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 * $FreeBSD$
28 */
29
30 #define DUMMYNET_DEBUG
31
32 #include "opt_inet6.h"
33
34 /*
35 * This module implements IP dummynet, a bandwidth limiter/delay emulator
36 * used in conjunction with the ipfw package.
37 * Description of the data structures used is in ip_dummynet.h
38 * Here you mainly find the following blocks of code:
39 * + variable declarations;
40 * + heap management functions;
41 * + scheduler and dummynet functions;
42 * + configuration and initialization.
43 *
44 * NOTA BENE: critical sections are protected by the "dummynet lock".
45 *
46 * Most important Changes:
47 *
48 * 011004: KLDable
49 * 010124: Fixed WF2Q behaviour
50 * 010122: Fixed spl protection.
51 * 000601: WF2Q support
52 * 000106: large rewrite, use heaps to handle very many pipes.
53 * 980513: initial release
54 *
55 * include files marked with XXX are probably not needed
56 */
57
58 #include <sys/limits.h>
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/malloc.h>
62 #include <sys/mbuf.h>
63 #include <sys/kernel.h>
64 #include <sys/module.h>
65 #include <sys/proc.h>
66 #include <sys/socket.h>
67 #include <sys/socketvar.h>
68 #include <sys/time.h>
69 #include <sys/sysctl.h>
70 #include <sys/taskqueue.h>
71 #include <net/if.h>
72 #include <net/netisr.h>
73 #include <net/route.h>
74 #include <netinet/in.h>
75 #include <netinet/in_systm.h>
76 #include <netinet/in_var.h>
77 #include <netinet/ip.h>
78 #include <netinet/ip_fw.h>
79 #include <netinet/ip_dummynet.h>
80 #include <netinet/ip_var.h>
81
82 #include <netinet/if_ether.h> /* for struct arpcom */
83 #include <net/bridge.h>
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_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW, 0, "Dummynet");
158 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
159 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size");
160 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, curr_time,
161 CTLFLAG_RD, &curr_time, 0, "Current tick");
162 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap,
163 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap");
164 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap,
165 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap");
166 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, searches,
167 CTLFLAG_RD, &searches, 0, "Number of queue searches");
168 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, search_steps,
169 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps");
170 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire,
171 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty");
172 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len,
173 CTLFLAG_RW, &dn_max_ratio, 0,
174 "Max ratio between dynamic queues and buckets");
175 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
176 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table");
177 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
178 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size");
179 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
180 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size");
181 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta,
182 CTLFLAG_RD, &tick_delta, 0, "Last vs standard tick difference (usec).");
183 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum,
184 CTLFLAG_RD, &tick_delta_sum, 0, "Accumulated tick difference (usec).");
185 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment,
186 CTLFLAG_RD, &tick_adjustment, 0, "Tick adjustments done.");
187 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff,
188 CTLFLAG_RD, &tick_diff, 0,
189 "Adjusted vs non-adjusted curr_time difference (ticks).");
190 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost,
191 CTLFLAG_RD, &tick_lost, 0,
192 "Number of ticks coalesced by dummynet taskqueue.");
193 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast,
194 CTLFLAG_RW, &io_fast, 0, "Enable fast dummynet io.");
195 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt,
196 CTLFLAG_RD, &io_pkt, 0,
197 "Number of packets passed to dummynet.");
198 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast,
199 CTLFLAG_RD, &io_pkt_fast, 0,
200 "Number of packets bypassed dummynet scheduler.");
201 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop,
202 CTLFLAG_RD, &io_pkt_drop, 0,
203 "Number of packets dropped by dummynet.");
204 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit,
205 CTLFLAG_RW, &pipe_slot_limit, 0, "Upper limit in slots for pipe queue.");
206 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit,
207 CTLFLAG_RW, &pipe_byte_limit, 0, "Upper limit in bytes for pipe queue.");
208 #endif
209
210 #ifdef DUMMYNET_DEBUG
211 int dummynet_debug = 0;
212 #ifdef SYSCTL_NODE
213 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug, CTLFLAG_RW, &dummynet_debug,
214 0, "control debugging printfs");
215 #endif
216 #define DPRINTF(X) if (dummynet_debug) printf X
217 #else
218 #define DPRINTF(X)
219 #endif
220
221 static struct task dn_task;
222 static struct taskqueue *dn_tq = NULL;
223 static void dummynet_task(void *, int);
224
225 static struct mtx dummynet_mtx;
226 #define DUMMYNET_LOCK_INIT() \
227 mtx_init(&dummynet_mtx, "dummynet", NULL, MTX_DEF)
228 #define DUMMYNET_LOCK_DESTROY() mtx_destroy(&dummynet_mtx)
229 #define DUMMYNET_LOCK() mtx_lock(&dummynet_mtx)
230 #define DUMMYNET_UNLOCK() mtx_unlock(&dummynet_mtx)
231 #define DUMMYNET_LOCK_ASSERT() do { \
232 mtx_assert(&dummynet_mtx, MA_OWNED); \
233 NET_ASSERT_GIANT(); \
234 } while (0)
235
236 static int config_pipe(struct dn_pipe *p);
237 static int ip_dn_ctl(struct sockopt *sopt);
238
239 static void dummynet(void *);
240 static void dummynet_flush(void);
241 static void dummynet_send(struct mbuf *);
242 void dummynet_drain(void);
243 static ip_dn_io_t dummynet_io;
244 static void dn_rule_delete(void *);
245
246 /*
247 * Heap management functions.
248 *
249 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2.
250 * Some macros help finding parent/children so we can optimize them.
251 *
252 * heap_init() is called to expand the heap when needed.
253 * Increment size in blocks of 16 entries.
254 * XXX failure to allocate a new element is a pretty bad failure
255 * as we basically stall a whole queue forever!!
256 * Returns 1 on error, 0 on success
257 */
258 #define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 )
259 #define HEAP_LEFT(x) ( 2*(x) + 1 )
260 #define HEAP_IS_LEFT(x) ( (x) & 1 )
261 #define HEAP_RIGHT(x) ( 2*(x) + 2 )
262 #define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; }
263 #define HEAP_INCREMENT 15
264
265 static int
266 heap_init(struct dn_heap *h, int new_size)
267 {
268 struct dn_heap_entry *p;
269
270 if (h->size >= new_size ) {
271 printf("dummynet: %s, Bogus call, have %d want %d\n", __func__,
272 h->size, new_size);
273 return 0 ;
274 }
275 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ;
276 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_NOWAIT);
277 if (p == NULL) {
278 printf("dummynet: %s, resize %d failed\n", __func__, new_size );
279 return 1 ; /* error */
280 }
281 if (h->size > 0) {
282 bcopy(h->p, p, h->size * sizeof(*p) );
283 free(h->p, M_DUMMYNET);
284 }
285 h->p = p ;
286 h->size = new_size ;
287 return 0 ;
288 }
289
290 /*
291 * Insert element in heap. Normally, p != NULL, we insert p in
292 * a new position and bubble up. If p == NULL, then the element is
293 * already in place, and key is the position where to start the
294 * bubble-up.
295 * Returns 1 on failure (cannot allocate new heap entry)
296 *
297 * If offset > 0 the position (index, int) of the element in the heap is
298 * also stored in the element itself at the given offset in bytes.
299 */
300 #define SET_OFFSET(heap, node) \
301 if (heap->offset > 0) \
302 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ;
303 /*
304 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value.
305 */
306 #define RESET_OFFSET(heap, node) \
307 if (heap->offset > 0) \
308 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ;
309 static int
310 heap_insert(struct dn_heap *h, dn_key key1, void *p)
311 {
312 int son = h->elements ;
313
314 if (p == NULL) /* data already there, set starting point */
315 son = key1 ;
316 else { /* insert new element at the end, possibly resize */
317 son = h->elements ;
318 if (son == h->size) /* need resize... */
319 if (heap_init(h, h->elements+1) )
320 return 1 ; /* failure... */
321 h->p[son].object = p ;
322 h->p[son].key = key1 ;
323 h->elements++ ;
324 }
325 while (son > 0) { /* bubble up */
326 int father = HEAP_FATHER(son) ;
327 struct dn_heap_entry tmp ;
328
329 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) )
330 break ; /* found right position */
331 /* son smaller than father, swap and repeat */
332 HEAP_SWAP(h->p[son], h->p[father], tmp) ;
333 SET_OFFSET(h, son);
334 son = father ;
335 }
336 SET_OFFSET(h, son);
337 return 0 ;
338 }
339
340 /*
341 * remove top element from heap, or obj if obj != NULL
342 */
343 static void
344 heap_extract(struct dn_heap *h, void *obj)
345 {
346 int child, father, max = h->elements - 1 ;
347
348 if (max < 0) {
349 printf("dummynet: warning, extract from empty heap 0x%p\n", h);
350 return ;
351 }
352 father = 0 ; /* default: move up smallest child */
353 if (obj != NULL) { /* extract specific element, index is at offset */
354 if (h->offset <= 0)
355 panic("dummynet: heap_extract from middle not supported on this heap!!!\n");
356 father = *((int *)((char *)obj + h->offset)) ;
357 if (father < 0 || father >= h->elements) {
358 printf("dummynet: heap_extract, father %d out of bound 0..%d\n",
359 father, h->elements);
360 panic("dummynet: heap_extract");
361 }
362 }
363 RESET_OFFSET(h, father);
364 child = HEAP_LEFT(father) ; /* left child */
365 while (child <= max) { /* valid entry */
366 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) )
367 child = child+1 ; /* take right child, otherwise left */
368 h->p[father] = h->p[child] ;
369 SET_OFFSET(h, father);
370 father = child ;
371 child = HEAP_LEFT(child) ; /* left child for next loop */
372 }
373 h->elements-- ;
374 if (father != max) {
375 /*
376 * Fill hole with last entry and bubble up, reusing the insert code
377 */
378 h->p[father] = h->p[max] ;
379 heap_insert(h, father, NULL); /* this one cannot fail */
380 }
381 }
382
383 #if 0
384 /*
385 * change object position and update references
386 * XXX this one is never used!
387 */
388 static void
389 heap_move(struct dn_heap *h, dn_key new_key, void *object)
390 {
391 int temp;
392 int i ;
393 int max = h->elements-1 ;
394 struct dn_heap_entry buf ;
395
396 if (h->offset <= 0)
397 panic("cannot move items on this heap");
398
399 i = *((int *)((char *)object + h->offset));
400 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */
401 h->p[i].key = new_key ;
402 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ;
403 i = temp ) { /* bubble up */
404 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
405 SET_OFFSET(h, i);
406 }
407 } else { /* must move down */
408 h->p[i].key = new_key ;
409 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */
410 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key))
411 temp++ ; /* select child with min key */
412 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */
413 HEAP_SWAP(h->p[i], h->p[temp], buf) ;
414 SET_OFFSET(h, i);
415 } else
416 break ;
417 i = temp ;
418 }
419 }
420 SET_OFFSET(h, i);
421 }
422 #endif /* heap_move, unused */
423
424 /*
425 * heapify() will reorganize data inside an array to maintain the
426 * heap property. It is needed when we delete a bunch of entries.
427 */
428 static void
429 heapify(struct dn_heap *h)
430 {
431 int i ;
432
433 for (i = 0 ; i < h->elements ; i++ )
434 heap_insert(h, i , NULL) ;
435 }
436
437 /*
438 * cleanup the heap and free data structure
439 */
440 static void
441 heap_free(struct dn_heap *h)
442 {
443 if (h->size >0 )
444 free(h->p, M_DUMMYNET);
445 bzero(h, sizeof(*h) );
446 }
447
448 /*
449 * --- end of heap management functions ---
450 */
451
452 /*
453 * Return the mbuf tag holding the dummynet state. As an optimization
454 * this is assumed to be the first tag on the list. If this turns out
455 * wrong we'll need to search the list.
456 */
457 static struct dn_pkt_tag *
458 dn_tag_get(struct mbuf *m)
459 {
460 struct m_tag *mtag = m_tag_first(m);
461 KASSERT(mtag != NULL &&
462 mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
463 mtag->m_tag_id == PACKET_TAG_DUMMYNET,
464 ("packet on dummynet queue w/o dummynet tag!"));
465 return (struct dn_pkt_tag *)(mtag+1);
466 }
467
468 /*
469 * Scheduler functions:
470 *
471 * transmit_event() is called when the delay-line needs to enter
472 * the scheduler, either because of existing pkts getting ready,
473 * or new packets entering the queue. The event handled is the delivery
474 * time of the packet.
475 *
476 * ready_event() does something similar with fixed-rate queues, and the
477 * event handled is the finish time of the head pkt.
478 *
479 * wfq_ready_event() does something similar with WF2Q queues, and the
480 * event handled is the start time of the head pkt.
481 *
482 * In all cases, we make sure that the data structures are consistent
483 * before passing pkts out, because this might trigger recursive
484 * invocations of the procedures.
485 */
486 static void
487 transmit_event(struct dn_pipe *pipe, struct mbuf **head, struct mbuf **tail)
488 {
489 struct mbuf *m;
490 struct dn_pkt_tag *pkt;
491
492 DUMMYNET_LOCK_ASSERT();
493
494 while ((m = pipe->head) != NULL) {
495 pkt = dn_tag_get(m);
496 if (!DN_KEY_LEQ(pkt->output_time, curr_time))
497 break;
498
499 pipe->head = m->m_nextpkt;
500 if (*tail != NULL)
501 (*tail)->m_nextpkt = m;
502 else
503 *head = m;
504 *tail = m;
505 }
506 if (*tail != NULL)
507 (*tail)->m_nextpkt = NULL;
508
509 /* If there are leftover packets, put into the heap for next event. */
510 if ((m = pipe->head) != NULL) {
511 pkt = dn_tag_get(m);
512 /*
513 * XXX Should check errors on heap_insert, by draining the
514 * whole pipe p and hoping in the future we are more successful.
515 */
516 heap_insert(&extract_heap, pkt->output_time, pipe);
517 }
518 }
519
520 /*
521 * the following macro computes how many ticks we have to wait
522 * before being able to transmit a packet. The credit is taken from
523 * either a pipe (WF2Q) or a flow_queue (per-flow queueing)
524 */
525 #define SET_TICKS(_m, q, p) \
526 ((_m)->m_pkthdr.len * 8 * hz - (q)->numbytes + p->bandwidth - 1) / \
527 p->bandwidth;
528
529 /*
530 * extract pkt from queue, compute output time (could be now)
531 * and put into delay line (p_queue)
532 */
533 static void
534 move_pkt(struct mbuf *pkt, struct dn_flow_queue *q,
535 struct dn_pipe *p, int len)
536 {
537 struct dn_pkt_tag *dt = dn_tag_get(pkt);
538
539 q->head = pkt->m_nextpkt ;
540 q->len-- ;
541 q->len_bytes -= len ;
542
543 dt->output_time = curr_time + p->delay ;
544
545 if (p->head == NULL)
546 p->head = pkt;
547 else
548 p->tail->m_nextpkt = pkt;
549 p->tail = pkt;
550 p->tail->m_nextpkt = NULL;
551 }
552
553 /*
554 * ready_event() is invoked every time the queue must enter the
555 * scheduler, either because the first packet arrives, or because
556 * a previously scheduled event fired.
557 * On invokation, drain as many pkts as possible (could be 0) and then
558 * if there are leftover packets reinsert the pkt in the scheduler.
559 */
560 static void
561 ready_event(struct dn_flow_queue *q, struct mbuf **head, struct mbuf **tail)
562 {
563 struct mbuf *pkt;
564 struct dn_pipe *p = q->fs->pipe;
565 int p_was_empty;
566
567 DUMMYNET_LOCK_ASSERT();
568
569 if (p == NULL) {
570 printf("dummynet: ready_event- pipe is gone\n");
571 return;
572 }
573 p_was_empty = (p->head == NULL);
574
575 /*
576 * Schedule fixed-rate queues linked to this pipe:
577 * account for the bw accumulated since last scheduling, then
578 * drain as many pkts as allowed by q->numbytes and move to
579 * the delay line (in p) computing output time.
580 * bandwidth==0 (no limit) means we can drain the whole queue,
581 * setting len_scaled = 0 does the job.
582 */
583 q->numbytes += (curr_time - q->sched_time) * p->bandwidth;
584 while ((pkt = q->head) != NULL) {
585 int len = pkt->m_pkthdr.len;
586 int len_scaled = p->bandwidth ? len * 8 * hz : 0;
587
588 if (len_scaled > q->numbytes)
589 break;
590 q->numbytes -= len_scaled;
591 move_pkt(pkt, q, p, len);
592 }
593 /*
594 * If we have more packets queued, schedule next ready event
595 * (can only occur when bandwidth != 0, otherwise we would have
596 * flushed the whole queue in the previous loop).
597 * To this purpose we record the current time and compute how many
598 * ticks to go for the finish time of the packet.
599 */
600 if ((pkt = q->head) != NULL) { /* this implies bandwidth != 0 */
601 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */
602
603 q->sched_time = curr_time;
604 heap_insert(&ready_heap, curr_time + t, (void *)q);
605 /*
606 * XXX Should check errors on heap_insert, and drain the whole
607 * queue on error hoping next time we are luckier.
608 */
609 } else /* RED needs to know when the queue becomes empty. */
610 q->q_time = curr_time;
611
612 /*
613 * If the delay line was empty call transmit_event() now.
614 * Otherwise, the scheduler will take care of it.
615 */
616 if (p_was_empty)
617 transmit_event(p, head, tail);
618 }
619
620 /*
621 * Called when we can transmit packets on WF2Q queues. Take pkts out of
622 * the queues at their start time, and enqueue into the delay line.
623 * Packets are drained until p->numbytes < 0. As long as
624 * len_scaled >= p->numbytes, the packet goes into the delay line
625 * with a deadline p->delay. For the last packet, if p->numbytes < 0,
626 * there is an additional delay.
627 */
628 static void
629 ready_event_wfq(struct dn_pipe *p, struct mbuf **head, struct mbuf **tail)
630 {
631 int p_was_empty = (p->head == NULL);
632 struct dn_heap *sch = &(p->scheduler_heap);
633 struct dn_heap *neh = &(p->not_eligible_heap);
634 int64_t p_numbytes = p->numbytes;
635
636 DUMMYNET_LOCK_ASSERT();
637
638 if (p->if_name[0] == 0) /* tx clock is simulated */
639 /*
640 * Since result may not fit into p->numbytes (32bit) we
641 * are using 64bit var here.
642 */
643 p_numbytes += (curr_time - p->sched_time) * p->bandwidth;
644 else { /*
645 * tx clock is for real,
646 * the ifq must be empty or this is a NOP.
647 */
648 if (p->ifp && p->ifp->if_snd.ifq_head != NULL)
649 return;
650 else {
651 DPRINTF(("dummynet: pipe %d ready from %s --\n",
652 p->pipe_nr, p->if_name));
653 }
654 }
655
656 /*
657 * While we have backlogged traffic AND credit, we need to do
658 * something on the queue.
659 */
660 while (p_numbytes >= 0 && (sch->elements > 0 || neh->elements > 0)) {
661 if (sch->elements > 0) {
662 /* Have some eligible pkts to send out. */
663 struct dn_flow_queue *q = sch->p[0].object;
664 struct mbuf *pkt = q->head;
665 struct dn_flow_set *fs = q->fs;
666 uint64_t len = pkt->m_pkthdr.len;
667 int len_scaled = p->bandwidth ? len * 8 * hz : 0;
668
669 heap_extract(sch, NULL); /* Remove queue from heap. */
670 p_numbytes -= len_scaled;
671 move_pkt(pkt, q, p, len);
672
673 p->V += (len << MY_M) / p->sum; /* Update V. */
674 q->S = q->F; /* Update start time. */
675 if (q->len == 0) {
676 /* Flow not backlogged any more. */
677 fs->backlogged--;
678 heap_insert(&(p->idle_heap), q->F, q);
679 } else {
680 /* Still backlogged. */
681
682 /*
683 * Update F and position in backlogged queue,
684 * then put flow in not_eligible_heap
685 * (we will fix this later).
686 */
687 len = (q->head)->m_pkthdr.len;
688 q->F += (len << MY_M) / (uint64_t)fs->weight;
689 if (DN_KEY_LEQ(q->S, p->V))
690 heap_insert(neh, q->S, q);
691 else
692 heap_insert(sch, q->F, q);
693 }
694 }
695 /*
696 * Now compute V = max(V, min(S_i)). Remember that all elements
697 * in sch have by definition S_i <= V so if sch is not empty,
698 * V is surely the max and we must not update it. Conversely,
699 * if sch is empty we only need to look at neh.
700 */
701 if (sch->elements == 0 && neh->elements > 0)
702 p->V = MAX64(p->V, neh->p[0].key);
703 /* Move from neh to sch any packets that have become eligible */
704 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V)) {
705 struct dn_flow_queue *q = neh->p[0].object;
706 heap_extract(neh, NULL);
707 heap_insert(sch, q->F, q);
708 }
709
710 if (p->if_name[0] != '\0') { /* Tx clock is from a real thing */
711 p_numbytes = -1; /* Mark not ready for I/O. */
712 break;
713 }
714 }
715 if (sch->elements == 0 && neh->elements == 0 && p_numbytes >= 0 &&
716 p->idle_heap.elements > 0) {
717 /*
718 * No traffic and no events scheduled.
719 * We can get rid of idle-heap.
720 */
721 int i;
722
723 for (i = 0; i < p->idle_heap.elements; i++) {
724 struct dn_flow_queue *q = p->idle_heap.p[i].object;
725
726 q->F = 0;
727 q->S = q->F + 1;
728 }
729 p->sum = 0;
730 p->V = 0;
731 p->idle_heap.elements = 0;
732 }
733 /*
734 * If we are getting clocks from dummynet (not a real interface) and
735 * If we are under credit, schedule the next ready event.
736 * Also fix the delivery time of the last packet.
737 */
738 if (p->if_name[0]==0 && p_numbytes < 0) { /* This implies bw > 0. */
739 dn_key t = 0; /* Number of ticks i have to wait. */
740
741 if (p->bandwidth > 0)
742 t = (p->bandwidth - 1 - p_numbytes) / p->bandwidth;
743 dn_tag_get(p->tail)->output_time += t;
744 p->sched_time = curr_time;
745 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p);
746 /*
747 * XXX Should check errors on heap_insert, and drain the whole
748 * queue on error hoping next time we are luckier.
749 */
750 }
751
752 /* Fit (adjust if necessary) 64bit result into 32bit variable. */
753 if (p_numbytes > INT_MAX)
754 p->numbytes = INT_MAX;
755 else if (p_numbytes < INT_MIN)
756 p->numbytes = INT_MIN;
757 else
758 p->numbytes = p_numbytes;
759
760 /*
761 * If the delay line was empty call transmit_event() now.
762 * Otherwise, the scheduler will take care of it.
763 */
764 if (p_was_empty)
765 transmit_event(p, head, tail);
766 }
767
768 /*
769 * This is called one tick, after previous run. It is used to
770 * schedule next run.
771 */
772 static void
773 dummynet(void * __unused unused)
774 {
775 taskqueue_enqueue(dn_tq, &dn_task);
776 }
777
778 /*
779 * The main dummynet processing function.
780 */
781 static void
782 dummynet_task(void *context, int pending)
783 {
784
785 struct mbuf *head = NULL, *tail = NULL;
786 struct dn_pipe *pipe;
787 struct dn_heap *heaps[3];
788 struct dn_heap *h;
789 void *p; /* generic parameter to handler */
790 int i;
791
792 NET_LOCK_GIANT();
793 DUMMYNET_LOCK();
794
795 heaps[0] = &ready_heap; /* fixed-rate queues */
796 heaps[1] = &wfq_ready_heap; /* wfq queues */
797 heaps[2] = &extract_heap; /* delay line */
798
799 /* Update number of lost(coalesced) ticks. */
800 tick_lost += pending - 1;
801
802 getmicrouptime(&t);
803 /* Last tick duration (usec). */
804 tick_last = (t.tv_sec - prev_t.tv_sec) * 1000000 +
805 (t.tv_usec - prev_t.tv_usec);
806 /* Last tick vs standard tick difference (usec). */
807 tick_delta = (tick_last * hz - 1000000) / hz;
808 /* Accumulated tick difference (usec). */
809 tick_delta_sum += tick_delta;
810
811 prev_t = t;
812
813 /*
814 * Adjust curr_time if accumulated tick difference greater than
815 * 'standard' tick. Since curr_time should be monotonically increasing,
816 * we do positive adjustment as required and throttle curr_time in
817 * case of negative adjustment.
818 */
819 curr_time++;
820 if (tick_delta_sum - tick >= 0) {
821 int diff = tick_delta_sum / tick;
822
823 curr_time += diff;
824 tick_diff += diff;
825 tick_delta_sum %= tick;
826 tick_adjustment++;
827 } else if (tick_delta_sum + tick <= 0) {
828 curr_time--;
829 tick_diff--;
830 tick_delta_sum += tick;
831 tick_adjustment++;
832 }
833
834 for (i = 0; i < 3; i++) {
835 h = heaps[i];
836 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time)) {
837 if (h->p[0].key > curr_time)
838 printf("dummynet: warning, "
839 "heap %d is %d ticks late\n",
840 i, (int)(curr_time - h->p[0].key));
841 /* store a copy before heap_extract */
842 p = h->p[0].object;
843 /* need to extract before processing */
844 heap_extract(h, NULL);
845 if (i == 0)
846 ready_event(p, &head, &tail);
847 else if (i == 1) {
848 struct dn_pipe *pipe = p;
849 if (pipe->if_name[0] != '\0')
850 printf("dummynet: bad ready_event_wfq "
851 "for pipe %s\n", pipe->if_name);
852 else
853 ready_event_wfq(p, &head, &tail);
854 } else
855 transmit_event(p, &head, &tail);
856 }
857 }
858
859 /* Sweep pipes trying to expire idle flow_queues. */
860 for (i = 0; i < HASHSIZE; i++)
861 SLIST_FOREACH(pipe, &pipehash[i], next)
862 if (pipe->idle_heap.elements > 0 &&
863 DN_KEY_LT(pipe->idle_heap.p[0].key, pipe->V)) {
864 struct dn_flow_queue *q =
865 pipe->idle_heap.p[0].object;
866
867 heap_extract(&(pipe->idle_heap), NULL);
868 /* Mark timestamp as invalid. */
869 q->S = q->F + 1;
870 pipe->sum -= q->fs->weight;
871 }
872
873 DUMMYNET_UNLOCK();
874
875 if (head != NULL)
876 dummynet_send(head);
877
878 callout_reset(&dn_timeout, 1, dummynet, NULL);
879
880 NET_UNLOCK_GIANT();
881 }
882
883 static void
884 dummynet_send(struct mbuf *m)
885 {
886 struct dn_pkt_tag *pkt;
887 struct mbuf *n;
888 struct ip *ip;
889
890 for (; m != NULL; m = n) {
891 n = m->m_nextpkt;
892 m->m_nextpkt = NULL;
893 pkt = dn_tag_get(m);
894 switch (pkt->dn_dir) {
895 case DN_TO_IP_OUT:
896 ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL);
897 break ;
898 case DN_TO_IP_IN :
899 ip = mtod(m, struct ip *);
900 ip->ip_len = htons(ip->ip_len);
901 ip->ip_off = htons(ip->ip_off);
902 netisr_dispatch(NETISR_IP, m);
903 break;
904 #ifdef INET6
905 case DN_TO_IP6_IN:
906 netisr_dispatch(NETISR_IPV6, m);
907 break;
908
909 case DN_TO_IP6_OUT:
910 ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL);
911 break;
912 #endif
913 case DN_TO_IFB_FWD:
914 if (bridge_dn_p != NULL)
915 ((*bridge_dn_p)(m, pkt->ifp));
916 else
917 printf("dummynet: if_bridge not loaded\n");
918
919 break;
920 case DN_TO_BDG_FWD :
921 /*
922 * The bridge requires/assumes the Ethernet header is
923 * contiguous in the first mbuf header. Ensure this
924 * is true.
925 */
926 if (BDG_LOADED) {
927 if (m->m_len < ETHER_HDR_LEN &&
928 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
929 printf("dummynet/bridge: pullup fail, "
930 "dropping pkt\n");
931 break;
932 }
933 m = bdg_forward_ptr(m, pkt->ifp);
934 } else {
935 /*
936 * somebody unloaded the bridge module.
937 * Drop pkt
938 */
939 /* XXX rate limit */
940 printf("dummynet: dropping bridged packet "
941 "trapped in pipe\n");
942 }
943 if (m)
944 m_freem(m);
945 break;
946 case DN_TO_ETH_DEMUX:
947 /*
948 * The Ethernet code assumes the Ethernet header is
949 * contiguous in the first mbuf header.
950 * Insure this is true.
951 */
952 if (m->m_len < ETHER_HDR_LEN &&
953 (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
954 printf("dummynet/ether: pullup failed, "
955 "dropping packet\n");
956 break;
957 }
958 ether_demux(m->m_pkthdr.rcvif, m);
959 break;
960 case DN_TO_ETH_OUT:
961 ether_output_frame(pkt->ifp, m);
962 break;
963 default:
964 printf("dummynet: bad switch %d!\n", pkt->dn_dir);
965 m_freem(m);
966 break;
967 }
968 }
969 }
970
971 /*
972 * Unconditionally expire empty queues in case of shortage.
973 * Returns the number of queues freed.
974 */
975 static int
976 expire_queues(struct dn_flow_set *fs)
977 {
978 struct dn_flow_queue *q, *prev ;
979 int i, initial_elements = fs->rq_elements ;
980
981 if (fs->last_expired == time_second)
982 return 0 ;
983 fs->last_expired = time_second ;
984 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */
985 for (prev=NULL, q = fs->rq[i] ; q != NULL ; )
986 if (q->head != NULL || q->S != q->F+1) {
987 prev = q ;
988 q = q->next ;
989 } else { /* entry is idle, expire it */
990 struct dn_flow_queue *old_q = q ;
991
992 if (prev != NULL)
993 prev->next = q = q->next ;
994 else
995 fs->rq[i] = q = q->next ;
996 fs->rq_elements-- ;
997 free(old_q, M_DUMMYNET);
998 }
999 return initial_elements - fs->rq_elements ;
1000 }
1001
1002 /*
1003 * If room, create a new queue and put at head of slot i;
1004 * otherwise, create or use the default queue.
1005 */
1006 static struct dn_flow_queue *
1007 create_queue(struct dn_flow_set *fs, int i)
1008 {
1009 struct dn_flow_queue *q;
1010
1011 if (fs->rq_elements > fs->rq_size * dn_max_ratio &&
1012 expire_queues(fs) == 0) {
1013 /* No way to get room, use or create overflow queue. */
1014 i = fs->rq_size;
1015 if (fs->rq[i] != NULL)
1016 return fs->rq[i];
1017 }
1018 q = malloc(sizeof(*q), M_DUMMYNET, M_NOWAIT | M_ZERO);
1019 if (q == NULL) {
1020 printf("dummynet: sorry, cannot allocate queue for new flow\n");
1021 return (NULL);
1022 }
1023 q->fs = fs;
1024 q->hash_slot = i;
1025 q->next = fs->rq[i];
1026 q->S = q->F + 1; /* hack - mark timestamp as invalid. */
1027 q->numbytes = io_fast ? fs->pipe->bandwidth : 0;
1028 fs->rq[i] = q;
1029 fs->rq_elements++;
1030 return (q);
1031 }
1032
1033 /*
1034 * Given a flow_set and a pkt in last_pkt, find a matching queue
1035 * after appropriate masking. The queue is moved to front
1036 * so that further searches take less time.
1037 */
1038 static struct dn_flow_queue *
1039 find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id)
1040 {
1041 int i = 0 ; /* we need i and q for new allocations */
1042 struct dn_flow_queue *q, *prev;
1043 int is_v6 = IS_IP6_FLOW_ID(id);
1044
1045 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) )
1046 q = fs->rq[0] ;
1047 else {
1048 /* first, do the masking, then hash */
1049 id->dst_port &= fs->flow_mask.dst_port ;
1050 id->src_port &= fs->flow_mask.src_port ;
1051 id->proto &= fs->flow_mask.proto ;
1052 id->flags = 0 ; /* we don't care about this one */
1053 if (is_v6) {
1054 APPLY_MASK(&id->dst_ip6, &fs->flow_mask.dst_ip6);
1055 APPLY_MASK(&id->src_ip6, &fs->flow_mask.src_ip6);
1056 id->flow_id6 &= fs->flow_mask.flow_id6;
1057
1058 i = ((id->dst_ip6.__u6_addr.__u6_addr32[0]) & 0xffff)^
1059 ((id->dst_ip6.__u6_addr.__u6_addr32[1]) & 0xffff)^
1060 ((id->dst_ip6.__u6_addr.__u6_addr32[2]) & 0xffff)^
1061 ((id->dst_ip6.__u6_addr.__u6_addr32[3]) & 0xffff)^
1062
1063 ((id->dst_ip6.__u6_addr.__u6_addr32[0] >> 15) & 0xffff)^
1064 ((id->dst_ip6.__u6_addr.__u6_addr32[1] >> 15) & 0xffff)^
1065 ((id->dst_ip6.__u6_addr.__u6_addr32[2] >> 15) & 0xffff)^
1066 ((id->dst_ip6.__u6_addr.__u6_addr32[3] >> 15) & 0xffff)^
1067
1068 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 1) & 0xfffff)^
1069 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 1) & 0xfffff)^
1070 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 1) & 0xfffff)^
1071 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 1) & 0xfffff)^
1072
1073 ((id->src_ip6.__u6_addr.__u6_addr32[0] << 16) & 0xffff)^
1074 ((id->src_ip6.__u6_addr.__u6_addr32[1] << 16) & 0xffff)^
1075 ((id->src_ip6.__u6_addr.__u6_addr32[2] << 16) & 0xffff)^
1076 ((id->src_ip6.__u6_addr.__u6_addr32[3] << 16) & 0xffff)^
1077
1078 (id->dst_port << 1) ^ (id->src_port) ^
1079 (id->proto ) ^
1080 (id->flow_id6);
1081 } else {
1082 id->dst_ip &= fs->flow_mask.dst_ip ;
1083 id->src_ip &= fs->flow_mask.src_ip ;
1084
1085 i = ( (id->dst_ip) & 0xffff ) ^
1086 ( (id->dst_ip >> 15) & 0xffff ) ^
1087 ( (id->src_ip << 1) & 0xffff ) ^
1088 ( (id->src_ip >> 16 ) & 0xffff ) ^
1089 (id->dst_port << 1) ^ (id->src_port) ^
1090 (id->proto );
1091 }
1092 i = i % fs->rq_size ;
1093 /* finally, scan the current list for a match */
1094 searches++ ;
1095 for (prev=NULL, q = fs->rq[i] ; q ; ) {
1096 search_steps++;
1097 if (is_v6 &&
1098 IN6_ARE_ADDR_EQUAL(&id->dst_ip6,&q->id.dst_ip6) &&
1099 IN6_ARE_ADDR_EQUAL(&id->src_ip6,&q->id.src_ip6) &&
1100 id->dst_port == q->id.dst_port &&
1101 id->src_port == q->id.src_port &&
1102 id->proto == q->id.proto &&
1103 id->flags == q->id.flags &&
1104 id->flow_id6 == q->id.flow_id6)
1105 break ; /* found */
1106
1107 if (!is_v6 && id->dst_ip == q->id.dst_ip &&
1108 id->src_ip == q->id.src_ip &&
1109 id->dst_port == q->id.dst_port &&
1110 id->src_port == q->id.src_port &&
1111 id->proto == q->id.proto &&
1112 id->flags == q->id.flags)
1113 break ; /* found */
1114
1115 /* No match. Check if we can expire the entry */
1116 if (pipe_expire && q->head == NULL && q->S == q->F+1 ) {
1117 /* entry is idle and not in any heap, expire it */
1118 struct dn_flow_queue *old_q = q ;
1119
1120 if (prev != NULL)
1121 prev->next = q = q->next ;
1122 else
1123 fs->rq[i] = q = q->next ;
1124 fs->rq_elements-- ;
1125 free(old_q, M_DUMMYNET);
1126 continue ;
1127 }
1128 prev = q ;
1129 q = q->next ;
1130 }
1131 if (q && prev != NULL) { /* found and not in front */
1132 prev->next = q->next ;
1133 q->next = fs->rq[i] ;
1134 fs->rq[i] = q ;
1135 }
1136 }
1137 if (q == NULL) { /* no match, need to allocate a new entry */
1138 q = create_queue(fs, i);
1139 if (q != NULL)
1140 q->id = *id ;
1141 }
1142 return q ;
1143 }
1144
1145 static int
1146 red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len)
1147 {
1148 /*
1149 * RED algorithm
1150 *
1151 * RED calculates the average queue size (avg) using a low-pass filter
1152 * with an exponential weighted (w_q) moving average:
1153 * avg <- (1-w_q) * avg + w_q * q_size
1154 * where q_size is the queue length (measured in bytes or * packets).
1155 *
1156 * If q_size == 0, we compute the idle time for the link, and set
1157 * avg = (1 - w_q)^(idle/s)
1158 * where s is the time needed for transmitting a medium-sized packet.
1159 *
1160 * Now, if avg < min_th the packet is enqueued.
1161 * If avg > max_th the packet is dropped. Otherwise, the packet is
1162 * dropped with probability P function of avg.
1163 */
1164
1165 int64_t p_b = 0;
1166
1167 /* Queue in bytes or packets? */
1168 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ?
1169 q->len_bytes : q->len;
1170
1171 DPRINTF(("\ndummynet: %d q: %2u ", (int)curr_time, q_size));
1172
1173 /* Average queue size estimation. */
1174 if (q_size != 0) {
1175 /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */
1176 int diff = SCALE(q_size) - q->avg;
1177 int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);
1178
1179 q->avg += (int)v;
1180 } else {
1181 /*
1182 * Queue is empty, find for how long the queue has been
1183 * empty and use a lookup table for computing
1184 * (1 - * w_q)^(idle_time/s) where s is the time to send a
1185 * (small) packet.
1186 * XXX check wraps...
1187 */
1188 if (q->avg) {
1189 u_int t = ((uint32_t)curr_time - q->q_time) /
1190 fs->lookup_step;
1191
1192 q->avg = (t < fs->lookup_depth) ?
1193 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
1194 }
1195 }
1196 DPRINTF(("dummynet: avg: %u ", SCALE_VAL(q->avg)));
1197
1198 /* Should i drop? */
1199 if (q->avg < fs->min_th) {
1200 q->count = -1;
1201 return (0); /* accept packet */
1202 }
1203 if (q->avg >= fs->max_th) { /* average queue >= max threshold */
1204 if (fs->flags_fs & DN_IS_GENTLE_RED) {
1205 /*
1206 * According to Gentle-RED, if avg is greater than
1207 * max_th the packet is dropped with a probability
1208 * p_b = c_3 * avg - c_4
1209 * where c_3 = (1 - max_p) / max_th
1210 * c_4 = 1 - 2 * max_p
1211 */
1212 p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) -
1213 fs->c_4;
1214 } else {
1215 q->count = -1;
1216 DPRINTF(("dummynet: - drop"));
1217 return (1);
1218 }
1219 } else if (q->avg > fs->min_th) {
1220 /*
1221 * We compute p_b using the linear dropping function
1222 * p_b = c_1 * avg - c_2
1223 * where c_1 = max_p / (max_th - min_th)
1224 * c_2 = max_p * min_th / (max_th - min_th)
1225 */
1226 p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
1227 }
1228
1229 if (fs->flags_fs & DN_QSIZE_IS_BYTES)
1230 p_b = (p_b * len) / fs->max_pkt_size;
1231 if (++q->count == 0)
1232 q->random = random() & 0xffff;
1233 else {
1234 /*
1235 * q->count counts packets arrived since last drop, so a greater
1236 * value of q->count means a greater packet drop probability.
1237 */
1238 if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
1239 q->count = 0;
1240 DPRINTF(("dummynet: - red drop"));
1241 /* After a drop we calculate a new random value. */
1242 q->random = random() & 0xffff;
1243 return (1); /* drop */
1244 }
1245 }
1246 /* End of RED algorithm. */
1247
1248 return (0); /* accept */
1249 }
1250
1251 static __inline struct dn_flow_set *
1252 locate_flowset(int fs_nr)
1253 {
1254 struct dn_flow_set *fs;
1255
1256 SLIST_FOREACH(fs, &flowsethash[HASH(fs_nr)], next)
1257 if (fs->fs_nr == fs_nr)
1258 return (fs);
1259
1260 return (NULL);
1261 }
1262
1263 static __inline struct dn_pipe *
1264 locate_pipe(int pipe_nr)
1265 {
1266 struct dn_pipe *pipe;
1267
1268 SLIST_FOREACH(pipe, &pipehash[HASH(pipe_nr)], next)
1269 if (pipe->pipe_nr == pipe_nr)
1270 return (pipe);
1271
1272 return (NULL);
1273 }
1274
1275 /*
1276 * dummynet hook for packets. Below 'pipe' is a pipe or a queue
1277 * depending on whether WF2Q or fixed bw is used.
1278 *
1279 * pipe_nr pipe or queue the packet is destined for.
1280 * dir where shall we send the packet after dummynet.
1281 * m the mbuf with the packet
1282 * ifp the 'ifp' parameter from the caller.
1283 * NULL in ip_input, destination interface in ip_output,
1284 * real_dst in bdg_forward
1285 * rule matching rule, in case of multiple passes
1286 */
1287 static int
1288 dummynet_io(struct mbuf **m0, int dir, struct ip_fw_args *fwa)
1289 {
1290 struct mbuf *m = *m0, *head = NULL, *tail = NULL;
1291 struct dn_pkt_tag *pkt;
1292 struct m_tag *mtag;
1293 struct dn_flow_set *fs = NULL;
1294 struct dn_pipe *pipe;
1295 uint64_t len = m->m_pkthdr.len;
1296 struct dn_flow_queue *q = NULL;
1297 int is_pipe;
1298 ipfw_insn *cmd = ACTION_PTR(fwa->rule);
1299
1300 KASSERT(m->m_nextpkt == NULL,
1301 ("dummynet_io: mbuf queue passed to dummynet"));
1302
1303 if (cmd->opcode == O_LOG)
1304 cmd += F_LEN(cmd);
1305 if (cmd->opcode == O_ALTQ)
1306 cmd += F_LEN(cmd);
1307 if (cmd->opcode == O_TAG)
1308 cmd += F_LEN(cmd);
1309 is_pipe = (cmd->opcode == O_PIPE);
1310
1311 DUMMYNET_LOCK();
1312 io_pkt++;
1313 /*
1314 * This is a dummynet rule, so we expect an O_PIPE or O_QUEUE rule.
1315 *
1316 * XXXGL: probably the pipe->fs and fs->pipe logic here
1317 * below can be simplified.
1318 */
1319 if (is_pipe) {
1320 pipe = locate_pipe(fwa->cookie);
1321 if (pipe != NULL)
1322 fs = &(pipe->fs);
1323 } else
1324 fs = locate_flowset(fwa->cookie);
1325
1326 if (fs == NULL)
1327 goto dropit; /* This queue/pipe does not exist! */
1328 pipe = fs->pipe;
1329 if (pipe == NULL) { /* Must be a queue, try find a matching pipe. */
1330 pipe = locate_pipe(fs->parent_nr);
1331 if (pipe != NULL)
1332 fs->pipe = pipe;
1333 else {
1334 printf("dummynet: no pipe %d for queue %d, drop pkt\n",
1335 fs->parent_nr, fs->fs_nr);
1336 goto dropit;
1337 }
1338 }
1339 q = find_queue(fs, &(fwa->f_id));
1340 if (q == NULL)
1341 goto dropit; /* Cannot allocate queue. */
1342
1343 /* Update statistics, then check reasons to drop pkt. */
1344 q->tot_bytes += len;
1345 q->tot_pkts++;
1346 if (fs->plr && random() < fs->plr)
1347 goto dropit; /* Random pkt drop. */
1348 if (fs->flags_fs & DN_QSIZE_IS_BYTES) {
1349 if (q->len_bytes > fs->qsize)
1350 goto dropit; /* Queue size overflow. */
1351 } else {
1352 if (q->len >= fs->qsize)
1353 goto dropit; /* Queue count overflow. */
1354 }
1355 if (fs->flags_fs & DN_IS_RED && red_drops(fs, q, len))
1356 goto dropit;
1357
1358 /* XXX expensive to zero, see if we can remove it. */
1359 mtag = m_tag_get(PACKET_TAG_DUMMYNET,
1360 sizeof(struct dn_pkt_tag), M_NOWAIT | M_ZERO);
1361 if (mtag == NULL)
1362 goto dropit; /* Cannot allocate packet header. */
1363 m_tag_prepend(m, mtag); /* Attach to mbuf chain. */
1364
1365 pkt = (struct dn_pkt_tag *)(mtag + 1);
1366 /*
1367 * Ok, i can handle the pkt now...
1368 * Build and enqueue packet + parameters.
1369 */
1370 pkt->rule = fwa->rule;
1371 pkt->dn_dir = dir;
1372
1373 pkt->ifp = fwa->oif;
1374
1375 if (q->head == NULL)
1376 q->head = m;
1377 else
1378 q->tail->m_nextpkt = m;
1379 q->tail = m;
1380 q->len++;
1381 q->len_bytes += len;
1382
1383 if (q->head != m) /* Flow was not idle, we are done. */
1384 goto done;
1385
1386 if (q->q_time < (uint32_t)curr_time)
1387 q->numbytes = io_fast ? fs->pipe->bandwidth : 0;
1388 q->q_time = curr_time;
1389
1390 /*
1391 * If we reach this point the flow was previously idle, so we need
1392 * to schedule it. This involves different actions for fixed-rate or
1393 * WF2Q queues.
1394 */
1395 if (is_pipe) {
1396 /* Fixed-rate queue: just insert into the ready_heap. */
1397 dn_key t = 0;
1398
1399 if (pipe->bandwidth && m->m_pkthdr.len * 8 * hz > q->numbytes)
1400 t = SET_TICKS(m, q, pipe);
1401 q->sched_time = curr_time;
1402 if (t == 0) /* Must process it now. */
1403 ready_event(q, &head, &tail);
1404 else
1405 heap_insert(&ready_heap, curr_time + t , q);
1406 } else {
1407 /*
1408 * WF2Q. First, compute start time S: if the flow was
1409 * idle (S = F + 1) set S to the virtual time V for the
1410 * controlling pipe, and update the sum of weights for the pipe;
1411 * otherwise, remove flow from idle_heap and set S to max(F,V).
1412 * Second, compute finish time F = S + len / weight.
1413 * Third, if pipe was idle, update V = max(S, V).
1414 * Fourth, count one more backlogged flow.
1415 */
1416 if (DN_KEY_GT(q->S, q->F)) { /* Means timestamps are invalid. */
1417 q->S = pipe->V;
1418 pipe->sum += fs->weight; /* Add weight of new queue. */
1419 } else {
1420 heap_extract(&(pipe->idle_heap), q);
1421 q->S = MAX64(q->F, pipe->V);
1422 }
1423 q->F = q->S + (len << MY_M) / (uint64_t)fs->weight;
1424
1425 if (pipe->not_eligible_heap.elements == 0 &&
1426 pipe->scheduler_heap.elements == 0)
1427 pipe->V = MAX64(q->S, pipe->V);
1428 fs->backlogged++;
1429 /*
1430 * Look at eligibility. A flow is not eligibile if S>V (when
1431 * this happens, it means that there is some other flow already
1432 * scheduled for the same pipe, so the scheduler_heap cannot be
1433 * empty). If the flow is not eligible we just store it in the
1434 * not_eligible_heap. Otherwise, we store in the scheduler_heap
1435 * and possibly invoke ready_event_wfq() right now if there is
1436 * leftover credit.
1437 * Note that for all flows in scheduler_heap (SCH), S_i <= V,
1438 * and for all flows in not_eligible_heap (NEH), S_i > V.
1439 * So when we need to compute max(V, min(S_i)) forall i in
1440 * SCH+NEH, we only need to look into NEH.
1441 */
1442 if (DN_KEY_GT(q->S, pipe->V)) { /* Not eligible. */
1443 if (pipe->scheduler_heap.elements == 0)
1444 printf("dummynet: ++ ouch! not eligible but empty scheduler!\n");
1445 heap_insert(&(pipe->not_eligible_heap), q->S, q);
1446 } else {
1447 heap_insert(&(pipe->scheduler_heap), q->F, q);
1448 if (pipe->numbytes >= 0) { /* Pipe is idle. */
1449 if (pipe->scheduler_heap.elements != 1)
1450 printf("dummynet: OUCH! pipe should have been idle!\n");
1451 DPRINTF(("dummynet: waking up pipe %d at %d\n",
1452 pipe->pipe_nr, (int)(q->F >> MY_M)));
1453 pipe->sched_time = curr_time;
1454 ready_event_wfq(pipe, &head, &tail);
1455 }
1456 }
1457 }
1458 done:
1459 if (head == m && dir != DN_TO_IFB_FWD && dir != DN_TO_ETH_DEMUX &&
1460 dir != DN_TO_ETH_OUT) { /* Fast io. */
1461 io_pkt_fast++;
1462 if (m->m_nextpkt != NULL)
1463 printf("dummynet: fast io: pkt chain detected!\n");
1464 head = m->m_nextpkt = NULL;
1465 } else
1466 *m0 = NULL; /* Normal io. */
1467
1468 DUMMYNET_UNLOCK();
1469 if (head != NULL)
1470 dummynet_send(head);
1471 return (0);
1472
1473 dropit:
1474 io_pkt_drop++;
1475 if (q)
1476 q->drops++;
1477 DUMMYNET_UNLOCK();
1478 m_freem(m);
1479 *m0 = NULL;
1480 return ((fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS);
1481 }
1482
1483 /*
1484 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT)
1485 * Doing this would probably save us the initial bzero of dn_pkt
1486 */
1487 #define DN_FREE_PKT(_m) do { \
1488 m_freem(_m); \
1489 } while (0)
1490
1491 /*
1492 * Dispose all packets and flow_queues on a flow_set.
1493 * If all=1, also remove red lookup table and other storage,
1494 * including the descriptor itself.
1495 * For the one in dn_pipe MUST also cleanup ready_heap...
1496 */
1497 static void
1498 purge_flow_set(struct dn_flow_set *fs, int all)
1499 {
1500 struct dn_flow_queue *q, *qn;
1501 int i;
1502
1503 DUMMYNET_LOCK_ASSERT();
1504
1505 for (i = 0; i <= fs->rq_size; i++) {
1506 for (q = fs->rq[i]; q != NULL; q = qn) {
1507 struct mbuf *m, *mnext;
1508
1509 mnext = q->head;
1510 while ((m = mnext) != NULL) {
1511 mnext = m->m_nextpkt;
1512 DN_FREE_PKT(m);
1513 }
1514 qn = q->next;
1515 free(q, M_DUMMYNET);
1516 }
1517 fs->rq[i] = NULL;
1518 }
1519
1520 fs->rq_elements = 0;
1521 if (all) {
1522 /* RED - free lookup table. */
1523 if (fs->w_q_lookup != NULL)
1524 free(fs->w_q_lookup, M_DUMMYNET);
1525 if (fs->rq != NULL)
1526 free(fs->rq, M_DUMMYNET);
1527 /* If this fs is not part of a pipe, free it. */
1528 if (fs->pipe == NULL || fs != &(fs->pipe->fs))
1529 free(fs, M_DUMMYNET);
1530 }
1531 }
1532
1533 /*
1534 * Dispose all packets queued on a pipe (not a flow_set).
1535 * Also free all resources associated to a pipe, which is about
1536 * to be deleted.
1537 */
1538 static void
1539 purge_pipe(struct dn_pipe *pipe)
1540 {
1541 struct mbuf *m, *mnext;
1542
1543 purge_flow_set( &(pipe->fs), 1 );
1544
1545 mnext = pipe->head;
1546 while ((m = mnext) != NULL) {
1547 mnext = m->m_nextpkt;
1548 DN_FREE_PKT(m);
1549 }
1550
1551 heap_free( &(pipe->scheduler_heap) );
1552 heap_free( &(pipe->not_eligible_heap) );
1553 heap_free( &(pipe->idle_heap) );
1554 }
1555
1556 /*
1557 * Delete all pipes and heaps returning memory. Must also
1558 * remove references from all ipfw rules to all pipes.
1559 */
1560 static void
1561 dummynet_flush(void)
1562 {
1563 struct dn_pipe *pipe, *pipe1;
1564 struct dn_flow_set *fs, *fs1;
1565 int i;
1566
1567 DUMMYNET_LOCK();
1568 /* Free heaps so we don't have unwanted events. */
1569 heap_free(&ready_heap);
1570 heap_free(&wfq_ready_heap);
1571 heap_free(&extract_heap);
1572
1573 /*
1574 * Now purge all queued pkts and delete all pipes.
1575 *
1576 * XXXGL: can we merge the for(;;) cycles into one or not?
1577 */
1578 for (i = 0; i < HASHSIZE; i++)
1579 SLIST_FOREACH_SAFE(fs, &flowsethash[i], next, fs1) {
1580 SLIST_REMOVE(&flowsethash[i], fs, dn_flow_set, next);
1581 purge_flow_set(fs, 1);
1582 }
1583 for (i = 0; i < HASHSIZE; i++)
1584 SLIST_FOREACH_SAFE(pipe, &pipehash[i], next, pipe1) {
1585 SLIST_REMOVE(&pipehash[i], pipe, dn_pipe, next);
1586 purge_pipe(pipe);
1587 free(pipe, M_DUMMYNET);
1588 }
1589 DUMMYNET_UNLOCK();
1590 }
1591
1592 extern struct ip_fw *ip_fw_default_rule ;
1593 static void
1594 dn_rule_delete_fs(struct dn_flow_set *fs, void *r)
1595 {
1596 int i ;
1597 struct dn_flow_queue *q ;
1598 struct mbuf *m ;
1599
1600 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */
1601 for (q = fs->rq[i] ; q ; q = q->next )
1602 for (m = q->head ; m ; m = m->m_nextpkt ) {
1603 struct dn_pkt_tag *pkt = dn_tag_get(m) ;
1604 if (pkt->rule == r)
1605 pkt->rule = ip_fw_default_rule ;
1606 }
1607 }
1608 /*
1609 * when a firewall rule is deleted, scan all queues and remove the flow-id
1610 * from packets matching this rule.
1611 */
1612 void
1613 dn_rule_delete(void *r)
1614 {
1615 struct dn_pipe *pipe;
1616 struct dn_flow_set *fs;
1617 struct dn_pkt_tag *pkt;
1618 struct mbuf *m;
1619 int i;
1620
1621 DUMMYNET_LOCK();
1622 /*
1623 * If the rule references a queue (dn_flow_set), then scan
1624 * the flow set, otherwise scan pipes. Should do either, but doing
1625 * both does not harm.
1626 */
1627 for (i = 0; i < HASHSIZE; i++)
1628 SLIST_FOREACH(fs, &flowsethash[i], next)
1629 dn_rule_delete_fs(fs, r);
1630
1631 for (i = 0; i < HASHSIZE; i++)
1632 SLIST_FOREACH(pipe, &pipehash[i], next) {
1633 fs = &(pipe->fs);
1634 dn_rule_delete_fs(fs, r);
1635 for (m = pipe->head ; m ; m = m->m_nextpkt ) {
1636 pkt = dn_tag_get(m);
1637 if (pkt->rule == r)
1638 pkt->rule = ip_fw_default_rule;
1639 }
1640 }
1641 DUMMYNET_UNLOCK();
1642 }
1643
1644 /*
1645 * setup RED parameters
1646 */
1647 static int
1648 config_red(struct dn_flow_set *p, struct dn_flow_set * x)
1649 {
1650 int i;
1651
1652 x->w_q = p->w_q;
1653 x->min_th = SCALE(p->min_th);
1654 x->max_th = SCALE(p->max_th);
1655 x->max_p = p->max_p;
1656
1657 x->c_1 = p->max_p / (p->max_th - p->min_th);
1658 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th));
1659
1660 if (x->flags_fs & DN_IS_GENTLE_RED) {
1661 x->c_3 = (SCALE(1) - p->max_p) / p->max_th;
1662 x->c_4 = SCALE(1) - 2 * p->max_p;
1663 }
1664
1665 /* If the lookup table already exist, free and create it again. */
1666 if (x->w_q_lookup) {
1667 free(x->w_q_lookup, M_DUMMYNET);
1668 x->w_q_lookup = NULL;
1669 }
1670 if (red_lookup_depth == 0) {
1671 printf("\ndummynet: net.inet.ip.dummynet.red_lookup_depth"
1672 "must be > 0\n");
1673 free(x, M_DUMMYNET);
1674 return (EINVAL);
1675 }
1676 x->lookup_depth = red_lookup_depth;
1677 x->w_q_lookup = (u_int *)malloc(x->lookup_depth * sizeof(int),
1678 M_DUMMYNET, M_NOWAIT);
1679 if (x->w_q_lookup == NULL) {
1680 printf("dummynet: sorry, cannot allocate red lookup table\n");
1681 free(x, M_DUMMYNET);
1682 return(ENOSPC);
1683 }
1684
1685 /* Fill the lookup table with (1 - w_q)^x */
1686 x->lookup_step = p->lookup_step;
1687 x->lookup_weight = p->lookup_weight;
1688 x->w_q_lookup[0] = SCALE(1) - x->w_q;
1689
1690 for (i = 1; i < x->lookup_depth; i++)
1691 x->w_q_lookup[i] =
1692 SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight);
1693
1694 if (red_avg_pkt_size < 1)
1695 red_avg_pkt_size = 512;
1696 x->avg_pkt_size = red_avg_pkt_size;
1697 if (red_max_pkt_size < 1)
1698 red_max_pkt_size = 1500;
1699 x->max_pkt_size = red_max_pkt_size;
1700 return (0);
1701 }
1702
1703 static int
1704 alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs)
1705 {
1706 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */
1707 int l = pfs->rq_size;
1708
1709 if (l == 0)
1710 l = dn_hash_size;
1711 if (l < 4)
1712 l = 4;
1713 else if (l > DN_MAX_HASH_SIZE)
1714 l = DN_MAX_HASH_SIZE;
1715 x->rq_size = l;
1716 } else /* one is enough for null mask */
1717 x->rq_size = 1;
1718 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *),
1719 M_DUMMYNET, M_NOWAIT | M_ZERO);
1720 if (x->rq == NULL) {
1721 printf("dummynet: sorry, cannot allocate queue\n");
1722 return (ENOMEM);
1723 }
1724 x->rq_elements = 0;
1725 return 0 ;
1726 }
1727
1728 static void
1729 set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src)
1730 {
1731 x->flags_fs = src->flags_fs;
1732 x->qsize = src->qsize;
1733 x->plr = src->plr;
1734 x->flow_mask = src->flow_mask;
1735 if (x->flags_fs & DN_QSIZE_IS_BYTES) {
1736 if (x->qsize > pipe_byte_limit)
1737 x->qsize = 1024 * 1024;
1738 } else {
1739 if (x->qsize == 0)
1740 x->qsize = 50;
1741 if (x->qsize > pipe_slot_limit)
1742 x->qsize = 50;
1743 }
1744 /* Configuring RED. */
1745 if (x->flags_fs & DN_IS_RED)
1746 config_red(src, x); /* XXX should check errors */
1747 }
1748
1749 /*
1750 * Setup pipe or queue parameters.
1751 */
1752 static int
1753 config_pipe(struct dn_pipe *p)
1754 {
1755 struct dn_flow_set *pfs = &(p->fs);
1756 struct dn_flow_queue *q;
1757 int i, error;
1758
1759 /*
1760 * The config program passes parameters as follows:
1761 * bw = bits/second (0 means no limits),
1762 * delay = ms, must be translated into ticks.
1763 * qsize = slots/bytes
1764 */
1765 p->delay = (p->delay * hz) / 1000;
1766 /* We need either a pipe number or a flow_set number. */
1767 if (p->pipe_nr == 0 && pfs->fs_nr == 0)
1768 return (EINVAL);
1769 if (p->pipe_nr != 0 && pfs->fs_nr != 0)
1770 return (EINVAL);
1771 if (p->pipe_nr != 0) { /* this is a pipe */
1772 struct dn_pipe *pipe;
1773
1774 DUMMYNET_LOCK();
1775 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1776
1777 if (pipe == NULL) { /* new pipe */
1778 pipe = malloc(sizeof(struct dn_pipe), M_DUMMYNET,
1779 M_NOWAIT | M_ZERO);
1780 if (pipe == NULL) {
1781 DUMMYNET_UNLOCK();
1782 printf("dummynet: no memory for new pipe\n");
1783 return (ENOMEM);
1784 }
1785 pipe->pipe_nr = p->pipe_nr;
1786 pipe->fs.pipe = pipe;
1787 /*
1788 * idle_heap is the only one from which
1789 * we extract from the middle.
1790 */
1791 pipe->idle_heap.size = pipe->idle_heap.elements = 0;
1792 pipe->idle_heap.offset =
1793 OFFSET_OF(struct dn_flow_queue, heap_pos);
1794 } else
1795 /* Flush accumulated credit for all queues. */
1796 for (i = 0; i <= pipe->fs.rq_size; i++)
1797 for (q = pipe->fs.rq[i]; q; q = q->next)
1798 q->numbytes = io_fast ? p->bandwidth : 0;
1799
1800 pipe->bandwidth = p->bandwidth;
1801 pipe->numbytes = 0; /* just in case... */
1802 bcopy(p->if_name, pipe->if_name, sizeof(p->if_name));
1803 pipe->ifp = NULL; /* reset interface ptr */
1804 pipe->delay = p->delay;
1805 set_fs_parms(&(pipe->fs), pfs);
1806
1807 if (pipe->fs.rq == NULL) { /* a new pipe */
1808 error = alloc_hash(&(pipe->fs), pfs);
1809 if (error) {
1810 DUMMYNET_UNLOCK();
1811 free(pipe, M_DUMMYNET);
1812 return (error);
1813 }
1814 SLIST_INSERT_HEAD(&pipehash[HASH(pipe->pipe_nr)],
1815 pipe, next);
1816 }
1817 DUMMYNET_UNLOCK();
1818 } else { /* config queue */
1819 struct dn_flow_set *fs;
1820
1821 DUMMYNET_LOCK();
1822 fs = locate_flowset(pfs->fs_nr); /* locate flow_set */
1823
1824 if (fs == NULL) { /* new */
1825 if (pfs->parent_nr == 0) { /* need link to a pipe */
1826 DUMMYNET_UNLOCK();
1827 return (EINVAL);
1828 }
1829 fs = malloc(sizeof(struct dn_flow_set), M_DUMMYNET,
1830 M_NOWAIT | M_ZERO);
1831 if (fs == NULL) {
1832 DUMMYNET_UNLOCK();
1833 printf(
1834 "dummynet: no memory for new flow_set\n");
1835 return (ENOMEM);
1836 }
1837 fs->fs_nr = pfs->fs_nr;
1838 fs->parent_nr = pfs->parent_nr;
1839 fs->weight = pfs->weight;
1840 if (fs->weight == 0)
1841 fs->weight = 1;
1842 else if (fs->weight > 100)
1843 fs->weight = 100;
1844 } else {
1845 /*
1846 * Change parent pipe not allowed;
1847 * must delete and recreate.
1848 */
1849 if (pfs->parent_nr != 0 &&
1850 fs->parent_nr != pfs->parent_nr) {
1851 DUMMYNET_UNLOCK();
1852 return (EINVAL);
1853 }
1854 }
1855
1856 set_fs_parms(fs, pfs);
1857
1858 if (fs->rq == NULL) { /* a new flow_set */
1859 error = alloc_hash(fs, pfs);
1860 if (error) {
1861 DUMMYNET_UNLOCK();
1862 free(fs, M_DUMMYNET);
1863 return (error);
1864 }
1865 SLIST_INSERT_HEAD(&flowsethash[HASH(fs->fs_nr)],
1866 fs, next);
1867 }
1868 DUMMYNET_UNLOCK();
1869 }
1870 return (0);
1871 }
1872
1873 /*
1874 * Helper function to remove from a heap queues which are linked to
1875 * a flow_set about to be deleted.
1876 */
1877 static void
1878 fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs)
1879 {
1880 int i = 0, found = 0 ;
1881 for (; i < h->elements ;)
1882 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) {
1883 h->elements-- ;
1884 h->p[i] = h->p[h->elements] ;
1885 found++ ;
1886 } else
1887 i++ ;
1888 if (found)
1889 heapify(h);
1890 }
1891
1892 /*
1893 * helper function to remove a pipe from a heap (can be there at most once)
1894 */
1895 static void
1896 pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p)
1897 {
1898 if (h->elements > 0) {
1899 int i = 0 ;
1900 for (i=0; i < h->elements ; i++ ) {
1901 if (h->p[i].object == p) { /* found it */
1902 h->elements-- ;
1903 h->p[i] = h->p[h->elements] ;
1904 heapify(h);
1905 break ;
1906 }
1907 }
1908 }
1909 }
1910
1911 /*
1912 * drain all queues. Called in case of severe mbuf shortage.
1913 */
1914 void
1915 dummynet_drain()
1916 {
1917 struct dn_flow_set *fs;
1918 struct dn_pipe *pipe;
1919 struct mbuf *m, *mnext;
1920 int i;
1921
1922 DUMMYNET_LOCK_ASSERT();
1923
1924 heap_free(&ready_heap);
1925 heap_free(&wfq_ready_heap);
1926 heap_free(&extract_heap);
1927 /* remove all references to this pipe from flow_sets */
1928 for (i = 0; i < HASHSIZE; i++)
1929 SLIST_FOREACH(fs, &flowsethash[i], next)
1930 purge_flow_set(fs, 0);
1931
1932 for (i = 0; i < HASHSIZE; i++) {
1933 SLIST_FOREACH(pipe, &pipehash[i], next) {
1934 purge_flow_set(&(pipe->fs), 0);
1935
1936 mnext = pipe->head;
1937 while ((m = mnext) != NULL) {
1938 mnext = m->m_nextpkt;
1939 DN_FREE_PKT(m);
1940 }
1941 pipe->head = pipe->tail = NULL;
1942 }
1943 }
1944 }
1945
1946 /*
1947 * Fully delete a pipe or a queue, cleaning up associated info.
1948 */
1949 static int
1950 delete_pipe(struct dn_pipe *p)
1951 {
1952 if (p->pipe_nr == 0 && p->fs.fs_nr == 0)
1953 return EINVAL ;
1954 if (p->pipe_nr != 0 && p->fs.fs_nr != 0)
1955 return EINVAL ;
1956 if (p->pipe_nr != 0) { /* this is an old-style pipe */
1957 struct dn_pipe *pipe;
1958 struct dn_flow_set *fs;
1959 int i;
1960
1961 DUMMYNET_LOCK();
1962 pipe = locate_pipe(p->pipe_nr); /* locate pipe */
1963
1964 if (pipe == NULL) {
1965 DUMMYNET_UNLOCK();
1966 return (ENOENT); /* not found */
1967 }
1968
1969 /* Unlink from list of pipes. */
1970 SLIST_REMOVE(&pipehash[HASH(pipe->pipe_nr)], pipe, dn_pipe, next);
1971
1972 /* Remove all references to this pipe from flow_sets. */
1973 for (i = 0; i < HASHSIZE; i++)
1974 SLIST_FOREACH(fs, &flowsethash[i], next)
1975 if (fs->pipe == pipe) {
1976 printf("dummynet: ++ ref to pipe %d from fs %d\n",
1977 p->pipe_nr, fs->fs_nr);
1978 fs->pipe = NULL ;
1979 purge_flow_set(fs, 0);
1980 }
1981 fs_remove_from_heap(&ready_heap, &(pipe->fs));
1982 purge_pipe(pipe); /* remove all data associated to this pipe */
1983 /* remove reference to here from extract_heap and wfq_ready_heap */
1984 pipe_remove_from_heap(&extract_heap, pipe);
1985 pipe_remove_from_heap(&wfq_ready_heap, pipe);
1986 DUMMYNET_UNLOCK();
1987
1988 free(pipe, M_DUMMYNET);
1989 } else { /* this is a WF2Q queue (dn_flow_set) */
1990 struct dn_flow_set *fs;
1991
1992 DUMMYNET_LOCK();
1993 fs = locate_flowset(p->fs.fs_nr); /* locate set */
1994
1995 if (fs == NULL) {
1996 DUMMYNET_UNLOCK();
1997 return (ENOENT); /* not found */
1998 }
1999
2000 /* Unlink from list of flowsets. */
2001 SLIST_REMOVE( &flowsethash[HASH(fs->fs_nr)], fs, dn_flow_set, next);
2002
2003 if (fs->pipe != NULL) {
2004 /* Update total weight on parent pipe and cleanup parent heaps. */
2005 fs->pipe->sum -= fs->weight * fs->backlogged ;
2006 fs_remove_from_heap(&(fs->pipe->not_eligible_heap), fs);
2007 fs_remove_from_heap(&(fs->pipe->scheduler_heap), fs);
2008 #if 1 /* XXX should i remove from idle_heap as well ? */
2009 fs_remove_from_heap(&(fs->pipe->idle_heap), fs);
2010 #endif
2011 }
2012 purge_flow_set(fs, 1);
2013 DUMMYNET_UNLOCK();
2014 }
2015 return 0 ;
2016 }
2017
2018 /*
2019 * helper function used to copy data from kernel in DUMMYNET_GET
2020 */
2021 static char *
2022 dn_copy_set(struct dn_flow_set *set, char *bp)
2023 {
2024 int i, copied = 0 ;
2025 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp;
2026
2027 DUMMYNET_LOCK_ASSERT();
2028
2029 for (i = 0 ; i <= set->rq_size ; i++)
2030 for (q = set->rq[i] ; q ; q = q->next, qp++ ) {
2031 if (q->hash_slot != i)
2032 printf("dummynet: ++ at %d: wrong slot (have %d, "
2033 "should be %d)\n", copied, q->hash_slot, i);
2034 if (q->fs != set)
2035 printf("dummynet: ++ at %d: wrong fs ptr (have %p, should be %p)\n",
2036 i, q->fs, set);
2037 copied++ ;
2038 bcopy(q, qp, sizeof( *q ) );
2039 /* cleanup pointers */
2040 qp->next = NULL ;
2041 qp->head = qp->tail = NULL ;
2042 qp->fs = NULL ;
2043 }
2044 if (copied != set->rq_elements)
2045 printf("dummynet: ++ wrong count, have %d should be %d\n",
2046 copied, set->rq_elements);
2047 return (char *)qp ;
2048 }
2049
2050 static size_t
2051 dn_calc_size(void)
2052 {
2053 struct dn_flow_set *fs;
2054 struct dn_pipe *pipe;
2055 size_t size = 0;
2056 int i;
2057
2058 DUMMYNET_LOCK_ASSERT();
2059 /*
2060 * Compute size of data structures: list of pipes and flow_sets.
2061 */
2062 for (i = 0; i < HASHSIZE; i++) {
2063 SLIST_FOREACH(pipe, &pipehash[i], next)
2064 size += sizeof(*pipe) +
2065 pipe->fs.rq_elements * sizeof(struct dn_flow_queue);
2066 SLIST_FOREACH(fs, &flowsethash[i], next)
2067 size += sizeof (*fs) +
2068 fs->rq_elements * sizeof(struct dn_flow_queue);
2069 }
2070 return size;
2071 }
2072
2073 static int
2074 dummynet_get(struct sockopt *sopt)
2075 {
2076 char *buf, *bp ; /* bp is the "copy-pointer" */
2077 size_t size ;
2078 struct dn_flow_set *fs;
2079 struct dn_pipe *pipe;
2080 int error=0, i ;
2081
2082 /* XXX lock held too long */
2083 DUMMYNET_LOCK();
2084 /*
2085 * XXX: Ugly, but we need to allocate memory with M_WAITOK flag and we
2086 * cannot use this flag while holding a mutex.
2087 */
2088 for (i = 0; i < 10; i++) {
2089 size = dn_calc_size();
2090 DUMMYNET_UNLOCK();
2091 buf = malloc(size, M_TEMP, M_WAITOK);
2092 DUMMYNET_LOCK();
2093 if (size == dn_calc_size())
2094 break;
2095 free(buf, M_TEMP);
2096 buf = NULL;
2097 }
2098 if (buf == NULL) {
2099 DUMMYNET_UNLOCK();
2100 return ENOBUFS ;
2101 }
2102 bp = buf;
2103 for (i = 0; i < HASHSIZE; i++)
2104 SLIST_FOREACH(pipe, &pipehash[i], next) {
2105 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp;
2106
2107 /*
2108 * Copy pipe descriptor into *bp, convert delay back to ms,
2109 * then copy the flow_set descriptor(s) one at a time.
2110 * After each flow_set, copy the queue descriptor it owns.
2111 */
2112 bcopy(pipe, bp, sizeof(*pipe));
2113 pipe_bp->delay = (pipe_bp->delay * 1000) / hz;
2114 /*
2115 * XXX the following is a hack based on ->next being the
2116 * first field in dn_pipe and dn_flow_set. The correct
2117 * solution would be to move the dn_flow_set to the beginning
2118 * of struct dn_pipe.
2119 */
2120 pipe_bp->next.sle_next = (struct dn_pipe *)DN_IS_PIPE;
2121 /* Clean pointers. */
2122 pipe_bp->head = pipe_bp->tail = NULL;
2123 pipe_bp->fs.next.sle_next = NULL;
2124 pipe_bp->fs.pipe = NULL;
2125 pipe_bp->fs.rq = NULL;
2126
2127 bp += sizeof(*pipe) ;
2128 bp = dn_copy_set(&(pipe->fs), bp);
2129 }
2130
2131 for (i = 0; i < HASHSIZE; i++)
2132 SLIST_FOREACH(fs, &flowsethash[i], next) {
2133 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp;
2134
2135 bcopy(fs, bp, sizeof(*fs));
2136 /* XXX same hack as above */
2137 fs_bp->next.sle_next = (struct dn_flow_set *)DN_IS_QUEUE;
2138 fs_bp->pipe = NULL;
2139 fs_bp->rq = NULL;
2140 bp += sizeof(*fs);
2141 bp = dn_copy_set(fs, bp);
2142 }
2143
2144 DUMMYNET_UNLOCK();
2145
2146 error = sooptcopyout(sopt, buf, size);
2147 free(buf, M_TEMP);
2148 return error ;
2149 }
2150
2151 /*
2152 * Handler for the various dummynet socket options (get, flush, config, del)
2153 */
2154 static int
2155 ip_dn_ctl(struct sockopt *sopt)
2156 {
2157 int error = 0 ;
2158 struct dn_pipe *p, tmp_pipe;
2159
2160 /* Disallow sets in really-really secure mode. */
2161 if (sopt->sopt_dir == SOPT_SET) {
2162 #if __FreeBSD_version >= 500034
2163 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
2164 if (error)
2165 return (error);
2166 #else
2167 if (securelevel >= 3)
2168 return (EPERM);
2169 #endif
2170 }
2171
2172 switch (sopt->sopt_name) {
2173 default :
2174 printf("dummynet: -- unknown option %d", sopt->sopt_name);
2175 return EINVAL ;
2176
2177 case IP_DUMMYNET_GET :
2178 error = dummynet_get(sopt);
2179 break ;
2180
2181 case IP_DUMMYNET_FLUSH :
2182 dummynet_flush() ;
2183 break ;
2184
2185 case IP_DUMMYNET_CONFIGURE :
2186 p = &tmp_pipe ;
2187 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2188 if (error)
2189 break ;
2190 error = config_pipe(p);
2191 break ;
2192
2193 case IP_DUMMYNET_DEL : /* remove a pipe or queue */
2194 p = &tmp_pipe ;
2195 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p);
2196 if (error)
2197 break ;
2198
2199 error = delete_pipe(p);
2200 break ;
2201 }
2202 return error ;
2203 }
2204
2205 static void
2206 ip_dn_init(void)
2207 {
2208 int i;
2209
2210 if (bootverbose)
2211 printf("DUMMYNET with IPv6 initialized (040826)\n");
2212
2213 DUMMYNET_LOCK_INIT();
2214
2215 for (i = 0; i < HASHSIZE; i++) {
2216 SLIST_INIT(&pipehash[i]);
2217 SLIST_INIT(&flowsethash[i]);
2218 }
2219 ready_heap.size = ready_heap.elements = 0;
2220 ready_heap.offset = 0;
2221
2222 wfq_ready_heap.size = wfq_ready_heap.elements = 0;
2223 wfq_ready_heap.offset = 0;
2224
2225 extract_heap.size = extract_heap.elements = 0;
2226 extract_heap.offset = 0;
2227
2228 ip_dn_ctl_ptr = ip_dn_ctl;
2229 ip_dn_io_ptr = dummynet_io;
2230 ip_dn_ruledel_ptr = dn_rule_delete;
2231
2232 TASK_INIT(&dn_task, 0, dummynet_task, NULL);
2233 dn_tq = taskqueue_create_fast("dummynet", M_NOWAIT,
2234 taskqueue_thread_enqueue, &dn_tq);
2235 taskqueue_start_threads(&dn_tq, 1, PI_NET, "dummynet");
2236
2237 callout_init(&dn_timeout, NET_CALLOUT_MPSAFE);
2238 callout_reset(&dn_timeout, 1, dummynet, NULL);
2239
2240 /* Initialize curr_time adjustment mechanics. */
2241 getmicrouptime(&prev_t);
2242 }
2243
2244 #ifdef KLD_MODULE
2245 static void
2246 ip_dn_destroy(void)
2247 {
2248 ip_dn_ctl_ptr = NULL;
2249 ip_dn_io_ptr = NULL;
2250 ip_dn_ruledel_ptr = NULL;
2251
2252 DUMMYNET_LOCK();
2253 callout_stop(&dn_timeout);
2254 DUMMYNET_UNLOCK();
2255 taskqueue_drain(dn_tq, &dn_task);
2256 taskqueue_free(dn_tq);
2257
2258 dummynet_flush();
2259
2260 DUMMYNET_LOCK_DESTROY();
2261 }
2262 #endif /* KLD_MODULE */
2263
2264 static int
2265 dummynet_modevent(module_t mod, int type, void *data)
2266 {
2267 switch (type) {
2268 case MOD_LOAD:
2269 if (DUMMYNET_LOADED) {
2270 printf("DUMMYNET already loaded\n");
2271 return EEXIST ;
2272 }
2273 ip_dn_init();
2274 break;
2275
2276 case MOD_UNLOAD:
2277 #if !defined(KLD_MODULE)
2278 printf("dummynet statically compiled, cannot unload\n");
2279 return EINVAL ;
2280 #else
2281 ip_dn_destroy();
2282 #endif
2283 break ;
2284 default:
2285 return EOPNOTSUPP;
2286 break ;
2287 }
2288 return 0 ;
2289 }
2290
2291 static moduledata_t dummynet_mod = {
2292 "dummynet",
2293 dummynet_modevent,
2294 NULL
2295 };
2296 DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY);
2297 MODULE_DEPEND(dummynet, ipfw, 2, 2, 2);
2298 MODULE_VERSION(dummynet, 1);
Cache object: 1240f9eeffb22b3eb4f019c327b9426c
|