2 * Copyright (c) 1998-2002,2010 Luigi Rizzo, Universita` di Pisa
3 * Portions Copyright (c) 2000 Akamba Corp.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
32 * Configuration and internal object management for dummynet.
35 #include "opt_inet6.h"
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/malloc.h>
41 #include <sys/kernel.h>
43 #include <sys/module.h>
46 #include <sys/rwlock.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
50 #include <sys/taskqueue.h>
51 #include <net/if.h> /* IFNAMSIZ, struct ifaddr, ifq head, lock.h mutex.h */
52 #include <netinet/in.h>
53 #include <netinet/ip_var.h> /* ip_output(), IP_FORWARDING */
54 #include <netinet/ip_fw.h>
55 #include <netinet/ip_dummynet.h>
57 #include <netpfil/ipfw/ip_fw_private.h>
58 #include <netpfil/ipfw/dn_heap.h>
59 #include <netpfil/ipfw/ip_dn_private.h>
60 #include <netpfil/ipfw/dn_sched.h>
62 /* which objects to copy */
63 #define DN_C_LINK 0x01
65 #define DN_C_FLOW 0x04
67 #define DN_C_QUEUE 0x10
69 /* we use this argument in case of a schk_new */
75 /*---- callout hooks. ----*/
76 static struct callout dn_timeout;
78 static struct task dn_task;
79 static struct taskqueue *dn_tq = NULL;
85 (void)arg; /* UNUSED */
86 taskqueue_enqueue_fast(dn_tq, &dn_task);
95 callout_reset_sbt(&dn_timeout, tick_sbt, 0, dummynet, NULL,
96 C_HARDCLOCK | C_DIRECT_EXEC);
98 /*----- end of callout hooks -----*/
100 /* Return a scheduler descriptor given the type or name. */
101 static struct dn_alg *
102 find_sched_type(int type, char *name)
106 SLIST_FOREACH(d, &dn_cfg.schedlist, next) {
107 if (d->type == type || (name && !strcasecmp(d->name, name)))
110 return NULL; /* not found */
114 ipdn_bound_var(int *v, int dflt, int lo, int hi, const char *msg)
117 const char *op = NULL;
125 } else if (oldv > hi) {
131 printf("%s %s to %d (was %d)\n", op, msg, *v, oldv);
135 /*---- flow_id mask, hash and compare functions ---*/
137 * The flow_id includes the 5-tuple, the queue/pipe number
138 * which we store in the extra area in host order,
139 * and for ipv6 also the flow_id6.
140 * XXX see if we want the tos byte (can store in 'flags')
142 static struct ipfw_flow_id *
143 flow_id_mask(struct ipfw_flow_id *mask, struct ipfw_flow_id *id)
145 int is_v6 = IS_IP6_FLOW_ID(id);
147 id->dst_port &= mask->dst_port;
148 id->src_port &= mask->src_port;
149 id->proto &= mask->proto;
150 id->extra &= mask->extra;
152 APPLY_MASK(&id->dst_ip6, &mask->dst_ip6);
153 APPLY_MASK(&id->src_ip6, &mask->src_ip6);
154 id->flow_id6 &= mask->flow_id6;
156 id->dst_ip &= mask->dst_ip;
157 id->src_ip &= mask->src_ip;
162 /* computes an OR of two masks, result in dst and also returned */
163 static struct ipfw_flow_id *
164 flow_id_or(struct ipfw_flow_id *src, struct ipfw_flow_id *dst)
166 int is_v6 = IS_IP6_FLOW_ID(dst);
168 dst->dst_port |= src->dst_port;
169 dst->src_port |= src->src_port;
170 dst->proto |= src->proto;
171 dst->extra |= src->extra;
173 #define OR_MASK(_d, _s) \
174 (_d)->__u6_addr.__u6_addr32[0] |= (_s)->__u6_addr.__u6_addr32[0]; \
175 (_d)->__u6_addr.__u6_addr32[1] |= (_s)->__u6_addr.__u6_addr32[1]; \
176 (_d)->__u6_addr.__u6_addr32[2] |= (_s)->__u6_addr.__u6_addr32[2]; \
177 (_d)->__u6_addr.__u6_addr32[3] |= (_s)->__u6_addr.__u6_addr32[3];
178 OR_MASK(&dst->dst_ip6, &src->dst_ip6);
179 OR_MASK(&dst->src_ip6, &src->src_ip6);
181 dst->flow_id6 |= src->flow_id6;
183 dst->dst_ip |= src->dst_ip;
184 dst->src_ip |= src->src_ip;
190 nonzero_mask(struct ipfw_flow_id *m)
192 if (m->dst_port || m->src_port || m->proto || m->extra)
194 if (IS_IP6_FLOW_ID(m)) {
196 m->dst_ip6.__u6_addr.__u6_addr32[0] ||
197 m->dst_ip6.__u6_addr.__u6_addr32[1] ||
198 m->dst_ip6.__u6_addr.__u6_addr32[2] ||
199 m->dst_ip6.__u6_addr.__u6_addr32[3] ||
200 m->src_ip6.__u6_addr.__u6_addr32[0] ||
201 m->src_ip6.__u6_addr.__u6_addr32[1] ||
202 m->src_ip6.__u6_addr.__u6_addr32[2] ||
203 m->src_ip6.__u6_addr.__u6_addr32[3] ||
206 return m->dst_ip || m->src_ip;
210 /* XXX we may want a better hash function */
212 flow_id_hash(struct ipfw_flow_id *id)
216 if (IS_IP6_FLOW_ID(id)) {
217 uint32_t *d = (uint32_t *)&id->dst_ip6;
218 uint32_t *s = (uint32_t *)&id->src_ip6;
219 i = (d[0] ) ^ (d[1]) ^
221 (d[0] >> 15) ^ (d[1] >> 15) ^
222 (d[2] >> 15) ^ (d[3] >> 15) ^
223 (s[0] << 1) ^ (s[1] << 1) ^
224 (s[2] << 1) ^ (s[3] << 1) ^
225 (s[0] << 16) ^ (s[1] << 16) ^
226 (s[2] << 16) ^ (s[3] << 16) ^
227 (id->dst_port << 1) ^ (id->src_port) ^
229 (id->proto ) ^ (id->flow_id6);
231 i = (id->dst_ip) ^ (id->dst_ip >> 15) ^
232 (id->src_ip << 1) ^ (id->src_ip >> 16) ^
234 (id->dst_port << 1) ^ (id->src_port) ^ (id->proto);
239 /* Like bcmp, returns 0 if ids match, 1 otherwise. */
241 flow_id_cmp(struct ipfw_flow_id *id1, struct ipfw_flow_id *id2)
243 int is_v6 = IS_IP6_FLOW_ID(id1);
246 if (IS_IP6_FLOW_ID(id2))
247 return 1; /* different address families */
249 return (id1->dst_ip == id2->dst_ip &&
250 id1->src_ip == id2->src_ip &&
251 id1->dst_port == id2->dst_port &&
252 id1->src_port == id2->src_port &&
253 id1->proto == id2->proto &&
254 id1->extra == id2->extra) ? 0 : 1;
258 !bcmp(&id1->dst_ip6,&id2->dst_ip6, sizeof(id1->dst_ip6)) &&
259 !bcmp(&id1->src_ip6,&id2->src_ip6, sizeof(id1->src_ip6)) &&
260 id1->dst_port == id2->dst_port &&
261 id1->src_port == id2->src_port &&
262 id1->proto == id2->proto &&
263 id1->extra == id2->extra &&
264 id1->flow_id6 == id2->flow_id6) ? 0 : 1;
266 /*--------- end of flow-id mask, hash and compare ---------*/
268 /*--- support functions for the qht hashtable ----
269 * Entries are hashed by flow-id
272 q_hash(uintptr_t key, int flags, void *arg)
274 /* compute the hash slot from the flow id */
275 struct ipfw_flow_id *id = (flags & DNHT_KEY_IS_OBJ) ?
276 &((struct dn_queue *)key)->ni.fid :
277 (struct ipfw_flow_id *)key;
279 return flow_id_hash(id);
283 q_match(void *obj, uintptr_t key, int flags, void *arg)
285 struct dn_queue *o = (struct dn_queue *)obj;
286 struct ipfw_flow_id *id2;
288 if (flags & DNHT_KEY_IS_OBJ) {
289 /* compare pointers */
290 id2 = &((struct dn_queue *)key)->ni.fid;
292 id2 = (struct ipfw_flow_id *)key;
294 return (0 == flow_id_cmp(&o->ni.fid, id2));
298 * create a new queue instance for the given 'key'.
301 q_new(uintptr_t key, int flags, void *arg)
303 struct dn_queue *q, *template = arg;
304 struct dn_fsk *fs = template->fs;
305 int size = sizeof(*q) + fs->sched->fp->q_datalen;
307 q = malloc(size, M_DUMMYNET, M_NOWAIT | M_ZERO);
309 D("no memory for new queue");
313 set_oid(&q->ni.oid, DN_QUEUE, size);
314 if (fs->fs.flags & DN_QHT_HASH)
315 q->ni.fid = *(struct ipfw_flow_id *)key;
317 q->_si = template->_si;
320 if (fs->sched->fp->new_queue)
321 fs->sched->fp->new_queue(q);
322 dn_cfg.queue_count++;
327 * Notify schedulers that a queue is going away.
328 * If (flags & DN_DESTROY), also free the packets.
329 * The version for callbacks is called q_delete_cb().
332 dn_delete_queue(struct dn_queue *q, int flags)
334 struct dn_fsk *fs = q->fs;
336 // D("fs %p si %p\n", fs, q->_si);
337 /* notify the parent scheduler that the queue is going away */
338 if (fs && fs->sched->fp->free_queue)
339 fs->sched->fp->free_queue(q);
342 if (flags & DN_DESTROY) {
344 dn_free_pkts(q->mq.head);
345 bzero(q, sizeof(*q)); // safety
347 dn_cfg.queue_count--;
352 q_delete_cb(void *q, void *arg)
354 int flags = (int)(uintptr_t)arg;
355 dn_delete_queue(q, flags);
356 return (flags & DN_DESTROY) ? DNHT_SCAN_DEL : 0;
360 * calls dn_delete_queue/q_delete_cb on all queues,
361 * which notifies the parent scheduler and possibly drains packets.
362 * flags & DN_DESTROY: drains queues and destroy qht;
365 qht_delete(struct dn_fsk *fs, int flags)
367 ND("fs %d start flags %d qht %p",
368 fs->fs.fs_nr, flags, fs->qht);
371 if (fs->fs.flags & DN_QHT_HASH) {
372 dn_ht_scan(fs->qht, q_delete_cb, (void *)(uintptr_t)flags);
373 if (flags & DN_DESTROY) {
374 dn_ht_free(fs->qht, 0);
378 dn_delete_queue((struct dn_queue *)(fs->qht), flags);
379 if (flags & DN_DESTROY)
385 * Find and possibly create the queue for a MULTIQUEUE scheduler.
386 * We never call it for !MULTIQUEUE (the queue is in the sch_inst).
389 ipdn_q_find(struct dn_fsk *fs, struct dn_sch_inst *si,
390 struct ipfw_flow_id *id)
392 struct dn_queue template;
397 if (fs->fs.flags & DN_QHT_HASH) {
398 struct ipfw_flow_id masked_id;
399 if (fs->qht == NULL) {
400 fs->qht = dn_ht_init(NULL, fs->fs.buckets,
401 offsetof(struct dn_queue, q_next),
402 q_hash, q_match, q_new);
407 flow_id_mask(&fs->fsk_mask, &masked_id);
408 return dn_ht_find(fs->qht, (uintptr_t)&masked_id,
409 DNHT_INSERT, &template);
412 fs->qht = q_new(0, 0, &template);
413 return (struct dn_queue *)fs->qht;
416 /*--- end of queue hash table ---*/
418 /*--- support functions for the sch_inst hashtable ----
420 * These are hashed by flow-id
423 si_hash(uintptr_t key, int flags, void *arg)
425 /* compute the hash slot from the flow id */
426 struct ipfw_flow_id *id = (flags & DNHT_KEY_IS_OBJ) ?
427 &((struct dn_sch_inst *)key)->ni.fid :
428 (struct ipfw_flow_id *)key;
430 return flow_id_hash(id);
434 si_match(void *obj, uintptr_t key, int flags, void *arg)
436 struct dn_sch_inst *o = obj;
437 struct ipfw_flow_id *id2;
439 id2 = (flags & DNHT_KEY_IS_OBJ) ?
440 &((struct dn_sch_inst *)key)->ni.fid :
441 (struct ipfw_flow_id *)key;
442 return flow_id_cmp(&o->ni.fid, id2) == 0;
446 * create a new instance for the given 'key'
447 * Allocate memory for instance, delay line and scheduler private data.
450 si_new(uintptr_t key, int flags, void *arg)
452 struct dn_schk *s = arg;
453 struct dn_sch_inst *si;
454 int l = sizeof(*si) + s->fp->si_datalen;
456 si = malloc(l, M_DUMMYNET, M_NOWAIT | M_ZERO);
460 /* Set length only for the part passed up to userland. */
461 set_oid(&si->ni.oid, DN_SCH_I, sizeof(struct dn_flow));
462 set_oid(&(si->dline.oid), DN_DELAY_LINE,
463 sizeof(struct delay_line));
464 /* mark si and dline as outside the event queue */
465 si->ni.oid.id = si->dline.oid.id = -1;
470 if (s->fp->new_sched && s->fp->new_sched(si)) {
471 D("new_sched error");
474 if (s->sch.flags & DN_HAVE_MASK)
475 si->ni.fid = *(struct ipfw_flow_id *)key;
482 bzero(si, sizeof(*si)); // safety
483 free(si, M_DUMMYNET);
489 * Callback from siht to delete all scheduler instances. Remove
490 * si and delay line from the system heap, destroy all queues.
491 * We assume that all flowset have been notified and do not
492 * point to us anymore.
495 si_destroy(void *_si, void *arg)
497 struct dn_sch_inst *si = _si;
498 struct dn_schk *s = si->sched;
499 struct delay_line *dl = &si->dline;
501 if (dl->oid.subtype) /* remove delay line from event heap */
502 heap_extract(&dn_cfg.evheap, dl);
503 dn_free_pkts(dl->mq.head); /* drain delay line */
504 if (si->kflags & DN_ACTIVE) /* remove si from event heap */
505 heap_extract(&dn_cfg.evheap, si);
506 if (s->fp->free_sched)
507 s->fp->free_sched(si);
508 bzero(si, sizeof(*si)); /* safety */
509 free(si, M_DUMMYNET);
511 return DNHT_SCAN_DEL;
515 * Find the scheduler instance for this packet. If we need to apply
516 * a mask, do on a local copy of the flow_id to preserve the original.
517 * Assume siht is always initialized if we have a mask.
520 ipdn_si_find(struct dn_schk *s, struct ipfw_flow_id *id)
523 if (s->sch.flags & DN_HAVE_MASK) {
524 struct ipfw_flow_id id_t = *id;
525 flow_id_mask(&s->sch.sched_mask, &id_t);
526 return dn_ht_find(s->siht, (uintptr_t)&id_t,
530 s->siht = si_new(0, 0, s);
531 return (struct dn_sch_inst *)s->siht;
534 /* callback to flush credit for the scheduler instance */
536 si_reset_credit(void *_si, void *arg)
538 struct dn_sch_inst *si = _si;
539 struct dn_link *p = &si->sched->link;
541 si->credit = p->burst + (dn_cfg.io_fast ? p->bandwidth : 0);
546 schk_reset_credit(struct dn_schk *s)
548 if (s->sch.flags & DN_HAVE_MASK)
549 dn_ht_scan(s->siht, si_reset_credit, NULL);
551 si_reset_credit(s->siht, NULL);
553 /*---- end of sch_inst hashtable ---------------------*/
555 /*-------------------------------------------------------
556 * flowset hash (fshash) support. Entries are hashed by fs_nr.
557 * New allocations are put in the fsunlinked list, from which
558 * they are removed when they point to a specific scheduler.
561 fsk_hash(uintptr_t key, int flags, void *arg)
563 uint32_t i = !(flags & DNHT_KEY_IS_OBJ) ? key :
564 ((struct dn_fsk *)key)->fs.fs_nr;
566 return ( (i>>8)^(i>>4)^i );
570 fsk_match(void *obj, uintptr_t key, int flags, void *arg)
572 struct dn_fsk *fs = obj;
573 int i = !(flags & DNHT_KEY_IS_OBJ) ? key :
574 ((struct dn_fsk *)key)->fs.fs_nr;
576 return (fs->fs.fs_nr == i);
580 fsk_new(uintptr_t key, int flags, void *arg)
584 fs = malloc(sizeof(*fs), M_DUMMYNET, M_NOWAIT | M_ZERO);
586 set_oid(&fs->fs.oid, DN_FS, sizeof(fs->fs));
588 fs->drain_bucket = 0;
589 SLIST_INSERT_HEAD(&dn_cfg.fsu, fs, sch_chain);
595 * detach flowset from its current scheduler. Flags as follows:
596 * DN_DETACH removes from the fsk_list
597 * DN_DESTROY deletes individual queues
598 * DN_DELETE_FS destroys the flowset (otherwise goes in unlinked).
601 fsk_detach(struct dn_fsk *fs, int flags)
603 if (flags & DN_DELETE_FS)
605 ND("fs %d from sched %d flags %s %s %s",
606 fs->fs.fs_nr, fs->fs.sched_nr,
607 (flags & DN_DELETE_FS) ? "DEL_FS":"",
608 (flags & DN_DESTROY) ? "DEL":"",
609 (flags & DN_DETACH) ? "DET":"");
610 if (flags & DN_DETACH) { /* detach from the list */
611 struct dn_fsk_head *h;
612 h = fs->sched ? &fs->sched->fsk_list : &dn_cfg.fsu;
613 SLIST_REMOVE(h, fs, dn_fsk, sch_chain);
615 /* Free the RED parameters, they will be recomputed on
616 * subsequent attach if needed.
619 free(fs->w_q_lookup, M_DUMMYNET);
620 fs->w_q_lookup = NULL;
621 qht_delete(fs, flags);
622 if (fs->sched && fs->sched->fp->free_fsk)
623 fs->sched->fp->free_fsk(fs);
625 if (flags & DN_DELETE_FS) {
626 bzero(fs, sizeof(*fs)); /* safety */
627 free(fs, M_DUMMYNET);
630 SLIST_INSERT_HEAD(&dn_cfg.fsu, fs, sch_chain);
635 * Detach or destroy all flowsets in a list.
636 * flags specifies what to do:
637 * DN_DESTROY: flush all queues
638 * DN_DELETE_FS: DN_DESTROY + destroy flowset
639 * DN_DELETE_FS implies DN_DESTROY
642 fsk_detach_list(struct dn_fsk_head *h, int flags)
645 int n = 0; /* only for stats */
647 ND("head %p flags %x", h, flags);
648 while ((fs = SLIST_FIRST(h))) {
649 SLIST_REMOVE_HEAD(h, sch_chain);
651 fsk_detach(fs, flags);
653 ND("done %d flowsets", n);
657 * called on 'queue X delete' -- removes the flowset from fshash,
658 * deletes all queues for the flowset, and removes the flowset.
661 delete_fs(int i, int locked)
668 fs = dn_ht_find(dn_cfg.fshash, i, DNHT_REMOVE, NULL);
669 ND("fs %d found %p", i, fs);
671 fsk_detach(fs, DN_DETACH | DN_DELETE_FS);
680 /*----- end of flowset hashtable support -------------*/
682 /*------------------------------------------------------------
683 * Scheduler hash. When searching by index we pass sched_nr,
684 * otherwise we pass struct dn_sch * which is the first field in
685 * struct dn_schk so we can cast between the two. We use this trick
686 * because in the create phase (but it should be fixed).
689 schk_hash(uintptr_t key, int flags, void *_arg)
691 uint32_t i = !(flags & DNHT_KEY_IS_OBJ) ? key :
692 ((struct dn_schk *)key)->sch.sched_nr;
693 return ( (i>>8)^(i>>4)^i );
697 schk_match(void *obj, uintptr_t key, int flags, void *_arg)
699 struct dn_schk *s = (struct dn_schk *)obj;
700 int i = !(flags & DNHT_KEY_IS_OBJ) ? key :
701 ((struct dn_schk *)key)->sch.sched_nr;
702 return (s->sch.sched_nr == i);
706 * Create the entry and intialize with the sched hash if needed.
707 * Leave s->fp unset so we can tell whether a dn_ht_find() returns
708 * a new object or a previously existing one.
711 schk_new(uintptr_t key, int flags, void *arg)
713 struct schk_new_arg *a = arg;
715 int l = sizeof(*s) +a->fp->schk_datalen;
717 s = malloc(l, M_DUMMYNET, M_NOWAIT | M_ZERO);
720 set_oid(&s->link.oid, DN_LINK, sizeof(s->link));
721 s->sch = *a->sch; // copy initial values
722 s->link.link_nr = s->sch.sched_nr;
723 SLIST_INIT(&s->fsk_list);
724 /* initialize the hash table or create the single instance */
725 s->fp = a->fp; /* si_new needs this */
727 if (s->sch.flags & DN_HAVE_MASK) {
728 s->siht = dn_ht_init(NULL, s->sch.buckets,
729 offsetof(struct dn_sch_inst, si_next),
730 si_hash, si_match, si_new);
731 if (s->siht == NULL) {
736 s->fp = NULL; /* mark as a new scheduler */
742 * Callback for sched delete. Notify all attached flowsets to
743 * detach from the scheduler, destroy the internal flowset, and
744 * all instances. The scheduler goes away too.
745 * arg is 0 (only detach flowsets and destroy instances)
746 * DN_DESTROY (detach & delete queues, delete schk)
747 * or DN_DELETE_FS (delete queues and flowsets, delete schk)
750 schk_delete_cb(void *obj, void *arg)
752 struct dn_schk *s = obj;
755 ND("sched %d arg %s%s",
757 a&DN_DESTROY ? "DEL ":"",
758 a&DN_DELETE_FS ? "DEL_FS":"");
760 fsk_detach_list(&s->fsk_list, arg ? DN_DESTROY : 0);
761 /* no more flowset pointing to us now */
762 if (s->sch.flags & DN_HAVE_MASK) {
763 dn_ht_scan(s->siht, si_destroy, NULL);
764 dn_ht_free(s->siht, 0);
766 si_destroy(s->siht, NULL);
768 free(s->profile, M_DUMMYNET);
774 bzero(s, sizeof(*s)); // safety
775 free(obj, M_DUMMYNET);
777 return DNHT_SCAN_DEL;
781 * called on a 'sched X delete' command. Deletes a single scheduler.
782 * This is done by removing from the schedhash, unlinking all
783 * flowsets and deleting their traffic.
790 s = dn_ht_find(dn_cfg.schedhash, i, DNHT_REMOVE, NULL);
794 delete_fs(i + DN_MAX_ID, 1); /* first delete internal fs */
795 /* then detach flowsets, delete traffic */
796 schk_delete_cb(s, (void*)(uintptr_t)DN_DESTROY);
799 /*--- end of schk hashtable support ---*/
802 copy_obj(char **start, char *end, void *_o, const char *msg, int i)
804 struct dn_id *o = _o;
805 int have = end - *start;
807 if (have < o->len || o->len == 0 || o->type == 0) {
808 D("(WARN) type %d %s %d have %d need %d",
809 o->type, msg, i, have, o->len);
812 ND("type %d %s %d len %d", o->type, msg, i, o->len);
813 bcopy(_o, *start, o->len);
814 if (o->type == DN_LINK) {
815 /* Adjust burst parameter for link */
816 struct dn_link *l = (struct dn_link *)*start;
817 l->burst = div64(l->burst, 8 * hz);
818 l->delay = l->delay * 1000 / hz;
819 } else if (o->type == DN_SCH) {
820 /* Set id->id to the number of instances */
821 struct dn_schk *s = _o;
822 struct dn_id *id = (struct dn_id *)(*start);
823 id->id = (s->sch.flags & DN_HAVE_MASK) ?
824 dn_ht_entries(s->siht) : (s->siht ? 1 : 0);
830 /* Specific function to copy a queue.
831 * Copies only the user-visible part of a queue (which is in
832 * a struct dn_flow), and sets len accordingly.
835 copy_obj_q(char **start, char *end, void *_o, const char *msg, int i)
837 struct dn_id *o = _o;
838 int have = end - *start;
839 int len = sizeof(struct dn_flow); /* see above comment */
841 if (have < len || o->len == 0 || o->type != DN_QUEUE) {
842 D("ERROR type %d %s %d have %d need %d",
843 o->type, msg, i, have, len);
846 ND("type %d %s %d len %d", o->type, msg, i, len);
847 bcopy(_o, *start, len);
848 ((struct dn_id*)(*start))->len = len;
854 copy_q_cb(void *obj, void *arg)
856 struct dn_queue *q = obj;
857 struct copy_args *a = arg;
858 struct dn_flow *ni = (struct dn_flow *)(*a->start);
859 if (copy_obj_q(a->start, a->end, &q->ni, "queue", -1))
860 return DNHT_SCAN_END;
861 ni->oid.type = DN_FLOW; /* override the DN_QUEUE */
862 ni->oid.id = si_hash((uintptr_t)&ni->fid, 0, NULL);
867 copy_q(struct copy_args *a, struct dn_fsk *fs, int flags)
871 if (fs->fs.flags & DN_QHT_HASH)
872 dn_ht_scan(fs->qht, copy_q_cb, a);
874 copy_q_cb(fs->qht, a);
879 * This routine only copies the initial part of a profile ? XXX
882 copy_profile(struct copy_args *a, struct dn_profile *p)
884 int have = a->end - *a->start;
885 /* XXX here we check for max length */
886 int profile_len = sizeof(struct dn_profile) -
887 ED_MAX_SAMPLES_NO*sizeof(int);
891 if (have < profile_len) {
892 D("error have %d need %d", have, profile_len);
895 bcopy(p, *a->start, profile_len);
896 ((struct dn_id *)(*a->start))->len = profile_len;
897 *a->start += profile_len;
902 copy_flowset(struct copy_args *a, struct dn_fsk *fs, int flags)
904 struct dn_fs *ufs = (struct dn_fs *)(*a->start);
907 ND("flowset %d", fs->fs.fs_nr);
908 if (copy_obj(a->start, a->end, &fs->fs, "flowset", fs->fs.fs_nr))
909 return DNHT_SCAN_END;
910 ufs->oid.id = (fs->fs.flags & DN_QHT_HASH) ?
911 dn_ht_entries(fs->qht) : (fs->qht ? 1 : 0);
912 if (flags) { /* copy queues */
919 copy_si_cb(void *obj, void *arg)
921 struct dn_sch_inst *si = obj;
922 struct copy_args *a = arg;
923 struct dn_flow *ni = (struct dn_flow *)(*a->start);
924 if (copy_obj(a->start, a->end, &si->ni, "inst",
925 si->sched->sch.sched_nr))
926 return DNHT_SCAN_END;
927 ni->oid.type = DN_FLOW; /* override the DN_SCH_I */
928 ni->oid.id = si_hash((uintptr_t)si, DNHT_KEY_IS_OBJ, NULL);
933 copy_si(struct copy_args *a, struct dn_schk *s, int flags)
935 if (s->sch.flags & DN_HAVE_MASK)
936 dn_ht_scan(s->siht, copy_si_cb, a);
938 copy_si_cb(s->siht, a);
943 * compute a list of children of a scheduler and copy up
946 copy_fsk_list(struct copy_args *a, struct dn_schk *s, int flags)
952 int n = 0, space = sizeof(*o);
953 SLIST_FOREACH(fs, &s->fsk_list, sch_chain) {
954 if (fs->fs.fs_nr < DN_MAX_ID)
957 space += n * sizeof(uint32_t);
958 DX(3, "sched %d has %d flowsets", s->sch.sched_nr, n);
959 if (a->end - *(a->start) < space)
960 return DNHT_SCAN_END;
961 o = (struct dn_id *)(*(a->start));
965 p = (uint32_t *)(o+1);
966 SLIST_FOREACH(fs, &s->fsk_list, sch_chain)
967 if (fs->fs.fs_nr < DN_MAX_ID)
973 copy_data_helper(void *_o, void *_arg)
975 struct copy_args *a = _arg;
976 uint32_t *r = a->extra->r; /* start of first range */
977 uint32_t *lim; /* first invalid pointer */
980 lim = (uint32_t *)((char *)(a->extra) + a->extra->o.len);
982 if (a->type == DN_LINK || a->type == DN_SCH) {
983 /* pipe|sched show, we receive a dn_schk */
984 struct dn_schk *s = _o;
987 if (a->type == DN_SCH && n >= DN_MAX_ID)
988 return 0; /* not a scheduler */
989 if (a->type == DN_LINK && n <= DN_MAX_ID)
990 return 0; /* not a pipe */
992 /* see if the object is within one of our ranges */
993 for (;r < lim; r += 2) {
994 if (n < r[0] || n > r[1])
996 /* Found a valid entry, copy and we are done */
997 if (a->flags & DN_C_LINK) {
998 if (copy_obj(a->start, a->end,
999 &s->link, "link", n))
1000 return DNHT_SCAN_END;
1001 if (copy_profile(a, s->profile))
1002 return DNHT_SCAN_END;
1003 if (copy_flowset(a, s->fs, 0))
1004 return DNHT_SCAN_END;
1006 if (a->flags & DN_C_SCH) {
1007 if (copy_obj(a->start, a->end,
1008 &s->sch, "sched", n))
1009 return DNHT_SCAN_END;
1010 /* list all attached flowsets */
1011 if (copy_fsk_list(a, s, 0))
1012 return DNHT_SCAN_END;
1014 if (a->flags & DN_C_FLOW)
1018 } else if (a->type == DN_FS) {
1019 /* queue show, skip internal flowsets */
1020 struct dn_fsk *fs = _o;
1025 /* see if the object is within one of our ranges */
1026 for (;r < lim; r += 2) {
1027 if (n < r[0] || n > r[1])
1029 if (copy_flowset(a, fs, 0))
1030 return DNHT_SCAN_END;
1032 break; /* we are done */
1038 static inline struct dn_schk *
1039 locate_scheduler(int i)
1041 return dn_ht_find(dn_cfg.schedhash, i, 0, NULL);
1045 * red parameters are in fixed point arithmetic.
1048 config_red(struct dn_fsk *fs)
1050 int64_t s, idle, weight, w0;
1053 fs->w_q = fs->fs.w_q;
1054 fs->max_p = fs->fs.max_p;
1056 /* Doing stuff that was in userland */
1057 i = fs->sched->link.bandwidth;
1059 hz * dn_cfg.red_avg_pkt_size * 8 * SCALE(1) / i;
1061 idle = div64((s * 3) , fs->w_q); /* s, fs->w_q scaled; idle not scaled */
1062 fs->lookup_step = div64(idle , dn_cfg.red_lookup_depth);
1063 /* fs->lookup_step not scaled, */
1064 if (!fs->lookup_step)
1065 fs->lookup_step = 1;
1066 w0 = weight = SCALE(1) - fs->w_q; //fs->w_q scaled
1068 for (t = fs->lookup_step; t > 1; --t)
1069 weight = SCALE_MUL(weight, w0);
1070 fs->lookup_weight = (int)(weight); // scaled
1072 /* Now doing stuff that was in kerneland */
1073 fs->min_th = SCALE(fs->fs.min_th);
1074 fs->max_th = SCALE(fs->fs.max_th);
1076 if (fs->fs.max_th == fs->fs.min_th)
1077 fs->c_1 = fs->max_p;
1079 fs->c_1 = SCALE((int64_t)(fs->max_p)) / (fs->fs.max_th - fs->fs.min_th);
1080 fs->c_2 = SCALE_MUL(fs->c_1, SCALE(fs->fs.min_th));
1082 if (fs->fs.flags & DN_IS_GENTLE_RED) {
1083 fs->c_3 = (SCALE(1) - fs->max_p) / fs->fs.max_th;
1084 fs->c_4 = SCALE(1) - 2 * fs->max_p;
1087 /* If the lookup table already exist, free and create it again. */
1088 if (fs->w_q_lookup) {
1089 free(fs->w_q_lookup, M_DUMMYNET);
1090 fs->w_q_lookup = NULL;
1092 if (dn_cfg.red_lookup_depth == 0) {
1093 printf("\ndummynet: net.inet.ip.dummynet.red_lookup_depth"
1095 fs->fs.flags &= ~DN_IS_RED;
1096 fs->fs.flags &= ~DN_IS_GENTLE_RED;
1099 fs->lookup_depth = dn_cfg.red_lookup_depth;
1100 fs->w_q_lookup = (u_int *)malloc(fs->lookup_depth * sizeof(int),
1101 M_DUMMYNET, M_NOWAIT);
1102 if (fs->w_q_lookup == NULL) {
1103 printf("dummynet: sorry, cannot allocate red lookup table\n");
1104 fs->fs.flags &= ~DN_IS_RED;
1105 fs->fs.flags &= ~DN_IS_GENTLE_RED;
1109 /* Fill the lookup table with (1 - w_q)^x */
1110 fs->w_q_lookup[0] = SCALE(1) - fs->w_q;
1112 for (i = 1; i < fs->lookup_depth; i++)
1114 SCALE_MUL(fs->w_q_lookup[i - 1], fs->lookup_weight);
1116 if (dn_cfg.red_avg_pkt_size < 1)
1117 dn_cfg.red_avg_pkt_size = 512;
1118 fs->avg_pkt_size = dn_cfg.red_avg_pkt_size;
1119 if (dn_cfg.red_max_pkt_size < 1)
1120 dn_cfg.red_max_pkt_size = 1500;
1121 fs->max_pkt_size = dn_cfg.red_max_pkt_size;
1126 /* Scan all flowset attached to this scheduler and update red */
1128 update_red(struct dn_schk *s)
1131 SLIST_FOREACH(fs, &s->fsk_list, sch_chain) {
1132 if (fs && (fs->fs.flags & DN_IS_RED))
1137 /* attach flowset to scheduler s, possibly requeue */
1139 fsk_attach(struct dn_fsk *fs, struct dn_schk *s)
1141 ND("remove fs %d from fsunlinked, link to sched %d",
1142 fs->fs.fs_nr, s->sch.sched_nr);
1143 SLIST_REMOVE(&dn_cfg.fsu, fs, dn_fsk, sch_chain);
1145 SLIST_INSERT_HEAD(&s->fsk_list, fs, sch_chain);
1148 /* XXX compute fsk_mask */
1149 fs->fsk_mask = fs->fs.flow_mask;
1150 if (fs->sched->sch.flags & DN_HAVE_MASK)
1151 flow_id_or(&fs->sched->sch.sched_mask, &fs->fsk_mask);
1154 * we must drain qht according to the old
1155 * type, and reinsert according to the new one.
1156 * The requeue is complex -- in general we need to
1157 * reclassify every single packet.
1158 * For the time being, let's hope qht is never set
1159 * when we reach this point.
1161 D("XXX TODO requeue from fs %d to sch %d",
1162 fs->fs.fs_nr, s->sch.sched_nr);
1165 /* set the new type for qht */
1166 if (nonzero_mask(&fs->fsk_mask))
1167 fs->fs.flags |= DN_QHT_HASH;
1169 fs->fs.flags &= ~DN_QHT_HASH;
1171 /* XXX config_red() can fail... */
1172 if (fs->fs.flags & DN_IS_RED)
1176 /* update all flowsets which may refer to this scheduler */
1178 update_fs(struct dn_schk *s)
1180 struct dn_fsk *fs, *tmp;
1182 SLIST_FOREACH_SAFE(fs, &dn_cfg.fsu, sch_chain, tmp) {
1183 if (s->sch.sched_nr != fs->fs.sched_nr) {
1184 D("fs %d for sch %d not %d still unlinked",
1185 fs->fs.fs_nr, fs->fs.sched_nr,
1194 * Configuration -- to preserve backward compatibility we use
1195 * the following scheme (N is 65536)
1196 * NUMBER SCHED LINK FLOWSET
1197 * 1 .. N-1 (1)WFQ (2)WFQ (3)queue
1198 * N+1 .. 2N-1 (4)FIFO (5)FIFO (6)FIFO for sched 1..N-1
1199 * 2N+1 .. 3N-1 -- -- (7)FIFO for sched N+1..2N-1
1201 * "pipe i config" configures #1, #2 and #3
1202 * "sched i config" configures #1 and possibly #6
1203 * "queue i config" configures #3
1204 * #1 is configured with 'pipe i config' or 'sched i config'
1205 * #2 is configured with 'pipe i config', and created if not
1206 * existing with 'sched i config'
1207 * #3 is configured with 'queue i config'
1208 * #4 is automatically configured after #1, can only be FIFO
1209 * #5 is automatically configured after #2
1210 * #6 is automatically created when #1 is !MULTIQUEUE,
1211 * and can be updated.
1212 * #7 is automatically configured after #2
1216 * configure a link (and its FIFO instance)
1219 config_link(struct dn_link *p, struct dn_id *arg)
1223 if (p->oid.len != sizeof(*p)) {
1224 D("invalid pipe len %d", p->oid.len);
1228 if (i <= 0 || i >= DN_MAX_ID)
1231 * The config program passes parameters as follows:
1232 * bw = bits/second (0 means no limits),
1233 * delay = ms, must be translated into ticks.
1234 * qsize = slots/bytes
1237 p->delay = (p->delay * hz) / 1000;
1238 /* Scale burst size: bytes -> bits * hz */
1242 /* do it twice, base link and FIFO link */
1243 for (; i < 2*DN_MAX_ID; i += DN_MAX_ID) {
1244 struct dn_schk *s = locate_scheduler(i);
1247 D("sched %d not found", i);
1250 /* remove profile if exists */
1252 free(s->profile, M_DUMMYNET);
1255 /* copy all parameters */
1256 s->link.oid = p->oid;
1257 s->link.link_nr = i;
1258 s->link.delay = p->delay;
1259 if (s->link.bandwidth != p->bandwidth) {
1260 /* XXX bandwidth changes, need to update red params */
1261 s->link.bandwidth = p->bandwidth;
1264 s->link.burst = p->burst;
1265 schk_reset_credit(s);
1273 * configure a flowset. Can be called from inside with locked=1,
1275 static struct dn_fsk *
1276 config_fs(struct dn_fs *nfs, struct dn_id *arg, int locked)
1281 if (nfs->oid.len != sizeof(*nfs)) {
1282 D("invalid flowset len %d", nfs->oid.len);
1286 if (i <= 0 || i >= 3*DN_MAX_ID)
1288 ND("flowset %d", i);
1289 /* XXX other sanity checks */
1290 if (nfs->flags & DN_QSIZE_BYTES) {
1291 ipdn_bound_var(&nfs->qsize, 16384,
1292 1500, dn_cfg.byte_limit, NULL); // "queue byte size");
1294 ipdn_bound_var(&nfs->qsize, 50,
1295 1, dn_cfg.slot_limit, NULL); // "queue slot size");
1297 if (nfs->flags & DN_HAVE_MASK) {
1298 /* make sure we have some buckets */
1299 ipdn_bound_var((int *)&nfs->buckets, dn_cfg.hash_size,
1300 1, dn_cfg.max_hash_size, "flowset buckets");
1302 nfs->buckets = 1; /* we only need 1 */
1306 do { /* exit with break when done */
1308 int flags = nfs->sched_nr ? DNHT_INSERT : 0;
1310 int oldc = dn_cfg.fsk_count;
1311 fs = dn_ht_find(dn_cfg.fshash, i, flags, NULL);
1313 D("missing sched for flowset %d", i);
1316 /* grab some defaults from the existing one */
1317 if (nfs->sched_nr == 0) /* reuse */
1318 nfs->sched_nr = fs->fs.sched_nr;
1319 for (j = 0; j < sizeof(nfs->par)/sizeof(nfs->par[0]); j++) {
1320 if (nfs->par[j] == -1) /* reuse */
1321 nfs->par[j] = fs->fs.par[j];
1323 if (bcmp(&fs->fs, nfs, sizeof(*nfs)) == 0) {
1324 ND("flowset %d unchanged", i);
1325 break; /* no change, nothing to do */
1327 if (oldc != dn_cfg.fsk_count) /* new item */
1329 s = locate_scheduler(nfs->sched_nr);
1330 /* detach from old scheduler if needed, preserving
1331 * queues if we need to reattach. Then update the
1332 * configuration, and possibly attach to the new sched.
1334 DX(2, "fs %d changed sched %d@%p to %d@%p",
1336 fs->fs.sched_nr, fs->sched, nfs->sched_nr, s);
1338 int flags = s ? DN_DETACH : (DN_DETACH | DN_DESTROY);
1339 flags |= DN_DESTROY; /* XXX temporary */
1340 fsk_detach(fs, flags);
1342 fs->fs = *nfs; /* copy configuration */
1352 * config/reconfig a scheduler and its FIFO variant.
1353 * For !MULTIQUEUE schedulers, also set up the flowset.
1355 * On reconfigurations (detected because s->fp is set),
1356 * detach existing flowsets preserving traffic, preserve link,
1357 * and delete the old scheduler creating a new one.
1360 config_sched(struct dn_sch *_nsch, struct dn_id *arg)
1363 struct schk_new_arg a; /* argument for schk_new */
1365 struct dn_link p; /* copy of oldlink */
1366 struct dn_profile *pf = NULL; /* copy of old link profile */
1367 /* Used to preserv mask parameter */
1368 struct ipfw_flow_id new_mask;
1369 int new_buckets = 0;
1375 if (a.sch->oid.len != sizeof(*a.sch)) {
1376 D("bad sched len %d", a.sch->oid.len);
1379 i = a.sch->sched_nr;
1380 if (i <= 0 || i >= DN_MAX_ID)
1382 /* make sure we have some buckets */
1383 if (a.sch->flags & DN_HAVE_MASK)
1384 ipdn_bound_var((int *)&a.sch->buckets, dn_cfg.hash_size,
1385 1, dn_cfg.max_hash_size, "sched buckets");
1386 /* XXX other sanity checks */
1387 bzero(&p, sizeof(p));
1389 pipe_cmd = a.sch->flags & DN_PIPE_CMD;
1390 a.sch->flags &= ~DN_PIPE_CMD; //XXX do it even if is not set?
1392 /* Copy mask parameter */
1393 new_mask = a.sch->sched_mask;
1394 new_buckets = a.sch->buckets;
1395 new_flags = a.sch->flags;
1398 again: /* run twice, for wfq and fifo */
1400 * lookup the type. If not supplied, use the previous one
1401 * or default to WF2Q+. Otherwise, return an error.
1404 a.fp = find_sched_type(a.sch->oid.subtype, a.sch->name);
1406 /* found. Lookup or create entry */
1407 s = dn_ht_find(dn_cfg.schedhash, i, DNHT_INSERT, &a);
1408 } else if (a.sch->oid.subtype == 0 && !a.sch->name[0]) {
1409 /* No type. search existing s* or retry with WF2Q+ */
1410 s = dn_ht_find(dn_cfg.schedhash, i, 0, &a);
1413 /* Scheduler exists, skip to FIFO scheduler
1414 * if command was pipe config...
1419 /* New scheduler, create a wf2q+ with no mask
1420 * if command was pipe config...
1423 /* clear mask parameter */
1424 bzero(&a.sch->sched_mask, sizeof(new_mask));
1426 a.sch->flags &= ~DN_HAVE_MASK;
1428 a.sch->oid.subtype = DN_SCHED_WF2QP;
1432 D("invalid scheduler type %d %s",
1433 a.sch->oid.subtype, a.sch->name);
1437 /* normalize name and subtype */
1438 a.sch->oid.subtype = a.fp->type;
1439 bzero(a.sch->name, sizeof(a.sch->name));
1440 strlcpy(a.sch->name, a.fp->name, sizeof(a.sch->name));
1442 D("cannot allocate scheduler %d", i);
1445 /* restore existing link if any */
1448 if (!pf || pf->link_nr != p.link_nr) { /* no saved value */
1449 s->profile = NULL; /* XXX maybe not needed */
1451 s->profile = malloc(sizeof(struct dn_profile),
1452 M_DUMMYNET, M_NOWAIT | M_ZERO);
1453 if (s->profile == NULL) {
1454 D("cannot allocate profile");
1457 bcopy(pf, s->profile, sizeof(*pf));
1461 if (s->fp == NULL) {
1462 DX(2, "sched %d new type %s", i, a.fp->name);
1463 } else if (s->fp != a.fp ||
1464 bcmp(a.sch, &s->sch, sizeof(*a.sch)) ) {
1465 /* already existing. */
1466 DX(2, "sched %d type changed from %s to %s",
1467 i, s->fp->name, a.fp->name);
1468 DX(4, " type/sub %d/%d -> %d/%d",
1469 s->sch.oid.type, s->sch.oid.subtype,
1470 a.sch->oid.type, a.sch->oid.subtype);
1471 if (s->link.link_nr == 0)
1472 D("XXX WARNING link 0 for sched %d", i);
1473 p = s->link; /* preserve link */
1474 if (s->profile) {/* preserve profile */
1476 pf = malloc(sizeof(*pf),
1477 M_DUMMYNET, M_NOWAIT | M_ZERO);
1478 if (pf) /* XXX should issue a warning otherwise */
1479 bcopy(s->profile, pf, sizeof(*pf));
1481 /* remove from the hash */
1482 dn_ht_find(dn_cfg.schedhash, i, DNHT_REMOVE, NULL);
1483 /* Detach flowsets, preserve queues. */
1484 // schk_delete_cb(s, NULL);
1485 // XXX temporarily, kill queues
1486 schk_delete_cb(s, (void *)DN_DESTROY);
1489 DX(4, "sched %d unchanged type %s", i, a.fp->name);
1491 /* complete initialization */
1495 // XXX schk_reset_credit(s);
1496 /* create the internal flowset if needed,
1497 * trying to reuse existing ones if available
1499 if (!(s->fp->flags & DN_MULTIQUEUE) && !s->fs) {
1500 s->fs = dn_ht_find(dn_cfg.fshash, i, 0, NULL);
1503 bzero(&fs, sizeof(fs));
1504 set_oid(&fs.oid, DN_FS, sizeof(fs));
1505 fs.fs_nr = i + DN_MAX_ID;
1507 s->fs = config_fs(&fs, NULL, 1 /* locked */);
1510 schk_delete_cb(s, (void *)DN_DESTROY);
1511 D("error creating internal fs for %d", i);
1515 /* call init function after the flowset is created */
1520 if (i < DN_MAX_ID) { /* now configure the FIFO instance */
1523 /* Restore mask parameter for FIFO */
1524 a.sch->sched_mask = new_mask;
1525 a.sch->buckets = new_buckets;
1526 a.sch->flags = new_flags;
1528 /* sched config shouldn't modify the FIFO scheduler */
1529 if (dn_ht_find(dn_cfg.schedhash, i, 0, &a) != NULL) {
1530 /* FIFO already exist, don't touch it */
1531 err = 0; /* and this is not an error */
1535 a.sch->sched_nr = i;
1536 a.sch->oid.subtype = DN_SCHED_FIFO;
1537 bzero(a.sch->name, sizeof(a.sch->name));
1544 free(pf, M_DUMMYNET);
1549 * attach a profile to a link
1552 config_profile(struct dn_profile *pf, struct dn_id *arg)
1555 int i, olen, err = 0;
1557 if (pf->oid.len < sizeof(*pf)) {
1558 D("short profile len %d", pf->oid.len);
1562 if (i <= 0 || i >= DN_MAX_ID)
1564 /* XXX other sanity checks */
1566 for (; i < 2*DN_MAX_ID; i += DN_MAX_ID) {
1567 s = locate_scheduler(i);
1575 * If we had a profile and the new one does not fit,
1576 * or it is deleted, then we need to free memory.
1578 if (s->profile && (pf->samples_no == 0 ||
1579 s->profile->oid.len < pf->oid.len)) {
1580 free(s->profile, M_DUMMYNET);
1583 if (pf->samples_no == 0)
1586 * new profile, possibly allocate memory
1589 if (s->profile == NULL)
1590 s->profile = malloc(pf->oid.len,
1591 M_DUMMYNET, M_NOWAIT | M_ZERO);
1592 if (s->profile == NULL) {
1593 D("no memory for profile %d", i);
1597 /* preserve larger length XXX double check */
1598 olen = s->profile->oid.len;
1599 if (olen < pf->oid.len)
1601 bcopy(pf, s->profile, pf->oid.len);
1602 s->profile->oid.len = olen;
1609 * Delete all objects:
1612 dummynet_flush(void)
1615 /* delete all schedulers and related links/queues/flowsets */
1616 dn_ht_scan(dn_cfg.schedhash, schk_delete_cb,
1617 (void *)(uintptr_t)DN_DELETE_FS);
1618 /* delete all remaining (unlinked) flowsets */
1619 DX(4, "still %d unlinked fs", dn_cfg.fsk_count);
1620 dn_ht_free(dn_cfg.fshash, DNHT_REMOVE);
1621 fsk_detach_list(&dn_cfg.fsu, DN_DELETE_FS);
1622 /* Reinitialize system heap... */
1623 heap_init(&dn_cfg.evheap, 16, offsetof(struct dn_id, id));
1627 * Main handler for configuration. We are guaranteed to be called
1628 * with an oid which is at least a dn_id.
1629 * - the first object is the command (config, delete, flush, ...)
1630 * - config_link must be issued after the corresponding config_sched
1631 * - parameters (DN_TXT) for an object must preceed the object
1632 * processed on a config_sched.
1635 do_config(void *p, int l)
1637 struct dn_id *next, *o;
1638 int err = 0, err2 = 0;
1639 struct dn_id *arg = NULL;
1643 if (o->id != DN_API_VERSION) {
1644 D("invalid api version got %d need %d",
1645 o->id, DN_API_VERSION);
1648 for (; l >= sizeof(*o); o = next) {
1649 struct dn_id *prev = arg;
1650 if (o->len < sizeof(*o) || l < o->len) {
1651 D("bad len o->len %d len %d", o->len, l);
1656 next = (struct dn_id *)((char *)o + o->len);
1660 D("cmd %d not implemented", o->type);
1663 #ifdef EMULATE_SYSCTL
1664 /* sysctl emulation.
1665 * if we recognize the command, jump to the correct
1666 * handler and return
1669 err = kesysctl_emu_set(p, l);
1673 case DN_CMD_CONFIG: /* simply a header */
1677 /* the argument is in the first uintptr_t after o */
1678 a = (uintptr_t *)(o+1);
1679 if (o->len < sizeof(*o) + sizeof(*a)) {
1683 switch (o->subtype) {
1685 /* delete base and derived schedulers */
1687 err = delete_schk(*a);
1688 err2 = delete_schk(*a + DN_MAX_ID);
1695 D("invalid delete type %d",
1701 err = (*a <1 || *a >= DN_MAX_ID) ?
1702 EINVAL : delete_fs(*a, 0) ;
1712 case DN_TEXT: /* store argument the next block */
1717 err = config_link((struct dn_link *)o, arg);
1720 err = config_profile((struct dn_profile *)o, arg);
1723 err = config_sched((struct dn_sch *)o, arg);
1726 err = (NULL==config_fs((struct dn_fs *)o, arg, 0));
1738 compute_space(struct dn_id *cmd, struct copy_args *a)
1740 int x = 0, need = 0;
1741 int profile_size = sizeof(struct dn_profile) -
1742 ED_MAX_SAMPLES_NO*sizeof(int);
1744 /* NOTE about compute space:
1745 * NP = dn_cfg.schk_count
1746 * NSI = dn_cfg.si_count
1747 * NF = dn_cfg.fsk_count
1748 * NQ = dn_cfg.queue_count
1750 * (NP/2)*(dn_link + dn_sch + dn_id + dn_fs) only half scheduler
1751 * link, scheduler template, flowset
1752 * integrated in scheduler and header
1754 * (NSI)*(dn_flow) all scheduler instance (includes
1755 * the queue instance)
1757 * (NP/2)*(dn_link + dn_sch + dn_id + dn_fs) only half scheduler
1758 * link, scheduler template, flowset
1759 * integrated in scheduler and header
1761 * (NSI * dn_flow) all scheduler instances
1762 * (NF * sizeof(uint_32)) space for flowset list linked to scheduler
1763 * (NQ * dn_queue) all queue [XXXfor now not listed]
1765 * (NF * dn_fs) all flowset
1766 * (NQ * dn_queue) all queues
1768 switch (cmd->subtype) {
1771 /* XXX where do LINK and SCH differ ? */
1772 /* 'ipfw sched show' could list all queues associated to
1773 * a scheduler. This feature for now is disabled
1775 case DN_LINK: /* pipe show */
1776 x = DN_C_LINK | DN_C_SCH | DN_C_FLOW;
1777 need += dn_cfg.schk_count *
1778 (sizeof(struct dn_fs) + profile_size) / 2;
1779 need += dn_cfg.fsk_count * sizeof(uint32_t);
1781 case DN_SCH: /* sched show */
1782 need += dn_cfg.schk_count *
1783 (sizeof(struct dn_fs) + profile_size) / 2;
1784 need += dn_cfg.fsk_count * sizeof(uint32_t);
1785 x = DN_C_SCH | DN_C_LINK | DN_C_FLOW;
1787 case DN_FS: /* queue show */
1788 x = DN_C_FS | DN_C_QUEUE;
1790 case DN_GET_COMPAT: /* compatibility mode */
1791 need = dn_compat_calc_size();
1796 need += dn_cfg.schk_count * sizeof(struct dn_sch) / 2;
1797 /* NOT also, each fs might be attached to a sched */
1798 need += dn_cfg.schk_count * sizeof(struct dn_id) / 2;
1801 need += dn_cfg.fsk_count * sizeof(struct dn_fs);
1802 if (x & DN_C_LINK) {
1803 need += dn_cfg.schk_count * sizeof(struct dn_link) / 2;
1806 * When exporting a queue to userland, only pass up the
1807 * struct dn_flow, which is the only visible part.
1811 need += dn_cfg.queue_count * sizeof(struct dn_flow);
1813 need += dn_cfg.si_count * (sizeof(struct dn_flow));
1818 * If compat != NULL dummynet_get is called in compatibility mode.
1819 * *compat will be the pointer to the buffer to pass to ipfw
1822 dummynet_get(struct sockopt *sopt, void **compat)
1824 int have, i, need, error;
1825 char *start = NULL, *buf;
1826 size_t sopt_valsize;
1829 struct copy_range r;
1830 int l = sizeof(struct dn_id);
1832 bzero(&a, sizeof(a));
1833 bzero(&r, sizeof(r));
1835 /* save and restore original sopt_valsize around copyin */
1836 sopt_valsize = sopt->sopt_valsize;
1841 /* copy at least an oid, and possibly a full object */
1842 error = sooptcopyin(sopt, cmd, sizeof(r), sizeof(*cmd));
1843 sopt->sopt_valsize = sopt_valsize;
1847 #ifdef EMULATE_SYSCTL
1848 /* sysctl emulation. */
1849 if (cmd->type == DN_SYSCTL_GET)
1850 return kesysctl_emu_get(sopt);
1852 if (l > sizeof(r)) {
1853 /* request larger than default, allocate buffer */
1854 cmd = malloc(l, M_DUMMYNET, M_WAITOK);
1855 error = sooptcopyin(sopt, cmd, l, l);
1856 sopt->sopt_valsize = sopt_valsize;
1860 } else { /* compatibility */
1862 cmd->type = DN_CMD_GET;
1863 cmd->len = sizeof(struct dn_id);
1864 cmd->subtype = DN_GET_COMPAT;
1865 // cmd->id = sopt_valsize;
1866 D("compatibility mode");
1868 a.extra = (struct copy_range *)cmd;
1869 if (cmd->len == sizeof(*cmd)) { /* no range, create a default */
1870 uint32_t *rp = (uint32_t *)(cmd + 1);
1871 cmd->len += 2* sizeof(uint32_t);
1873 rp[1] = DN_MAX_ID - 1;
1874 if (cmd->subtype == DN_LINK) {
1879 /* Count space (under lock) and allocate (outside lock).
1880 * Exit with lock held if we manage to get enough buffer.
1881 * Try a few times then give up.
1883 for (have = 0, i = 0; i < 10; i++) {
1885 need = compute_space(cmd, &a);
1887 /* if there is a range, ignore value from compute_space() */
1888 if (l > sizeof(*cmd))
1889 need = sopt_valsize - sizeof(*cmd);
1896 need += sizeof(*cmd);
1903 free(start, M_DUMMYNET);
1905 if (need > sopt_valsize)
1909 start = malloc(have, M_DUMMYNET, M_WAITOK | M_ZERO);
1912 if (start == NULL) {
1917 error = sooptcopyout(sopt, cmd, sizeof(*cmd));
1921 ND("have %d:%d sched %d, %d:%d links %d, %d:%d flowsets %d, "
1922 "%d:%d si %d, %d:%d queues %d",
1923 dn_cfg.schk_count, sizeof(struct dn_sch), DN_SCH,
1924 dn_cfg.schk_count, sizeof(struct dn_link), DN_LINK,
1925 dn_cfg.fsk_count, sizeof(struct dn_fs), DN_FS,
1926 dn_cfg.si_count, sizeof(struct dn_flow), DN_SCH_I,
1927 dn_cfg.queue_count, sizeof(struct dn_queue), DN_QUEUE);
1928 sopt->sopt_valsize = sopt_valsize;
1929 a.type = cmd->subtype;
1931 if (compat == NULL) {
1932 bcopy(cmd, start, sizeof(*cmd));
1933 ((struct dn_id*)(start))->len = sizeof(struct dn_id);
1934 buf = start + sizeof(*cmd);
1938 a.end = start + have;
1939 /* start copying other objects */
1941 a.type = DN_COMPAT_PIPE;
1942 dn_ht_scan(dn_cfg.schedhash, copy_data_helper_compat, &a);
1943 a.type = DN_COMPAT_QUEUE;
1944 dn_ht_scan(dn_cfg.fshash, copy_data_helper_compat, &a);
1945 } else if (a.type == DN_FS) {
1946 dn_ht_scan(dn_cfg.fshash, copy_data_helper, &a);
1948 dn_ht_scan(dn_cfg.schedhash, copy_data_helper, &a);
1954 sopt->sopt_valsize = buf - start;
1955 /* free() is done by ip_dummynet_compat() */
1956 start = NULL; //XXX hack
1958 error = sooptcopyout(sopt, start, buf - start);
1961 if (cmd && cmd != &r.o)
1962 free(cmd, M_DUMMYNET);
1964 free(start, M_DUMMYNET);
1968 /* Callback called on scheduler instance to delete it if idle */
1970 drain_scheduler_cb(void *_si, void *arg)
1972 struct dn_sch_inst *si = _si;
1974 if ((si->kflags & DN_ACTIVE) || si->dline.mq.head != NULL)
1977 if (si->sched->fp->flags & DN_MULTIQUEUE) {
1978 if (si->q_count == 0)
1979 return si_destroy(si, NULL);
1982 } else { /* !DN_MULTIQUEUE */
1983 if ((si+1)->ni.length == 0)
1984 return si_destroy(si, NULL);
1988 return 0; /* unreachable */
1991 /* Callback called on scheduler to check if it has instances */
1993 drain_scheduler_sch_cb(void *_s, void *arg)
1995 struct dn_schk *s = _s;
1997 if (s->sch.flags & DN_HAVE_MASK) {
1998 dn_ht_scan_bucket(s->siht, &s->drain_bucket,
1999 drain_scheduler_cb, NULL);
2003 if (drain_scheduler_cb(s->siht, NULL) == DNHT_SCAN_DEL)
2010 /* Called every tick, try to delete a 'bucket' of scheduler */
2012 dn_drain_scheduler(void)
2014 dn_ht_scan_bucket(dn_cfg.schedhash, &dn_cfg.drain_sch,
2015 drain_scheduler_sch_cb, NULL);
2019 /* Callback called on queue to delete if it is idle */
2021 drain_queue_cb(void *_q, void *arg)
2023 struct dn_queue *q = _q;
2025 if (q->ni.length == 0) {
2026 dn_delete_queue(q, DN_DESTROY);
2027 return DNHT_SCAN_DEL; /* queue is deleted */
2030 return 0; /* queue isn't deleted */
2033 /* Callback called on flowset used to check if it has queues */
2035 drain_queue_fs_cb(void *_fs, void *arg)
2037 struct dn_fsk *fs = _fs;
2039 if (fs->fs.flags & DN_QHT_HASH) {
2040 /* Flowset has a hash table for queues */
2041 dn_ht_scan_bucket(fs->qht, &fs->drain_bucket,
2042 drain_queue_cb, NULL);
2045 /* No hash table for this flowset, null the pointer
2046 * if the queue is deleted
2049 if (drain_queue_cb(fs->qht, NULL) == DNHT_SCAN_DEL)
2056 /* Called every tick, try to delete a 'bucket' of queue */
2058 dn_drain_queue(void)
2060 /* scan a bucket of flowset */
2061 dn_ht_scan_bucket(dn_cfg.fshash, &dn_cfg.drain_fs,
2062 drain_queue_fs_cb, NULL);
2067 * Handler for the various dummynet socket options
2070 ip_dn_ctl(struct sockopt *sopt)
2075 error = priv_check(sopt->sopt_td, PRIV_NETINET_DUMMYNET);
2079 /* Disallow sets in really-really secure mode. */
2080 if (sopt->sopt_dir == SOPT_SET) {
2081 error = securelevel_ge(sopt->sopt_td->td_ucred, 3);
2086 switch (sopt->sopt_name) {
2088 D("dummynet: unknown option %d", sopt->sopt_name);
2092 case IP_DUMMYNET_FLUSH:
2093 case IP_DUMMYNET_CONFIGURE:
2094 case IP_DUMMYNET_DEL: /* remove a pipe or queue */
2095 case IP_DUMMYNET_GET:
2096 D("dummynet: compat option %d", sopt->sopt_name);
2097 error = ip_dummynet_compat(sopt);
2101 if (sopt->sopt_dir == SOPT_GET) {
2102 error = dummynet_get(sopt, NULL);
2105 l = sopt->sopt_valsize;
2106 if (l < sizeof(struct dn_id) || l > 12000) {
2107 D("argument len %d invalid", l);
2110 p = malloc(l, M_TEMP, M_WAITOK); // XXX can it fail ?
2111 error = sooptcopyin(sopt, p, l, l);
2114 error = do_config(p, l);
2128 if (dn_cfg.init_done)
2130 printf("DUMMYNET %p with IPv6 initialized (100409)\n", curvnet);
2131 dn_cfg.init_done = 1;
2132 /* Set defaults here. MSVC does not accept initializers,
2133 * and this is also useful for vimages
2136 dn_cfg.slot_limit = 100; /* Foot shooting limit for queues. */
2137 dn_cfg.byte_limit = 1024 * 1024;
2140 /* RED parameters */
2141 dn_cfg.red_lookup_depth = 256; /* default lookup table depth */
2142 dn_cfg.red_avg_pkt_size = 512; /* default medium packet size */
2143 dn_cfg.red_max_pkt_size = 1500; /* default max packet size */
2146 dn_cfg.max_hash_size = 65536; /* max in the hash tables */
2147 dn_cfg.hash_size = 64; /* default hash size */
2149 /* create hash tables for schedulers and flowsets.
2150 * In both we search by key and by pointer.
2152 dn_cfg.schedhash = dn_ht_init(NULL, dn_cfg.hash_size,
2153 offsetof(struct dn_schk, schk_next),
2154 schk_hash, schk_match, schk_new);
2155 dn_cfg.fshash = dn_ht_init(NULL, dn_cfg.hash_size,
2156 offsetof(struct dn_fsk, fsk_next),
2157 fsk_hash, fsk_match, fsk_new);
2159 /* bucket index to drain object */
2160 dn_cfg.drain_fs = 0;
2161 dn_cfg.drain_sch = 0;
2163 heap_init(&dn_cfg.evheap, 16, offsetof(struct dn_id, id));
2164 SLIST_INIT(&dn_cfg.fsu);
2165 SLIST_INIT(&dn_cfg.schedlist);
2169 TASK_INIT(&dn_task, 0, dummynet_task, curvnet);
2170 dn_tq = taskqueue_create_fast("dummynet", M_WAITOK,
2171 taskqueue_thread_enqueue, &dn_tq);
2172 taskqueue_start_threads(&dn_tq, 1, PI_NET, "dummynet");
2174 callout_init(&dn_timeout, CALLOUT_MPSAFE);
2177 /* Initialize curr_time adjustment mechanics. */
2178 getmicrouptime(&dn_cfg.prev_t);
2182 ip_dn_destroy(int last)
2185 /* ensure no more callouts are started */
2188 /* check for last */
2190 ND("removing last instance\n");
2191 ip_dn_ctl_ptr = NULL;
2192 ip_dn_io_ptr = NULL;
2198 callout_drain(&dn_timeout);
2199 taskqueue_drain(dn_tq, &dn_task);
2200 taskqueue_free(dn_tq);
2202 dn_ht_free(dn_cfg.schedhash, 0);
2203 dn_ht_free(dn_cfg.fshash, 0);
2204 heap_free(&dn_cfg.evheap);
2210 dummynet_modevent(module_t mod, int type, void *data)
2213 if (type == MOD_LOAD) {
2215 printf("DUMMYNET already loaded\n");
2219 ip_dn_ctl_ptr = ip_dn_ctl;
2220 ip_dn_io_ptr = dummynet_io;
2222 } else if (type == MOD_UNLOAD) {
2223 ip_dn_destroy(1 /* last */);
2229 /* modevent helpers for the modules */
2231 load_dn_sched(struct dn_alg *d)
2236 return 1; /* error */
2237 ip_dn_init(); /* just in case, we need the lock */
2239 /* Check that mandatory funcs exists */
2240 if (d->enqueue == NULL || d->dequeue == NULL) {
2241 D("missing enqueue or dequeue for %s", d->name);
2245 /* Search if scheduler already exists */
2247 SLIST_FOREACH(s, &dn_cfg.schedlist, next) {
2248 if (strcmp(s->name, d->name) == 0) {
2249 D("%s already loaded", d->name);
2250 break; /* scheduler already exists */
2254 SLIST_INSERT_HEAD(&dn_cfg.schedlist, d, next);
2256 D("dn_sched %s %sloaded", d->name, s ? "not ":"");
2261 unload_dn_sched(struct dn_alg *s)
2263 struct dn_alg *tmp, *r;
2266 ND("called for %s", s->name);
2269 SLIST_FOREACH_SAFE(r, &dn_cfg.schedlist, next, tmp) {
2270 if (strcmp(s->name, r->name) != 0)
2272 ND("ref_count = %d", r->ref_count);
2273 err = (r->ref_count != 0) ? EBUSY : 0;
2275 SLIST_REMOVE(&dn_cfg.schedlist, r, dn_alg, next);
2279 D("dn_sched %s %sunloaded", s->name, err ? "not ":"");
2284 dn_sched_modevent(module_t mod, int cmd, void *arg)
2286 struct dn_alg *sch = arg;
2288 if (cmd == MOD_LOAD)
2289 return load_dn_sched(sch);
2290 else if (cmd == MOD_UNLOAD)
2291 return unload_dn_sched(sch);
2296 static moduledata_t dummynet_mod = {
2297 "dummynet", dummynet_modevent, NULL
2300 #define DN_SI_SUB SI_SUB_PROTO_IFATTACHDOMAIN
2301 #define DN_MODEV_ORD (SI_ORDER_ANY - 128) /* after ipfw */
2302 DECLARE_MODULE(dummynet, dummynet_mod, DN_SI_SUB, DN_MODEV_ORD);
2303 MODULE_DEPEND(dummynet, ipfw, 2, 2, 2);
2304 MODULE_VERSION(dummynet, 3);
2307 * Starting up. Done in order after dummynet_modevent() has been called.
2308 * VNET_SYSINIT is also called for each existing vnet and each new vnet.
2310 //VNET_SYSINIT(vnet_dn_init, DN_SI_SUB, DN_MODEV_ORD+2, ip_dn_init, NULL);
2313 * Shutdown handlers up shop. These are done in REVERSE ORDER, but still
2314 * after dummynet_modevent() has been called. Not called on reboot.
2315 * VNET_SYSUNINIT is also called for each exiting vnet as it exits.
2316 * or when the module is unloaded.
2318 //VNET_SYSUNINIT(vnet_dn_uninit, DN_SI_SUB, DN_MODEV_ORD+2, ip_dn_destroy, NULL);