Commit | Line | Data |
---|---|---|
25db26a9 VCG |
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> | |
4b15c707 | 14 | #include <linux/errqueue.h> |
25db26a9 VCG |
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) | |
88cab771 | 24 | #define OFFLOAD_IS_ON(x) ((x)->flags & TC_ETF_OFFLOAD_ON) |
25db26a9 VCG |
25 | |
26 | struct etf_sched_data { | |
88cab771 | 27 | bool offload; |
25db26a9 VCG |
28 | bool deadline_mode; |
29 | int clockid; | |
30 | int queue; | |
31 | s32 delta; /* in ns */ | |
32 | ktime_t last; /* The txtime of the last skb sent to the netdevice. */ | |
09fd4860 | 33 | struct rb_root_cached head; |
25db26a9 VCG |
34 | struct qdisc_watchdog watchdog; |
35 | ktime_t (*get_time)(void); | |
36 | }; | |
37 | ||
38 | static const struct nla_policy etf_policy[TCA_ETF_MAX + 1] = { | |
39 | [TCA_ETF_PARMS] = { .len = sizeof(struct tc_etf_qopt) }, | |
40 | }; | |
41 | ||
42 | static inline int validate_input_params(struct tc_etf_qopt *qopt, | |
43 | struct netlink_ext_ack *extack) | |
44 | { | |
45 | /* Check if params comply to the following rules: | |
46 | * * Clockid and delta must be valid. | |
47 | * | |
48 | * * Dynamic clockids are not supported. | |
49 | * | |
50 | * * Delta must be a positive integer. | |
88cab771 JSP |
51 | * |
52 | * Also note that for the HW offload case, we must | |
53 | * expect that system clocks have been synchronized to PHC. | |
25db26a9 VCG |
54 | */ |
55 | if (qopt->clockid < 0) { | |
56 | NL_SET_ERR_MSG(extack, "Dynamic clockids are not supported"); | |
57 | return -ENOTSUPP; | |
58 | } | |
59 | ||
60 | if (qopt->clockid != CLOCK_TAI) { | |
61 | NL_SET_ERR_MSG(extack, "Invalid clockid. CLOCK_TAI must be used"); | |
62 | return -EINVAL; | |
63 | } | |
64 | ||
65 | if (qopt->delta < 0) { | |
66 | NL_SET_ERR_MSG(extack, "Delta must be positive"); | |
67 | return -EINVAL; | |
68 | } | |
69 | ||
70 | return 0; | |
71 | } | |
72 | ||
73 | static bool is_packet_valid(struct Qdisc *sch, struct sk_buff *nskb) | |
74 | { | |
75 | struct etf_sched_data *q = qdisc_priv(sch); | |
76 | ktime_t txtime = nskb->tstamp; | |
77 | struct sock *sk = nskb->sk; | |
78 | ktime_t now; | |
79 | ||
80 | if (!sk) | |
81 | return false; | |
82 | ||
83 | if (!sock_flag(sk, SOCK_TXTIME)) | |
84 | return false; | |
85 | ||
86 | /* We don't perform crosstimestamping. | |
87 | * Drop if packet's clockid differs from qdisc's. | |
88 | */ | |
89 | if (sk->sk_clockid != q->clockid) | |
90 | return false; | |
91 | ||
92 | if (sk->sk_txtime_deadline_mode != q->deadline_mode) | |
93 | return false; | |
94 | ||
95 | now = q->get_time(); | |
96 | if (ktime_before(txtime, now) || ktime_before(txtime, q->last)) | |
97 | return false; | |
98 | ||
99 | return true; | |
100 | } | |
101 | ||
102 | static struct sk_buff *etf_peek_timesortedlist(struct Qdisc *sch) | |
103 | { | |
104 | struct etf_sched_data *q = qdisc_priv(sch); | |
105 | struct rb_node *p; | |
106 | ||
09fd4860 | 107 | p = rb_first_cached(&q->head); |
25db26a9 VCG |
108 | if (!p) |
109 | return NULL; | |
110 | ||
111 | return rb_to_skb(p); | |
112 | } | |
113 | ||
114 | static void reset_watchdog(struct Qdisc *sch) | |
115 | { | |
116 | struct etf_sched_data *q = qdisc_priv(sch); | |
117 | struct sk_buff *skb = etf_peek_timesortedlist(sch); | |
118 | ktime_t next; | |
119 | ||
3fcbdaee JSP |
120 | if (!skb) { |
121 | qdisc_watchdog_cancel(&q->watchdog); | |
25db26a9 | 122 | return; |
3fcbdaee | 123 | } |
25db26a9 VCG |
124 | |
125 | next = ktime_sub_ns(skb->tstamp, q->delta); | |
126 | qdisc_watchdog_schedule_ns(&q->watchdog, ktime_to_ns(next)); | |
127 | } | |
128 | ||
4b15c707 JSP |
129 | static void report_sock_error(struct sk_buff *skb, u32 err, u8 code) |
130 | { | |
131 | struct sock_exterr_skb *serr; | |
132 | struct sk_buff *clone; | |
133 | ktime_t txtime = skb->tstamp; | |
134 | ||
135 | if (!skb->sk || !(skb->sk->sk_txtime_report_errors)) | |
136 | return; | |
137 | ||
138 | clone = skb_clone(skb, GFP_ATOMIC); | |
139 | if (!clone) | |
140 | return; | |
141 | ||
142 | serr = SKB_EXT_ERR(clone); | |
143 | serr->ee.ee_errno = err; | |
144 | serr->ee.ee_origin = SO_EE_ORIGIN_TXTIME; | |
145 | serr->ee.ee_type = 0; | |
146 | serr->ee.ee_code = code; | |
147 | serr->ee.ee_pad = 0; | |
148 | serr->ee.ee_data = (txtime >> 32); /* high part of tstamp */ | |
149 | serr->ee.ee_info = txtime; /* low part of tstamp */ | |
150 | ||
151 | if (sock_queue_err_skb(skb->sk, clone)) | |
152 | kfree_skb(clone); | |
153 | } | |
154 | ||
25db26a9 VCG |
155 | static int etf_enqueue_timesortedlist(struct sk_buff *nskb, struct Qdisc *sch, |
156 | struct sk_buff **to_free) | |
157 | { | |
158 | struct etf_sched_data *q = qdisc_priv(sch); | |
09fd4860 | 159 | struct rb_node **p = &q->head.rb_root.rb_node, *parent = NULL; |
25db26a9 | 160 | ktime_t txtime = nskb->tstamp; |
09fd4860 | 161 | bool leftmost = true; |
25db26a9 | 162 | |
4b15c707 JSP |
163 | if (!is_packet_valid(sch, nskb)) { |
164 | report_sock_error(nskb, EINVAL, | |
165 | SO_EE_CODE_TXTIME_INVALID_PARAM); | |
25db26a9 | 166 | return qdisc_drop(nskb, sch, to_free); |
4b15c707 | 167 | } |
25db26a9 VCG |
168 | |
169 | while (*p) { | |
170 | struct sk_buff *skb; | |
171 | ||
172 | parent = *p; | |
173 | skb = rb_to_skb(parent); | |
09fd4860 | 174 | if (ktime_after(txtime, skb->tstamp)) { |
25db26a9 | 175 | p = &parent->rb_right; |
09fd4860 JSP |
176 | leftmost = false; |
177 | } else { | |
25db26a9 | 178 | p = &parent->rb_left; |
09fd4860 | 179 | } |
25db26a9 VCG |
180 | } |
181 | rb_link_node(&nskb->rbnode, parent, p); | |
09fd4860 | 182 | rb_insert_color_cached(&nskb->rbnode, &q->head, leftmost); |
25db26a9 VCG |
183 | |
184 | qdisc_qstats_backlog_inc(sch, nskb); | |
185 | sch->q.qlen++; | |
186 | ||
187 | /* Now we may need to re-arm the qdisc watchdog for the next packet. */ | |
188 | reset_watchdog(sch); | |
189 | ||
190 | return NET_XMIT_SUCCESS; | |
191 | } | |
192 | ||
193 | static void timesortedlist_erase(struct Qdisc *sch, struct sk_buff *skb, | |
194 | bool drop) | |
195 | { | |
196 | struct etf_sched_data *q = qdisc_priv(sch); | |
197 | ||
09fd4860 | 198 | rb_erase_cached(&skb->rbnode, &q->head); |
25db26a9 VCG |
199 | |
200 | /* The rbnode field in the skb re-uses these fields, now that | |
201 | * we are done with the rbnode, reset them. | |
202 | */ | |
203 | skb->next = NULL; | |
204 | skb->prev = NULL; | |
205 | skb->dev = qdisc_dev(sch); | |
206 | ||
207 | qdisc_qstats_backlog_dec(sch, skb); | |
208 | ||
209 | if (drop) { | |
210 | struct sk_buff *to_free = NULL; | |
211 | ||
4b15c707 JSP |
212 | report_sock_error(skb, ECANCELED, SO_EE_CODE_TXTIME_MISSED); |
213 | ||
25db26a9 VCG |
214 | qdisc_drop(skb, sch, &to_free); |
215 | kfree_skb_list(to_free); | |
216 | qdisc_qstats_overlimit(sch); | |
217 | } else { | |
218 | qdisc_bstats_update(sch, skb); | |
219 | ||
220 | q->last = skb->tstamp; | |
221 | } | |
222 | ||
223 | sch->q.qlen--; | |
224 | } | |
225 | ||
226 | static struct sk_buff *etf_dequeue_timesortedlist(struct Qdisc *sch) | |
227 | { | |
228 | struct etf_sched_data *q = qdisc_priv(sch); | |
229 | struct sk_buff *skb; | |
230 | ktime_t now, next; | |
231 | ||
232 | skb = etf_peek_timesortedlist(sch); | |
233 | if (!skb) | |
234 | return NULL; | |
235 | ||
236 | now = q->get_time(); | |
237 | ||
238 | /* Drop if packet has expired while in queue. */ | |
25db26a9 VCG |
239 | if (ktime_before(skb->tstamp, now)) { |
240 | timesortedlist_erase(sch, skb, true); | |
241 | skb = NULL; | |
242 | goto out; | |
243 | } | |
244 | ||
245 | /* When in deadline mode, dequeue as soon as possible and change the | |
246 | * txtime from deadline to (now + delta). | |
247 | */ | |
248 | if (q->deadline_mode) { | |
249 | timesortedlist_erase(sch, skb, false); | |
250 | skb->tstamp = now; | |
251 | goto out; | |
252 | } | |
253 | ||
254 | next = ktime_sub_ns(skb->tstamp, q->delta); | |
255 | ||
256 | /* Dequeue only if now is within the [txtime - delta, txtime] range. */ | |
257 | if (ktime_after(now, next)) | |
258 | timesortedlist_erase(sch, skb, false); | |
259 | else | |
260 | skb = NULL; | |
261 | ||
262 | out: | |
263 | /* Now we may need to re-arm the qdisc watchdog for the next packet. */ | |
264 | reset_watchdog(sch); | |
265 | ||
266 | return skb; | |
267 | } | |
268 | ||
88cab771 JSP |
269 | static void etf_disable_offload(struct net_device *dev, |
270 | struct etf_sched_data *q) | |
271 | { | |
272 | struct tc_etf_qopt_offload etf = { }; | |
273 | const struct net_device_ops *ops; | |
274 | int err; | |
275 | ||
276 | if (!q->offload) | |
277 | return; | |
278 | ||
279 | ops = dev->netdev_ops; | |
280 | if (!ops->ndo_setup_tc) | |
281 | return; | |
282 | ||
283 | etf.queue = q->queue; | |
284 | etf.enable = 0; | |
285 | ||
286 | err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETF, &etf); | |
287 | if (err < 0) | |
288 | pr_warn("Couldn't disable ETF offload for queue %d\n", | |
289 | etf.queue); | |
290 | } | |
291 | ||
292 | static int etf_enable_offload(struct net_device *dev, struct etf_sched_data *q, | |
293 | struct netlink_ext_ack *extack) | |
294 | { | |
295 | const struct net_device_ops *ops = dev->netdev_ops; | |
296 | struct tc_etf_qopt_offload etf = { }; | |
297 | int err; | |
298 | ||
299 | if (q->offload) | |
300 | return 0; | |
301 | ||
302 | if (!ops->ndo_setup_tc) { | |
303 | NL_SET_ERR_MSG(extack, "Specified device does not support ETF offload"); | |
304 | return -EOPNOTSUPP; | |
305 | } | |
306 | ||
307 | etf.queue = q->queue; | |
308 | etf.enable = 1; | |
309 | ||
310 | err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETF, &etf); | |
311 | if (err < 0) { | |
312 | NL_SET_ERR_MSG(extack, "Specified device failed to setup ETF hardware offload"); | |
313 | return err; | |
314 | } | |
315 | ||
316 | return 0; | |
317 | } | |
318 | ||
25db26a9 VCG |
319 | static int etf_init(struct Qdisc *sch, struct nlattr *opt, |
320 | struct netlink_ext_ack *extack) | |
321 | { | |
322 | struct etf_sched_data *q = qdisc_priv(sch); | |
323 | struct net_device *dev = qdisc_dev(sch); | |
324 | struct nlattr *tb[TCA_ETF_MAX + 1]; | |
325 | struct tc_etf_qopt *qopt; | |
326 | int err; | |
327 | ||
328 | if (!opt) { | |
329 | NL_SET_ERR_MSG(extack, | |
330 | "Missing ETF qdisc options which are mandatory"); | |
331 | return -EINVAL; | |
332 | } | |
333 | ||
334 | err = nla_parse_nested(tb, TCA_ETF_MAX, opt, etf_policy, extack); | |
335 | if (err < 0) | |
336 | return err; | |
337 | ||
338 | if (!tb[TCA_ETF_PARMS]) { | |
339 | NL_SET_ERR_MSG(extack, "Missing mandatory ETF parameters"); | |
340 | return -EINVAL; | |
341 | } | |
342 | ||
343 | qopt = nla_data(tb[TCA_ETF_PARMS]); | |
344 | ||
88cab771 | 345 | pr_debug("delta %d clockid %d offload %s deadline %s\n", |
25db26a9 | 346 | qopt->delta, qopt->clockid, |
88cab771 | 347 | OFFLOAD_IS_ON(qopt) ? "on" : "off", |
25db26a9 VCG |
348 | DEADLINE_MODE_IS_ON(qopt) ? "on" : "off"); |
349 | ||
350 | err = validate_input_params(qopt, extack); | |
351 | if (err < 0) | |
352 | return err; | |
353 | ||
354 | q->queue = sch->dev_queue - netdev_get_tx_queue(dev, 0); | |
355 | ||
88cab771 JSP |
356 | if (OFFLOAD_IS_ON(qopt)) { |
357 | err = etf_enable_offload(dev, q, extack); | |
358 | if (err < 0) | |
359 | return err; | |
360 | } | |
361 | ||
25db26a9 VCG |
362 | /* Everything went OK, save the parameters used. */ |
363 | q->delta = qopt->delta; | |
364 | q->clockid = qopt->clockid; | |
88cab771 | 365 | q->offload = OFFLOAD_IS_ON(qopt); |
25db26a9 VCG |
366 | q->deadline_mode = DEADLINE_MODE_IS_ON(qopt); |
367 | ||
368 | switch (q->clockid) { | |
369 | case CLOCK_REALTIME: | |
370 | q->get_time = ktime_get_real; | |
371 | break; | |
372 | case CLOCK_MONOTONIC: | |
373 | q->get_time = ktime_get; | |
374 | break; | |
375 | case CLOCK_BOOTTIME: | |
376 | q->get_time = ktime_get_boottime; | |
377 | break; | |
378 | case CLOCK_TAI: | |
379 | q->get_time = ktime_get_clocktai; | |
380 | break; | |
381 | default: | |
382 | NL_SET_ERR_MSG(extack, "Clockid is not supported"); | |
383 | return -ENOTSUPP; | |
384 | } | |
385 | ||
386 | qdisc_watchdog_init_clockid(&q->watchdog, sch, q->clockid); | |
387 | ||
388 | return 0; | |
389 | } | |
390 | ||
391 | static void timesortedlist_clear(struct Qdisc *sch) | |
392 | { | |
393 | struct etf_sched_data *q = qdisc_priv(sch); | |
09fd4860 | 394 | struct rb_node *p = rb_first_cached(&q->head); |
25db26a9 VCG |
395 | |
396 | while (p) { | |
397 | struct sk_buff *skb = rb_to_skb(p); | |
398 | ||
399 | p = rb_next(p); | |
400 | ||
09fd4860 | 401 | rb_erase_cached(&skb->rbnode, &q->head); |
25db26a9 VCG |
402 | rtnl_kfree_skbs(skb, skb); |
403 | sch->q.qlen--; | |
404 | } | |
405 | } | |
406 | ||
407 | static void etf_reset(struct Qdisc *sch) | |
408 | { | |
409 | struct etf_sched_data *q = qdisc_priv(sch); | |
410 | ||
411 | /* Only cancel watchdog if it's been initialized. */ | |
412 | if (q->watchdog.qdisc == sch) | |
413 | qdisc_watchdog_cancel(&q->watchdog); | |
414 | ||
415 | /* No matter which mode we are on, it's safe to clear both lists. */ | |
416 | timesortedlist_clear(sch); | |
417 | __qdisc_reset_queue(&sch->q); | |
418 | ||
419 | sch->qstats.backlog = 0; | |
420 | sch->q.qlen = 0; | |
421 | ||
422 | q->last = 0; | |
423 | } | |
424 | ||
425 | static void etf_destroy(struct Qdisc *sch) | |
426 | { | |
427 | struct etf_sched_data *q = qdisc_priv(sch); | |
88cab771 | 428 | struct net_device *dev = qdisc_dev(sch); |
25db26a9 VCG |
429 | |
430 | /* Only cancel watchdog if it's been initialized. */ | |
431 | if (q->watchdog.qdisc == sch) | |
432 | qdisc_watchdog_cancel(&q->watchdog); | |
88cab771 JSP |
433 | |
434 | etf_disable_offload(dev, q); | |
25db26a9 VCG |
435 | } |
436 | ||
437 | static int etf_dump(struct Qdisc *sch, struct sk_buff *skb) | |
438 | { | |
439 | struct etf_sched_data *q = qdisc_priv(sch); | |
440 | struct tc_etf_qopt opt = { }; | |
441 | struct nlattr *nest; | |
442 | ||
443 | nest = nla_nest_start(skb, TCA_OPTIONS); | |
444 | if (!nest) | |
445 | goto nla_put_failure; | |
446 | ||
447 | opt.delta = q->delta; | |
448 | opt.clockid = q->clockid; | |
88cab771 JSP |
449 | if (q->offload) |
450 | opt.flags |= TC_ETF_OFFLOAD_ON; | |
451 | ||
25db26a9 VCG |
452 | if (q->deadline_mode) |
453 | opt.flags |= TC_ETF_DEADLINE_MODE_ON; | |
454 | ||
455 | if (nla_put(skb, TCA_ETF_PARMS, sizeof(opt), &opt)) | |
456 | goto nla_put_failure; | |
457 | ||
458 | return nla_nest_end(skb, nest); | |
459 | ||
460 | nla_put_failure: | |
461 | nla_nest_cancel(skb, nest); | |
462 | return -1; | |
463 | } | |
464 | ||
465 | static struct Qdisc_ops etf_qdisc_ops __read_mostly = { | |
466 | .id = "etf", | |
467 | .priv_size = sizeof(struct etf_sched_data), | |
468 | .enqueue = etf_enqueue_timesortedlist, | |
469 | .dequeue = etf_dequeue_timesortedlist, | |
470 | .peek = etf_peek_timesortedlist, | |
471 | .init = etf_init, | |
472 | .reset = etf_reset, | |
473 | .destroy = etf_destroy, | |
474 | .dump = etf_dump, | |
475 | .owner = THIS_MODULE, | |
476 | }; | |
477 | ||
478 | static int __init etf_module_init(void) | |
479 | { | |
480 | return register_qdisc(&etf_qdisc_ops); | |
481 | } | |
482 | ||
483 | static void __exit etf_module_exit(void) | |
484 | { | |
485 | unregister_qdisc(&etf_qdisc_ops); | |
486 | } | |
487 | module_init(etf_module_init) | |
488 | module_exit(etf_module_exit) | |
489 | MODULE_LICENSE("GPL"); |