1 /* net/sched/sch_atm.c - ATM VC selection "queueing discipline" */
3 /* Written 1998-2000 by Werner Almesberger, EPFL ICA */
6 #include <linux/config.h>
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/string.h>
10 #include <linux/errno.h>
11 #include <linux/skbuff.h>
12 #include <linux/interrupt.h>
13 #include <linux/atmdev.h>
14 #include <linux/atmclip.h>
15 #include <linux/netdevice.h>
16 #include <linux/rtnetlink.h>
17 #include <linux/file.h> /* for fput */
18 #include <net/pkt_sched.h>
22 extern struct socket *sockfd_lookup(int fd, int *err); /* @@@ fix this */
25 #define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
27 #define DPRINTK(format,args...)
31 #define D2PRINTK(format,args...) printk(KERN_DEBUG format,##args)
33 #define D2PRINTK(format,args...)
38 * The ATM queuing discipline provides a framework for invoking classifiers
39 * (aka "filters"), which in turn select classes of this queuing discipline.
40 * Each class maps the flow(s) it is handling to a given VC. Multiple classes
41 * may share the same VC.
43 * When creating a class, VCs are specified by passing the number of the open
44 * socket descriptor by which the calling process references the VC. The kernel
45 * keeps the VC open at least until all classes using it are removed.
47 * In this file, most functions are named atm_tc_* to avoid confusion with all
48 * the atm_* in net/atm. This naming convention differs from what's used in the
52 * - sometimes messes up the IP stack
53 * - any manipulations besides the few operations described in the README, are
54 * untested and likely to crash the system
55 * - should lock the flow while there is data in the queue (?)
59 #define PRIV(sch) qdisc_priv(sch)
60 #define VCC2FLOW(vcc) ((struct atm_flow_data *) ((vcc)->user_back))
63 struct atm_flow_data {
64 struct Qdisc *q; /* FIFO, TBF, etc. */
65 struct tcf_proto *filter_list;
66 struct atm_vcc *vcc; /* VCC; NULL if VCC is closed */
67 void (*old_pop)(struct atm_vcc *vcc,struct sk_buff *skb); /* chaining */
68 struct atm_qdisc_data *parent; /* parent qdisc */
69 struct socket *sock; /* for closing */
70 u32 classid; /* x:y type ID */
71 int ref; /* reference count */
72 struct gnet_stats_basic bstats;
73 struct gnet_stats_queue qstats;
74 spinlock_t *stats_lock;
75 struct atm_flow_data *next;
76 struct atm_flow_data *excess; /* flow for excess traffic;
77 NULL to set CLP instead */
79 unsigned char hdr[0]; /* header data; MUST BE LAST */
82 struct atm_qdisc_data {
83 struct atm_flow_data link; /* unclassified skbs go here */
84 struct atm_flow_data *flows; /* NB: "link" is also on this
86 struct tasklet_struct task; /* requeue tasklet */
90 /* ------------------------- Class/flow operations ------------------------- */
93 static int find_flow(struct atm_qdisc_data *qdisc,struct atm_flow_data *flow)
95 struct atm_flow_data *walk;
97 DPRINTK("find_flow(qdisc %p,flow %p)\n",qdisc,flow);
98 for (walk = qdisc->flows; walk; walk = walk->next)
99 if (walk == flow) return 1;
100 DPRINTK("find_flow: not found\n");
105 static __inline__ struct atm_flow_data *lookup_flow(struct Qdisc *sch,
108 struct atm_qdisc_data *p = PRIV(sch);
109 struct atm_flow_data *flow;
111 for (flow = p->flows; flow; flow = flow->next)
112 if (flow->classid == classid) break;
117 static int atm_tc_graft(struct Qdisc *sch,unsigned long arg,
118 struct Qdisc *new,struct Qdisc **old)
120 struct atm_qdisc_data *p = PRIV(sch);
121 struct atm_flow_data *flow = (struct atm_flow_data *) arg;
123 DPRINTK("atm_tc_graft(sch %p,[qdisc %p],flow %p,new %p,old %p)\n",sch,
125 if (!find_flow(p,flow)) return -EINVAL;
126 if (!new) new = &noop_qdisc;
127 *old = xchg(&flow->q,new);
128 if (*old) qdisc_reset(*old);
133 static struct Qdisc *atm_tc_leaf(struct Qdisc *sch,unsigned long cl)
135 struct atm_flow_data *flow = (struct atm_flow_data *) cl;
137 DPRINTK("atm_tc_leaf(sch %p,flow %p)\n",sch,flow);
138 return flow ? flow->q : NULL;
142 static unsigned long atm_tc_get(struct Qdisc *sch,u32 classid)
144 struct atm_qdisc_data *p __attribute__((unused)) = PRIV(sch);
145 struct atm_flow_data *flow;
147 DPRINTK("atm_tc_get(sch %p,[qdisc %p],classid %x)\n",sch,p,classid);
148 flow = lookup_flow(sch,classid);
149 if (flow) flow->ref++;
150 DPRINTK("atm_tc_get: flow %p\n",flow);
151 return (unsigned long) flow;
155 static unsigned long atm_tc_bind_filter(struct Qdisc *sch,
156 unsigned long parent, u32 classid)
158 return atm_tc_get(sch,classid);
162 static void destroy_filters(struct atm_flow_data *flow)
164 struct tcf_proto *filter;
166 while ((filter = flow->filter_list)) {
167 DPRINTK("destroy_filters: destroying filter %p\n",filter);
168 flow->filter_list = filter->next;
175 * atm_tc_put handles all destructions, including the ones that are explicitly
176 * requested (atm_tc_destroy, etc.). The assumption here is that we never drop
177 * anything that still seems to be in use.
180 static void atm_tc_put(struct Qdisc *sch, unsigned long cl)
182 struct atm_qdisc_data *p = PRIV(sch);
183 struct atm_flow_data *flow = (struct atm_flow_data *) cl;
184 struct atm_flow_data **prev;
186 DPRINTK("atm_tc_put(sch %p,[qdisc %p],flow %p)\n",sch,p,flow);
187 if (--flow->ref) return;
188 DPRINTK("atm_tc_put: destroying\n");
189 for (prev = &p->flows; *prev; prev = &(*prev)->next)
190 if (*prev == flow) break;
192 printk(KERN_CRIT "atm_tc_put: class %p not found\n",flow);
196 DPRINTK("atm_tc_put: qdisc %p\n",flow->q);
197 qdisc_destroy(flow->q);
198 destroy_filters(flow);
200 DPRINTK("atm_tc_put: f_count %d\n",
201 file_count(flow->sock->file));
202 flow->vcc->pop = flow->old_pop;
203 sockfd_put(flow->sock);
205 if (flow->excess) atm_tc_put(sch,(unsigned long) flow->excess);
206 if (flow != &p->link) kfree(flow);
208 * If flow == &p->link, the qdisc no longer works at this point and
209 * needs to be removed. (By the caller of atm_tc_put.)
214 static void sch_atm_pop(struct atm_vcc *vcc,struct sk_buff *skb)
216 struct atm_qdisc_data *p = VCC2FLOW(vcc)->parent;
218 D2PRINTK("sch_atm_pop(vcc %p,skb %p,[qdisc %p])\n",vcc,skb,p);
219 VCC2FLOW(vcc)->old_pop(vcc,skb);
220 tasklet_schedule(&p->task);
223 static const u8 llc_oui_ip[] = {
224 0xaa, /* DSAP: non-ISO */
225 0xaa, /* SSAP: non-ISO */
226 0x03, /* Ctrl: Unnumbered Information Command PDU */
227 0x00, /* OUI: EtherType */
229 0x08, 0x00 }; /* Ethertype IP (0800) */
231 static int atm_tc_change(struct Qdisc *sch, u32 classid, u32 parent,
232 struct rtattr **tca, unsigned long *arg)
234 struct atm_qdisc_data *p = PRIV(sch);
235 struct atm_flow_data *flow = (struct atm_flow_data *) *arg;
236 struct atm_flow_data *excess = NULL;
237 struct rtattr *opt = tca[TCA_OPTIONS-1];
238 struct rtattr *tb[TCA_ATM_MAX];
240 int fd,error,hdr_len;
243 DPRINTK("atm_tc_change(sch %p,[qdisc %p],classid %x,parent %x,"
244 "flow %p,opt %p)\n",sch,p,classid,parent,flow,opt);
246 * The concept of parents doesn't apply for this qdisc.
248 if (parent && parent != TC_H_ROOT && parent != sch->handle)
251 * ATM classes cannot be changed. In order to change properties of the
252 * ATM connection, that socket needs to be modified directly (via the
253 * native ATM API. In order to send a flow to a different VC, the old
254 * class needs to be removed and a new one added. (This may be changed
257 if (flow) return -EBUSY;
258 if (opt == NULL || rtattr_parse_nested(tb, TCA_ATM_MAX, opt))
260 if (!tb[TCA_ATM_FD-1] || RTA_PAYLOAD(tb[TCA_ATM_FD-1]) < sizeof(fd))
262 fd = *(int *) RTA_DATA(tb[TCA_ATM_FD-1]);
263 DPRINTK("atm_tc_change: fd %d\n",fd);
264 if (tb[TCA_ATM_HDR-1]) {
265 hdr_len = RTA_PAYLOAD(tb[TCA_ATM_HDR-1]);
266 hdr = RTA_DATA(tb[TCA_ATM_HDR-1]);
269 hdr_len = RFC1483LLC_LEN;
270 hdr = NULL; /* default LLC/SNAP for IP */
272 if (!tb[TCA_ATM_EXCESS-1]) excess = NULL;
274 if (RTA_PAYLOAD(tb[TCA_ATM_EXCESS-1]) != sizeof(u32))
276 excess = (struct atm_flow_data *) atm_tc_get(sch,
277 *(u32 *) RTA_DATA(tb[TCA_ATM_EXCESS-1]));
278 if (!excess) return -ENOENT;
280 DPRINTK("atm_tc_change: type %d, payload %d, hdr_len %d\n",
281 opt->rta_type,RTA_PAYLOAD(opt),hdr_len);
282 if (!(sock = sockfd_lookup(fd,&error))) return error; /* f_count++ */
283 DPRINTK("atm_tc_change: f_count %d\n",file_count(sock->file));
284 if (sock->ops->family != PF_ATMSVC && sock->ops->family != PF_ATMPVC) {
288 /* @@@ should check if the socket is really operational or we'll crash
291 if (TC_H_MAJ(classid ^ sch->handle)) {
292 DPRINTK("atm_tc_change: classid mismatch\n");
296 if (find_flow(p,flow)) {
305 for (i = 1; i < 0x8000; i++) {
306 classid = TC_H_MAKE(sch->handle,0x8000 | i);
307 if (!(cl = atm_tc_get(sch,classid))) break;
311 DPRINTK("atm_tc_change: new id %x\n",classid);
312 flow = kmalloc(sizeof(struct atm_flow_data)+hdr_len,GFP_KERNEL);
313 DPRINTK("atm_tc_change: flow %p\n",flow);
318 memset(flow,0,sizeof(*flow));
319 flow->filter_list = NULL;
320 if (!(flow->q = qdisc_create_dflt(sch->dev,&pfifo_qdisc_ops)))
321 flow->q = &noop_qdisc;
322 DPRINTK("atm_tc_change: qdisc %p\n",flow->q);
324 flow->vcc = ATM_SD(sock); /* speedup */
325 flow->vcc->user_back = flow;
326 DPRINTK("atm_tc_change: vcc %p\n",flow->vcc);
327 flow->old_pop = flow->vcc->pop;
329 flow->vcc->pop = sch_atm_pop;
330 flow->classid = classid;
332 flow->excess = excess;
333 flow->next = p->link.next;
335 flow->hdr_len = hdr_len;
337 memcpy(flow->hdr,hdr,hdr_len);
339 memcpy(flow->hdr,llc_oui_ip,sizeof(llc_oui_ip));
340 *arg = (unsigned long) flow;
343 if (excess) atm_tc_put(sch,(unsigned long) excess);
349 static int atm_tc_delete(struct Qdisc *sch,unsigned long arg)
351 struct atm_qdisc_data *p = PRIV(sch);
352 struct atm_flow_data *flow = (struct atm_flow_data *) arg;
354 DPRINTK("atm_tc_delete(sch %p,[qdisc %p],flow %p)\n",sch,p,flow);
355 if (!find_flow(PRIV(sch),flow)) return -EINVAL;
356 if (flow->filter_list || flow == &p->link) return -EBUSY;
358 * Reference count must be 2: one for "keepalive" (set at class
359 * creation), and one for the reference held when calling delete.
362 printk(KERN_ERR "atm_tc_delete: flow->ref == %d\n",flow->ref);
365 if (flow->ref > 2) return -EBUSY; /* catch references via excess, etc.*/
371 static void atm_tc_walk(struct Qdisc *sch,struct qdisc_walker *walker)
373 struct atm_qdisc_data *p = PRIV(sch);
374 struct atm_flow_data *flow;
376 DPRINTK("atm_tc_walk(sch %p,[qdisc %p],walker %p)\n",sch,p,walker);
377 if (walker->stop) return;
378 for (flow = p->flows; flow; flow = flow->next) {
379 if (walker->count >= walker->skip)
380 if (walker->fn(sch,(unsigned long) flow,walker) < 0) {
389 static struct tcf_proto **atm_tc_find_tcf(struct Qdisc *sch,unsigned long cl)
391 struct atm_qdisc_data *p = PRIV(sch);
392 struct atm_flow_data *flow = (struct atm_flow_data *) cl;
394 DPRINTK("atm_tc_find_tcf(sch %p,[qdisc %p],flow %p)\n",sch,p,flow);
395 return flow ? &flow->filter_list : &p->link.filter_list;
399 /* --------------------------- Qdisc operations ---------------------------- */
402 static int atm_tc_enqueue(struct sk_buff *skb,struct Qdisc *sch)
404 struct atm_qdisc_data *p = PRIV(sch);
405 struct atm_flow_data *flow = NULL ; /* @@@ */
406 struct tcf_result res;
408 int ret = NET_XMIT_POLICED;
410 D2PRINTK("atm_tc_enqueue(skb %p,sch %p,[qdisc %p])\n",skb,sch,p);
411 result = TC_POLICE_OK; /* be nice to gcc */
412 if (TC_H_MAJ(skb->priority) != sch->handle ||
413 !(flow = (struct atm_flow_data *) atm_tc_get(sch,skb->priority)))
414 for (flow = p->flows; flow; flow = flow->next)
415 if (flow->filter_list) {
416 result = tc_classify(skb,flow->filter_list,
418 if (result < 0) continue;
419 flow = (struct atm_flow_data *) res.class;
420 if (!flow) flow = lookup_flow(sch,res.classid);
423 if (!flow) flow = &p->link;
426 ATM_SKB(skb)->atm_options = flow->vcc->atm_options;
427 /*@@@ looks good ... but it's not supposed to work :-)*/
428 #ifdef CONFIG_NET_CLS_POLICE
433 case TC_POLICE_RECLASSIFY:
434 if (flow->excess) flow = flow->excess;
436 ATM_SKB(skb)->atm_options |=
449 #ifdef CONFIG_NET_CLS_POLICE
450 result == TC_POLICE_SHOT ||
452 (ret = flow->q->enqueue(skb,flow->q)) != 0) {
454 if (flow) flow->qstats.drops++;
457 sch->bstats.bytes += skb->len;
458 sch->bstats.packets++;
459 flow->bstats.bytes += skb->len;
460 flow->bstats.packets++;
462 * Okay, this may seem weird. We pretend we've dropped the packet if
463 * it goes via ATM. The reason for this is that the outer qdisc
464 * expects to be able to q->dequeue the packet later on if we return
465 * success at this place. Also, sch->q.qdisc needs to reflect whether
466 * there is a packet egligible for dequeuing or not. Note that the
467 * statistics of the outer qdisc are necessarily wrong because of all
468 * this. There's currently no correct solution for this.
470 if (flow == &p->link) {
474 tasklet_schedule(&p->task);
475 return NET_XMIT_BYPASS;
480 * Dequeue packets and send them over ATM. Note that we quite deliberately
481 * avoid checking net_device's flow control here, simply because sch_atm
482 * uses its own channels, which have nothing to do with any CLIP/LANE/or
483 * non-ATM interfaces.
487 static void sch_atm_dequeue(unsigned long data)
489 struct Qdisc *sch = (struct Qdisc *) data;
490 struct atm_qdisc_data *p = PRIV(sch);
491 struct atm_flow_data *flow;
494 D2PRINTK("sch_atm_dequeue(sch %p,[qdisc %p])\n",sch,p);
495 for (flow = p->link.next; flow; flow = flow->next)
497 * If traffic is properly shaped, this won't generate nasty
498 * little bursts. Otherwise, it may ... (but that's okay)
500 while ((skb = flow->q->dequeue(flow->q))) {
501 if (!atm_may_send(flow->vcc,skb->truesize)) {
502 (void) flow->q->ops->requeue(skb,flow->q);
505 D2PRINTK("atm_tc_dequeue: sending on class %p\n",flow);
506 /* remove any LL header somebody else has attached */
507 skb_pull(skb,(char *) skb->nh.iph-(char *) skb->data);
508 if (skb_headroom(skb) < flow->hdr_len) {
511 new = skb_realloc_headroom(skb,flow->hdr_len);
516 D2PRINTK("sch_atm_dequeue: ip %p, data %p\n",
517 skb->nh.iph,skb->data);
518 ATM_SKB(skb)->vcc = flow->vcc;
519 memcpy(skb_push(skb,flow->hdr_len),flow->hdr,
521 atomic_add(skb->truesize,
522 &sk_atm(flow->vcc)->sk_wmem_alloc);
523 /* atm.atm_options are already set by atm_tc_enqueue */
524 (void) flow->vcc->send(flow->vcc,skb);
529 static struct sk_buff *atm_tc_dequeue(struct Qdisc *sch)
531 struct atm_qdisc_data *p = PRIV(sch);
534 D2PRINTK("atm_tc_dequeue(sch %p,[qdisc %p])\n",sch,p);
535 tasklet_schedule(&p->task);
536 skb = p->link.q->dequeue(p->link.q);
537 if (skb) sch->q.qlen--;
542 static int atm_tc_requeue(struct sk_buff *skb,struct Qdisc *sch)
544 struct atm_qdisc_data *p = PRIV(sch);
547 D2PRINTK("atm_tc_requeue(skb %p,sch %p,[qdisc %p])\n",skb,sch,p);
548 ret = p->link.q->ops->requeue(skb,p->link.q);
551 sch->qstats.requeues++;
554 p->link.qstats.drops++;
560 static unsigned int atm_tc_drop(struct Qdisc *sch)
562 struct atm_qdisc_data *p = PRIV(sch);
563 struct atm_flow_data *flow;
566 DPRINTK("atm_tc_drop(sch %p,[qdisc %p])\n",sch,p);
567 for (flow = p->flows; flow; flow = flow->next)
568 if (flow->q->ops->drop && (len = flow->q->ops->drop(flow->q)))
574 static int atm_tc_init(struct Qdisc *sch,struct rtattr *opt)
576 struct atm_qdisc_data *p = PRIV(sch);
578 DPRINTK("atm_tc_init(sch %p,[qdisc %p],opt %p)\n",sch,p,opt);
580 if(!(p->link.q = qdisc_create_dflt(sch->dev,&pfifo_qdisc_ops)))
581 p->link.q = &noop_qdisc;
582 DPRINTK("atm_tc_init: link (%p) qdisc %p\n",&p->link,p->link.q);
583 p->link.filter_list = NULL;
586 p->link.classid = sch->handle;
589 tasklet_init(&p->task,sch_atm_dequeue,(unsigned long) sch);
594 static void atm_tc_reset(struct Qdisc *sch)
596 struct atm_qdisc_data *p = PRIV(sch);
597 struct atm_flow_data *flow;
599 DPRINTK("atm_tc_reset(sch %p,[qdisc %p])\n",sch,p);
600 for (flow = p->flows; flow; flow = flow->next) qdisc_reset(flow->q);
605 static void atm_tc_destroy(struct Qdisc *sch)
607 struct atm_qdisc_data *p = PRIV(sch);
608 struct atm_flow_data *flow;
610 DPRINTK("atm_tc_destroy(sch %p,[qdisc %p])\n",sch,p);
612 while ((flow = p->flows)) {
613 destroy_filters(flow);
615 printk(KERN_ERR "atm_destroy: %p->ref = %d\n",flow,
617 atm_tc_put(sch,(unsigned long) flow);
618 if (p->flows == flow) {
619 printk(KERN_ERR "atm_destroy: putting flow %p didn't "
621 p->flows = flow->next; /* brute force */
625 tasklet_kill(&p->task);
629 static int atm_tc_dump_class(struct Qdisc *sch, unsigned long cl,
630 struct sk_buff *skb, struct tcmsg *tcm)
632 struct atm_qdisc_data *p = PRIV(sch);
633 struct atm_flow_data *flow = (struct atm_flow_data *) cl;
634 unsigned char *b = skb->tail;
637 DPRINTK("atm_tc_dump_class(sch %p,[qdisc %p],flow %p,skb %p,tcm %p)\n",
639 if (!find_flow(p,flow)) return -EINVAL;
640 tcm->tcm_handle = flow->classid;
641 rta = (struct rtattr *) b;
642 RTA_PUT(skb,TCA_OPTIONS,0,NULL);
643 RTA_PUT(skb,TCA_ATM_HDR,flow->hdr_len,flow->hdr);
645 struct sockaddr_atmpvc pvc;
648 pvc.sap_family = AF_ATMPVC;
649 pvc.sap_addr.itf = flow->vcc->dev ? flow->vcc->dev->number : -1;
650 pvc.sap_addr.vpi = flow->vcc->vpi;
651 pvc.sap_addr.vci = flow->vcc->vci;
652 RTA_PUT(skb,TCA_ATM_ADDR,sizeof(pvc),&pvc);
653 state = ATM_VF2VS(flow->vcc->flags);
654 RTA_PUT(skb,TCA_ATM_STATE,sizeof(state),&state);
657 RTA_PUT(skb,TCA_ATM_EXCESS,sizeof(u32),&flow->classid);
661 RTA_PUT(skb,TCA_ATM_EXCESS,sizeof(zero),&zero);
663 rta->rta_len = skb->tail-b;
667 skb_trim(skb,b-skb->data);
671 atm_tc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
674 struct atm_flow_data *flow = (struct atm_flow_data *) arg;
676 flow->qstats.qlen = flow->q->q.qlen;
678 if (gnet_stats_copy_basic(d, &flow->bstats) < 0 ||
679 gnet_stats_copy_queue(d, &flow->qstats) < 0)
685 static int atm_tc_dump(struct Qdisc *sch, struct sk_buff *skb)
690 static struct Qdisc_class_ops atm_class_ops = {
691 .graft = atm_tc_graft,
695 .change = atm_tc_change,
696 .delete = atm_tc_delete,
698 .tcf_chain = atm_tc_find_tcf,
699 .bind_tcf = atm_tc_bind_filter,
700 .unbind_tcf = atm_tc_put,
701 .dump = atm_tc_dump_class,
702 .dump_stats = atm_tc_dump_class_stats,
705 static struct Qdisc_ops atm_qdisc_ops = {
707 .cl_ops = &atm_class_ops,
709 .priv_size = sizeof(struct atm_qdisc_data),
710 .enqueue = atm_tc_enqueue,
711 .dequeue = atm_tc_dequeue,
712 .requeue = atm_tc_requeue,
715 .reset = atm_tc_reset,
716 .destroy = atm_tc_destroy,
719 .owner = THIS_MODULE,
723 static int __init atm_init(void)
725 return register_qdisc(&atm_qdisc_ops);
728 static void __exit atm_exit(void)
730 unregister_qdisc(&atm_qdisc_ops);
733 module_init(atm_init)
734 module_exit(atm_exit)
735 MODULE_LICENSE("GPL");