1 // SPDX-License-Identifier: GPL-2.0
2
3 /* net/sched/sch_etf.c Earliest TxTime First queueing discipline.
4 *
5 * Authors: Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
6 * Vinicius Costa Gomes <vinicius.gomes@intel.com>
7 */
8
9 #include <linux/module.h>
10 #include <linux/types.h>
11 #include <linux/kernel.h>
12 #include <linux/string.h>
13 #include <linux/errno.h>
14 #include <linux/errqueue.h>
15 #include <linux/rbtree.h>
16 #include <linux/skbuff.h>
17 #include <linux/posix-timers.h>
18 #include <net/netlink.h>
19 #include <net/sch_generic.h>
20 #include <net/pkt_sched.h>
21 #include <net/sock.h>
22
23 #define DEADLINE_MODE_IS_ON(x) ((x)->flags & TC_ETF_DEADLINE_MODE_ON)
24 #define OFFLOAD_IS_ON(x) ((x)->flags & TC_ETF_OFFLOAD_ON)
25 #define SKIP_SOCK_CHECK_IS_SET(x) ((x)->flags & TC_ETF_SKIP_SOCK_CHECK)
26
27 struct etf_sched_data {
28 bool offload;
29 bool deadline_mode;
30 bool skip_sock_check;
31 int clockid;
32 int queue;
33 s32 delta; /* in ns */
34 ktime_t last; /* The txtime of the last skb sent to the netdevice. */
35 struct rb_root_cached head;
36 struct qdisc_watchdog watchdog;
37 ktime_t (*get_time)(void);
38 };
39
40 static const struct nla_policy etf_policy[TCA_ETF_MAX + 1] = {
41 [TCA_ETF_PARMS] = { .len = sizeof(struct tc_etf_qopt) },
42 };
43
44 static inline int validate_input_params(struct tc_etf_qopt *qopt,
45 struct netlink_ext_ack *extack)
46 {
47 /* Check if params comply to the following rules:
48 * * Clockid and delta must be valid.
49 *
50 * * Dynamic clockids are not supported.
51 *
52 * * Delta must be a positive integer.
53 *
54 * Also note that for the HW offload case, we must
55 * expect that system clocks have been synchronized to PHC.
56 */
57 if (qopt->clockid < 0) {
58 NL_SET_ERR_MSG(extack, "Dynamic clockids are not supported");
59 return -ENOTSUPP;
60 }
61
62 if (qopt->clockid != CLOCK_TAI) {
63 NL_SET_ERR_MSG(extack, "Invalid clockid. CLOCK_TAI must be used");
64 return -EINVAL;
65 }
66
67 if (qopt->delta < 0) {
68 NL_SET_ERR_MSG(extack, "Delta must be positive");
69 return -EINVAL;
70 }
71
72 return 0;
73 }
74
75 static bool is_packet_valid(struct Qdisc *sch, struct sk_buff *nskb)
76 {
77 struct etf_sched_data *q = qdisc_priv(sch);
78 ktime_t txtime = nskb->tstamp;
79 struct sock *sk = nskb->sk;
80 ktime_t now;
81
82 if (q->skip_sock_check)
83 goto skip;
84
85 if (!sk || !sk_fullsock(sk))
86 return false;
87
88 if (!sock_flag(sk, SOCK_TXTIME))
89 return false;
90
91 /* We don't perform crosstimestamping.
92 * Drop if packet's clockid differs from qdisc's.
93 */
94 if (sk->sk_clockid != q->clockid)
95 return false;
96
97 if (sk->sk_txtime_deadline_mode != q->deadline_mode)
98 return false;
99
100 skip:
101 now = q->get_time();
102 if (ktime_before(txtime, now) || ktime_before(txtime, q->last))
103 return false;
104
105 return true;
106 }
107
108 static struct sk_buff *etf_peek_timesortedlist(struct Qdisc *sch)
109 {
110 struct etf_sched_data *q = qdisc_priv(sch);
111 struct rb_node *p;
112
113 p = rb_first_cached(&q->head);
114 if (!p)
115 return NULL;
116
117 return rb_to_skb(p);
118 }
119
120 static void reset_watchdog(struct Qdisc *sch)
121 {
122 struct etf_sched_data *q = qdisc_priv(sch);
123 struct sk_buff *skb = etf_peek_timesortedlist(sch);
124 ktime_t next;
125
126 if (!skb) {
127 qdisc_watchdog_cancel(&q->watchdog);
128 return;
129 }
130
131 next = ktime_sub_ns(skb->tstamp, q->delta);
132 qdisc_watchdog_schedule_ns(&q->watchdog, ktime_to_ns(next));
133 }
134
135 static void report_sock_error(struct sk_buff *skb, u32 err, u8 code)
136 {
137 struct sock_exterr_skb *serr;
138 struct sk_buff *clone;
139 ktime_t txtime = skb->tstamp;
140 struct sock *sk = skb->sk;
141
142 if (!sk || !sk_fullsock(sk) || !(sk->sk_txtime_report_errors))
143 return;
144
145 clone = skb_clone(skb, GFP_ATOMIC);
146 if (!clone)
147 return;
148
149 serr = SKB_EXT_ERR(clone);
150 serr->ee.ee_errno = err;
151 serr->ee.ee_origin = SO_EE_ORIGIN_TXTIME;
152 serr->ee.ee_type = 0;
153 serr->ee.ee_code = code;
154 serr->ee.ee_pad = 0;
155 serr->ee.ee_data = (txtime >> 32); /* high part of tstamp */
156 serr->ee.ee_info = txtime; /* low part of tstamp */
157
158 if (sock_queue_err_skb(sk, clone))
159 kfree_skb(clone);
160 }
161
162 static int etf_enqueue_timesortedlist(struct sk_buff *nskb, struct Qdisc *sch,
163 struct sk_buff **to_free)
164 {
165 struct etf_sched_data *q = qdisc_priv(sch);
166 struct rb_node **p = &q->head.rb_root.rb_node, *parent = NULL;
167 ktime_t txtime = nskb->tstamp;
168 bool leftmost = true;
169
170 if (!is_packet_valid(sch, nskb)) {
171 report_sock_error(nskb, EINVAL,
172 SO_EE_CODE_TXTIME_INVALID_PARAM);
173 return qdisc_drop(nskb, sch, to_free);
174 }
175
176 while (*p) {
177 struct sk_buff *skb;
178
179 parent = *p;
180 skb = rb_to_skb(parent);
181 if (ktime_compare(txtime, skb->tstamp) >= 0) {
182 p = &parent->rb_right;
183 leftmost = false;
184 } else {
185 p = &parent->rb_left;
186 }
187 }
188 rb_link_node(&nskb->rbnode, parent, p);
189 rb_insert_color_cached(&nskb->rbnode, &q->head, leftmost);
190
191 qdisc_qstats_backlog_inc(sch, nskb);
192 sch->q.qlen++;
193
194 /* Now we may need to re-arm the qdisc watchdog for the next packet. */
195 reset_watchdog(sch);
196
197 return NET_XMIT_SUCCESS;
198 }
199
200 static void timesortedlist_drop(struct Qdisc *sch, struct sk_buff *skb,
201 ktime_t now)
202 {
203 struct etf_sched_data *q = qdisc_priv(sch);
204 struct sk_buff *to_free = NULL;
205 struct sk_buff *tmp = NULL;
206
207 skb_rbtree_walk_from_safe(skb, tmp) {
208 if (ktime_after(skb->tstamp, now))
209 break;
210
211 rb_erase_cached(&skb->rbnode, &q->head);
212
213 /* The rbnode field in the skb re-uses these fields, now that
214 * we are done with the rbnode, reset them.
215 */
216 skb->next = NULL;
217 skb->prev = NULL;
218 skb->dev = qdisc_dev(sch);
219
220 report_sock_error(skb, ECANCELED, SO_EE_CODE_TXTIME_MISSED);
221
222 qdisc_qstats_backlog_dec(sch, skb);
223 qdisc_drop(skb, sch, &to_free);
224 qdisc_qstats_overlimit(sch);
225 sch->q.qlen--;
226 }
227
228 kfree_skb_list(to_free);
229 }
230
231 static void timesortedlist_remove(struct Qdisc *sch, struct sk_buff *skb)
232 {
233 struct etf_sched_data *q = qdisc_priv(sch);
234
235 rb_erase_cached(&skb->rbnode, &q->head);
236
237 /* The rbnode field in the skb re-uses these fields, now that
238 * we are done with the rbnode, reset them.
239 */
240 skb->next = NULL;
241 skb->prev = NULL;
242 skb->dev = qdisc_dev(sch);
243
244 qdisc_qstats_backlog_dec(sch, skb);
245
246 qdisc_bstats_update(sch, skb);
247
248 q->last = skb->tstamp;
249
250 sch->q.qlen--;
251 }
252
253 static struct sk_buff *etf_dequeue_timesortedlist(struct Qdisc *sch)
254 {
255 struct etf_sched_data *q = qdisc_priv(sch);
256 struct sk_buff *skb;
257 ktime_t now, next;
258
259 skb = etf_peek_timesortedlist(sch);
260 if (!skb)
261 return NULL;
262
263 now = q->get_time();
264
265 /* Drop if packet has expired while in queue. */
266 if (ktime_before(skb->tstamp, now)) {
267 timesortedlist_drop(sch, skb, now);
268 skb = NULL;
269 goto out;
270 }
271
272 /* When in deadline mode, dequeue as soon as possible and change the
273 * txtime from deadline to (now + delta).
274 */
275 if (q->deadline_mode) {
276 timesortedlist_remove(sch, skb);
277 skb->tstamp = now;
278 goto out;
279 }
280
281 next = ktime_sub_ns(skb->tstamp, q->delta);
282
283 /* Dequeue only if now is within the [txtime - delta, txtime] range. */
284 if (ktime_after(now, next))
285 timesortedlist_remove(sch, skb);
286 else
287 skb = NULL;
288
289 out:
290 /* Now we may need to re-arm the qdisc watchdog for the next packet. */
291 reset_watchdog(sch);
292
293 return skb;
294 }
295
296 static void etf_disable_offload(struct net_device *dev,
297 struct etf_sched_data *q)
298 {
299 struct tc_etf_qopt_offload etf = { };
300 const struct net_device_ops *ops;
301 int err;
302
303 if (!q->offload)
304 return;
305
306 ops = dev->netdev_ops;
307 if (!ops->ndo_setup_tc)
308 return;
309
310 etf.queue = q->queue;
311 etf.enable = 0;
312
313 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETF, &etf);
314 if (err < 0)
315 pr_warn("Couldn't disable ETF offload for queue %d\n",
316 etf.queue);
317 }
318
319 static int etf_enable_offload(struct net_device *dev, struct etf_sched_data *q,
320 struct netlink_ext_ack *extack)
321 {
322 const struct net_device_ops *ops = dev->netdev_ops;
323 struct tc_etf_qopt_offload etf = { };
324 int err;
325
326 if (!ops->ndo_setup_tc) {
327 NL_SET_ERR_MSG(extack, "Specified device does not support ETF offload");
328 return -EOPNOTSUPP;
329 }
330
331 etf.queue = q->queue;
332 etf.enable = 1;
333
334 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETF, &etf);
335 if (err < 0) {
336 NL_SET_ERR_MSG(extack, "Specified device failed to setup ETF hardware offload");
337 return err;
338 }
339
340 return 0;
341 }
342
343 static int etf_init(struct Qdisc *sch, struct nlattr *opt,
344 struct netlink_ext_ack *extack)
345 {
346 struct etf_sched_data *q = qdisc_priv(sch);
347 struct net_device *dev = qdisc_dev(sch);
348 struct nlattr *tb[TCA_ETF_MAX + 1];
349 struct tc_etf_qopt *qopt;
350 int err;
351
352 if (!opt) {
353 NL_SET_ERR_MSG(extack,
354 "Missing ETF qdisc options which are mandatory");
355 return -EINVAL;
356 }
357
358 err = nla_parse_nested_deprecated(tb, TCA_ETF_MAX, opt, etf_policy,
359 extack);
360 if (err < 0)
361 return err;
362
363 if (!tb[TCA_ETF_PARMS]) {
364 NL_SET_ERR_MSG(extack, "Missing mandatory ETF parameters");
365 return -EINVAL;
366 }
367
368 qopt = nla_data(tb[TCA_ETF_PARMS]);
369
370 pr_debug("delta %d clockid %d offload %s deadline %s\n",
371 qopt->delta, qopt->clockid,
372 OFFLOAD_IS_ON(qopt) ? "on" : "off",
373 DEADLINE_MODE_IS_ON(qopt) ? "on" : "off");
374
375 err = validate_input_params(qopt, extack);
376 if (err < 0)
377 return err;
378
379 q->queue = sch->dev_queue - netdev_get_tx_queue(dev, 0);
380
381 if (OFFLOAD_IS_ON(qopt)) {
382 err = etf_enable_offload(dev, q, extack);
383 if (err < 0)
384 return err;
385 }
386
387 /* Everything went OK, save the parameters used. */
388 q->delta = qopt->delta;
389 q->clockid = qopt->clockid;
390 q->offload = OFFLOAD_IS_ON(qopt);
391 q->deadline_mode = DEADLINE_MODE_IS_ON(qopt);
392 q->skip_sock_check = SKIP_SOCK_CHECK_IS_SET(qopt);
393
394 switch (q->clockid) {
395 case CLOCK_REALTIME:
396 q->get_time = ktime_get_real;
397 break;
398 case CLOCK_MONOTONIC:
399 q->get_time = ktime_get;
400 break;
401 case CLOCK_BOOTTIME:
402 q->get_time = ktime_get_boottime;
403 break;
404 case CLOCK_TAI:
405 q->get_time = ktime_get_clocktai;
406 break;
407 default:
408 NL_SET_ERR_MSG(extack, "Clockid is not supported");
409 return -ENOTSUPP;
410 }
411
412 qdisc_watchdog_init_clockid(&q->watchdog, sch, q->clockid);
413
414 return 0;
415 }
416
417 static void timesortedlist_clear(struct Qdisc *sch)
418 {
419 struct etf_sched_data *q = qdisc_priv(sch);
420 struct rb_node *p = rb_first_cached(&q->head);
421
422 while (p) {
423 struct sk_buff *skb = rb_to_skb(p);
424
425 p = rb_next(p);
426
427 rb_erase_cached(&skb->rbnode, &q->head);
428 rtnl_kfree_skbs(skb, skb);
429 sch->q.qlen--;
430 }
431 }
432
433 static void etf_reset(struct Qdisc *sch)
434 {
435 struct etf_sched_data *q = qdisc_priv(sch);
436
437 /* Only cancel watchdog if it's been initialized. */
438 if (q->watchdog.qdisc == sch)
439 qdisc_watchdog_cancel(&q->watchdog);
440
441 /* No matter which mode we are on, it's safe to clear both lists. */
442 timesortedlist_clear(sch);
443 __qdisc_reset_queue(&sch->q);
444
445 q->last = 0;
446 }
447
448 static void etf_destroy(struct Qdisc *sch)
449 {
450 struct etf_sched_data *q = qdisc_priv(sch);
451 struct net_device *dev = qdisc_dev(sch);
452
453 /* Only cancel watchdog if it's been initialized. */
454 if (q->watchdog.qdisc == sch)
455 qdisc_watchdog_cancel(&q->watchdog);
456
457 etf_disable_offload(dev, q);
458 }
459
460 static int etf_dump(struct Qdisc *sch, struct sk_buff *skb)
461 {
462 struct etf_sched_data *q = qdisc_priv(sch);
463 struct tc_etf_qopt opt = { };
464 struct nlattr *nest;
465
466 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
467 if (!nest)
468 goto nla_put_failure;
469
470 opt.delta = READ_ONCE(q->delta);
471 opt.clockid = READ_ONCE(q->clockid);
472 if (READ_ONCE(q->offload))
473 opt.flags |= TC_ETF_OFFLOAD_ON;
474
475 if (READ_ONCE(q->deadline_mode))
476 opt.flags |= TC_ETF_DEADLINE_MODE_ON;
477
478 if (READ_ONCE(q->skip_sock_check))
479 opt.flags |= TC_ETF_SKIP_SOCK_CHECK;
480
481 if (nla_put(skb, TCA_ETF_PARMS, sizeof(opt), &opt))
482 goto nla_put_failure;
483
484 return nla_nest_end(skb, nest);
485
486 nla_put_failure:
487 nla_nest_cancel(skb, nest);
488 return -1;
489 }
490
491 static struct Qdisc_ops etf_qdisc_ops __read_mostly = {
492 .id = "etf",
493 .priv_size = sizeof(struct etf_sched_data),
494 .enqueue = etf_enqueue_timesortedlist,
495 .dequeue = etf_dequeue_timesortedlist,
496 .peek = etf_peek_timesortedlist,
497 .init = etf_init,
498 .reset = etf_reset,
499 .destroy = etf_destroy,
500 .dump = etf_dump,
501 .owner = THIS_MODULE,
502 };
503 MODULE_ALIAS_NET_SCH("etf");
504
505 static int __init etf_module_init(void)
506 {
507 return register_qdisc(&etf_qdisc_ops);
508 }
509
510 static void __exit etf_module_exit(void)
511 {
512 unregister_qdisc(&etf_qdisc_ops);
513 }
514 module_init(etf_module_init)
515 module_exit(etf_module_exit)
516 MODULE_LICENSE("GPL");
517 MODULE_DESCRIPTION("Earliest TxTime First (ETF) qdisc");
518
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.