071 bp = buf;
2072 for (i = 0; i < HASHSIZE; i++)
2073 SLIST_FOREACH(pipe, &pipehash[i], next) {
2074 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp;
2075
2076 /*
2077 * Copy pipe descriptor into *bp, convert delay back to ms,
2078 * then copy the flow_set descriptor(s) one at a time.
2079 * After each flow_set, copy the queue descriptor it owns.
2080 */
2081 bcopy(pipe, bp, sizeof(*pipe));
2082 pipe_bp->delay = (pipe_bp->delay * 1000) / hz;
2083 /*
2084 * XXX the following is a hack based on ->next being the
2085 * first field in dn_pipe and dn_flow_set. The correct
2086 * solution would be to move the dn_flow_set to the beginning
2087 * of struct dn_pipe.
2088 */
2089 pipe_bp->next.sle_next = (struct dn_pipe *)DN_IS_PIPE;
2090 /* Clean pointers. */
2091 pipe_bp->head = pipe_bp->tail = NULL;
2092 pipe_bp->fs.next.sle_next = NULL;
2093 pipe_bp->fs.pipe = NULL;
2094 pipe_bp->fs.rq = NULL;
2095
2096 bp += sizeof(*pipe) ;
2097 bp = dn_copy_set(&(pipe->fs), bp);
2098 }
2099
2100 for (i = 0; i < HASHSIZE; i++)
2101 SLIST_FOREACH(fs, &flowsethash[i], next) {
2102 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp;
2103
2104 bcopy(fs, bp, sizeof(*fs));
2105 /* XXX same hack as above */
2106 fs_bp->next.sle_next = (struct dn_flow_set *)DN_IS_QUEUE;
2107 fs_bp->pipe = NULL;
2108 fs_bp->rq = NULL;
2109 bp += sizeof(*fs);
2110 bp = dn_copy_set(fs, bp);
2111 }
2112
2113 DUMMYNET_UNLOCK();
2114
2115 error = sooptcopyout(sopt, buf, size);
2116 free(buf, M_TEMP);
2117 return error ;
2118 }
2119
2120 /*
2121 * Handler for the various dummynet socket options (get, flush, config, del)
2122 */
2123 static int
2124 ip_dn_ctl(struct sockopt *sopt)
2125 {
2126 int error = 0 ;
2127 struct dn_pipe *p, tmp_pipe;
2128
2129 error = priv_check(sopt->sopt_td, PRIV_NETINET_DUMMYNET);
2130 if (error)
2131 return (error);
2132
2133 /* Disallow sets in really-really secure mode. */
2134 if (sopt->sopt_dir == SOPT_SET) {
2135 #if __FreeBSD_version >= 500034
2136 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
2137 if (error)
2138 return (error);
2139 #else
2140 if (securelevel >= 3)
2141 return (EPERM);
2142 #endif
2143 }
2144
2145 switch (sopt->sopt_name) {
2146 default :
2147 printf("dummynet: -- unknown option %d", sopt->sopt_name);
2148 return EINVAL ;
2149
2150 case IP_DUMMYNET_GET :
2151 error = dummynet_get(sopt);
2152 break ;
2153
2154 case IP_DUMMYNET_FLUSH :
2155 dummynet_flush() ;
2156 break ;
2157
2158 case IP_DUMMYNET_CONFIGURE :
2159 p = &tmp_pipe ;
2160 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2161 if (error)
2162 break ;
2163 error = config_pipe(p);
2164 break ;
2165
2166 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
2167 p = &tmp_pipe ;
2168 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2169 if (error)
2170 break ;
2171
2172 error = delete_pipe(p);
2173 break ;
2174 }
2175 return error ;
2176 }
2177
2178 static void
2179 ip_dn_init(void)
2180 {
2181 int i;
2182
2183 if (bootverbose)
2184 printf("DUMMYNET with IPv6 initialized (040826)\n");
2185
2186 DUMMYNET_LOCK_INIT();
2187
2188 for (i = 0; i < HASHSIZE; i++) {
2189 SLIST_INIT(&pipehash[i]);
2190 SLIST_INIT(&flowsethash[i]);
2191 }
2192 ready_heap.size = ready_heap.elements = 0;
2193 ready_heap.offset = 0;
2194
2195 wfq_ready_heap.size = wfq_ready_heap.elements = 0;
2196 wfq_ready_heap.offset = 0;
2197
2198 extract_heap.size = extract_heap.elements = 0;
2199 extract_heap.offset = 0;
2200
2201 ip_dn_ctl_ptr = ip_dn_ctl;
2202 ip_dn_io_ptr = dummynet_io;
2203 ip_dn_ruledel_ptr = dn_rule_delete;
2204
2205 TASK_INIT(&dn_task, 0, dummynet_task, NULL);
2206 dn_tq = taskqueue_create_fast("dummynet", M_NOWAIT,
2207 taskqueue_thread_enqueue, &dn_tq);
2208 taskqueue_start_threads(&dn_tq, 1, PI_NET, "dummynet");
2209
2210 callout_init(&dn_timeout, CALLOUT_MPSAFE);
2211 callout_reset(&dn_timeout, 1, dummynet, NULL);
2212
2213 /* Initialize curr_time adjustment mechanics. */
2214 getmicrouptime(&prev_t);
2215 }
2216
2217 #ifdef KLD_MODULE
2218 static void
2219 ip_dn_destroy(void)
2220 {
2221 ip_dn_ctl_ptr = NULL;
2222 ip_dn_io_ptr = NULL;
2223 ip_dn_ruledel_ptr = NULL;
2224
2225 DUMMYNET_LOCK();
2226 callout_stop(&dn_timeout);
2227 DUMMYNET_UNLOCK();
2228 taskqueue_drain(dn_tq, &dn_task);
2229 taskqueue_free(dn_tq);
2230
2231 dummynet_flush();
2232
2233 DUMMYNET_LOCK_DESTROY();
2234 }
2235 #endif /* KLD_MODULE */
2236
2237 static int
2238 dummynet_modevent(module_t mod, int type, void *data)
2239 {
2240
2241 switch (type) {
2242 case MOD_LOAD:
2243 if (DUMMYNET_LOADED) {
2244 printf("DUMMYNET already loaded\n");
2245 return EEXIST ;
2246 }
2247 ip_dn_init();
2248 break;
2249
2250 case MOD_UNLOAD:
2251 #if !defined(KLD_MODULE)
2252 printf("dummynet statically compiled, cannot unload\n");
2253 return EINVAL ;
2254 #else
2255 ip_dn_destroy();
2256 #endif
2257 break ;
2258 default:
2259 return EOPNOTSUPP;
2260 break ;
2261 }
2262 return 0 ;
2263 }
2264
2265 static moduledata_t dummynet_mod = {
2266 "dummynet",
2267 dummynet_modevent,
2268 NULL
2269 };
2270 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY);
2271 MODULE_DEPEND(dummynet, ipfw, 2, 2, 2);
2272 MODULE_VERSION(dummynet, 1);
Cache object: b7981b27773d085304b9cb1f6cbf374e
|