Loading...
Note: File does not exist in v4.17.
1// SPDX-License-Identifier: GPL-2.0
2
3/* net/sched/sch_taprio.c Time Aware Priority Scheduler
4 *
5 * Authors: Vinicius Costa Gomes <vinicius.gomes@intel.com>
6 *
7 */
8
9#include <linux/types.h>
10#include <linux/slab.h>
11#include <linux/kernel.h>
12#include <linux/string.h>
13#include <linux/list.h>
14#include <linux/errno.h>
15#include <linux/skbuff.h>
16#include <linux/math64.h>
17#include <linux/module.h>
18#include <linux/spinlock.h>
19#include <linux/rcupdate.h>
20#include <net/netlink.h>
21#include <net/pkt_sched.h>
22#include <net/pkt_cls.h>
23#include <net/sch_generic.h>
24#include <net/sock.h>
25#include <net/tcp.h>
26
27static LIST_HEAD(taprio_list);
28static DEFINE_SPINLOCK(taprio_list_lock);
29
30#define TAPRIO_ALL_GATES_OPEN -1
31
32#define TXTIME_ASSIST_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST)
33#define FULL_OFFLOAD_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)
34
35struct sched_entry {
36 struct list_head list;
37
38 /* The instant that this entry "closes" and the next one
39 * should open, the qdisc will make some effort so that no
40 * packet leaves after this time.
41 */
42 ktime_t close_time;
43 ktime_t next_txtime;
44 atomic_t budget;
45 int index;
46 u32 gate_mask;
47 u32 interval;
48 u8 command;
49};
50
51struct sched_gate_list {
52 struct rcu_head rcu;
53 struct list_head entries;
54 size_t num_entries;
55 ktime_t cycle_close_time;
56 s64 cycle_time;
57 s64 cycle_time_extension;
58 s64 base_time;
59};
60
61struct taprio_sched {
62 struct Qdisc **qdiscs;
63 struct Qdisc *root;
64 u32 flags;
65 enum tk_offsets tk_offset;
66 int clockid;
67 atomic64_t picos_per_byte; /* Using picoseconds because for 10Gbps+
68 * speeds it's sub-nanoseconds per byte
69 */
70
71 /* Protects the update side of the RCU protected current_entry */
72 spinlock_t current_entry_lock;
73 struct sched_entry __rcu *current_entry;
74 struct sched_gate_list __rcu *oper_sched;
75 struct sched_gate_list __rcu *admin_sched;
76 struct hrtimer advance_timer;
77 struct list_head taprio_list;
78 struct sk_buff *(*dequeue)(struct Qdisc *sch);
79 struct sk_buff *(*peek)(struct Qdisc *sch);
80 u32 txtime_delay;
81};
82
83struct __tc_taprio_qopt_offload {
84 refcount_t users;
85 struct tc_taprio_qopt_offload offload;
86};
87
88static ktime_t sched_base_time(const struct sched_gate_list *sched)
89{
90 if (!sched)
91 return KTIME_MAX;
92
93 return ns_to_ktime(sched->base_time);
94}
95
96static ktime_t taprio_get_time(struct taprio_sched *q)
97{
98 ktime_t mono = ktime_get();
99
100 switch (q->tk_offset) {
101 case TK_OFFS_MAX:
102 return mono;
103 default:
104 return ktime_mono_to_any(mono, q->tk_offset);
105 }
106
107 return KTIME_MAX;
108}
109
110static void taprio_free_sched_cb(struct rcu_head *head)
111{
112 struct sched_gate_list *sched = container_of(head, struct sched_gate_list, rcu);
113 struct sched_entry *entry, *n;
114
115 if (!sched)
116 return;
117
118 list_for_each_entry_safe(entry, n, &sched->entries, list) {
119 list_del(&entry->list);
120 kfree(entry);
121 }
122
123 kfree(sched);
124}
125
126static void switch_schedules(struct taprio_sched *q,
127 struct sched_gate_list **admin,
128 struct sched_gate_list **oper)
129{
130 rcu_assign_pointer(q->oper_sched, *admin);
131 rcu_assign_pointer(q->admin_sched, NULL);
132
133 if (*oper)
134 call_rcu(&(*oper)->rcu, taprio_free_sched_cb);
135
136 *oper = *admin;
137 *admin = NULL;
138}
139
140/* Get how much time has been already elapsed in the current cycle. */
141static s32 get_cycle_time_elapsed(struct sched_gate_list *sched, ktime_t time)
142{
143 ktime_t time_since_sched_start;
144 s32 time_elapsed;
145
146 time_since_sched_start = ktime_sub(time, sched->base_time);
147 div_s64_rem(time_since_sched_start, sched->cycle_time, &time_elapsed);
148
149 return time_elapsed;
150}
151
152static ktime_t get_interval_end_time(struct sched_gate_list *sched,
153 struct sched_gate_list *admin,
154 struct sched_entry *entry,
155 ktime_t intv_start)
156{
157 s32 cycle_elapsed = get_cycle_time_elapsed(sched, intv_start);
158 ktime_t intv_end, cycle_ext_end, cycle_end;
159
160 cycle_end = ktime_add_ns(intv_start, sched->cycle_time - cycle_elapsed);
161 intv_end = ktime_add_ns(intv_start, entry->interval);
162 cycle_ext_end = ktime_add(cycle_end, sched->cycle_time_extension);
163
164 if (ktime_before(intv_end, cycle_end))
165 return intv_end;
166 else if (admin && admin != sched &&
167 ktime_after(admin->base_time, cycle_end) &&
168 ktime_before(admin->base_time, cycle_ext_end))
169 return admin->base_time;
170 else
171 return cycle_end;
172}
173
174static int length_to_duration(struct taprio_sched *q, int len)
175{
176 return div_u64(len * atomic64_read(&q->picos_per_byte), 1000);
177}
178
179/* Returns the entry corresponding to next available interval. If
180 * validate_interval is set, it only validates whether the timestamp occurs
181 * when the gate corresponding to the skb's traffic class is open.
182 */
183static struct sched_entry *find_entry_to_transmit(struct sk_buff *skb,
184 struct Qdisc *sch,
185 struct sched_gate_list *sched,
186 struct sched_gate_list *admin,
187 ktime_t time,
188 ktime_t *interval_start,
189 ktime_t *interval_end,
190 bool validate_interval)
191{
192 ktime_t curr_intv_start, curr_intv_end, cycle_end, packet_transmit_time;
193 ktime_t earliest_txtime = KTIME_MAX, txtime, cycle, transmit_end_time;
194 struct sched_entry *entry = NULL, *entry_found = NULL;
195 struct taprio_sched *q = qdisc_priv(sch);
196 struct net_device *dev = qdisc_dev(sch);
197 bool entry_available = false;
198 s32 cycle_elapsed;
199 int tc, n;
200
201 tc = netdev_get_prio_tc_map(dev, skb->priority);
202 packet_transmit_time = length_to_duration(q, qdisc_pkt_len(skb));
203
204 *interval_start = 0;
205 *interval_end = 0;
206
207 if (!sched)
208 return NULL;
209
210 cycle = sched->cycle_time;
211 cycle_elapsed = get_cycle_time_elapsed(sched, time);
212 curr_intv_end = ktime_sub_ns(time, cycle_elapsed);
213 cycle_end = ktime_add_ns(curr_intv_end, cycle);
214
215 list_for_each_entry(entry, &sched->entries, list) {
216 curr_intv_start = curr_intv_end;
217 curr_intv_end = get_interval_end_time(sched, admin, entry,
218 curr_intv_start);
219
220 if (ktime_after(curr_intv_start, cycle_end))
221 break;
222
223 if (!(entry->gate_mask & BIT(tc)) ||
224 packet_transmit_time > entry->interval)
225 continue;
226
227 txtime = entry->next_txtime;
228
229 if (ktime_before(txtime, time) || validate_interval) {
230 transmit_end_time = ktime_add_ns(time, packet_transmit_time);
231 if ((ktime_before(curr_intv_start, time) &&
232 ktime_before(transmit_end_time, curr_intv_end)) ||
233 (ktime_after(curr_intv_start, time) && !validate_interval)) {
234 entry_found = entry;
235 *interval_start = curr_intv_start;
236 *interval_end = curr_intv_end;
237 break;
238 } else if (!entry_available && !validate_interval) {
239 /* Here, we are just trying to find out the
240 * first available interval in the next cycle.
241 */
242 entry_available = 1;
243 entry_found = entry;
244 *interval_start = ktime_add_ns(curr_intv_start, cycle);
245 *interval_end = ktime_add_ns(curr_intv_end, cycle);
246 }
247 } else if (ktime_before(txtime, earliest_txtime) &&
248 !entry_available) {
249 earliest_txtime = txtime;
250 entry_found = entry;
251 n = div_s64(ktime_sub(txtime, curr_intv_start), cycle);
252 *interval_start = ktime_add(curr_intv_start, n * cycle);
253 *interval_end = ktime_add(curr_intv_end, n * cycle);
254 }
255 }
256
257 return entry_found;
258}
259
260static bool is_valid_interval(struct sk_buff *skb, struct Qdisc *sch)
261{
262 struct taprio_sched *q = qdisc_priv(sch);
263 struct sched_gate_list *sched, *admin;
264 ktime_t interval_start, interval_end;
265 struct sched_entry *entry;
266
267 rcu_read_lock();
268 sched = rcu_dereference(q->oper_sched);
269 admin = rcu_dereference(q->admin_sched);
270
271 entry = find_entry_to_transmit(skb, sch, sched, admin, skb->tstamp,
272 &interval_start, &interval_end, true);
273 rcu_read_unlock();
274
275 return entry;
276}
277
278static bool taprio_flags_valid(u32 flags)
279{
280 /* Make sure no other flag bits are set. */
281 if (flags & ~(TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST |
282 TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
283 return false;
284 /* txtime-assist and full offload are mutually exclusive */
285 if ((flags & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST) &&
286 (flags & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
287 return false;
288 return true;
289}
290
291/* This returns the tstamp value set by TCP in terms of the set clock. */
292static ktime_t get_tcp_tstamp(struct taprio_sched *q, struct sk_buff *skb)
293{
294 unsigned int offset = skb_network_offset(skb);
295 const struct ipv6hdr *ipv6h;
296 const struct iphdr *iph;
297 struct ipv6hdr _ipv6h;
298
299 ipv6h = skb_header_pointer(skb, offset, sizeof(_ipv6h), &_ipv6h);
300 if (!ipv6h)
301 return 0;
302
303 if (ipv6h->version == 4) {
304 iph = (struct iphdr *)ipv6h;
305 offset += iph->ihl * 4;
306
307 /* special-case 6in4 tunnelling, as that is a common way to get
308 * v6 connectivity in the home
309 */
310 if (iph->protocol == IPPROTO_IPV6) {
311 ipv6h = skb_header_pointer(skb, offset,
312 sizeof(_ipv6h), &_ipv6h);
313
314 if (!ipv6h || ipv6h->nexthdr != IPPROTO_TCP)
315 return 0;
316 } else if (iph->protocol != IPPROTO_TCP) {
317 return 0;
318 }
319 } else if (ipv6h->version == 6 && ipv6h->nexthdr != IPPROTO_TCP) {
320 return 0;
321 }
322
323 return ktime_mono_to_any(skb->skb_mstamp_ns, q->tk_offset);
324}
325
326/* There are a few scenarios where we will have to modify the txtime from
327 * what is read from next_txtime in sched_entry. They are:
328 * 1. If txtime is in the past,
329 * a. The gate for the traffic class is currently open and packet can be
330 * transmitted before it closes, schedule the packet right away.
331 * b. If the gate corresponding to the traffic class is going to open later
332 * in the cycle, set the txtime of packet to the interval start.
333 * 2. If txtime is in the future, there are packets corresponding to the
334 * current traffic class waiting to be transmitted. So, the following
335 * possibilities exist:
336 * a. We can transmit the packet before the window containing the txtime
337 * closes.
338 * b. The window might close before the transmission can be completed
339 * successfully. So, schedule the packet in the next open window.
340 */
341static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)
342{
343 ktime_t transmit_end_time, interval_end, interval_start, tcp_tstamp;
344 struct taprio_sched *q = qdisc_priv(sch);
345 struct sched_gate_list *sched, *admin;
346 ktime_t minimum_time, now, txtime;
347 int len, packet_transmit_time;
348 struct sched_entry *entry;
349 bool sched_changed;
350
351 now = taprio_get_time(q);
352 minimum_time = ktime_add_ns(now, q->txtime_delay);
353
354 tcp_tstamp = get_tcp_tstamp(q, skb);
355 minimum_time = max_t(ktime_t, minimum_time, tcp_tstamp);
356
357 rcu_read_lock();
358 admin = rcu_dereference(q->admin_sched);
359 sched = rcu_dereference(q->oper_sched);
360 if (admin && ktime_after(minimum_time, admin->base_time))
361 switch_schedules(q, &admin, &sched);
362
363 /* Until the schedule starts, all the queues are open */
364 if (!sched || ktime_before(minimum_time, sched->base_time)) {
365 txtime = minimum_time;
366 goto done;
367 }
368
369 len = qdisc_pkt_len(skb);
370 packet_transmit_time = length_to_duration(q, len);
371
372 do {
373 sched_changed = 0;
374
375 entry = find_entry_to_transmit(skb, sch, sched, admin,
376 minimum_time,
377 &interval_start, &interval_end,
378 false);
379 if (!entry) {
380 txtime = 0;
381 goto done;
382 }
383
384 txtime = entry->next_txtime;
385 txtime = max_t(ktime_t, txtime, minimum_time);
386 txtime = max_t(ktime_t, txtime, interval_start);
387
388 if (admin && admin != sched &&
389 ktime_after(txtime, admin->base_time)) {
390 sched = admin;
391 sched_changed = 1;
392 continue;
393 }
394
395 transmit_end_time = ktime_add(txtime, packet_transmit_time);
396 minimum_time = transmit_end_time;
397
398 /* Update the txtime of current entry to the next time it's
399 * interval starts.
400 */
401 if (ktime_after(transmit_end_time, interval_end))
402 entry->next_txtime = ktime_add(interval_start, sched->cycle_time);
403 } while (sched_changed || ktime_after(transmit_end_time, interval_end));
404
405 entry->next_txtime = transmit_end_time;
406
407done:
408 rcu_read_unlock();
409 return txtime;
410}
411
412static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,
413 struct sk_buff **to_free)
414{
415 struct taprio_sched *q = qdisc_priv(sch);
416 struct Qdisc *child;
417 int queue;
418
419 queue = skb_get_queue_mapping(skb);
420
421 child = q->qdiscs[queue];
422 if (unlikely(!child))
423 return qdisc_drop(skb, sch, to_free);
424
425 if (skb->sk && sock_flag(skb->sk, SOCK_TXTIME)) {
426 if (!is_valid_interval(skb, sch))
427 return qdisc_drop(skb, sch, to_free);
428 } else if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
429 skb->tstamp = get_packet_txtime(skb, sch);
430 if (!skb->tstamp)
431 return qdisc_drop(skb, sch, to_free);
432 }
433
434 qdisc_qstats_backlog_inc(sch, skb);
435 sch->q.qlen++;
436
437 return qdisc_enqueue(skb, child, to_free);
438}
439
440static struct sk_buff *taprio_peek_soft(struct Qdisc *sch)
441{
442 struct taprio_sched *q = qdisc_priv(sch);
443 struct net_device *dev = qdisc_dev(sch);
444 struct sched_entry *entry;
445 struct sk_buff *skb;
446 u32 gate_mask;
447 int i;
448
449 rcu_read_lock();
450 entry = rcu_dereference(q->current_entry);
451 gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
452 rcu_read_unlock();
453
454 if (!gate_mask)
455 return NULL;
456
457 for (i = 0; i < dev->num_tx_queues; i++) {
458 struct Qdisc *child = q->qdiscs[i];
459 int prio;
460 u8 tc;
461
462 if (unlikely(!child))
463 continue;
464
465 skb = child->ops->peek(child);
466 if (!skb)
467 continue;
468
469 if (TXTIME_ASSIST_IS_ENABLED(q->flags))
470 return skb;
471
472 prio = skb->priority;
473 tc = netdev_get_prio_tc_map(dev, prio);
474
475 if (!(gate_mask & BIT(tc)))
476 continue;
477
478 return skb;
479 }
480
481 return NULL;
482}
483
484static struct sk_buff *taprio_peek_offload(struct Qdisc *sch)
485{
486 struct taprio_sched *q = qdisc_priv(sch);
487 struct net_device *dev = qdisc_dev(sch);
488 struct sk_buff *skb;
489 int i;
490
491 for (i = 0; i < dev->num_tx_queues; i++) {
492 struct Qdisc *child = q->qdiscs[i];
493
494 if (unlikely(!child))
495 continue;
496
497 skb = child->ops->peek(child);
498 if (!skb)
499 continue;
500
501 return skb;
502 }
503
504 return NULL;
505}
506
507static struct sk_buff *taprio_peek(struct Qdisc *sch)
508{
509 struct taprio_sched *q = qdisc_priv(sch);
510
511 return q->peek(sch);
512}
513
514static void taprio_set_budget(struct taprio_sched *q, struct sched_entry *entry)
515{
516 atomic_set(&entry->budget,
517 div64_u64((u64)entry->interval * 1000,
518 atomic64_read(&q->picos_per_byte)));
519}
520
521static struct sk_buff *taprio_dequeue_soft(struct Qdisc *sch)
522{
523 struct taprio_sched *q = qdisc_priv(sch);
524 struct net_device *dev = qdisc_dev(sch);
525 struct sk_buff *skb = NULL;
526 struct sched_entry *entry;
527 u32 gate_mask;
528 int i;
529
530 rcu_read_lock();
531 entry = rcu_dereference(q->current_entry);
532 /* if there's no entry, it means that the schedule didn't
533 * start yet, so force all gates to be open, this is in
534 * accordance to IEEE 802.1Qbv-2015 Section 8.6.9.4.5
535 * "AdminGateSates"
536 */
537 gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
538
539 if (!gate_mask)
540 goto done;
541
542 for (i = 0; i < dev->num_tx_queues; i++) {
543 struct Qdisc *child = q->qdiscs[i];
544 ktime_t guard;
545 int prio;
546 int len;
547 u8 tc;
548
549 if (unlikely(!child))
550 continue;
551
552 if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
553 skb = child->ops->dequeue(child);
554 if (!skb)
555 continue;
556 goto skb_found;
557 }
558
559 skb = child->ops->peek(child);
560 if (!skb)
561 continue;
562
563 prio = skb->priority;
564 tc = netdev_get_prio_tc_map(dev, prio);
565
566 if (!(gate_mask & BIT(tc)))
567 continue;
568
569 len = qdisc_pkt_len(skb);
570 guard = ktime_add_ns(taprio_get_time(q),
571 length_to_duration(q, len));
572
573 /* In the case that there's no gate entry, there's no
574 * guard band ...
575 */
576 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
577 ktime_after(guard, entry->close_time))
578 continue;
579
580 /* ... and no budget. */
581 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
582 atomic_sub_return(len, &entry->budget) < 0)
583 continue;
584
585 skb = child->ops->dequeue(child);
586 if (unlikely(!skb))
587 goto done;
588
589skb_found:
590 qdisc_bstats_update(sch, skb);
591 qdisc_qstats_backlog_dec(sch, skb);
592 sch->q.qlen--;
593
594 goto done;
595 }
596
597done:
598 rcu_read_unlock();
599
600 return skb;
601}
602
603static struct sk_buff *taprio_dequeue_offload(struct Qdisc *sch)
604{
605 struct taprio_sched *q = qdisc_priv(sch);
606 struct net_device *dev = qdisc_dev(sch);
607 struct sk_buff *skb;
608 int i;
609
610 for (i = 0; i < dev->num_tx_queues; i++) {
611 struct Qdisc *child = q->qdiscs[i];
612
613 if (unlikely(!child))
614 continue;
615
616 skb = child->ops->dequeue(child);
617 if (unlikely(!skb))
618 continue;
619
620 qdisc_bstats_update(sch, skb);
621 qdisc_qstats_backlog_dec(sch, skb);
622 sch->q.qlen--;
623
624 return skb;
625 }
626
627 return NULL;
628}
629
630static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
631{
632 struct taprio_sched *q = qdisc_priv(sch);
633
634 return q->dequeue(sch);
635}
636
637static bool should_restart_cycle(const struct sched_gate_list *oper,
638 const struct sched_entry *entry)
639{
640 if (list_is_last(&entry->list, &oper->entries))
641 return true;
642
643 if (ktime_compare(entry->close_time, oper->cycle_close_time) == 0)
644 return true;
645
646 return false;
647}
648
649static bool should_change_schedules(const struct sched_gate_list *admin,
650 const struct sched_gate_list *oper,
651 ktime_t close_time)
652{
653 ktime_t next_base_time, extension_time;
654
655 if (!admin)
656 return false;
657
658 next_base_time = sched_base_time(admin);
659
660 /* This is the simple case, the close_time would fall after
661 * the next schedule base_time.
662 */
663 if (ktime_compare(next_base_time, close_time) <= 0)
664 return true;
665
666 /* This is the cycle_time_extension case, if the close_time
667 * plus the amount that can be extended would fall after the
668 * next schedule base_time, we can extend the current schedule
669 * for that amount.
670 */
671 extension_time = ktime_add_ns(close_time, oper->cycle_time_extension);
672
673 /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
674 * how precisely the extension should be made. So after
675 * conformance testing, this logic may change.
676 */
677 if (ktime_compare(next_base_time, extension_time) <= 0)
678 return true;
679
680 return false;
681}
682
683static enum hrtimer_restart advance_sched(struct hrtimer *timer)
684{
685 struct taprio_sched *q = container_of(timer, struct taprio_sched,
686 advance_timer);
687 struct sched_gate_list *oper, *admin;
688 struct sched_entry *entry, *next;
689 struct Qdisc *sch = q->root;
690 ktime_t close_time;
691
692 spin_lock(&q->current_entry_lock);
693 entry = rcu_dereference_protected(q->current_entry,
694 lockdep_is_held(&q->current_entry_lock));
695 oper = rcu_dereference_protected(q->oper_sched,
696 lockdep_is_held(&q->current_entry_lock));
697 admin = rcu_dereference_protected(q->admin_sched,
698 lockdep_is_held(&q->current_entry_lock));
699
700 if (!oper)
701 switch_schedules(q, &admin, &oper);
702
703 /* This can happen in two cases: 1. this is the very first run
704 * of this function (i.e. we weren't running any schedule
705 * previously); 2. The previous schedule just ended. The first
706 * entry of all schedules are pre-calculated during the
707 * schedule initialization.
708 */
709 if (unlikely(!entry || entry->close_time == oper->base_time)) {
710 next = list_first_entry(&oper->entries, struct sched_entry,
711 list);
712 close_time = next->close_time;
713 goto first_run;
714 }
715
716 if (should_restart_cycle(oper, entry)) {
717 next = list_first_entry(&oper->entries, struct sched_entry,
718 list);
719 oper->cycle_close_time = ktime_add_ns(oper->cycle_close_time,
720 oper->cycle_time);
721 } else {
722 next = list_next_entry(entry, list);
723 }
724
725 close_time = ktime_add_ns(entry->close_time, next->interval);
726 close_time = min_t(ktime_t, close_time, oper->cycle_close_time);
727
728 if (should_change_schedules(admin, oper, close_time)) {
729 /* Set things so the next time this runs, the new
730 * schedule runs.
731 */
732 close_time = sched_base_time(admin);
733 switch_schedules(q, &admin, &oper);
734 }
735
736 next->close_time = close_time;
737 taprio_set_budget(q, next);
738
739first_run:
740 rcu_assign_pointer(q->current_entry, next);
741 spin_unlock(&q->current_entry_lock);
742
743 hrtimer_set_expires(&q->advance_timer, close_time);
744
745 rcu_read_lock();
746 __netif_schedule(sch);
747 rcu_read_unlock();
748
749 return HRTIMER_RESTART;
750}
751
752static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
753 [TCA_TAPRIO_SCHED_ENTRY_INDEX] = { .type = NLA_U32 },
754 [TCA_TAPRIO_SCHED_ENTRY_CMD] = { .type = NLA_U8 },
755 [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
756 [TCA_TAPRIO_SCHED_ENTRY_INTERVAL] = { .type = NLA_U32 },
757};
758
759static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
760 [TCA_TAPRIO_ATTR_PRIOMAP] = {
761 .len = sizeof(struct tc_mqprio_qopt)
762 },
763 [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST] = { .type = NLA_NESTED },
764 [TCA_TAPRIO_ATTR_SCHED_BASE_TIME] = { .type = NLA_S64 },
765 [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY] = { .type = NLA_NESTED },
766 [TCA_TAPRIO_ATTR_SCHED_CLOCKID] = { .type = NLA_S32 },
767 [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME] = { .type = NLA_S64 },
768 [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
769};
770
771static int fill_sched_entry(struct nlattr **tb, struct sched_entry *entry,
772 struct netlink_ext_ack *extack)
773{
774 u32 interval = 0;
775
776 if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
777 entry->command = nla_get_u8(
778 tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
779
780 if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
781 entry->gate_mask = nla_get_u32(
782 tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
783
784 if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
785 interval = nla_get_u32(
786 tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
787
788 if (interval == 0) {
789 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
790 return -EINVAL;
791 }
792
793 entry->interval = interval;
794
795 return 0;
796}
797
798static int parse_sched_entry(struct nlattr *n, struct sched_entry *entry,
799 int index, struct netlink_ext_ack *extack)
800{
801 struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
802 int err;
803
804 err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
805 entry_policy, NULL);
806 if (err < 0) {
807 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
808 return -EINVAL;
809 }
810
811 entry->index = index;
812
813 return fill_sched_entry(tb, entry, extack);
814}
815
816static int parse_sched_list(struct nlattr *list,
817 struct sched_gate_list *sched,
818 struct netlink_ext_ack *extack)
819{
820 struct nlattr *n;
821 int err, rem;
822 int i = 0;
823
824 if (!list)
825 return -EINVAL;
826
827 nla_for_each_nested(n, list, rem) {
828 struct sched_entry *entry;
829
830 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
831 NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
832 continue;
833 }
834
835 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
836 if (!entry) {
837 NL_SET_ERR_MSG(extack, "Not enough memory for entry");
838 return -ENOMEM;
839 }
840
841 err = parse_sched_entry(n, entry, i, extack);
842 if (err < 0) {
843 kfree(entry);
844 return err;
845 }
846
847 list_add_tail(&entry->list, &sched->entries);
848 i++;
849 }
850
851 sched->num_entries = i;
852
853 return i;
854}
855
856static int parse_taprio_schedule(struct nlattr **tb,
857 struct sched_gate_list *new,
858 struct netlink_ext_ack *extack)
859{
860 int err = 0;
861
862 if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
863 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
864 return -ENOTSUPP;
865 }
866
867 if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
868 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
869
870 if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
871 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
872
873 if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
874 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
875
876 if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
877 err = parse_sched_list(
878 tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST], new, extack);
879 if (err < 0)
880 return err;
881
882 if (!new->cycle_time) {
883 struct sched_entry *entry;
884 ktime_t cycle = 0;
885
886 list_for_each_entry(entry, &new->entries, list)
887 cycle = ktime_add_ns(cycle, entry->interval);
888 new->cycle_time = cycle;
889 }
890
891 return 0;
892}
893
894static int taprio_parse_mqprio_opt(struct net_device *dev,
895 struct tc_mqprio_qopt *qopt,
896 struct netlink_ext_ack *extack,
897 u32 taprio_flags)
898{
899 int i, j;
900
901 if (!qopt && !dev->num_tc) {
902 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
903 return -EINVAL;
904 }
905
906 /* If num_tc is already set, it means that the user already
907 * configured the mqprio part
908 */
909 if (dev->num_tc)
910 return 0;
911
912 /* Verify num_tc is not out of max range */
913 if (qopt->num_tc > TC_MAX_QUEUE) {
914 NL_SET_ERR_MSG(extack, "Number of traffic classes is outside valid range");
915 return -EINVAL;
916 }
917
918 /* taprio imposes that traffic classes map 1:n to tx queues */
919 if (qopt->num_tc > dev->num_tx_queues) {
920 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
921 return -EINVAL;
922 }
923
924 /* Verify priority mapping uses valid tcs */
925 for (i = 0; i <= TC_BITMASK; i++) {
926 if (qopt->prio_tc_map[i] >= qopt->num_tc) {
927 NL_SET_ERR_MSG(extack, "Invalid traffic class in priority to traffic class mapping");
928 return -EINVAL;
929 }
930 }
931
932 for (i = 0; i < qopt->num_tc; i++) {
933 unsigned int last = qopt->offset[i] + qopt->count[i];
934
935 /* Verify the queue count is in tx range being equal to the
936 * real_num_tx_queues indicates the last queue is in use.
937 */
938 if (qopt->offset[i] >= dev->num_tx_queues ||
939 !qopt->count[i] ||
940 last > dev->real_num_tx_queues) {
941 NL_SET_ERR_MSG(extack, "Invalid queue in traffic class to queue mapping");
942 return -EINVAL;
943 }
944
945 if (TXTIME_ASSIST_IS_ENABLED(taprio_flags))
946 continue;
947
948 /* Verify that the offset and counts do not overlap */
949 for (j = i + 1; j < qopt->num_tc; j++) {
950 if (last > qopt->offset[j]) {
951 NL_SET_ERR_MSG(extack, "Detected overlap in the traffic class to queue mapping");
952 return -EINVAL;
953 }
954 }
955 }
956
957 return 0;
958}
959
960static int taprio_get_start_time(struct Qdisc *sch,
961 struct sched_gate_list *sched,
962 ktime_t *start)
963{
964 struct taprio_sched *q = qdisc_priv(sch);
965 ktime_t now, base, cycle;
966 s64 n;
967
968 base = sched_base_time(sched);
969 now = taprio_get_time(q);
970
971 if (ktime_after(base, now)) {
972 *start = base;
973 return 0;
974 }
975
976 cycle = sched->cycle_time;
977
978 /* The qdisc is expected to have at least one sched_entry. Moreover,
979 * any entry must have 'interval' > 0. Thus if the cycle time is zero,
980 * something went really wrong. In that case, we should warn about this
981 * inconsistent state and return error.
982 */
983 if (WARN_ON(!cycle))
984 return -EFAULT;
985
986 /* Schedule the start time for the beginning of the next
987 * cycle.
988 */
989 n = div64_s64(ktime_sub_ns(now, base), cycle);
990 *start = ktime_add_ns(base, (n + 1) * cycle);
991 return 0;
992}
993
994static void setup_first_close_time(struct taprio_sched *q,
995 struct sched_gate_list *sched, ktime_t base)
996{
997 struct sched_entry *first;
998 ktime_t cycle;
999
1000 first = list_first_entry(&sched->entries,
1001 struct sched_entry, list);
1002
1003 cycle = sched->cycle_time;
1004
1005 /* FIXME: find a better place to do this */
1006 sched->cycle_close_time = ktime_add_ns(base, cycle);
1007
1008 first->close_time = ktime_add_ns(base, first->interval);
1009 taprio_set_budget(q, first);
1010 rcu_assign_pointer(q->current_entry, NULL);
1011}
1012
1013static void taprio_start_sched(struct Qdisc *sch,
1014 ktime_t start, struct sched_gate_list *new)
1015{
1016 struct taprio_sched *q = qdisc_priv(sch);
1017 ktime_t expires;
1018
1019 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1020 return;
1021
1022 expires = hrtimer_get_expires(&q->advance_timer);
1023 if (expires == 0)
1024 expires = KTIME_MAX;
1025
1026 /* If the new schedule starts before the next expiration, we
1027 * reprogram it to the earliest one, so we change the admin
1028 * schedule to the operational one at the right time.
1029 */
1030 start = min_t(ktime_t, start, expires);
1031
1032 hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1033}
1034
1035static void taprio_set_picos_per_byte(struct net_device *dev,
1036 struct taprio_sched *q)
1037{
1038 struct ethtool_link_ksettings ecmd;
1039 int speed = SPEED_10;
1040 int picos_per_byte;
1041 int err;
1042
1043 err = __ethtool_get_link_ksettings(dev, &ecmd);
1044 if (err < 0)
1045 goto skip;
1046
1047 if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1048 speed = ecmd.base.speed;
1049
1050skip:
1051 picos_per_byte = (USEC_PER_SEC * 8) / speed;
1052
1053 atomic64_set(&q->picos_per_byte, picos_per_byte);
1054 netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1055 dev->name, (long long)atomic64_read(&q->picos_per_byte),
1056 ecmd.base.speed);
1057}
1058
1059static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1060 void *ptr)
1061{
1062 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1063 struct net_device *qdev;
1064 struct taprio_sched *q;
1065 bool found = false;
1066
1067 ASSERT_RTNL();
1068
1069 if (event != NETDEV_UP && event != NETDEV_CHANGE)
1070 return NOTIFY_DONE;
1071
1072 spin_lock(&taprio_list_lock);
1073 list_for_each_entry(q, &taprio_list, taprio_list) {
1074 qdev = qdisc_dev(q->root);
1075 if (qdev == dev) {
1076 found = true;
1077 break;
1078 }
1079 }
1080 spin_unlock(&taprio_list_lock);
1081
1082 if (found)
1083 taprio_set_picos_per_byte(dev, q);
1084
1085 return NOTIFY_DONE;
1086}
1087
1088static void setup_txtime(struct taprio_sched *q,
1089 struct sched_gate_list *sched, ktime_t base)
1090{
1091 struct sched_entry *entry;
1092 u32 interval = 0;
1093
1094 list_for_each_entry(entry, &sched->entries, list) {
1095 entry->next_txtime = ktime_add_ns(base, interval);
1096 interval += entry->interval;
1097 }
1098}
1099
1100static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1101{
1102 size_t size = sizeof(struct tc_taprio_sched_entry) * num_entries +
1103 sizeof(struct __tc_taprio_qopt_offload);
1104 struct __tc_taprio_qopt_offload *__offload;
1105
1106 __offload = kzalloc(size, GFP_KERNEL);
1107 if (!__offload)
1108 return NULL;
1109
1110 refcount_set(&__offload->users, 1);
1111
1112 return &__offload->offload;
1113}
1114
1115struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1116 *offload)
1117{
1118 struct __tc_taprio_qopt_offload *__offload;
1119
1120 __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1121 offload);
1122
1123 refcount_inc(&__offload->users);
1124
1125 return offload;
1126}
1127EXPORT_SYMBOL_GPL(taprio_offload_get);
1128
1129void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1130{
1131 struct __tc_taprio_qopt_offload *__offload;
1132
1133 __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1134 offload);
1135
1136 if (!refcount_dec_and_test(&__offload->users))
1137 return;
1138
1139 kfree(__offload);
1140}
1141EXPORT_SYMBOL_GPL(taprio_offload_free);
1142
1143/* The function will only serve to keep the pointers to the "oper" and "admin"
1144 * schedules valid in relation to their base times, so when calling dump() the
1145 * users looks at the right schedules.
1146 * When using full offload, the admin configuration is promoted to oper at the
1147 * base_time in the PHC time domain. But because the system time is not
1148 * necessarily in sync with that, we can't just trigger a hrtimer to call
1149 * switch_schedules at the right hardware time.
1150 * At the moment we call this by hand right away from taprio, but in the future
1151 * it will be useful to create a mechanism for drivers to notify taprio of the
1152 * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1153 * This is left as TODO.
1154 */
1155static void taprio_offload_config_changed(struct taprio_sched *q)
1156{
1157 struct sched_gate_list *oper, *admin;
1158
1159 spin_lock(&q->current_entry_lock);
1160
1161 oper = rcu_dereference_protected(q->oper_sched,
1162 lockdep_is_held(&q->current_entry_lock));
1163 admin = rcu_dereference_protected(q->admin_sched,
1164 lockdep_is_held(&q->current_entry_lock));
1165
1166 switch_schedules(q, &admin, &oper);
1167
1168 spin_unlock(&q->current_entry_lock);
1169}
1170
1171static void taprio_sched_to_offload(struct taprio_sched *q,
1172 struct sched_gate_list *sched,
1173 const struct tc_mqprio_qopt *mqprio,
1174 struct tc_taprio_qopt_offload *offload)
1175{
1176 struct sched_entry *entry;
1177 int i = 0;
1178
1179 offload->base_time = sched->base_time;
1180 offload->cycle_time = sched->cycle_time;
1181 offload->cycle_time_extension = sched->cycle_time_extension;
1182
1183 list_for_each_entry(entry, &sched->entries, list) {
1184 struct tc_taprio_sched_entry *e = &offload->entries[i];
1185
1186 e->command = entry->command;
1187 e->interval = entry->interval;
1188 e->gate_mask = entry->gate_mask;
1189 i++;
1190 }
1191
1192 offload->num_entries = i;
1193}
1194
1195static int taprio_enable_offload(struct net_device *dev,
1196 struct tc_mqprio_qopt *mqprio,
1197 struct taprio_sched *q,
1198 struct sched_gate_list *sched,
1199 struct netlink_ext_ack *extack)
1200{
1201 const struct net_device_ops *ops = dev->netdev_ops;
1202 struct tc_taprio_qopt_offload *offload;
1203 int err = 0;
1204
1205 if (!ops->ndo_setup_tc) {
1206 NL_SET_ERR_MSG(extack,
1207 "Device does not support taprio offload");
1208 return -EOPNOTSUPP;
1209 }
1210
1211 offload = taprio_offload_alloc(sched->num_entries);
1212 if (!offload) {
1213 NL_SET_ERR_MSG(extack,
1214 "Not enough memory for enabling offload mode");
1215 return -ENOMEM;
1216 }
1217 offload->enable = 1;
1218 taprio_sched_to_offload(q, sched, mqprio, offload);
1219
1220 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1221 if (err < 0) {
1222 NL_SET_ERR_MSG(extack,
1223 "Device failed to setup taprio offload");
1224 goto done;
1225 }
1226
1227done:
1228 taprio_offload_free(offload);
1229
1230 return err;
1231}
1232
1233static int taprio_disable_offload(struct net_device *dev,
1234 struct taprio_sched *q,
1235 struct netlink_ext_ack *extack)
1236{
1237 const struct net_device_ops *ops = dev->netdev_ops;
1238 struct tc_taprio_qopt_offload *offload;
1239 int err;
1240
1241 if (!FULL_OFFLOAD_IS_ENABLED(q->flags))
1242 return 0;
1243
1244 if (!ops->ndo_setup_tc)
1245 return -EOPNOTSUPP;
1246
1247 offload = taprio_offload_alloc(0);
1248 if (!offload) {
1249 NL_SET_ERR_MSG(extack,
1250 "Not enough memory to disable offload mode");
1251 return -ENOMEM;
1252 }
1253 offload->enable = 0;
1254
1255 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1256 if (err < 0) {
1257 NL_SET_ERR_MSG(extack,
1258 "Device failed to disable offload");
1259 goto out;
1260 }
1261
1262out:
1263 taprio_offload_free(offload);
1264
1265 return err;
1266}
1267
1268/* If full offload is enabled, the only possible clockid is the net device's
1269 * PHC. For that reason, specifying a clockid through netlink is incorrect.
1270 * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1271 * in sync with the specified clockid via a user space daemon such as phc2sys.
1272 * For both software taprio and txtime-assist, the clockid is used for the
1273 * hrtimer that advances the schedule and hence mandatory.
1274 */
1275static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1276 struct netlink_ext_ack *extack)
1277{
1278 struct taprio_sched *q = qdisc_priv(sch);
1279 struct net_device *dev = qdisc_dev(sch);
1280 int err = -EINVAL;
1281
1282 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1283 const struct ethtool_ops *ops = dev->ethtool_ops;
1284 struct ethtool_ts_info info = {
1285 .cmd = ETHTOOL_GET_TS_INFO,
1286 .phc_index = -1,
1287 };
1288
1289 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1290 NL_SET_ERR_MSG(extack,
1291 "The 'clockid' cannot be specified for full offload");
1292 goto out;
1293 }
1294
1295 if (ops && ops->get_ts_info)
1296 err = ops->get_ts_info(dev, &info);
1297
1298 if (err || info.phc_index < 0) {
1299 NL_SET_ERR_MSG(extack,
1300 "Device does not have a PTP clock");
1301 err = -ENOTSUPP;
1302 goto out;
1303 }
1304 } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1305 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1306
1307 /* We only support static clockids and we don't allow
1308 * for it to be modified after the first init.
1309 */
1310 if (clockid < 0 ||
1311 (q->clockid != -1 && q->clockid != clockid)) {
1312 NL_SET_ERR_MSG(extack,
1313 "Changing the 'clockid' of a running schedule is not supported");
1314 err = -ENOTSUPP;
1315 goto out;
1316 }
1317
1318 switch (clockid) {
1319 case CLOCK_REALTIME:
1320 q->tk_offset = TK_OFFS_REAL;
1321 break;
1322 case CLOCK_MONOTONIC:
1323 q->tk_offset = TK_OFFS_MAX;
1324 break;
1325 case CLOCK_BOOTTIME:
1326 q->tk_offset = TK_OFFS_BOOT;
1327 break;
1328 case CLOCK_TAI:
1329 q->tk_offset = TK_OFFS_TAI;
1330 break;
1331 default:
1332 NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1333 err = -EINVAL;
1334 goto out;
1335 }
1336
1337 q->clockid = clockid;
1338 } else {
1339 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1340 goto out;
1341 }
1342
1343 /* Everything went ok, return success. */
1344 err = 0;
1345
1346out:
1347 return err;
1348}
1349
1350static int taprio_mqprio_cmp(const struct net_device *dev,
1351 const struct tc_mqprio_qopt *mqprio)
1352{
1353 int i;
1354
1355 if (!mqprio || mqprio->num_tc != dev->num_tc)
1356 return -1;
1357
1358 for (i = 0; i < mqprio->num_tc; i++)
1359 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1360 dev->tc_to_txq[i].offset != mqprio->offset[i])
1361 return -1;
1362
1363 for (i = 0; i <= TC_BITMASK; i++)
1364 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1365 return -1;
1366
1367 return 0;
1368}
1369
1370static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1371 struct netlink_ext_ack *extack)
1372{
1373 struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1374 struct sched_gate_list *oper, *admin, *new_admin;
1375 struct taprio_sched *q = qdisc_priv(sch);
1376 struct net_device *dev = qdisc_dev(sch);
1377 struct tc_mqprio_qopt *mqprio = NULL;
1378 u32 taprio_flags = 0;
1379 unsigned long flags;
1380 ktime_t start;
1381 int i, err;
1382
1383 err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1384 taprio_policy, extack);
1385 if (err < 0)
1386 return err;
1387
1388 if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1389 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1390
1391 if (tb[TCA_TAPRIO_ATTR_FLAGS]) {
1392 taprio_flags = nla_get_u32(tb[TCA_TAPRIO_ATTR_FLAGS]);
1393
1394 if (q->flags != 0 && q->flags != taprio_flags) {
1395 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1396 return -EOPNOTSUPP;
1397 } else if (!taprio_flags_valid(taprio_flags)) {
1398 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1399 return -EINVAL;
1400 }
1401
1402 q->flags = taprio_flags;
1403 }
1404
1405 err = taprio_parse_mqprio_opt(dev, mqprio, extack, taprio_flags);
1406 if (err < 0)
1407 return err;
1408
1409 new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1410 if (!new_admin) {
1411 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1412 return -ENOMEM;
1413 }
1414 INIT_LIST_HEAD(&new_admin->entries);
1415
1416 rcu_read_lock();
1417 oper = rcu_dereference(q->oper_sched);
1418 admin = rcu_dereference(q->admin_sched);
1419 rcu_read_unlock();
1420
1421 /* no changes - no new mqprio settings */
1422 if (!taprio_mqprio_cmp(dev, mqprio))
1423 mqprio = NULL;
1424
1425 if (mqprio && (oper || admin)) {
1426 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1427 err = -ENOTSUPP;
1428 goto free_sched;
1429 }
1430
1431 err = parse_taprio_schedule(tb, new_admin, extack);
1432 if (err < 0)
1433 goto free_sched;
1434
1435 if (new_admin->num_entries == 0) {
1436 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1437 err = -EINVAL;
1438 goto free_sched;
1439 }
1440
1441 err = taprio_parse_clockid(sch, tb, extack);
1442 if (err < 0)
1443 goto free_sched;
1444
1445 taprio_set_picos_per_byte(dev, q);
1446
1447 if (FULL_OFFLOAD_IS_ENABLED(taprio_flags))
1448 err = taprio_enable_offload(dev, mqprio, q, new_admin, extack);
1449 else
1450 err = taprio_disable_offload(dev, q, extack);
1451 if (err)
1452 goto free_sched;
1453
1454 /* Protects against enqueue()/dequeue() */
1455 spin_lock_bh(qdisc_lock(sch));
1456
1457 if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1458 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1459 NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1460 err = -EINVAL;
1461 goto unlock;
1462 }
1463
1464 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1465 }
1466
1467 if (!TXTIME_ASSIST_IS_ENABLED(taprio_flags) &&
1468 !FULL_OFFLOAD_IS_ENABLED(taprio_flags) &&
1469 !hrtimer_active(&q->advance_timer)) {
1470 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1471 q->advance_timer.function = advance_sched;
1472 }
1473
1474 if (mqprio) {
1475 netdev_set_num_tc(dev, mqprio->num_tc);
1476 for (i = 0; i < mqprio->num_tc; i++)
1477 netdev_set_tc_queue(dev, i,
1478 mqprio->count[i],
1479 mqprio->offset[i]);
1480
1481 /* Always use supplied priority mappings */
1482 for (i = 0; i <= TC_BITMASK; i++)
1483 netdev_set_prio_tc_map(dev, i,
1484 mqprio->prio_tc_map[i]);
1485 }
1486
1487 if (FULL_OFFLOAD_IS_ENABLED(taprio_flags)) {
1488 q->dequeue = taprio_dequeue_offload;
1489 q->peek = taprio_peek_offload;
1490 } else {
1491 /* Be sure to always keep the function pointers
1492 * in a consistent state.
1493 */
1494 q->dequeue = taprio_dequeue_soft;
1495 q->peek = taprio_peek_soft;
1496 }
1497
1498 err = taprio_get_start_time(sch, new_admin, &start);
1499 if (err < 0) {
1500 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1501 goto unlock;
1502 }
1503
1504 if (TXTIME_ASSIST_IS_ENABLED(taprio_flags)) {
1505 setup_txtime(q, new_admin, start);
1506
1507 if (!oper) {
1508 rcu_assign_pointer(q->oper_sched, new_admin);
1509 err = 0;
1510 new_admin = NULL;
1511 goto unlock;
1512 }
1513
1514 rcu_assign_pointer(q->admin_sched, new_admin);
1515 if (admin)
1516 call_rcu(&admin->rcu, taprio_free_sched_cb);
1517 } else {
1518 setup_first_close_time(q, new_admin, start);
1519
1520 /* Protects against advance_sched() */
1521 spin_lock_irqsave(&q->current_entry_lock, flags);
1522
1523 taprio_start_sched(sch, start, new_admin);
1524
1525 rcu_assign_pointer(q->admin_sched, new_admin);
1526 if (admin)
1527 call_rcu(&admin->rcu, taprio_free_sched_cb);
1528
1529 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1530
1531 if (FULL_OFFLOAD_IS_ENABLED(taprio_flags))
1532 taprio_offload_config_changed(q);
1533 }
1534
1535 new_admin = NULL;
1536 err = 0;
1537
1538unlock:
1539 spin_unlock_bh(qdisc_lock(sch));
1540
1541free_sched:
1542 if (new_admin)
1543 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
1544
1545 return err;
1546}
1547
1548static void taprio_destroy(struct Qdisc *sch)
1549{
1550 struct taprio_sched *q = qdisc_priv(sch);
1551 struct net_device *dev = qdisc_dev(sch);
1552 unsigned int i;
1553
1554 spin_lock(&taprio_list_lock);
1555 list_del(&q->taprio_list);
1556 spin_unlock(&taprio_list_lock);
1557
1558 hrtimer_cancel(&q->advance_timer);
1559
1560 taprio_disable_offload(dev, q, NULL);
1561
1562 if (q->qdiscs) {
1563 for (i = 0; i < dev->num_tx_queues && q->qdiscs[i]; i++)
1564 qdisc_put(q->qdiscs[i]);
1565
1566 kfree(q->qdiscs);
1567 }
1568 q->qdiscs = NULL;
1569
1570 netdev_set_num_tc(dev, 0);
1571
1572 if (q->oper_sched)
1573 call_rcu(&q->oper_sched->rcu, taprio_free_sched_cb);
1574
1575 if (q->admin_sched)
1576 call_rcu(&q->admin_sched->rcu, taprio_free_sched_cb);
1577}
1578
1579static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
1580 struct netlink_ext_ack *extack)
1581{
1582 struct taprio_sched *q = qdisc_priv(sch);
1583 struct net_device *dev = qdisc_dev(sch);
1584 int i;
1585
1586 spin_lock_init(&q->current_entry_lock);
1587
1588 hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
1589 q->advance_timer.function = advance_sched;
1590
1591 q->dequeue = taprio_dequeue_soft;
1592 q->peek = taprio_peek_soft;
1593
1594 q->root = sch;
1595
1596 /* We only support static clockids. Use an invalid value as default
1597 * and get the valid one on taprio_change().
1598 */
1599 q->clockid = -1;
1600
1601 spin_lock(&taprio_list_lock);
1602 list_add(&q->taprio_list, &taprio_list);
1603 spin_unlock(&taprio_list_lock);
1604
1605 if (sch->parent != TC_H_ROOT)
1606 return -EOPNOTSUPP;
1607
1608 if (!netif_is_multiqueue(dev))
1609 return -EOPNOTSUPP;
1610
1611 /* pre-allocate qdisc, attachment can't fail */
1612 q->qdiscs = kcalloc(dev->num_tx_queues,
1613 sizeof(q->qdiscs[0]),
1614 GFP_KERNEL);
1615
1616 if (!q->qdiscs)
1617 return -ENOMEM;
1618
1619 if (!opt)
1620 return -EINVAL;
1621
1622 for (i = 0; i < dev->num_tx_queues; i++) {
1623 struct netdev_queue *dev_queue;
1624 struct Qdisc *qdisc;
1625
1626 dev_queue = netdev_get_tx_queue(dev, i);
1627 qdisc = qdisc_create_dflt(dev_queue,
1628 &pfifo_qdisc_ops,
1629 TC_H_MAKE(TC_H_MAJ(sch->handle),
1630 TC_H_MIN(i + 1)),
1631 extack);
1632 if (!qdisc)
1633 return -ENOMEM;
1634
1635 if (i < dev->real_num_tx_queues)
1636 qdisc_hash_add(qdisc, false);
1637
1638 q->qdiscs[i] = qdisc;
1639 }
1640
1641 return taprio_change(sch, opt, extack);
1642}
1643
1644static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
1645 unsigned long cl)
1646{
1647 struct net_device *dev = qdisc_dev(sch);
1648 unsigned long ntx = cl - 1;
1649
1650 if (ntx >= dev->num_tx_queues)
1651 return NULL;
1652
1653 return netdev_get_tx_queue(dev, ntx);
1654}
1655
1656static int taprio_graft(struct Qdisc *sch, unsigned long cl,
1657 struct Qdisc *new, struct Qdisc **old,
1658 struct netlink_ext_ack *extack)
1659{
1660 struct taprio_sched *q = qdisc_priv(sch);
1661 struct net_device *dev = qdisc_dev(sch);
1662 struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1663
1664 if (!dev_queue)
1665 return -EINVAL;
1666
1667 if (dev->flags & IFF_UP)
1668 dev_deactivate(dev);
1669
1670 *old = q->qdiscs[cl - 1];
1671 q->qdiscs[cl - 1] = new;
1672
1673 if (new)
1674 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1675
1676 if (dev->flags & IFF_UP)
1677 dev_activate(dev);
1678
1679 return 0;
1680}
1681
1682static int dump_entry(struct sk_buff *msg,
1683 const struct sched_entry *entry)
1684{
1685 struct nlattr *item;
1686
1687 item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
1688 if (!item)
1689 return -ENOSPC;
1690
1691 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
1692 goto nla_put_failure;
1693
1694 if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
1695 goto nla_put_failure;
1696
1697 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
1698 entry->gate_mask))
1699 goto nla_put_failure;
1700
1701 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
1702 entry->interval))
1703 goto nla_put_failure;
1704
1705 return nla_nest_end(msg, item);
1706
1707nla_put_failure:
1708 nla_nest_cancel(msg, item);
1709 return -1;
1710}
1711
1712static int dump_schedule(struct sk_buff *msg,
1713 const struct sched_gate_list *root)
1714{
1715 struct nlattr *entry_list;
1716 struct sched_entry *entry;
1717
1718 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
1719 root->base_time, TCA_TAPRIO_PAD))
1720 return -1;
1721
1722 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
1723 root->cycle_time, TCA_TAPRIO_PAD))
1724 return -1;
1725
1726 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
1727 root->cycle_time_extension, TCA_TAPRIO_PAD))
1728 return -1;
1729
1730 entry_list = nla_nest_start_noflag(msg,
1731 TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
1732 if (!entry_list)
1733 goto error_nest;
1734
1735 list_for_each_entry(entry, &root->entries, list) {
1736 if (dump_entry(msg, entry) < 0)
1737 goto error_nest;
1738 }
1739
1740 nla_nest_end(msg, entry_list);
1741 return 0;
1742
1743error_nest:
1744 nla_nest_cancel(msg, entry_list);
1745 return -1;
1746}
1747
1748static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
1749{
1750 struct taprio_sched *q = qdisc_priv(sch);
1751 struct net_device *dev = qdisc_dev(sch);
1752 struct sched_gate_list *oper, *admin;
1753 struct tc_mqprio_qopt opt = { 0 };
1754 struct nlattr *nest, *sched_nest;
1755 unsigned int i;
1756
1757 rcu_read_lock();
1758 oper = rcu_dereference(q->oper_sched);
1759 admin = rcu_dereference(q->admin_sched);
1760
1761 opt.num_tc = netdev_get_num_tc(dev);
1762 memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
1763
1764 for (i = 0; i < netdev_get_num_tc(dev); i++) {
1765 opt.count[i] = dev->tc_to_txq[i].count;
1766 opt.offset[i] = dev->tc_to_txq[i].offset;
1767 }
1768
1769 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1770 if (!nest)
1771 goto start_error;
1772
1773 if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
1774 goto options_error;
1775
1776 if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1777 nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
1778 goto options_error;
1779
1780 if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
1781 goto options_error;
1782
1783 if (q->txtime_delay &&
1784 nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
1785 goto options_error;
1786
1787 if (oper && dump_schedule(skb, oper))
1788 goto options_error;
1789
1790 if (!admin)
1791 goto done;
1792
1793 sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
1794 if (!sched_nest)
1795 goto options_error;
1796
1797 if (dump_schedule(skb, admin))
1798 goto admin_error;
1799
1800 nla_nest_end(skb, sched_nest);
1801
1802done:
1803 rcu_read_unlock();
1804
1805 return nla_nest_end(skb, nest);
1806
1807admin_error:
1808 nla_nest_cancel(skb, sched_nest);
1809
1810options_error:
1811 nla_nest_cancel(skb, nest);
1812
1813start_error:
1814 rcu_read_unlock();
1815 return -ENOSPC;
1816}
1817
1818static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
1819{
1820 struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1821
1822 if (!dev_queue)
1823 return NULL;
1824
1825 return dev_queue->qdisc_sleeping;
1826}
1827
1828static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
1829{
1830 unsigned int ntx = TC_H_MIN(classid);
1831
1832 if (!taprio_queue_get(sch, ntx))
1833 return 0;
1834 return ntx;
1835}
1836
1837static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
1838 struct sk_buff *skb, struct tcmsg *tcm)
1839{
1840 struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1841
1842 tcm->tcm_parent = TC_H_ROOT;
1843 tcm->tcm_handle |= TC_H_MIN(cl);
1844 tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
1845
1846 return 0;
1847}
1848
1849static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
1850 struct gnet_dump *d)
1851 __releases(d->lock)
1852 __acquires(d->lock)
1853{
1854 struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1855
1856 sch = dev_queue->qdisc_sleeping;
1857 if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
1858 qdisc_qstats_copy(d, sch) < 0)
1859 return -1;
1860 return 0;
1861}
1862
1863static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1864{
1865 struct net_device *dev = qdisc_dev(sch);
1866 unsigned long ntx;
1867
1868 if (arg->stop)
1869 return;
1870
1871 arg->count = arg->skip;
1872 for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
1873 if (arg->fn(sch, ntx + 1, arg) < 0) {
1874 arg->stop = 1;
1875 break;
1876 }
1877 arg->count++;
1878 }
1879}
1880
1881static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
1882 struct tcmsg *tcm)
1883{
1884 return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
1885}
1886
1887static const struct Qdisc_class_ops taprio_class_ops = {
1888 .graft = taprio_graft,
1889 .leaf = taprio_leaf,
1890 .find = taprio_find,
1891 .walk = taprio_walk,
1892 .dump = taprio_dump_class,
1893 .dump_stats = taprio_dump_class_stats,
1894 .select_queue = taprio_select_queue,
1895};
1896
1897static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
1898 .cl_ops = &taprio_class_ops,
1899 .id = "taprio",
1900 .priv_size = sizeof(struct taprio_sched),
1901 .init = taprio_init,
1902 .change = taprio_change,
1903 .destroy = taprio_destroy,
1904 .peek = taprio_peek,
1905 .dequeue = taprio_dequeue,
1906 .enqueue = taprio_enqueue,
1907 .dump = taprio_dump,
1908 .owner = THIS_MODULE,
1909};
1910
1911static struct notifier_block taprio_device_notifier = {
1912 .notifier_call = taprio_dev_notifier,
1913};
1914
1915static int __init taprio_module_init(void)
1916{
1917 int err = register_netdevice_notifier(&taprio_device_notifier);
1918
1919 if (err)
1920 return err;
1921
1922 return register_qdisc(&taprio_qdisc_ops);
1923}
1924
1925static void __exit taprio_module_exit(void)
1926{
1927 unregister_qdisc(&taprio_qdisc_ops);
1928 unregister_netdevice_notifier(&taprio_device_notifier);
1929}
1930
1931module_init(taprio_module_init);
1932module_exit(taprio_module_exit);
1933MODULE_LICENSE("GPL");