Loading...
Note: File does not exist in v3.5.6.
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/ethtool.h>
10#include <linux/ethtool_netlink.h>
11#include <linux/types.h>
12#include <linux/slab.h>
13#include <linux/kernel.h>
14#include <linux/string.h>
15#include <linux/list.h>
16#include <linux/errno.h>
17#include <linux/skbuff.h>
18#include <linux/math64.h>
19#include <linux/module.h>
20#include <linux/spinlock.h>
21#include <linux/rcupdate.h>
22#include <linux/time.h>
23#include <net/gso.h>
24#include <net/netlink.h>
25#include <net/pkt_sched.h>
26#include <net/pkt_cls.h>
27#include <net/sch_generic.h>
28#include <net/sock.h>
29#include <net/tcp.h>
30
31#define TAPRIO_STAT_NOT_SET (~0ULL)
32
33#include "sch_mqprio_lib.h"
34
35static LIST_HEAD(taprio_list);
36static struct static_key_false taprio_have_broken_mqprio;
37static struct static_key_false taprio_have_working_mqprio;
38
39#define TAPRIO_ALL_GATES_OPEN -1
40
41#define TXTIME_ASSIST_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST)
42#define FULL_OFFLOAD_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)
43#define TAPRIO_FLAGS_INVALID U32_MAX
44
45struct sched_entry {
46 /* Durations between this GCL entry and the GCL entry where the
47 * respective traffic class gate closes
48 */
49 u64 gate_duration[TC_MAX_QUEUE];
50 atomic_t budget[TC_MAX_QUEUE];
51 /* The qdisc makes some effort so that no packet leaves
52 * after this time
53 */
54 ktime_t gate_close_time[TC_MAX_QUEUE];
55 struct list_head list;
56 /* Used to calculate when to advance the schedule */
57 ktime_t end_time;
58 ktime_t next_txtime;
59 int index;
60 u32 gate_mask;
61 u32 interval;
62 u8 command;
63};
64
65struct sched_gate_list {
66 /* Longest non-zero contiguous gate durations per traffic class,
67 * or 0 if a traffic class gate never opens during the schedule.
68 */
69 u64 max_open_gate_duration[TC_MAX_QUEUE];
70 u32 max_frm_len[TC_MAX_QUEUE]; /* for the fast path */
71 u32 max_sdu[TC_MAX_QUEUE]; /* for dump */
72 struct rcu_head rcu;
73 struct list_head entries;
74 size_t num_entries;
75 ktime_t cycle_end_time;
76 s64 cycle_time;
77 s64 cycle_time_extension;
78 s64 base_time;
79};
80
81struct taprio_sched {
82 struct Qdisc **qdiscs;
83 struct Qdisc *root;
84 u32 flags;
85 enum tk_offsets tk_offset;
86 int clockid;
87 bool offloaded;
88 bool detected_mqprio;
89 bool broken_mqprio;
90 atomic64_t picos_per_byte; /* Using picoseconds because for 10Gbps+
91 * speeds it's sub-nanoseconds per byte
92 */
93
94 /* Protects the update side of the RCU protected current_entry */
95 spinlock_t current_entry_lock;
96 struct sched_entry __rcu *current_entry;
97 struct sched_gate_list __rcu *oper_sched;
98 struct sched_gate_list __rcu *admin_sched;
99 struct hrtimer advance_timer;
100 struct list_head taprio_list;
101 int cur_txq[TC_MAX_QUEUE];
102 u32 max_sdu[TC_MAX_QUEUE]; /* save info from the user */
103 u32 fp[TC_QOPT_MAX_QUEUE]; /* only for dump and offloading */
104 u32 txtime_delay;
105};
106
107struct __tc_taprio_qopt_offload {
108 refcount_t users;
109 struct tc_taprio_qopt_offload offload;
110};
111
112static void taprio_calculate_gate_durations(struct taprio_sched *q,
113 struct sched_gate_list *sched)
114{
115 struct net_device *dev = qdisc_dev(q->root);
116 int num_tc = netdev_get_num_tc(dev);
117 struct sched_entry *entry, *cur;
118 int tc;
119
120 list_for_each_entry(entry, &sched->entries, list) {
121 u32 gates_still_open = entry->gate_mask;
122
123 /* For each traffic class, calculate each open gate duration,
124 * starting at this schedule entry and ending at the schedule
125 * entry containing a gate close event for that TC.
126 */
127 cur = entry;
128
129 do {
130 if (!gates_still_open)
131 break;
132
133 for (tc = 0; tc < num_tc; tc++) {
134 if (!(gates_still_open & BIT(tc)))
135 continue;
136
137 if (cur->gate_mask & BIT(tc))
138 entry->gate_duration[tc] += cur->interval;
139 else
140 gates_still_open &= ~BIT(tc);
141 }
142
143 cur = list_next_entry_circular(cur, &sched->entries, list);
144 } while (cur != entry);
145
146 /* Keep track of the maximum gate duration for each traffic
147 * class, taking care to not confuse a traffic class which is
148 * temporarily closed with one that is always closed.
149 */
150 for (tc = 0; tc < num_tc; tc++)
151 if (entry->gate_duration[tc] &&
152 sched->max_open_gate_duration[tc] < entry->gate_duration[tc])
153 sched->max_open_gate_duration[tc] = entry->gate_duration[tc];
154 }
155}
156
157static bool taprio_entry_allows_tx(ktime_t skb_end_time,
158 struct sched_entry *entry, int tc)
159{
160 return ktime_before(skb_end_time, entry->gate_close_time[tc]);
161}
162
163static ktime_t sched_base_time(const struct sched_gate_list *sched)
164{
165 if (!sched)
166 return KTIME_MAX;
167
168 return ns_to_ktime(sched->base_time);
169}
170
171static ktime_t taprio_mono_to_any(const struct taprio_sched *q, ktime_t mono)
172{
173 /* This pairs with WRITE_ONCE() in taprio_parse_clockid() */
174 enum tk_offsets tk_offset = READ_ONCE(q->tk_offset);
175
176 switch (tk_offset) {
177 case TK_OFFS_MAX:
178 return mono;
179 default:
180 return ktime_mono_to_any(mono, tk_offset);
181 }
182}
183
184static ktime_t taprio_get_time(const struct taprio_sched *q)
185{
186 return taprio_mono_to_any(q, ktime_get());
187}
188
189static void taprio_free_sched_cb(struct rcu_head *head)
190{
191 struct sched_gate_list *sched = container_of(head, struct sched_gate_list, rcu);
192 struct sched_entry *entry, *n;
193
194 list_for_each_entry_safe(entry, n, &sched->entries, list) {
195 list_del(&entry->list);
196 kfree(entry);
197 }
198
199 kfree(sched);
200}
201
202static void switch_schedules(struct taprio_sched *q,
203 struct sched_gate_list **admin,
204 struct sched_gate_list **oper)
205{
206 rcu_assign_pointer(q->oper_sched, *admin);
207 rcu_assign_pointer(q->admin_sched, NULL);
208
209 if (*oper)
210 call_rcu(&(*oper)->rcu, taprio_free_sched_cb);
211
212 *oper = *admin;
213 *admin = NULL;
214}
215
216/* Get how much time has been already elapsed in the current cycle. */
217static s32 get_cycle_time_elapsed(struct sched_gate_list *sched, ktime_t time)
218{
219 ktime_t time_since_sched_start;
220 s32 time_elapsed;
221
222 time_since_sched_start = ktime_sub(time, sched->base_time);
223 div_s64_rem(time_since_sched_start, sched->cycle_time, &time_elapsed);
224
225 return time_elapsed;
226}
227
228static ktime_t get_interval_end_time(struct sched_gate_list *sched,
229 struct sched_gate_list *admin,
230 struct sched_entry *entry,
231 ktime_t intv_start)
232{
233 s32 cycle_elapsed = get_cycle_time_elapsed(sched, intv_start);
234 ktime_t intv_end, cycle_ext_end, cycle_end;
235
236 cycle_end = ktime_add_ns(intv_start, sched->cycle_time - cycle_elapsed);
237 intv_end = ktime_add_ns(intv_start, entry->interval);
238 cycle_ext_end = ktime_add(cycle_end, sched->cycle_time_extension);
239
240 if (ktime_before(intv_end, cycle_end))
241 return intv_end;
242 else if (admin && admin != sched &&
243 ktime_after(admin->base_time, cycle_end) &&
244 ktime_before(admin->base_time, cycle_ext_end))
245 return admin->base_time;
246 else
247 return cycle_end;
248}
249
250static int length_to_duration(struct taprio_sched *q, int len)
251{
252 return div_u64(len * atomic64_read(&q->picos_per_byte), PSEC_PER_NSEC);
253}
254
255static int duration_to_length(struct taprio_sched *q, u64 duration)
256{
257 return div_u64(duration * PSEC_PER_NSEC, atomic64_read(&q->picos_per_byte));
258}
259
260/* Sets sched->max_sdu[] and sched->max_frm_len[] to the minimum between the
261 * q->max_sdu[] requested by the user and the max_sdu dynamically determined by
262 * the maximum open gate durations at the given link speed.
263 */
264static void taprio_update_queue_max_sdu(struct taprio_sched *q,
265 struct sched_gate_list *sched,
266 struct qdisc_size_table *stab)
267{
268 struct net_device *dev = qdisc_dev(q->root);
269 int num_tc = netdev_get_num_tc(dev);
270 u32 max_sdu_from_user;
271 u32 max_sdu_dynamic;
272 u32 max_sdu;
273 int tc;
274
275 for (tc = 0; tc < num_tc; tc++) {
276 max_sdu_from_user = q->max_sdu[tc] ?: U32_MAX;
277
278 /* TC gate never closes => keep the queueMaxSDU
279 * selected by the user
280 */
281 if (sched->max_open_gate_duration[tc] == sched->cycle_time) {
282 max_sdu_dynamic = U32_MAX;
283 } else {
284 u32 max_frm_len;
285
286 max_frm_len = duration_to_length(q, sched->max_open_gate_duration[tc]);
287 /* Compensate for L1 overhead from size table,
288 * but don't let the frame size go negative
289 */
290 if (stab) {
291 max_frm_len -= stab->szopts.overhead;
292 max_frm_len = max_t(int, max_frm_len,
293 dev->hard_header_len + 1);
294 }
295 max_sdu_dynamic = max_frm_len - dev->hard_header_len;
296 if (max_sdu_dynamic > dev->max_mtu)
297 max_sdu_dynamic = U32_MAX;
298 }
299
300 max_sdu = min(max_sdu_dynamic, max_sdu_from_user);
301
302 if (max_sdu != U32_MAX) {
303 sched->max_frm_len[tc] = max_sdu + dev->hard_header_len;
304 sched->max_sdu[tc] = max_sdu;
305 } else {
306 sched->max_frm_len[tc] = U32_MAX; /* never oversized */
307 sched->max_sdu[tc] = 0;
308 }
309 }
310}
311
312/* Returns the entry corresponding to next available interval. If
313 * validate_interval is set, it only validates whether the timestamp occurs
314 * when the gate corresponding to the skb's traffic class is open.
315 */
316static struct sched_entry *find_entry_to_transmit(struct sk_buff *skb,
317 struct Qdisc *sch,
318 struct sched_gate_list *sched,
319 struct sched_gate_list *admin,
320 ktime_t time,
321 ktime_t *interval_start,
322 ktime_t *interval_end,
323 bool validate_interval)
324{
325 ktime_t curr_intv_start, curr_intv_end, cycle_end, packet_transmit_time;
326 ktime_t earliest_txtime = KTIME_MAX, txtime, cycle, transmit_end_time;
327 struct sched_entry *entry = NULL, *entry_found = NULL;
328 struct taprio_sched *q = qdisc_priv(sch);
329 struct net_device *dev = qdisc_dev(sch);
330 bool entry_available = false;
331 s32 cycle_elapsed;
332 int tc, n;
333
334 tc = netdev_get_prio_tc_map(dev, skb->priority);
335 packet_transmit_time = length_to_duration(q, qdisc_pkt_len(skb));
336
337 *interval_start = 0;
338 *interval_end = 0;
339
340 if (!sched)
341 return NULL;
342
343 cycle = sched->cycle_time;
344 cycle_elapsed = get_cycle_time_elapsed(sched, time);
345 curr_intv_end = ktime_sub_ns(time, cycle_elapsed);
346 cycle_end = ktime_add_ns(curr_intv_end, cycle);
347
348 list_for_each_entry(entry, &sched->entries, list) {
349 curr_intv_start = curr_intv_end;
350 curr_intv_end = get_interval_end_time(sched, admin, entry,
351 curr_intv_start);
352
353 if (ktime_after(curr_intv_start, cycle_end))
354 break;
355
356 if (!(entry->gate_mask & BIT(tc)) ||
357 packet_transmit_time > entry->interval)
358 continue;
359
360 txtime = entry->next_txtime;
361
362 if (ktime_before(txtime, time) || validate_interval) {
363 transmit_end_time = ktime_add_ns(time, packet_transmit_time);
364 if ((ktime_before(curr_intv_start, time) &&
365 ktime_before(transmit_end_time, curr_intv_end)) ||
366 (ktime_after(curr_intv_start, time) && !validate_interval)) {
367 entry_found = entry;
368 *interval_start = curr_intv_start;
369 *interval_end = curr_intv_end;
370 break;
371 } else if (!entry_available && !validate_interval) {
372 /* Here, we are just trying to find out the
373 * first available interval in the next cycle.
374 */
375 entry_available = true;
376 entry_found = entry;
377 *interval_start = ktime_add_ns(curr_intv_start, cycle);
378 *interval_end = ktime_add_ns(curr_intv_end, cycle);
379 }
380 } else if (ktime_before(txtime, earliest_txtime) &&
381 !entry_available) {
382 earliest_txtime = txtime;
383 entry_found = entry;
384 n = div_s64(ktime_sub(txtime, curr_intv_start), cycle);
385 *interval_start = ktime_add(curr_intv_start, n * cycle);
386 *interval_end = ktime_add(curr_intv_end, n * cycle);
387 }
388 }
389
390 return entry_found;
391}
392
393static bool is_valid_interval(struct sk_buff *skb, struct Qdisc *sch)
394{
395 struct taprio_sched *q = qdisc_priv(sch);
396 struct sched_gate_list *sched, *admin;
397 ktime_t interval_start, interval_end;
398 struct sched_entry *entry;
399
400 rcu_read_lock();
401 sched = rcu_dereference(q->oper_sched);
402 admin = rcu_dereference(q->admin_sched);
403
404 entry = find_entry_to_transmit(skb, sch, sched, admin, skb->tstamp,
405 &interval_start, &interval_end, true);
406 rcu_read_unlock();
407
408 return entry;
409}
410
411static bool taprio_flags_valid(u32 flags)
412{
413 /* Make sure no other flag bits are set. */
414 if (flags & ~(TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST |
415 TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
416 return false;
417 /* txtime-assist and full offload are mutually exclusive */
418 if ((flags & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST) &&
419 (flags & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
420 return false;
421 return true;
422}
423
424/* This returns the tstamp value set by TCP in terms of the set clock. */
425static ktime_t get_tcp_tstamp(struct taprio_sched *q, struct sk_buff *skb)
426{
427 unsigned int offset = skb_network_offset(skb);
428 const struct ipv6hdr *ipv6h;
429 const struct iphdr *iph;
430 struct ipv6hdr _ipv6h;
431
432 ipv6h = skb_header_pointer(skb, offset, sizeof(_ipv6h), &_ipv6h);
433 if (!ipv6h)
434 return 0;
435
436 if (ipv6h->version == 4) {
437 iph = (struct iphdr *)ipv6h;
438 offset += iph->ihl * 4;
439
440 /* special-case 6in4 tunnelling, as that is a common way to get
441 * v6 connectivity in the home
442 */
443 if (iph->protocol == IPPROTO_IPV6) {
444 ipv6h = skb_header_pointer(skb, offset,
445 sizeof(_ipv6h), &_ipv6h);
446
447 if (!ipv6h || ipv6h->nexthdr != IPPROTO_TCP)
448 return 0;
449 } else if (iph->protocol != IPPROTO_TCP) {
450 return 0;
451 }
452 } else if (ipv6h->version == 6 && ipv6h->nexthdr != IPPROTO_TCP) {
453 return 0;
454 }
455
456 return taprio_mono_to_any(q, skb->skb_mstamp_ns);
457}
458
459/* There are a few scenarios where we will have to modify the txtime from
460 * what is read from next_txtime in sched_entry. They are:
461 * 1. If txtime is in the past,
462 * a. The gate for the traffic class is currently open and packet can be
463 * transmitted before it closes, schedule the packet right away.
464 * b. If the gate corresponding to the traffic class is going to open later
465 * in the cycle, set the txtime of packet to the interval start.
466 * 2. If txtime is in the future, there are packets corresponding to the
467 * current traffic class waiting to be transmitted. So, the following
468 * possibilities exist:
469 * a. We can transmit the packet before the window containing the txtime
470 * closes.
471 * b. The window might close before the transmission can be completed
472 * successfully. So, schedule the packet in the next open window.
473 */
474static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)
475{
476 ktime_t transmit_end_time, interval_end, interval_start, tcp_tstamp;
477 struct taprio_sched *q = qdisc_priv(sch);
478 struct sched_gate_list *sched, *admin;
479 ktime_t minimum_time, now, txtime;
480 int len, packet_transmit_time;
481 struct sched_entry *entry;
482 bool sched_changed;
483
484 now = taprio_get_time(q);
485 minimum_time = ktime_add_ns(now, q->txtime_delay);
486
487 tcp_tstamp = get_tcp_tstamp(q, skb);
488 minimum_time = max_t(ktime_t, minimum_time, tcp_tstamp);
489
490 rcu_read_lock();
491 admin = rcu_dereference(q->admin_sched);
492 sched = rcu_dereference(q->oper_sched);
493 if (admin && ktime_after(minimum_time, admin->base_time))
494 switch_schedules(q, &admin, &sched);
495
496 /* Until the schedule starts, all the queues are open */
497 if (!sched || ktime_before(minimum_time, sched->base_time)) {
498 txtime = minimum_time;
499 goto done;
500 }
501
502 len = qdisc_pkt_len(skb);
503 packet_transmit_time = length_to_duration(q, len);
504
505 do {
506 sched_changed = false;
507
508 entry = find_entry_to_transmit(skb, sch, sched, admin,
509 minimum_time,
510 &interval_start, &interval_end,
511 false);
512 if (!entry) {
513 txtime = 0;
514 goto done;
515 }
516
517 txtime = entry->next_txtime;
518 txtime = max_t(ktime_t, txtime, minimum_time);
519 txtime = max_t(ktime_t, txtime, interval_start);
520
521 if (admin && admin != sched &&
522 ktime_after(txtime, admin->base_time)) {
523 sched = admin;
524 sched_changed = true;
525 continue;
526 }
527
528 transmit_end_time = ktime_add(txtime, packet_transmit_time);
529 minimum_time = transmit_end_time;
530
531 /* Update the txtime of current entry to the next time it's
532 * interval starts.
533 */
534 if (ktime_after(transmit_end_time, interval_end))
535 entry->next_txtime = ktime_add(interval_start, sched->cycle_time);
536 } while (sched_changed || ktime_after(transmit_end_time, interval_end));
537
538 entry->next_txtime = transmit_end_time;
539
540done:
541 rcu_read_unlock();
542 return txtime;
543}
544
545/* Devices with full offload are expected to honor this in hardware */
546static bool taprio_skb_exceeds_queue_max_sdu(struct Qdisc *sch,
547 struct sk_buff *skb)
548{
549 struct taprio_sched *q = qdisc_priv(sch);
550 struct net_device *dev = qdisc_dev(sch);
551 struct sched_gate_list *sched;
552 int prio = skb->priority;
553 bool exceeds = false;
554 u8 tc;
555
556 tc = netdev_get_prio_tc_map(dev, prio);
557
558 rcu_read_lock();
559 sched = rcu_dereference(q->oper_sched);
560 if (sched && skb->len > sched->max_frm_len[tc])
561 exceeds = true;
562 rcu_read_unlock();
563
564 return exceeds;
565}
566
567static int taprio_enqueue_one(struct sk_buff *skb, struct Qdisc *sch,
568 struct Qdisc *child, struct sk_buff **to_free)
569{
570 struct taprio_sched *q = qdisc_priv(sch);
571
572 /* sk_flags are only safe to use on full sockets. */
573 if (skb->sk && sk_fullsock(skb->sk) && sock_flag(skb->sk, SOCK_TXTIME)) {
574 if (!is_valid_interval(skb, sch))
575 return qdisc_drop(skb, sch, to_free);
576 } else if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
577 skb->tstamp = get_packet_txtime(skb, sch);
578 if (!skb->tstamp)
579 return qdisc_drop(skb, sch, to_free);
580 }
581
582 qdisc_qstats_backlog_inc(sch, skb);
583 sch->q.qlen++;
584
585 return qdisc_enqueue(skb, child, to_free);
586}
587
588static int taprio_enqueue_segmented(struct sk_buff *skb, struct Qdisc *sch,
589 struct Qdisc *child,
590 struct sk_buff **to_free)
591{
592 unsigned int slen = 0, numsegs = 0, len = qdisc_pkt_len(skb);
593 netdev_features_t features = netif_skb_features(skb);
594 struct sk_buff *segs, *nskb;
595 int ret;
596
597 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
598 if (IS_ERR_OR_NULL(segs))
599 return qdisc_drop(skb, sch, to_free);
600
601 skb_list_walk_safe(segs, segs, nskb) {
602 skb_mark_not_on_list(segs);
603 qdisc_skb_cb(segs)->pkt_len = segs->len;
604 slen += segs->len;
605
606 /* FIXME: we should be segmenting to a smaller size
607 * rather than dropping these
608 */
609 if (taprio_skb_exceeds_queue_max_sdu(sch, segs))
610 ret = qdisc_drop(segs, sch, to_free);
611 else
612 ret = taprio_enqueue_one(segs, sch, child, to_free);
613
614 if (ret != NET_XMIT_SUCCESS) {
615 if (net_xmit_drop_count(ret))
616 qdisc_qstats_drop(sch);
617 } else {
618 numsegs++;
619 }
620 }
621
622 if (numsegs > 1)
623 qdisc_tree_reduce_backlog(sch, 1 - numsegs, len - slen);
624 consume_skb(skb);
625
626 return numsegs > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
627}
628
629/* Will not be called in the full offload case, since the TX queues are
630 * attached to the Qdisc created using qdisc_create_dflt()
631 */
632static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,
633 struct sk_buff **to_free)
634{
635 struct taprio_sched *q = qdisc_priv(sch);
636 struct Qdisc *child;
637 int queue;
638
639 queue = skb_get_queue_mapping(skb);
640
641 child = q->qdiscs[queue];
642 if (unlikely(!child))
643 return qdisc_drop(skb, sch, to_free);
644
645 if (taprio_skb_exceeds_queue_max_sdu(sch, skb)) {
646 /* Large packets might not be transmitted when the transmission
647 * duration exceeds any configured interval. Therefore, segment
648 * the skb into smaller chunks. Drivers with full offload are
649 * expected to handle this in hardware.
650 */
651 if (skb_is_gso(skb))
652 return taprio_enqueue_segmented(skb, sch, child,
653 to_free);
654
655 return qdisc_drop(skb, sch, to_free);
656 }
657
658 return taprio_enqueue_one(skb, sch, child, to_free);
659}
660
661static struct sk_buff *taprio_peek(struct Qdisc *sch)
662{
663 WARN_ONCE(1, "taprio only supports operating as root qdisc, peek() not implemented");
664 return NULL;
665}
666
667static void taprio_set_budgets(struct taprio_sched *q,
668 struct sched_gate_list *sched,
669 struct sched_entry *entry)
670{
671 struct net_device *dev = qdisc_dev(q->root);
672 int num_tc = netdev_get_num_tc(dev);
673 int tc, budget;
674
675 for (tc = 0; tc < num_tc; tc++) {
676 /* Traffic classes which never close have infinite budget */
677 if (entry->gate_duration[tc] == sched->cycle_time)
678 budget = INT_MAX;
679 else
680 budget = div64_u64((u64)entry->gate_duration[tc] * PSEC_PER_NSEC,
681 atomic64_read(&q->picos_per_byte));
682
683 atomic_set(&entry->budget[tc], budget);
684 }
685}
686
687/* When an skb is sent, it consumes from the budget of all traffic classes */
688static int taprio_update_budgets(struct sched_entry *entry, size_t len,
689 int tc_consumed, int num_tc)
690{
691 int tc, budget, new_budget = 0;
692
693 for (tc = 0; tc < num_tc; tc++) {
694 budget = atomic_read(&entry->budget[tc]);
695 /* Don't consume from infinite budget */
696 if (budget == INT_MAX) {
697 if (tc == tc_consumed)
698 new_budget = budget;
699 continue;
700 }
701
702 if (tc == tc_consumed)
703 new_budget = atomic_sub_return(len, &entry->budget[tc]);
704 else
705 atomic_sub(len, &entry->budget[tc]);
706 }
707
708 return new_budget;
709}
710
711static struct sk_buff *taprio_dequeue_from_txq(struct Qdisc *sch, int txq,
712 struct sched_entry *entry,
713 u32 gate_mask)
714{
715 struct taprio_sched *q = qdisc_priv(sch);
716 struct net_device *dev = qdisc_dev(sch);
717 struct Qdisc *child = q->qdiscs[txq];
718 int num_tc = netdev_get_num_tc(dev);
719 struct sk_buff *skb;
720 ktime_t guard;
721 int prio;
722 int len;
723 u8 tc;
724
725 if (unlikely(!child))
726 return NULL;
727
728 if (TXTIME_ASSIST_IS_ENABLED(q->flags))
729 goto skip_peek_checks;
730
731 skb = child->ops->peek(child);
732 if (!skb)
733 return NULL;
734
735 prio = skb->priority;
736 tc = netdev_get_prio_tc_map(dev, prio);
737
738 if (!(gate_mask & BIT(tc)))
739 return NULL;
740
741 len = qdisc_pkt_len(skb);
742 guard = ktime_add_ns(taprio_get_time(q), length_to_duration(q, len));
743
744 /* In the case that there's no gate entry, there's no
745 * guard band ...
746 */
747 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
748 !taprio_entry_allows_tx(guard, entry, tc))
749 return NULL;
750
751 /* ... and no budget. */
752 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
753 taprio_update_budgets(entry, len, tc, num_tc) < 0)
754 return NULL;
755
756skip_peek_checks:
757 skb = child->ops->dequeue(child);
758 if (unlikely(!skb))
759 return NULL;
760
761 qdisc_bstats_update(sch, skb);
762 qdisc_qstats_backlog_dec(sch, skb);
763 sch->q.qlen--;
764
765 return skb;
766}
767
768static void taprio_next_tc_txq(struct net_device *dev, int tc, int *txq)
769{
770 int offset = dev->tc_to_txq[tc].offset;
771 int count = dev->tc_to_txq[tc].count;
772
773 (*txq)++;
774 if (*txq == offset + count)
775 *txq = offset;
776}
777
778/* Prioritize higher traffic classes, and select among TXQs belonging to the
779 * same TC using round robin
780 */
781static struct sk_buff *taprio_dequeue_tc_priority(struct Qdisc *sch,
782 struct sched_entry *entry,
783 u32 gate_mask)
784{
785 struct taprio_sched *q = qdisc_priv(sch);
786 struct net_device *dev = qdisc_dev(sch);
787 int num_tc = netdev_get_num_tc(dev);
788 struct sk_buff *skb;
789 int tc;
790
791 for (tc = num_tc - 1; tc >= 0; tc--) {
792 int first_txq = q->cur_txq[tc];
793
794 if (!(gate_mask & BIT(tc)))
795 continue;
796
797 do {
798 skb = taprio_dequeue_from_txq(sch, q->cur_txq[tc],
799 entry, gate_mask);
800
801 taprio_next_tc_txq(dev, tc, &q->cur_txq[tc]);
802
803 if (q->cur_txq[tc] >= dev->num_tx_queues)
804 q->cur_txq[tc] = first_txq;
805
806 if (skb)
807 return skb;
808 } while (q->cur_txq[tc] != first_txq);
809 }
810
811 return NULL;
812}
813
814/* Broken way of prioritizing smaller TXQ indices and ignoring the traffic
815 * class other than to determine whether the gate is open or not
816 */
817static struct sk_buff *taprio_dequeue_txq_priority(struct Qdisc *sch,
818 struct sched_entry *entry,
819 u32 gate_mask)
820{
821 struct net_device *dev = qdisc_dev(sch);
822 struct sk_buff *skb;
823 int i;
824
825 for (i = 0; i < dev->num_tx_queues; i++) {
826 skb = taprio_dequeue_from_txq(sch, i, entry, gate_mask);
827 if (skb)
828 return skb;
829 }
830
831 return NULL;
832}
833
834/* Will not be called in the full offload case, since the TX queues are
835 * attached to the Qdisc created using qdisc_create_dflt()
836 */
837static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
838{
839 struct taprio_sched *q = qdisc_priv(sch);
840 struct sk_buff *skb = NULL;
841 struct sched_entry *entry;
842 u32 gate_mask;
843
844 rcu_read_lock();
845 entry = rcu_dereference(q->current_entry);
846 /* if there's no entry, it means that the schedule didn't
847 * start yet, so force all gates to be open, this is in
848 * accordance to IEEE 802.1Qbv-2015 Section 8.6.9.4.5
849 * "AdminGateStates"
850 */
851 gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
852 if (!gate_mask)
853 goto done;
854
855 if (static_branch_unlikely(&taprio_have_broken_mqprio) &&
856 !static_branch_likely(&taprio_have_working_mqprio)) {
857 /* Single NIC kind which is broken */
858 skb = taprio_dequeue_txq_priority(sch, entry, gate_mask);
859 } else if (static_branch_likely(&taprio_have_working_mqprio) &&
860 !static_branch_unlikely(&taprio_have_broken_mqprio)) {
861 /* Single NIC kind which prioritizes properly */
862 skb = taprio_dequeue_tc_priority(sch, entry, gate_mask);
863 } else {
864 /* Mixed NIC kinds present in system, need dynamic testing */
865 if (q->broken_mqprio)
866 skb = taprio_dequeue_txq_priority(sch, entry, gate_mask);
867 else
868 skb = taprio_dequeue_tc_priority(sch, entry, gate_mask);
869 }
870
871done:
872 rcu_read_unlock();
873
874 return skb;
875}
876
877static bool should_restart_cycle(const struct sched_gate_list *oper,
878 const struct sched_entry *entry)
879{
880 if (list_is_last(&entry->list, &oper->entries))
881 return true;
882
883 if (ktime_compare(entry->end_time, oper->cycle_end_time) == 0)
884 return true;
885
886 return false;
887}
888
889static bool should_change_schedules(const struct sched_gate_list *admin,
890 const struct sched_gate_list *oper,
891 ktime_t end_time)
892{
893 ktime_t next_base_time, extension_time;
894
895 if (!admin)
896 return false;
897
898 next_base_time = sched_base_time(admin);
899
900 /* This is the simple case, the end_time would fall after
901 * the next schedule base_time.
902 */
903 if (ktime_compare(next_base_time, end_time) <= 0)
904 return true;
905
906 /* This is the cycle_time_extension case, if the end_time
907 * plus the amount that can be extended would fall after the
908 * next schedule base_time, we can extend the current schedule
909 * for that amount.
910 */
911 extension_time = ktime_add_ns(end_time, oper->cycle_time_extension);
912
913 /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
914 * how precisely the extension should be made. So after
915 * conformance testing, this logic may change.
916 */
917 if (ktime_compare(next_base_time, extension_time) <= 0)
918 return true;
919
920 return false;
921}
922
923static enum hrtimer_restart advance_sched(struct hrtimer *timer)
924{
925 struct taprio_sched *q = container_of(timer, struct taprio_sched,
926 advance_timer);
927 struct net_device *dev = qdisc_dev(q->root);
928 struct sched_gate_list *oper, *admin;
929 int num_tc = netdev_get_num_tc(dev);
930 struct sched_entry *entry, *next;
931 struct Qdisc *sch = q->root;
932 ktime_t end_time;
933 int tc;
934
935 spin_lock(&q->current_entry_lock);
936 entry = rcu_dereference_protected(q->current_entry,
937 lockdep_is_held(&q->current_entry_lock));
938 oper = rcu_dereference_protected(q->oper_sched,
939 lockdep_is_held(&q->current_entry_lock));
940 admin = rcu_dereference_protected(q->admin_sched,
941 lockdep_is_held(&q->current_entry_lock));
942
943 if (!oper)
944 switch_schedules(q, &admin, &oper);
945
946 /* This can happen in two cases: 1. this is the very first run
947 * of this function (i.e. we weren't running any schedule
948 * previously); 2. The previous schedule just ended. The first
949 * entry of all schedules are pre-calculated during the
950 * schedule initialization.
951 */
952 if (unlikely(!entry || entry->end_time == oper->base_time)) {
953 next = list_first_entry(&oper->entries, struct sched_entry,
954 list);
955 end_time = next->end_time;
956 goto first_run;
957 }
958
959 if (should_restart_cycle(oper, entry)) {
960 next = list_first_entry(&oper->entries, struct sched_entry,
961 list);
962 oper->cycle_end_time = ktime_add_ns(oper->cycle_end_time,
963 oper->cycle_time);
964 } else {
965 next = list_next_entry(entry, list);
966 }
967
968 end_time = ktime_add_ns(entry->end_time, next->interval);
969 end_time = min_t(ktime_t, end_time, oper->cycle_end_time);
970
971 for (tc = 0; tc < num_tc; tc++) {
972 if (next->gate_duration[tc] == oper->cycle_time)
973 next->gate_close_time[tc] = KTIME_MAX;
974 else
975 next->gate_close_time[tc] = ktime_add_ns(entry->end_time,
976 next->gate_duration[tc]);
977 }
978
979 if (should_change_schedules(admin, oper, end_time)) {
980 /* Set things so the next time this runs, the new
981 * schedule runs.
982 */
983 end_time = sched_base_time(admin);
984 switch_schedules(q, &admin, &oper);
985 }
986
987 next->end_time = end_time;
988 taprio_set_budgets(q, oper, next);
989
990first_run:
991 rcu_assign_pointer(q->current_entry, next);
992 spin_unlock(&q->current_entry_lock);
993
994 hrtimer_set_expires(&q->advance_timer, end_time);
995
996 rcu_read_lock();
997 __netif_schedule(sch);
998 rcu_read_unlock();
999
1000 return HRTIMER_RESTART;
1001}
1002
1003static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
1004 [TCA_TAPRIO_SCHED_ENTRY_INDEX] = { .type = NLA_U32 },
1005 [TCA_TAPRIO_SCHED_ENTRY_CMD] = { .type = NLA_U8 },
1006 [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
1007 [TCA_TAPRIO_SCHED_ENTRY_INTERVAL] = { .type = NLA_U32 },
1008};
1009
1010static const struct nla_policy taprio_tc_policy[TCA_TAPRIO_TC_ENTRY_MAX + 1] = {
1011 [TCA_TAPRIO_TC_ENTRY_INDEX] = { .type = NLA_U32 },
1012 [TCA_TAPRIO_TC_ENTRY_MAX_SDU] = { .type = NLA_U32 },
1013 [TCA_TAPRIO_TC_ENTRY_FP] = NLA_POLICY_RANGE(NLA_U32,
1014 TC_FP_EXPRESS,
1015 TC_FP_PREEMPTIBLE),
1016};
1017
1018static const struct netlink_range_validation_signed taprio_cycle_time_range = {
1019 .min = 0,
1020 .max = INT_MAX,
1021};
1022
1023static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
1024 [TCA_TAPRIO_ATTR_PRIOMAP] = {
1025 .len = sizeof(struct tc_mqprio_qopt)
1026 },
1027 [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST] = { .type = NLA_NESTED },
1028 [TCA_TAPRIO_ATTR_SCHED_BASE_TIME] = { .type = NLA_S64 },
1029 [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY] = { .type = NLA_NESTED },
1030 [TCA_TAPRIO_ATTR_SCHED_CLOCKID] = { .type = NLA_S32 },
1031 [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME] =
1032 NLA_POLICY_FULL_RANGE_SIGNED(NLA_S64, &taprio_cycle_time_range),
1033 [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
1034 [TCA_TAPRIO_ATTR_FLAGS] = { .type = NLA_U32 },
1035 [TCA_TAPRIO_ATTR_TXTIME_DELAY] = { .type = NLA_U32 },
1036 [TCA_TAPRIO_ATTR_TC_ENTRY] = { .type = NLA_NESTED },
1037};
1038
1039static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
1040 struct sched_entry *entry,
1041 struct netlink_ext_ack *extack)
1042{
1043 int min_duration = length_to_duration(q, ETH_ZLEN);
1044 u32 interval = 0;
1045
1046 if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
1047 entry->command = nla_get_u8(
1048 tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
1049
1050 if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
1051 entry->gate_mask = nla_get_u32(
1052 tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
1053
1054 if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
1055 interval = nla_get_u32(
1056 tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
1057
1058 /* The interval should allow at least the minimum ethernet
1059 * frame to go out.
1060 */
1061 if (interval < min_duration) {
1062 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
1063 return -EINVAL;
1064 }
1065
1066 entry->interval = interval;
1067
1068 return 0;
1069}
1070
1071static int parse_sched_entry(struct taprio_sched *q, struct nlattr *n,
1072 struct sched_entry *entry, int index,
1073 struct netlink_ext_ack *extack)
1074{
1075 struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
1076 int err;
1077
1078 err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
1079 entry_policy, NULL);
1080 if (err < 0) {
1081 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
1082 return -EINVAL;
1083 }
1084
1085 entry->index = index;
1086
1087 return fill_sched_entry(q, tb, entry, extack);
1088}
1089
1090static int parse_sched_list(struct taprio_sched *q, struct nlattr *list,
1091 struct sched_gate_list *sched,
1092 struct netlink_ext_ack *extack)
1093{
1094 struct nlattr *n;
1095 int err, rem;
1096 int i = 0;
1097
1098 if (!list)
1099 return -EINVAL;
1100
1101 nla_for_each_nested(n, list, rem) {
1102 struct sched_entry *entry;
1103
1104 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
1105 NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
1106 continue;
1107 }
1108
1109 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1110 if (!entry) {
1111 NL_SET_ERR_MSG(extack, "Not enough memory for entry");
1112 return -ENOMEM;
1113 }
1114
1115 err = parse_sched_entry(q, n, entry, i, extack);
1116 if (err < 0) {
1117 kfree(entry);
1118 return err;
1119 }
1120
1121 list_add_tail(&entry->list, &sched->entries);
1122 i++;
1123 }
1124
1125 sched->num_entries = i;
1126
1127 return i;
1128}
1129
1130static int parse_taprio_schedule(struct taprio_sched *q, struct nlattr **tb,
1131 struct sched_gate_list *new,
1132 struct netlink_ext_ack *extack)
1133{
1134 int err = 0;
1135
1136 if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
1137 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
1138 return -ENOTSUPP;
1139 }
1140
1141 if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
1142 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
1143
1144 if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
1145 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
1146
1147 if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
1148 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
1149
1150 if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
1151 err = parse_sched_list(q, tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST],
1152 new, extack);
1153 if (err < 0)
1154 return err;
1155
1156 if (!new->cycle_time) {
1157 struct sched_entry *entry;
1158 ktime_t cycle = 0;
1159
1160 list_for_each_entry(entry, &new->entries, list)
1161 cycle = ktime_add_ns(cycle, entry->interval);
1162
1163 if (!cycle) {
1164 NL_SET_ERR_MSG(extack, "'cycle_time' can never be 0");
1165 return -EINVAL;
1166 }
1167
1168 if (cycle < 0 || cycle > INT_MAX) {
1169 NL_SET_ERR_MSG(extack, "'cycle_time' is too big");
1170 return -EINVAL;
1171 }
1172
1173 new->cycle_time = cycle;
1174 }
1175
1176 taprio_calculate_gate_durations(q, new);
1177
1178 return 0;
1179}
1180
1181static int taprio_parse_mqprio_opt(struct net_device *dev,
1182 struct tc_mqprio_qopt *qopt,
1183 struct netlink_ext_ack *extack,
1184 u32 taprio_flags)
1185{
1186 bool allow_overlapping_txqs = TXTIME_ASSIST_IS_ENABLED(taprio_flags);
1187
1188 if (!qopt && !dev->num_tc) {
1189 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
1190 return -EINVAL;
1191 }
1192
1193 /* If num_tc is already set, it means that the user already
1194 * configured the mqprio part
1195 */
1196 if (dev->num_tc)
1197 return 0;
1198
1199 /* taprio imposes that traffic classes map 1:n to tx queues */
1200 if (qopt->num_tc > dev->num_tx_queues) {
1201 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
1202 return -EINVAL;
1203 }
1204
1205 /* For some reason, in txtime-assist mode, we allow TXQ ranges for
1206 * different TCs to overlap, and just validate the TXQ ranges.
1207 */
1208 return mqprio_validate_qopt(dev, qopt, true, allow_overlapping_txqs,
1209 extack);
1210}
1211
1212static int taprio_get_start_time(struct Qdisc *sch,
1213 struct sched_gate_list *sched,
1214 ktime_t *start)
1215{
1216 struct taprio_sched *q = qdisc_priv(sch);
1217 ktime_t now, base, cycle;
1218 s64 n;
1219
1220 base = sched_base_time(sched);
1221 now = taprio_get_time(q);
1222
1223 if (ktime_after(base, now)) {
1224 *start = base;
1225 return 0;
1226 }
1227
1228 cycle = sched->cycle_time;
1229
1230 /* The qdisc is expected to have at least one sched_entry. Moreover,
1231 * any entry must have 'interval' > 0. Thus if the cycle time is zero,
1232 * something went really wrong. In that case, we should warn about this
1233 * inconsistent state and return error.
1234 */
1235 if (WARN_ON(!cycle))
1236 return -EFAULT;
1237
1238 /* Schedule the start time for the beginning of the next
1239 * cycle.
1240 */
1241 n = div64_s64(ktime_sub_ns(now, base), cycle);
1242 *start = ktime_add_ns(base, (n + 1) * cycle);
1243 return 0;
1244}
1245
1246static void setup_first_end_time(struct taprio_sched *q,
1247 struct sched_gate_list *sched, ktime_t base)
1248{
1249 struct net_device *dev = qdisc_dev(q->root);
1250 int num_tc = netdev_get_num_tc(dev);
1251 struct sched_entry *first;
1252 ktime_t cycle;
1253 int tc;
1254
1255 first = list_first_entry(&sched->entries,
1256 struct sched_entry, list);
1257
1258 cycle = sched->cycle_time;
1259
1260 /* FIXME: find a better place to do this */
1261 sched->cycle_end_time = ktime_add_ns(base, cycle);
1262
1263 first->end_time = ktime_add_ns(base, first->interval);
1264 taprio_set_budgets(q, sched, first);
1265
1266 for (tc = 0; tc < num_tc; tc++) {
1267 if (first->gate_duration[tc] == sched->cycle_time)
1268 first->gate_close_time[tc] = KTIME_MAX;
1269 else
1270 first->gate_close_time[tc] = ktime_add_ns(base, first->gate_duration[tc]);
1271 }
1272
1273 rcu_assign_pointer(q->current_entry, NULL);
1274}
1275
1276static void taprio_start_sched(struct Qdisc *sch,
1277 ktime_t start, struct sched_gate_list *new)
1278{
1279 struct taprio_sched *q = qdisc_priv(sch);
1280 ktime_t expires;
1281
1282 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1283 return;
1284
1285 expires = hrtimer_get_expires(&q->advance_timer);
1286 if (expires == 0)
1287 expires = KTIME_MAX;
1288
1289 /* If the new schedule starts before the next expiration, we
1290 * reprogram it to the earliest one, so we change the admin
1291 * schedule to the operational one at the right time.
1292 */
1293 start = min_t(ktime_t, start, expires);
1294
1295 hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1296}
1297
1298static void taprio_set_picos_per_byte(struct net_device *dev,
1299 struct taprio_sched *q)
1300{
1301 struct ethtool_link_ksettings ecmd;
1302 int speed = SPEED_10;
1303 int picos_per_byte;
1304 int err;
1305
1306 err = __ethtool_get_link_ksettings(dev, &ecmd);
1307 if (err < 0)
1308 goto skip;
1309
1310 if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1311 speed = ecmd.base.speed;
1312
1313skip:
1314 picos_per_byte = (USEC_PER_SEC * 8) / speed;
1315
1316 atomic64_set(&q->picos_per_byte, picos_per_byte);
1317 netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1318 dev->name, (long long)atomic64_read(&q->picos_per_byte),
1319 ecmd.base.speed);
1320}
1321
1322static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1323 void *ptr)
1324{
1325 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1326 struct sched_gate_list *oper, *admin;
1327 struct qdisc_size_table *stab;
1328 struct taprio_sched *q;
1329
1330 ASSERT_RTNL();
1331
1332 if (event != NETDEV_UP && event != NETDEV_CHANGE)
1333 return NOTIFY_DONE;
1334
1335 list_for_each_entry(q, &taprio_list, taprio_list) {
1336 if (dev != qdisc_dev(q->root))
1337 continue;
1338
1339 taprio_set_picos_per_byte(dev, q);
1340
1341 stab = rtnl_dereference(q->root->stab);
1342
1343 oper = rtnl_dereference(q->oper_sched);
1344 if (oper)
1345 taprio_update_queue_max_sdu(q, oper, stab);
1346
1347 admin = rtnl_dereference(q->admin_sched);
1348 if (admin)
1349 taprio_update_queue_max_sdu(q, admin, stab);
1350
1351 break;
1352 }
1353
1354 return NOTIFY_DONE;
1355}
1356
1357static void setup_txtime(struct taprio_sched *q,
1358 struct sched_gate_list *sched, ktime_t base)
1359{
1360 struct sched_entry *entry;
1361 u64 interval = 0;
1362
1363 list_for_each_entry(entry, &sched->entries, list) {
1364 entry->next_txtime = ktime_add_ns(base, interval);
1365 interval += entry->interval;
1366 }
1367}
1368
1369static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1370{
1371 struct __tc_taprio_qopt_offload *__offload;
1372
1373 __offload = kzalloc(struct_size(__offload, offload.entries, num_entries),
1374 GFP_KERNEL);
1375 if (!__offload)
1376 return NULL;
1377
1378 refcount_set(&__offload->users, 1);
1379
1380 return &__offload->offload;
1381}
1382
1383struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1384 *offload)
1385{
1386 struct __tc_taprio_qopt_offload *__offload;
1387
1388 __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1389 offload);
1390
1391 refcount_inc(&__offload->users);
1392
1393 return offload;
1394}
1395EXPORT_SYMBOL_GPL(taprio_offload_get);
1396
1397void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1398{
1399 struct __tc_taprio_qopt_offload *__offload;
1400
1401 __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1402 offload);
1403
1404 if (!refcount_dec_and_test(&__offload->users))
1405 return;
1406
1407 kfree(__offload);
1408}
1409EXPORT_SYMBOL_GPL(taprio_offload_free);
1410
1411/* The function will only serve to keep the pointers to the "oper" and "admin"
1412 * schedules valid in relation to their base times, so when calling dump() the
1413 * users looks at the right schedules.
1414 * When using full offload, the admin configuration is promoted to oper at the
1415 * base_time in the PHC time domain. But because the system time is not
1416 * necessarily in sync with that, we can't just trigger a hrtimer to call
1417 * switch_schedules at the right hardware time.
1418 * At the moment we call this by hand right away from taprio, but in the future
1419 * it will be useful to create a mechanism for drivers to notify taprio of the
1420 * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1421 * This is left as TODO.
1422 */
1423static void taprio_offload_config_changed(struct taprio_sched *q)
1424{
1425 struct sched_gate_list *oper, *admin;
1426
1427 oper = rtnl_dereference(q->oper_sched);
1428 admin = rtnl_dereference(q->admin_sched);
1429
1430 switch_schedules(q, &admin, &oper);
1431}
1432
1433static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
1434{
1435 u32 i, queue_mask = 0;
1436
1437 for (i = 0; i < dev->num_tc; i++) {
1438 u32 offset, count;
1439
1440 if (!(tc_mask & BIT(i)))
1441 continue;
1442
1443 offset = dev->tc_to_txq[i].offset;
1444 count = dev->tc_to_txq[i].count;
1445
1446 queue_mask |= GENMASK(offset + count - 1, offset);
1447 }
1448
1449 return queue_mask;
1450}
1451
1452static void taprio_sched_to_offload(struct net_device *dev,
1453 struct sched_gate_list *sched,
1454 struct tc_taprio_qopt_offload *offload,
1455 const struct tc_taprio_caps *caps)
1456{
1457 struct sched_entry *entry;
1458 int i = 0;
1459
1460 offload->base_time = sched->base_time;
1461 offload->cycle_time = sched->cycle_time;
1462 offload->cycle_time_extension = sched->cycle_time_extension;
1463
1464 list_for_each_entry(entry, &sched->entries, list) {
1465 struct tc_taprio_sched_entry *e = &offload->entries[i];
1466
1467 e->command = entry->command;
1468 e->interval = entry->interval;
1469 if (caps->gate_mask_per_txq)
1470 e->gate_mask = tc_map_to_queue_mask(dev,
1471 entry->gate_mask);
1472 else
1473 e->gate_mask = entry->gate_mask;
1474
1475 i++;
1476 }
1477
1478 offload->num_entries = i;
1479}
1480
1481static void taprio_detect_broken_mqprio(struct taprio_sched *q)
1482{
1483 struct net_device *dev = qdisc_dev(q->root);
1484 struct tc_taprio_caps caps;
1485
1486 qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1487 &caps, sizeof(caps));
1488
1489 q->broken_mqprio = caps.broken_mqprio;
1490 if (q->broken_mqprio)
1491 static_branch_inc(&taprio_have_broken_mqprio);
1492 else
1493 static_branch_inc(&taprio_have_working_mqprio);
1494
1495 q->detected_mqprio = true;
1496}
1497
1498static void taprio_cleanup_broken_mqprio(struct taprio_sched *q)
1499{
1500 if (!q->detected_mqprio)
1501 return;
1502
1503 if (q->broken_mqprio)
1504 static_branch_dec(&taprio_have_broken_mqprio);
1505 else
1506 static_branch_dec(&taprio_have_working_mqprio);
1507}
1508
1509static int taprio_enable_offload(struct net_device *dev,
1510 struct taprio_sched *q,
1511 struct sched_gate_list *sched,
1512 struct netlink_ext_ack *extack)
1513{
1514 const struct net_device_ops *ops = dev->netdev_ops;
1515 struct tc_taprio_qopt_offload *offload;
1516 struct tc_taprio_caps caps;
1517 int tc, err = 0;
1518
1519 if (!ops->ndo_setup_tc) {
1520 NL_SET_ERR_MSG(extack,
1521 "Device does not support taprio offload");
1522 return -EOPNOTSUPP;
1523 }
1524
1525 qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1526 &caps, sizeof(caps));
1527
1528 if (!caps.supports_queue_max_sdu) {
1529 for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
1530 if (q->max_sdu[tc]) {
1531 NL_SET_ERR_MSG_MOD(extack,
1532 "Device does not handle queueMaxSDU");
1533 return -EOPNOTSUPP;
1534 }
1535 }
1536 }
1537
1538 offload = taprio_offload_alloc(sched->num_entries);
1539 if (!offload) {
1540 NL_SET_ERR_MSG(extack,
1541 "Not enough memory for enabling offload mode");
1542 return -ENOMEM;
1543 }
1544 offload->cmd = TAPRIO_CMD_REPLACE;
1545 offload->extack = extack;
1546 mqprio_qopt_reconstruct(dev, &offload->mqprio.qopt);
1547 offload->mqprio.extack = extack;
1548 taprio_sched_to_offload(dev, sched, offload, &caps);
1549 mqprio_fp_to_offload(q->fp, &offload->mqprio);
1550
1551 for (tc = 0; tc < TC_MAX_QUEUE; tc++)
1552 offload->max_sdu[tc] = q->max_sdu[tc];
1553
1554 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1555 if (err < 0) {
1556 NL_SET_ERR_MSG_WEAK(extack,
1557 "Device failed to setup taprio offload");
1558 goto done;
1559 }
1560
1561 q->offloaded = true;
1562
1563done:
1564 /* The offload structure may linger around via a reference taken by the
1565 * device driver, so clear up the netlink extack pointer so that the
1566 * driver isn't tempted to dereference data which stopped being valid
1567 */
1568 offload->extack = NULL;
1569 offload->mqprio.extack = NULL;
1570 taprio_offload_free(offload);
1571
1572 return err;
1573}
1574
1575static int taprio_disable_offload(struct net_device *dev,
1576 struct taprio_sched *q,
1577 struct netlink_ext_ack *extack)
1578{
1579 const struct net_device_ops *ops = dev->netdev_ops;
1580 struct tc_taprio_qopt_offload *offload;
1581 int err;
1582
1583 if (!q->offloaded)
1584 return 0;
1585
1586 offload = taprio_offload_alloc(0);
1587 if (!offload) {
1588 NL_SET_ERR_MSG(extack,
1589 "Not enough memory to disable offload mode");
1590 return -ENOMEM;
1591 }
1592 offload->cmd = TAPRIO_CMD_DESTROY;
1593
1594 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1595 if (err < 0) {
1596 NL_SET_ERR_MSG(extack,
1597 "Device failed to disable offload");
1598 goto out;
1599 }
1600
1601 q->offloaded = false;
1602
1603out:
1604 taprio_offload_free(offload);
1605
1606 return err;
1607}
1608
1609/* If full offload is enabled, the only possible clockid is the net device's
1610 * PHC. For that reason, specifying a clockid through netlink is incorrect.
1611 * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1612 * in sync with the specified clockid via a user space daemon such as phc2sys.
1613 * For both software taprio and txtime-assist, the clockid is used for the
1614 * hrtimer that advances the schedule and hence mandatory.
1615 */
1616static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1617 struct netlink_ext_ack *extack)
1618{
1619 struct taprio_sched *q = qdisc_priv(sch);
1620 struct net_device *dev = qdisc_dev(sch);
1621 int err = -EINVAL;
1622
1623 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1624 const struct ethtool_ops *ops = dev->ethtool_ops;
1625 struct ethtool_ts_info info = {
1626 .cmd = ETHTOOL_GET_TS_INFO,
1627 .phc_index = -1,
1628 };
1629
1630 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1631 NL_SET_ERR_MSG(extack,
1632 "The 'clockid' cannot be specified for full offload");
1633 goto out;
1634 }
1635
1636 if (ops && ops->get_ts_info)
1637 err = ops->get_ts_info(dev, &info);
1638
1639 if (err || info.phc_index < 0) {
1640 NL_SET_ERR_MSG(extack,
1641 "Device does not have a PTP clock");
1642 err = -ENOTSUPP;
1643 goto out;
1644 }
1645 } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1646 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1647 enum tk_offsets tk_offset;
1648
1649 /* We only support static clockids and we don't allow
1650 * for it to be modified after the first init.
1651 */
1652 if (clockid < 0 ||
1653 (q->clockid != -1 && q->clockid != clockid)) {
1654 NL_SET_ERR_MSG(extack,
1655 "Changing the 'clockid' of a running schedule is not supported");
1656 err = -ENOTSUPP;
1657 goto out;
1658 }
1659
1660 switch (clockid) {
1661 case CLOCK_REALTIME:
1662 tk_offset = TK_OFFS_REAL;
1663 break;
1664 case CLOCK_MONOTONIC:
1665 tk_offset = TK_OFFS_MAX;
1666 break;
1667 case CLOCK_BOOTTIME:
1668 tk_offset = TK_OFFS_BOOT;
1669 break;
1670 case CLOCK_TAI:
1671 tk_offset = TK_OFFS_TAI;
1672 break;
1673 default:
1674 NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1675 err = -EINVAL;
1676 goto out;
1677 }
1678 /* This pairs with READ_ONCE() in taprio_mono_to_any */
1679 WRITE_ONCE(q->tk_offset, tk_offset);
1680
1681 q->clockid = clockid;
1682 } else {
1683 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1684 goto out;
1685 }
1686
1687 /* Everything went ok, return success. */
1688 err = 0;
1689
1690out:
1691 return err;
1692}
1693
1694static int taprio_parse_tc_entry(struct Qdisc *sch,
1695 struct nlattr *opt,
1696 u32 max_sdu[TC_QOPT_MAX_QUEUE],
1697 u32 fp[TC_QOPT_MAX_QUEUE],
1698 unsigned long *seen_tcs,
1699 struct netlink_ext_ack *extack)
1700{
1701 struct nlattr *tb[TCA_TAPRIO_TC_ENTRY_MAX + 1] = { };
1702 struct net_device *dev = qdisc_dev(sch);
1703 int err, tc;
1704 u32 val;
1705
1706 err = nla_parse_nested(tb, TCA_TAPRIO_TC_ENTRY_MAX, opt,
1707 taprio_tc_policy, extack);
1708 if (err < 0)
1709 return err;
1710
1711 if (!tb[TCA_TAPRIO_TC_ENTRY_INDEX]) {
1712 NL_SET_ERR_MSG_MOD(extack, "TC entry index missing");
1713 return -EINVAL;
1714 }
1715
1716 tc = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_INDEX]);
1717 if (tc >= TC_QOPT_MAX_QUEUE) {
1718 NL_SET_ERR_MSG_MOD(extack, "TC entry index out of range");
1719 return -ERANGE;
1720 }
1721
1722 if (*seen_tcs & BIT(tc)) {
1723 NL_SET_ERR_MSG_MOD(extack, "Duplicate TC entry");
1724 return -EINVAL;
1725 }
1726
1727 *seen_tcs |= BIT(tc);
1728
1729 if (tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]) {
1730 val = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]);
1731 if (val > dev->max_mtu) {
1732 NL_SET_ERR_MSG_MOD(extack, "TC max SDU exceeds device max MTU");
1733 return -ERANGE;
1734 }
1735
1736 max_sdu[tc] = val;
1737 }
1738
1739 if (tb[TCA_TAPRIO_TC_ENTRY_FP])
1740 fp[tc] = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_FP]);
1741
1742 return 0;
1743}
1744
1745static int taprio_parse_tc_entries(struct Qdisc *sch,
1746 struct nlattr *opt,
1747 struct netlink_ext_ack *extack)
1748{
1749 struct taprio_sched *q = qdisc_priv(sch);
1750 struct net_device *dev = qdisc_dev(sch);
1751 u32 max_sdu[TC_QOPT_MAX_QUEUE];
1752 bool have_preemption = false;
1753 unsigned long seen_tcs = 0;
1754 u32 fp[TC_QOPT_MAX_QUEUE];
1755 struct nlattr *n;
1756 int tc, rem;
1757 int err = 0;
1758
1759 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1760 max_sdu[tc] = q->max_sdu[tc];
1761 fp[tc] = q->fp[tc];
1762 }
1763
1764 nla_for_each_nested(n, opt, rem) {
1765 if (nla_type(n) != TCA_TAPRIO_ATTR_TC_ENTRY)
1766 continue;
1767
1768 err = taprio_parse_tc_entry(sch, n, max_sdu, fp, &seen_tcs,
1769 extack);
1770 if (err)
1771 return err;
1772 }
1773
1774 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1775 q->max_sdu[tc] = max_sdu[tc];
1776 q->fp[tc] = fp[tc];
1777 if (fp[tc] != TC_FP_EXPRESS)
1778 have_preemption = true;
1779 }
1780
1781 if (have_preemption) {
1782 if (!FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1783 NL_SET_ERR_MSG(extack,
1784 "Preemption only supported with full offload");
1785 return -EOPNOTSUPP;
1786 }
1787
1788 if (!ethtool_dev_mm_supported(dev)) {
1789 NL_SET_ERR_MSG(extack,
1790 "Device does not support preemption");
1791 return -EOPNOTSUPP;
1792 }
1793 }
1794
1795 return err;
1796}
1797
1798static int taprio_mqprio_cmp(const struct net_device *dev,
1799 const struct tc_mqprio_qopt *mqprio)
1800{
1801 int i;
1802
1803 if (!mqprio || mqprio->num_tc != dev->num_tc)
1804 return -1;
1805
1806 for (i = 0; i < mqprio->num_tc; i++)
1807 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1808 dev->tc_to_txq[i].offset != mqprio->offset[i])
1809 return -1;
1810
1811 for (i = 0; i <= TC_BITMASK; i++)
1812 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1813 return -1;
1814
1815 return 0;
1816}
1817
1818/* The semantics of the 'flags' argument in relation to 'change()'
1819 * requests, are interpreted following two rules (which are applied in
1820 * this order): (1) an omitted 'flags' argument is interpreted as
1821 * zero; (2) the 'flags' of a "running" taprio instance cannot be
1822 * changed.
1823 */
1824static int taprio_new_flags(const struct nlattr *attr, u32 old,
1825 struct netlink_ext_ack *extack)
1826{
1827 u32 new = 0;
1828
1829 if (attr)
1830 new = nla_get_u32(attr);
1831
1832 if (old != TAPRIO_FLAGS_INVALID && old != new) {
1833 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1834 return -EOPNOTSUPP;
1835 }
1836
1837 if (!taprio_flags_valid(new)) {
1838 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1839 return -EINVAL;
1840 }
1841
1842 return new;
1843}
1844
1845static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1846 struct netlink_ext_ack *extack)
1847{
1848 struct qdisc_size_table *stab = rtnl_dereference(sch->stab);
1849 struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1850 struct sched_gate_list *oper, *admin, *new_admin;
1851 struct taprio_sched *q = qdisc_priv(sch);
1852 struct net_device *dev = qdisc_dev(sch);
1853 struct tc_mqprio_qopt *mqprio = NULL;
1854 unsigned long flags;
1855 ktime_t start;
1856 int i, err;
1857
1858 err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1859 taprio_policy, extack);
1860 if (err < 0)
1861 return err;
1862
1863 if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1864 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1865
1866 err = taprio_new_flags(tb[TCA_TAPRIO_ATTR_FLAGS],
1867 q->flags, extack);
1868 if (err < 0)
1869 return err;
1870
1871 q->flags = err;
1872
1873 err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1874 if (err < 0)
1875 return err;
1876
1877 err = taprio_parse_tc_entries(sch, opt, extack);
1878 if (err)
1879 return err;
1880
1881 new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1882 if (!new_admin) {
1883 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1884 return -ENOMEM;
1885 }
1886 INIT_LIST_HEAD(&new_admin->entries);
1887
1888 oper = rtnl_dereference(q->oper_sched);
1889 admin = rtnl_dereference(q->admin_sched);
1890
1891 /* no changes - no new mqprio settings */
1892 if (!taprio_mqprio_cmp(dev, mqprio))
1893 mqprio = NULL;
1894
1895 if (mqprio && (oper || admin)) {
1896 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1897 err = -ENOTSUPP;
1898 goto free_sched;
1899 }
1900
1901 if (mqprio) {
1902 err = netdev_set_num_tc(dev, mqprio->num_tc);
1903 if (err)
1904 goto free_sched;
1905 for (i = 0; i < mqprio->num_tc; i++) {
1906 netdev_set_tc_queue(dev, i,
1907 mqprio->count[i],
1908 mqprio->offset[i]);
1909 q->cur_txq[i] = mqprio->offset[i];
1910 }
1911
1912 /* Always use supplied priority mappings */
1913 for (i = 0; i <= TC_BITMASK; i++)
1914 netdev_set_prio_tc_map(dev, i,
1915 mqprio->prio_tc_map[i]);
1916 }
1917
1918 err = parse_taprio_schedule(q, tb, new_admin, extack);
1919 if (err < 0)
1920 goto free_sched;
1921
1922 if (new_admin->num_entries == 0) {
1923 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1924 err = -EINVAL;
1925 goto free_sched;
1926 }
1927
1928 err = taprio_parse_clockid(sch, tb, extack);
1929 if (err < 0)
1930 goto free_sched;
1931
1932 taprio_set_picos_per_byte(dev, q);
1933 taprio_update_queue_max_sdu(q, new_admin, stab);
1934
1935 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1936 err = taprio_enable_offload(dev, q, new_admin, extack);
1937 else
1938 err = taprio_disable_offload(dev, q, extack);
1939 if (err)
1940 goto free_sched;
1941
1942 /* Protects against enqueue()/dequeue() */
1943 spin_lock_bh(qdisc_lock(sch));
1944
1945 if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1946 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1947 NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1948 err = -EINVAL;
1949 goto unlock;
1950 }
1951
1952 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1953 }
1954
1955 if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1956 !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1957 !hrtimer_active(&q->advance_timer)) {
1958 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1959 q->advance_timer.function = advance_sched;
1960 }
1961
1962 err = taprio_get_start_time(sch, new_admin, &start);
1963 if (err < 0) {
1964 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1965 goto unlock;
1966 }
1967
1968 setup_txtime(q, new_admin, start);
1969
1970 if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1971 if (!oper) {
1972 rcu_assign_pointer(q->oper_sched, new_admin);
1973 err = 0;
1974 new_admin = NULL;
1975 goto unlock;
1976 }
1977
1978 rcu_assign_pointer(q->admin_sched, new_admin);
1979 if (admin)
1980 call_rcu(&admin->rcu, taprio_free_sched_cb);
1981 } else {
1982 setup_first_end_time(q, new_admin, start);
1983
1984 /* Protects against advance_sched() */
1985 spin_lock_irqsave(&q->current_entry_lock, flags);
1986
1987 taprio_start_sched(sch, start, new_admin);
1988
1989 rcu_assign_pointer(q->admin_sched, new_admin);
1990 if (admin)
1991 call_rcu(&admin->rcu, taprio_free_sched_cb);
1992
1993 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1994
1995 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1996 taprio_offload_config_changed(q);
1997 }
1998
1999 new_admin = NULL;
2000 err = 0;
2001
2002 if (!stab)
2003 NL_SET_ERR_MSG_MOD(extack,
2004 "Size table not specified, frame length estimations may be inaccurate");
2005
2006unlock:
2007 spin_unlock_bh(qdisc_lock(sch));
2008
2009free_sched:
2010 if (new_admin)
2011 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
2012
2013 return err;
2014}
2015
2016static void taprio_reset(struct Qdisc *sch)
2017{
2018 struct taprio_sched *q = qdisc_priv(sch);
2019 struct net_device *dev = qdisc_dev(sch);
2020 int i;
2021
2022 hrtimer_cancel(&q->advance_timer);
2023
2024 if (q->qdiscs) {
2025 for (i = 0; i < dev->num_tx_queues; i++)
2026 if (q->qdiscs[i])
2027 qdisc_reset(q->qdiscs[i]);
2028 }
2029}
2030
2031static void taprio_destroy(struct Qdisc *sch)
2032{
2033 struct taprio_sched *q = qdisc_priv(sch);
2034 struct net_device *dev = qdisc_dev(sch);
2035 struct sched_gate_list *oper, *admin;
2036 unsigned int i;
2037
2038 list_del(&q->taprio_list);
2039
2040 /* Note that taprio_reset() might not be called if an error
2041 * happens in qdisc_create(), after taprio_init() has been called.
2042 */
2043 hrtimer_cancel(&q->advance_timer);
2044 qdisc_synchronize(sch);
2045
2046 taprio_disable_offload(dev, q, NULL);
2047
2048 if (q->qdiscs) {
2049 for (i = 0; i < dev->num_tx_queues; i++)
2050 qdisc_put(q->qdiscs[i]);
2051
2052 kfree(q->qdiscs);
2053 }
2054 q->qdiscs = NULL;
2055
2056 netdev_reset_tc(dev);
2057
2058 oper = rtnl_dereference(q->oper_sched);
2059 admin = rtnl_dereference(q->admin_sched);
2060
2061 if (oper)
2062 call_rcu(&oper->rcu, taprio_free_sched_cb);
2063
2064 if (admin)
2065 call_rcu(&admin->rcu, taprio_free_sched_cb);
2066
2067 taprio_cleanup_broken_mqprio(q);
2068}
2069
2070static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
2071 struct netlink_ext_ack *extack)
2072{
2073 struct taprio_sched *q = qdisc_priv(sch);
2074 struct net_device *dev = qdisc_dev(sch);
2075 int i, tc;
2076
2077 spin_lock_init(&q->current_entry_lock);
2078
2079 hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
2080 q->advance_timer.function = advance_sched;
2081
2082 q->root = sch;
2083
2084 /* We only support static clockids. Use an invalid value as default
2085 * and get the valid one on taprio_change().
2086 */
2087 q->clockid = -1;
2088 q->flags = TAPRIO_FLAGS_INVALID;
2089
2090 list_add(&q->taprio_list, &taprio_list);
2091
2092 if (sch->parent != TC_H_ROOT) {
2093 NL_SET_ERR_MSG_MOD(extack, "Can only be attached as root qdisc");
2094 return -EOPNOTSUPP;
2095 }
2096
2097 if (!netif_is_multiqueue(dev)) {
2098 NL_SET_ERR_MSG_MOD(extack, "Multi-queue device is required");
2099 return -EOPNOTSUPP;
2100 }
2101
2102 q->qdiscs = kcalloc(dev->num_tx_queues, sizeof(q->qdiscs[0]),
2103 GFP_KERNEL);
2104 if (!q->qdiscs)
2105 return -ENOMEM;
2106
2107 if (!opt)
2108 return -EINVAL;
2109
2110 for (i = 0; i < dev->num_tx_queues; i++) {
2111 struct netdev_queue *dev_queue;
2112 struct Qdisc *qdisc;
2113
2114 dev_queue = netdev_get_tx_queue(dev, i);
2115 qdisc = qdisc_create_dflt(dev_queue,
2116 &pfifo_qdisc_ops,
2117 TC_H_MAKE(TC_H_MAJ(sch->handle),
2118 TC_H_MIN(i + 1)),
2119 extack);
2120 if (!qdisc)
2121 return -ENOMEM;
2122
2123 if (i < dev->real_num_tx_queues)
2124 qdisc_hash_add(qdisc, false);
2125
2126 q->qdiscs[i] = qdisc;
2127 }
2128
2129 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++)
2130 q->fp[tc] = TC_FP_EXPRESS;
2131
2132 taprio_detect_broken_mqprio(q);
2133
2134 return taprio_change(sch, opt, extack);
2135}
2136
2137static void taprio_attach(struct Qdisc *sch)
2138{
2139 struct taprio_sched *q = qdisc_priv(sch);
2140 struct net_device *dev = qdisc_dev(sch);
2141 unsigned int ntx;
2142
2143 /* Attach underlying qdisc */
2144 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
2145 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
2146 struct Qdisc *old, *dev_queue_qdisc;
2147
2148 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2149 struct Qdisc *qdisc = q->qdiscs[ntx];
2150
2151 /* In offload mode, the root taprio qdisc is bypassed
2152 * and the netdev TX queues see the children directly
2153 */
2154 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2155 dev_queue_qdisc = qdisc;
2156 } else {
2157 /* In software mode, attach the root taprio qdisc
2158 * to all netdev TX queues, so that dev_qdisc_enqueue()
2159 * goes through taprio_enqueue().
2160 */
2161 dev_queue_qdisc = sch;
2162 }
2163 old = dev_graft_qdisc(dev_queue, dev_queue_qdisc);
2164 /* The qdisc's refcount requires to be elevated once
2165 * for each netdev TX queue it is grafted onto
2166 */
2167 qdisc_refcount_inc(dev_queue_qdisc);
2168 if (old)
2169 qdisc_put(old);
2170 }
2171}
2172
2173static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
2174 unsigned long cl)
2175{
2176 struct net_device *dev = qdisc_dev(sch);
2177 unsigned long ntx = cl - 1;
2178
2179 if (ntx >= dev->num_tx_queues)
2180 return NULL;
2181
2182 return netdev_get_tx_queue(dev, ntx);
2183}
2184
2185static int taprio_graft(struct Qdisc *sch, unsigned long cl,
2186 struct Qdisc *new, struct Qdisc **old,
2187 struct netlink_ext_ack *extack)
2188{
2189 struct taprio_sched *q = qdisc_priv(sch);
2190 struct net_device *dev = qdisc_dev(sch);
2191 struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2192
2193 if (!dev_queue)
2194 return -EINVAL;
2195
2196 if (dev->flags & IFF_UP)
2197 dev_deactivate(dev);
2198
2199 /* In offload mode, the child Qdisc is directly attached to the netdev
2200 * TX queue, and thus, we need to keep its refcount elevated in order
2201 * to counteract qdisc_graft()'s call to qdisc_put() once per TX queue.
2202 * However, save the reference to the new qdisc in the private array in
2203 * both software and offload cases, to have an up-to-date reference to
2204 * our children.
2205 */
2206 *old = q->qdiscs[cl - 1];
2207 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2208 WARN_ON_ONCE(dev_graft_qdisc(dev_queue, new) != *old);
2209 if (new)
2210 qdisc_refcount_inc(new);
2211 if (*old)
2212 qdisc_put(*old);
2213 }
2214
2215 q->qdiscs[cl - 1] = new;
2216 if (new)
2217 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2218
2219 if (dev->flags & IFF_UP)
2220 dev_activate(dev);
2221
2222 return 0;
2223}
2224
2225static int dump_entry(struct sk_buff *msg,
2226 const struct sched_entry *entry)
2227{
2228 struct nlattr *item;
2229
2230 item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
2231 if (!item)
2232 return -ENOSPC;
2233
2234 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
2235 goto nla_put_failure;
2236
2237 if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
2238 goto nla_put_failure;
2239
2240 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
2241 entry->gate_mask))
2242 goto nla_put_failure;
2243
2244 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
2245 entry->interval))
2246 goto nla_put_failure;
2247
2248 return nla_nest_end(msg, item);
2249
2250nla_put_failure:
2251 nla_nest_cancel(msg, item);
2252 return -1;
2253}
2254
2255static int dump_schedule(struct sk_buff *msg,
2256 const struct sched_gate_list *root)
2257{
2258 struct nlattr *entry_list;
2259 struct sched_entry *entry;
2260
2261 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
2262 root->base_time, TCA_TAPRIO_PAD))
2263 return -1;
2264
2265 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
2266 root->cycle_time, TCA_TAPRIO_PAD))
2267 return -1;
2268
2269 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
2270 root->cycle_time_extension, TCA_TAPRIO_PAD))
2271 return -1;
2272
2273 entry_list = nla_nest_start_noflag(msg,
2274 TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
2275 if (!entry_list)
2276 goto error_nest;
2277
2278 list_for_each_entry(entry, &root->entries, list) {
2279 if (dump_entry(msg, entry) < 0)
2280 goto error_nest;
2281 }
2282
2283 nla_nest_end(msg, entry_list);
2284 return 0;
2285
2286error_nest:
2287 nla_nest_cancel(msg, entry_list);
2288 return -1;
2289}
2290
2291static int taprio_dump_tc_entries(struct sk_buff *skb,
2292 struct taprio_sched *q,
2293 struct sched_gate_list *sched)
2294{
2295 struct nlattr *n;
2296 int tc;
2297
2298 for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
2299 n = nla_nest_start(skb, TCA_TAPRIO_ATTR_TC_ENTRY);
2300 if (!n)
2301 return -EMSGSIZE;
2302
2303 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_INDEX, tc))
2304 goto nla_put_failure;
2305
2306 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_MAX_SDU,
2307 sched->max_sdu[tc]))
2308 goto nla_put_failure;
2309
2310 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_FP, q->fp[tc]))
2311 goto nla_put_failure;
2312
2313 nla_nest_end(skb, n);
2314 }
2315
2316 return 0;
2317
2318nla_put_failure:
2319 nla_nest_cancel(skb, n);
2320 return -EMSGSIZE;
2321}
2322
2323static int taprio_put_stat(struct sk_buff *skb, u64 val, u16 attrtype)
2324{
2325 if (val == TAPRIO_STAT_NOT_SET)
2326 return 0;
2327 if (nla_put_u64_64bit(skb, attrtype, val, TCA_TAPRIO_OFFLOAD_STATS_PAD))
2328 return -EMSGSIZE;
2329 return 0;
2330}
2331
2332static int taprio_dump_xstats(struct Qdisc *sch, struct gnet_dump *d,
2333 struct tc_taprio_qopt_offload *offload,
2334 struct tc_taprio_qopt_stats *stats)
2335{
2336 struct net_device *dev = qdisc_dev(sch);
2337 const struct net_device_ops *ops;
2338 struct sk_buff *skb = d->skb;
2339 struct nlattr *xstats;
2340 int err;
2341
2342 ops = qdisc_dev(sch)->netdev_ops;
2343
2344 /* FIXME I could use qdisc_offload_dump_helper(), but that messes
2345 * with sch->flags depending on whether the device reports taprio
2346 * stats, and I'm not sure whether that's a good idea, considering
2347 * that stats are optional to the offload itself
2348 */
2349 if (!ops->ndo_setup_tc)
2350 return 0;
2351
2352 memset(stats, 0xff, sizeof(*stats));
2353
2354 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
2355 if (err == -EOPNOTSUPP)
2356 return 0;
2357 if (err)
2358 return err;
2359
2360 xstats = nla_nest_start(skb, TCA_STATS_APP);
2361 if (!xstats)
2362 goto err;
2363
2364 if (taprio_put_stat(skb, stats->window_drops,
2365 TCA_TAPRIO_OFFLOAD_STATS_WINDOW_DROPS) ||
2366 taprio_put_stat(skb, stats->tx_overruns,
2367 TCA_TAPRIO_OFFLOAD_STATS_TX_OVERRUNS))
2368 goto err_cancel;
2369
2370 nla_nest_end(skb, xstats);
2371
2372 return 0;
2373
2374err_cancel:
2375 nla_nest_cancel(skb, xstats);
2376err:
2377 return -EMSGSIZE;
2378}
2379
2380static int taprio_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
2381{
2382 struct tc_taprio_qopt_offload offload = {
2383 .cmd = TAPRIO_CMD_STATS,
2384 };
2385
2386 return taprio_dump_xstats(sch, d, &offload, &offload.stats);
2387}
2388
2389static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
2390{
2391 struct taprio_sched *q = qdisc_priv(sch);
2392 struct net_device *dev = qdisc_dev(sch);
2393 struct sched_gate_list *oper, *admin;
2394 struct tc_mqprio_qopt opt = { 0 };
2395 struct nlattr *nest, *sched_nest;
2396
2397 oper = rtnl_dereference(q->oper_sched);
2398 admin = rtnl_dereference(q->admin_sched);
2399
2400 mqprio_qopt_reconstruct(dev, &opt);
2401
2402 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2403 if (!nest)
2404 goto start_error;
2405
2406 if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
2407 goto options_error;
2408
2409 if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
2410 nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
2411 goto options_error;
2412
2413 if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
2414 goto options_error;
2415
2416 if (q->txtime_delay &&
2417 nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
2418 goto options_error;
2419
2420 if (oper && taprio_dump_tc_entries(skb, q, oper))
2421 goto options_error;
2422
2423 if (oper && dump_schedule(skb, oper))
2424 goto options_error;
2425
2426 if (!admin)
2427 goto done;
2428
2429 sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
2430 if (!sched_nest)
2431 goto options_error;
2432
2433 if (dump_schedule(skb, admin))
2434 goto admin_error;
2435
2436 nla_nest_end(skb, sched_nest);
2437
2438done:
2439 return nla_nest_end(skb, nest);
2440
2441admin_error:
2442 nla_nest_cancel(skb, sched_nest);
2443
2444options_error:
2445 nla_nest_cancel(skb, nest);
2446
2447start_error:
2448 return -ENOSPC;
2449}
2450
2451static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
2452{
2453 struct taprio_sched *q = qdisc_priv(sch);
2454 struct net_device *dev = qdisc_dev(sch);
2455 unsigned int ntx = cl - 1;
2456
2457 if (ntx >= dev->num_tx_queues)
2458 return NULL;
2459
2460 return q->qdiscs[ntx];
2461}
2462
2463static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
2464{
2465 unsigned int ntx = TC_H_MIN(classid);
2466
2467 if (!taprio_queue_get(sch, ntx))
2468 return 0;
2469 return ntx;
2470}
2471
2472static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
2473 struct sk_buff *skb, struct tcmsg *tcm)
2474{
2475 struct Qdisc *child = taprio_leaf(sch, cl);
2476
2477 tcm->tcm_parent = TC_H_ROOT;
2478 tcm->tcm_handle |= TC_H_MIN(cl);
2479 tcm->tcm_info = child->handle;
2480
2481 return 0;
2482}
2483
2484static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
2485 struct gnet_dump *d)
2486 __releases(d->lock)
2487 __acquires(d->lock)
2488{
2489 struct Qdisc *child = taprio_leaf(sch, cl);
2490 struct tc_taprio_qopt_offload offload = {
2491 .cmd = TAPRIO_CMD_QUEUE_STATS,
2492 .queue_stats = {
2493 .queue = cl - 1,
2494 },
2495 };
2496
2497 if (gnet_stats_copy_basic(d, NULL, &child->bstats, true) < 0 ||
2498 qdisc_qstats_copy(d, child) < 0)
2499 return -1;
2500
2501 return taprio_dump_xstats(sch, d, &offload, &offload.queue_stats.stats);
2502}
2503
2504static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2505{
2506 struct net_device *dev = qdisc_dev(sch);
2507 unsigned long ntx;
2508
2509 if (arg->stop)
2510 return;
2511
2512 arg->count = arg->skip;
2513 for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
2514 if (!tc_qdisc_stats_dump(sch, ntx + 1, arg))
2515 break;
2516 }
2517}
2518
2519static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
2520 struct tcmsg *tcm)
2521{
2522 return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
2523}
2524
2525static const struct Qdisc_class_ops taprio_class_ops = {
2526 .graft = taprio_graft,
2527 .leaf = taprio_leaf,
2528 .find = taprio_find,
2529 .walk = taprio_walk,
2530 .dump = taprio_dump_class,
2531 .dump_stats = taprio_dump_class_stats,
2532 .select_queue = taprio_select_queue,
2533};
2534
2535static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
2536 .cl_ops = &taprio_class_ops,
2537 .id = "taprio",
2538 .priv_size = sizeof(struct taprio_sched),
2539 .init = taprio_init,
2540 .change = taprio_change,
2541 .destroy = taprio_destroy,
2542 .reset = taprio_reset,
2543 .attach = taprio_attach,
2544 .peek = taprio_peek,
2545 .dequeue = taprio_dequeue,
2546 .enqueue = taprio_enqueue,
2547 .dump = taprio_dump,
2548 .dump_stats = taprio_dump_stats,
2549 .owner = THIS_MODULE,
2550};
2551
2552static struct notifier_block taprio_device_notifier = {
2553 .notifier_call = taprio_dev_notifier,
2554};
2555
2556static int __init taprio_module_init(void)
2557{
2558 int err = register_netdevice_notifier(&taprio_device_notifier);
2559
2560 if (err)
2561 return err;
2562
2563 return register_qdisc(&taprio_qdisc_ops);
2564}
2565
2566static void __exit taprio_module_exit(void)
2567{
2568 unregister_qdisc(&taprio_qdisc_ops);
2569 unregister_netdevice_notifier(&taprio_device_notifier);
2570}
2571
2572module_init(taprio_module_init);
2573module_exit(taprio_module_exit);
2574MODULE_LICENSE("GPL");
2575MODULE_DESCRIPTION("Time Aware Priority qdisc");