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_SUPPORTED_FLAGS \
44 (TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST | TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)
45#define TAPRIO_FLAGS_INVALID U32_MAX
46
47struct sched_entry {
48 /* Durations between this GCL entry and the GCL entry where the
49 * respective traffic class gate closes
50 */
51 u64 gate_duration[TC_MAX_QUEUE];
52 atomic_t budget[TC_MAX_QUEUE];
53 /* The qdisc makes some effort so that no packet leaves
54 * after this time
55 */
56 ktime_t gate_close_time[TC_MAX_QUEUE];
57 struct list_head list;
58 /* Used to calculate when to advance the schedule */
59 ktime_t end_time;
60 ktime_t next_txtime;
61 int index;
62 u32 gate_mask;
63 u32 interval;
64 u8 command;
65};
66
67struct sched_gate_list {
68 /* Longest non-zero contiguous gate durations per traffic class,
69 * or 0 if a traffic class gate never opens during the schedule.
70 */
71 u64 max_open_gate_duration[TC_MAX_QUEUE];
72 u32 max_frm_len[TC_MAX_QUEUE]; /* for the fast path */
73 u32 max_sdu[TC_MAX_QUEUE]; /* for dump */
74 struct rcu_head rcu;
75 struct list_head entries;
76 size_t num_entries;
77 ktime_t cycle_end_time;
78 s64 cycle_time;
79 s64 cycle_time_extension;
80 s64 base_time;
81};
82
83struct taprio_sched {
84 struct Qdisc **qdiscs;
85 struct Qdisc *root;
86 u32 flags;
87 enum tk_offsets tk_offset;
88 int clockid;
89 bool offloaded;
90 bool detected_mqprio;
91 bool broken_mqprio;
92 atomic64_t picos_per_byte; /* Using picoseconds because for 10Gbps+
93 * speeds it's sub-nanoseconds per byte
94 */
95
96 /* Protects the update side of the RCU protected current_entry */
97 spinlock_t current_entry_lock;
98 struct sched_entry __rcu *current_entry;
99 struct sched_gate_list __rcu *oper_sched;
100 struct sched_gate_list __rcu *admin_sched;
101 struct hrtimer advance_timer;
102 struct list_head taprio_list;
103 int cur_txq[TC_MAX_QUEUE];
104 u32 max_sdu[TC_MAX_QUEUE]; /* save info from the user */
105 u32 fp[TC_QOPT_MAX_QUEUE]; /* only for dump and offloading */
106 u32 txtime_delay;
107};
108
109struct __tc_taprio_qopt_offload {
110 refcount_t users;
111 struct tc_taprio_qopt_offload offload;
112};
113
114static void taprio_calculate_gate_durations(struct taprio_sched *q,
115 struct sched_gate_list *sched)
116{
117 struct net_device *dev = qdisc_dev(q->root);
118 int num_tc = netdev_get_num_tc(dev);
119 struct sched_entry *entry, *cur;
120 int tc;
121
122 list_for_each_entry(entry, &sched->entries, list) {
123 u32 gates_still_open = entry->gate_mask;
124
125 /* For each traffic class, calculate each open gate duration,
126 * starting at this schedule entry and ending at the schedule
127 * entry containing a gate close event for that TC.
128 */
129 cur = entry;
130
131 do {
132 if (!gates_still_open)
133 break;
134
135 for (tc = 0; tc < num_tc; tc++) {
136 if (!(gates_still_open & BIT(tc)))
137 continue;
138
139 if (cur->gate_mask & BIT(tc))
140 entry->gate_duration[tc] += cur->interval;
141 else
142 gates_still_open &= ~BIT(tc);
143 }
144
145 cur = list_next_entry_circular(cur, &sched->entries, list);
146 } while (cur != entry);
147
148 /* Keep track of the maximum gate duration for each traffic
149 * class, taking care to not confuse a traffic class which is
150 * temporarily closed with one that is always closed.
151 */
152 for (tc = 0; tc < num_tc; tc++)
153 if (entry->gate_duration[tc] &&
154 sched->max_open_gate_duration[tc] < entry->gate_duration[tc])
155 sched->max_open_gate_duration[tc] = entry->gate_duration[tc];
156 }
157}
158
159static bool taprio_entry_allows_tx(ktime_t skb_end_time,
160 struct sched_entry *entry, int tc)
161{
162 return ktime_before(skb_end_time, entry->gate_close_time[tc]);
163}
164
165static ktime_t sched_base_time(const struct sched_gate_list *sched)
166{
167 if (!sched)
168 return KTIME_MAX;
169
170 return ns_to_ktime(sched->base_time);
171}
172
173static ktime_t taprio_mono_to_any(const struct taprio_sched *q, ktime_t mono)
174{
175 /* This pairs with WRITE_ONCE() in taprio_parse_clockid() */
176 enum tk_offsets tk_offset = READ_ONCE(q->tk_offset);
177
178 switch (tk_offset) {
179 case TK_OFFS_MAX:
180 return mono;
181 default:
182 return ktime_mono_to_any(mono, tk_offset);
183 }
184}
185
186static ktime_t taprio_get_time(const struct taprio_sched *q)
187{
188 return taprio_mono_to_any(q, ktime_get());
189}
190
191static void taprio_free_sched_cb(struct rcu_head *head)
192{
193 struct sched_gate_list *sched = container_of(head, struct sched_gate_list, rcu);
194 struct sched_entry *entry, *n;
195
196 list_for_each_entry_safe(entry, n, &sched->entries, list) {
197 list_del(&entry->list);
198 kfree(entry);
199 }
200
201 kfree(sched);
202}
203
204static void switch_schedules(struct taprio_sched *q,
205 struct sched_gate_list **admin,
206 struct sched_gate_list **oper)
207{
208 rcu_assign_pointer(q->oper_sched, *admin);
209 rcu_assign_pointer(q->admin_sched, NULL);
210
211 if (*oper)
212 call_rcu(&(*oper)->rcu, taprio_free_sched_cb);
213
214 *oper = *admin;
215 *admin = NULL;
216}
217
218/* Get how much time has been already elapsed in the current cycle. */
219static s32 get_cycle_time_elapsed(struct sched_gate_list *sched, ktime_t time)
220{
221 ktime_t time_since_sched_start;
222 s32 time_elapsed;
223
224 time_since_sched_start = ktime_sub(time, sched->base_time);
225 div_s64_rem(time_since_sched_start, sched->cycle_time, &time_elapsed);
226
227 return time_elapsed;
228}
229
230static ktime_t get_interval_end_time(struct sched_gate_list *sched,
231 struct sched_gate_list *admin,
232 struct sched_entry *entry,
233 ktime_t intv_start)
234{
235 s32 cycle_elapsed = get_cycle_time_elapsed(sched, intv_start);
236 ktime_t intv_end, cycle_ext_end, cycle_end;
237
238 cycle_end = ktime_add_ns(intv_start, sched->cycle_time - cycle_elapsed);
239 intv_end = ktime_add_ns(intv_start, entry->interval);
240 cycle_ext_end = ktime_add(cycle_end, sched->cycle_time_extension);
241
242 if (ktime_before(intv_end, cycle_end))
243 return intv_end;
244 else if (admin && admin != sched &&
245 ktime_after(admin->base_time, cycle_end) &&
246 ktime_before(admin->base_time, cycle_ext_end))
247 return admin->base_time;
248 else
249 return cycle_end;
250}
251
252static int length_to_duration(struct taprio_sched *q, int len)
253{
254 return div_u64(len * atomic64_read(&q->picos_per_byte), PSEC_PER_NSEC);
255}
256
257static int duration_to_length(struct taprio_sched *q, u64 duration)
258{
259 return div_u64(duration * PSEC_PER_NSEC, atomic64_read(&q->picos_per_byte));
260}
261
262/* Sets sched->max_sdu[] and sched->max_frm_len[] to the minimum between the
263 * q->max_sdu[] requested by the user and the max_sdu dynamically determined by
264 * the maximum open gate durations at the given link speed.
265 */
266static void taprio_update_queue_max_sdu(struct taprio_sched *q,
267 struct sched_gate_list *sched,
268 struct qdisc_size_table *stab)
269{
270 struct net_device *dev = qdisc_dev(q->root);
271 int num_tc = netdev_get_num_tc(dev);
272 u32 max_sdu_from_user;
273 u32 max_sdu_dynamic;
274 u32 max_sdu;
275 int tc;
276
277 for (tc = 0; tc < num_tc; tc++) {
278 max_sdu_from_user = q->max_sdu[tc] ?: U32_MAX;
279
280 /* TC gate never closes => keep the queueMaxSDU
281 * selected by the user
282 */
283 if (sched->max_open_gate_duration[tc] == sched->cycle_time) {
284 max_sdu_dynamic = U32_MAX;
285 } else {
286 u32 max_frm_len;
287
288 max_frm_len = duration_to_length(q, sched->max_open_gate_duration[tc]);
289 /* Compensate for L1 overhead from size table,
290 * but don't let the frame size go negative
291 */
292 if (stab) {
293 max_frm_len -= stab->szopts.overhead;
294 max_frm_len = max_t(int, max_frm_len,
295 dev->hard_header_len + 1);
296 }
297 max_sdu_dynamic = max_frm_len - dev->hard_header_len;
298 if (max_sdu_dynamic > dev->max_mtu)
299 max_sdu_dynamic = U32_MAX;
300 }
301
302 max_sdu = min(max_sdu_dynamic, max_sdu_from_user);
303
304 if (max_sdu != U32_MAX) {
305 sched->max_frm_len[tc] = max_sdu + dev->hard_header_len;
306 sched->max_sdu[tc] = max_sdu;
307 } else {
308 sched->max_frm_len[tc] = U32_MAX; /* never oversized */
309 sched->max_sdu[tc] = 0;
310 }
311 }
312}
313
314/* Returns the entry corresponding to next available interval. If
315 * validate_interval is set, it only validates whether the timestamp occurs
316 * when the gate corresponding to the skb's traffic class is open.
317 */
318static struct sched_entry *find_entry_to_transmit(struct sk_buff *skb,
319 struct Qdisc *sch,
320 struct sched_gate_list *sched,
321 struct sched_gate_list *admin,
322 ktime_t time,
323 ktime_t *interval_start,
324 ktime_t *interval_end,
325 bool validate_interval)
326{
327 ktime_t curr_intv_start, curr_intv_end, cycle_end, packet_transmit_time;
328 ktime_t earliest_txtime = KTIME_MAX, txtime, cycle, transmit_end_time;
329 struct sched_entry *entry = NULL, *entry_found = NULL;
330 struct taprio_sched *q = qdisc_priv(sch);
331 struct net_device *dev = qdisc_dev(sch);
332 bool entry_available = false;
333 s32 cycle_elapsed;
334 int tc, n;
335
336 tc = netdev_get_prio_tc_map(dev, skb->priority);
337 packet_transmit_time = length_to_duration(q, qdisc_pkt_len(skb));
338
339 *interval_start = 0;
340 *interval_end = 0;
341
342 if (!sched)
343 return NULL;
344
345 cycle = sched->cycle_time;
346 cycle_elapsed = get_cycle_time_elapsed(sched, time);
347 curr_intv_end = ktime_sub_ns(time, cycle_elapsed);
348 cycle_end = ktime_add_ns(curr_intv_end, cycle);
349
350 list_for_each_entry(entry, &sched->entries, list) {
351 curr_intv_start = curr_intv_end;
352 curr_intv_end = get_interval_end_time(sched, admin, entry,
353 curr_intv_start);
354
355 if (ktime_after(curr_intv_start, cycle_end))
356 break;
357
358 if (!(entry->gate_mask & BIT(tc)) ||
359 packet_transmit_time > entry->interval)
360 continue;
361
362 txtime = entry->next_txtime;
363
364 if (ktime_before(txtime, time) || validate_interval) {
365 transmit_end_time = ktime_add_ns(time, packet_transmit_time);
366 if ((ktime_before(curr_intv_start, time) &&
367 ktime_before(transmit_end_time, curr_intv_end)) ||
368 (ktime_after(curr_intv_start, time) && !validate_interval)) {
369 entry_found = entry;
370 *interval_start = curr_intv_start;
371 *interval_end = curr_intv_end;
372 break;
373 } else if (!entry_available && !validate_interval) {
374 /* Here, we are just trying to find out the
375 * first available interval in the next cycle.
376 */
377 entry_available = true;
378 entry_found = entry;
379 *interval_start = ktime_add_ns(curr_intv_start, cycle);
380 *interval_end = ktime_add_ns(curr_intv_end, cycle);
381 }
382 } else if (ktime_before(txtime, earliest_txtime) &&
383 !entry_available) {
384 earliest_txtime = txtime;
385 entry_found = entry;
386 n = div_s64(ktime_sub(txtime, curr_intv_start), cycle);
387 *interval_start = ktime_add(curr_intv_start, n * cycle);
388 *interval_end = ktime_add(curr_intv_end, n * cycle);
389 }
390 }
391
392 return entry_found;
393}
394
395static bool is_valid_interval(struct sk_buff *skb, struct Qdisc *sch)
396{
397 struct taprio_sched *q = qdisc_priv(sch);
398 struct sched_gate_list *sched, *admin;
399 ktime_t interval_start, interval_end;
400 struct sched_entry *entry;
401
402 rcu_read_lock();
403 sched = rcu_dereference(q->oper_sched);
404 admin = rcu_dereference(q->admin_sched);
405
406 entry = find_entry_to_transmit(skb, sch, sched, admin, skb->tstamp,
407 &interval_start, &interval_end, true);
408 rcu_read_unlock();
409
410 return entry;
411}
412
413/* This returns the tstamp value set by TCP in terms of the set clock. */
414static ktime_t get_tcp_tstamp(struct taprio_sched *q, struct sk_buff *skb)
415{
416 unsigned int offset = skb_network_offset(skb);
417 const struct ipv6hdr *ipv6h;
418 const struct iphdr *iph;
419 struct ipv6hdr _ipv6h;
420
421 ipv6h = skb_header_pointer(skb, offset, sizeof(_ipv6h), &_ipv6h);
422 if (!ipv6h)
423 return 0;
424
425 if (ipv6h->version == 4) {
426 iph = (struct iphdr *)ipv6h;
427 offset += iph->ihl * 4;
428
429 /* special-case 6in4 tunnelling, as that is a common way to get
430 * v6 connectivity in the home
431 */
432 if (iph->protocol == IPPROTO_IPV6) {
433 ipv6h = skb_header_pointer(skb, offset,
434 sizeof(_ipv6h), &_ipv6h);
435
436 if (!ipv6h || ipv6h->nexthdr != IPPROTO_TCP)
437 return 0;
438 } else if (iph->protocol != IPPROTO_TCP) {
439 return 0;
440 }
441 } else if (ipv6h->version == 6 && ipv6h->nexthdr != IPPROTO_TCP) {
442 return 0;
443 }
444
445 return taprio_mono_to_any(q, skb->skb_mstamp_ns);
446}
447
448/* There are a few scenarios where we will have to modify the txtime from
449 * what is read from next_txtime in sched_entry. They are:
450 * 1. If txtime is in the past,
451 * a. The gate for the traffic class is currently open and packet can be
452 * transmitted before it closes, schedule the packet right away.
453 * b. If the gate corresponding to the traffic class is going to open later
454 * in the cycle, set the txtime of packet to the interval start.
455 * 2. If txtime is in the future, there are packets corresponding to the
456 * current traffic class waiting to be transmitted. So, the following
457 * possibilities exist:
458 * a. We can transmit the packet before the window containing the txtime
459 * closes.
460 * b. The window might close before the transmission can be completed
461 * successfully. So, schedule the packet in the next open window.
462 */
463static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)
464{
465 ktime_t transmit_end_time, interval_end, interval_start, tcp_tstamp;
466 struct taprio_sched *q = qdisc_priv(sch);
467 struct sched_gate_list *sched, *admin;
468 ktime_t minimum_time, now, txtime;
469 int len, packet_transmit_time;
470 struct sched_entry *entry;
471 bool sched_changed;
472
473 now = taprio_get_time(q);
474 minimum_time = ktime_add_ns(now, q->txtime_delay);
475
476 tcp_tstamp = get_tcp_tstamp(q, skb);
477 minimum_time = max_t(ktime_t, minimum_time, tcp_tstamp);
478
479 rcu_read_lock();
480 admin = rcu_dereference(q->admin_sched);
481 sched = rcu_dereference(q->oper_sched);
482 if (admin && ktime_after(minimum_time, admin->base_time))
483 switch_schedules(q, &admin, &sched);
484
485 /* Until the schedule starts, all the queues are open */
486 if (!sched || ktime_before(minimum_time, sched->base_time)) {
487 txtime = minimum_time;
488 goto done;
489 }
490
491 len = qdisc_pkt_len(skb);
492 packet_transmit_time = length_to_duration(q, len);
493
494 do {
495 sched_changed = false;
496
497 entry = find_entry_to_transmit(skb, sch, sched, admin,
498 minimum_time,
499 &interval_start, &interval_end,
500 false);
501 if (!entry) {
502 txtime = 0;
503 goto done;
504 }
505
506 txtime = entry->next_txtime;
507 txtime = max_t(ktime_t, txtime, minimum_time);
508 txtime = max_t(ktime_t, txtime, interval_start);
509
510 if (admin && admin != sched &&
511 ktime_after(txtime, admin->base_time)) {
512 sched = admin;
513 sched_changed = true;
514 continue;
515 }
516
517 transmit_end_time = ktime_add(txtime, packet_transmit_time);
518 minimum_time = transmit_end_time;
519
520 /* Update the txtime of current entry to the next time it's
521 * interval starts.
522 */
523 if (ktime_after(transmit_end_time, interval_end))
524 entry->next_txtime = ktime_add(interval_start, sched->cycle_time);
525 } while (sched_changed || ktime_after(transmit_end_time, interval_end));
526
527 entry->next_txtime = transmit_end_time;
528
529done:
530 rcu_read_unlock();
531 return txtime;
532}
533
534/* Devices with full offload are expected to honor this in hardware */
535static bool taprio_skb_exceeds_queue_max_sdu(struct Qdisc *sch,
536 struct sk_buff *skb)
537{
538 struct taprio_sched *q = qdisc_priv(sch);
539 struct net_device *dev = qdisc_dev(sch);
540 struct sched_gate_list *sched;
541 int prio = skb->priority;
542 bool exceeds = false;
543 u8 tc;
544
545 tc = netdev_get_prio_tc_map(dev, prio);
546
547 rcu_read_lock();
548 sched = rcu_dereference(q->oper_sched);
549 if (sched && skb->len > sched->max_frm_len[tc])
550 exceeds = true;
551 rcu_read_unlock();
552
553 return exceeds;
554}
555
556static int taprio_enqueue_one(struct sk_buff *skb, struct Qdisc *sch,
557 struct Qdisc *child, struct sk_buff **to_free)
558{
559 struct taprio_sched *q = qdisc_priv(sch);
560
561 /* sk_flags are only safe to use on full sockets. */
562 if (skb->sk && sk_fullsock(skb->sk) && sock_flag(skb->sk, SOCK_TXTIME)) {
563 if (!is_valid_interval(skb, sch))
564 return qdisc_drop(skb, sch, to_free);
565 } else if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
566 skb->tstamp = get_packet_txtime(skb, sch);
567 if (!skb->tstamp)
568 return qdisc_drop(skb, sch, to_free);
569 }
570
571 qdisc_qstats_backlog_inc(sch, skb);
572 sch->q.qlen++;
573
574 return qdisc_enqueue(skb, child, to_free);
575}
576
577static int taprio_enqueue_segmented(struct sk_buff *skb, struct Qdisc *sch,
578 struct Qdisc *child,
579 struct sk_buff **to_free)
580{
581 unsigned int slen = 0, numsegs = 0, len = qdisc_pkt_len(skb);
582 netdev_features_t features = netif_skb_features(skb);
583 struct sk_buff *segs, *nskb;
584 int ret;
585
586 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
587 if (IS_ERR_OR_NULL(segs))
588 return qdisc_drop(skb, sch, to_free);
589
590 skb_list_walk_safe(segs, segs, nskb) {
591 skb_mark_not_on_list(segs);
592 qdisc_skb_cb(segs)->pkt_len = segs->len;
593 slen += segs->len;
594
595 /* FIXME: we should be segmenting to a smaller size
596 * rather than dropping these
597 */
598 if (taprio_skb_exceeds_queue_max_sdu(sch, segs))
599 ret = qdisc_drop(segs, sch, to_free);
600 else
601 ret = taprio_enqueue_one(segs, sch, child, to_free);
602
603 if (ret != NET_XMIT_SUCCESS) {
604 if (net_xmit_drop_count(ret))
605 qdisc_qstats_drop(sch);
606 } else {
607 numsegs++;
608 }
609 }
610
611 if (numsegs > 1)
612 qdisc_tree_reduce_backlog(sch, 1 - numsegs, len - slen);
613 consume_skb(skb);
614
615 return numsegs > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
616}
617
618/* Will not be called in the full offload case, since the TX queues are
619 * attached to the Qdisc created using qdisc_create_dflt()
620 */
621static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,
622 struct sk_buff **to_free)
623{
624 struct taprio_sched *q = qdisc_priv(sch);
625 struct Qdisc *child;
626 int queue;
627
628 queue = skb_get_queue_mapping(skb);
629
630 child = q->qdiscs[queue];
631 if (unlikely(!child))
632 return qdisc_drop(skb, sch, to_free);
633
634 if (taprio_skb_exceeds_queue_max_sdu(sch, skb)) {
635 /* Large packets might not be transmitted when the transmission
636 * duration exceeds any configured interval. Therefore, segment
637 * the skb into smaller chunks. Drivers with full offload are
638 * expected to handle this in hardware.
639 */
640 if (skb_is_gso(skb))
641 return taprio_enqueue_segmented(skb, sch, child,
642 to_free);
643
644 return qdisc_drop(skb, sch, to_free);
645 }
646
647 return taprio_enqueue_one(skb, sch, child, to_free);
648}
649
650static struct sk_buff *taprio_peek(struct Qdisc *sch)
651{
652 WARN_ONCE(1, "taprio only supports operating as root qdisc, peek() not implemented");
653 return NULL;
654}
655
656static void taprio_set_budgets(struct taprio_sched *q,
657 struct sched_gate_list *sched,
658 struct sched_entry *entry)
659{
660 struct net_device *dev = qdisc_dev(q->root);
661 int num_tc = netdev_get_num_tc(dev);
662 int tc, budget;
663
664 for (tc = 0; tc < num_tc; tc++) {
665 /* Traffic classes which never close have infinite budget */
666 if (entry->gate_duration[tc] == sched->cycle_time)
667 budget = INT_MAX;
668 else
669 budget = div64_u64((u64)entry->gate_duration[tc] * PSEC_PER_NSEC,
670 atomic64_read(&q->picos_per_byte));
671
672 atomic_set(&entry->budget[tc], budget);
673 }
674}
675
676/* When an skb is sent, it consumes from the budget of all traffic classes */
677static int taprio_update_budgets(struct sched_entry *entry, size_t len,
678 int tc_consumed, int num_tc)
679{
680 int tc, budget, new_budget = 0;
681
682 for (tc = 0; tc < num_tc; tc++) {
683 budget = atomic_read(&entry->budget[tc]);
684 /* Don't consume from infinite budget */
685 if (budget == INT_MAX) {
686 if (tc == tc_consumed)
687 new_budget = budget;
688 continue;
689 }
690
691 if (tc == tc_consumed)
692 new_budget = atomic_sub_return(len, &entry->budget[tc]);
693 else
694 atomic_sub(len, &entry->budget[tc]);
695 }
696
697 return new_budget;
698}
699
700static struct sk_buff *taprio_dequeue_from_txq(struct Qdisc *sch, int txq,
701 struct sched_entry *entry,
702 u32 gate_mask)
703{
704 struct taprio_sched *q = qdisc_priv(sch);
705 struct net_device *dev = qdisc_dev(sch);
706 struct Qdisc *child = q->qdiscs[txq];
707 int num_tc = netdev_get_num_tc(dev);
708 struct sk_buff *skb;
709 ktime_t guard;
710 int prio;
711 int len;
712 u8 tc;
713
714 if (unlikely(!child))
715 return NULL;
716
717 if (TXTIME_ASSIST_IS_ENABLED(q->flags))
718 goto skip_peek_checks;
719
720 skb = child->ops->peek(child);
721 if (!skb)
722 return NULL;
723
724 prio = skb->priority;
725 tc = netdev_get_prio_tc_map(dev, prio);
726
727 if (!(gate_mask & BIT(tc)))
728 return NULL;
729
730 len = qdisc_pkt_len(skb);
731 guard = ktime_add_ns(taprio_get_time(q), length_to_duration(q, len));
732
733 /* In the case that there's no gate entry, there's no
734 * guard band ...
735 */
736 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
737 !taprio_entry_allows_tx(guard, entry, tc))
738 return NULL;
739
740 /* ... and no budget. */
741 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
742 taprio_update_budgets(entry, len, tc, num_tc) < 0)
743 return NULL;
744
745skip_peek_checks:
746 skb = child->ops->dequeue(child);
747 if (unlikely(!skb))
748 return NULL;
749
750 qdisc_bstats_update(sch, skb);
751 qdisc_qstats_backlog_dec(sch, skb);
752 sch->q.qlen--;
753
754 return skb;
755}
756
757static void taprio_next_tc_txq(struct net_device *dev, int tc, int *txq)
758{
759 int offset = dev->tc_to_txq[tc].offset;
760 int count = dev->tc_to_txq[tc].count;
761
762 (*txq)++;
763 if (*txq == offset + count)
764 *txq = offset;
765}
766
767/* Prioritize higher traffic classes, and select among TXQs belonging to the
768 * same TC using round robin
769 */
770static struct sk_buff *taprio_dequeue_tc_priority(struct Qdisc *sch,
771 struct sched_entry *entry,
772 u32 gate_mask)
773{
774 struct taprio_sched *q = qdisc_priv(sch);
775 struct net_device *dev = qdisc_dev(sch);
776 int num_tc = netdev_get_num_tc(dev);
777 struct sk_buff *skb;
778 int tc;
779
780 for (tc = num_tc - 1; tc >= 0; tc--) {
781 int first_txq = q->cur_txq[tc];
782
783 if (!(gate_mask & BIT(tc)))
784 continue;
785
786 do {
787 skb = taprio_dequeue_from_txq(sch, q->cur_txq[tc],
788 entry, gate_mask);
789
790 taprio_next_tc_txq(dev, tc, &q->cur_txq[tc]);
791
792 if (q->cur_txq[tc] >= dev->num_tx_queues)
793 q->cur_txq[tc] = first_txq;
794
795 if (skb)
796 return skb;
797 } while (q->cur_txq[tc] != first_txq);
798 }
799
800 return NULL;
801}
802
803/* Broken way of prioritizing smaller TXQ indices and ignoring the traffic
804 * class other than to determine whether the gate is open or not
805 */
806static struct sk_buff *taprio_dequeue_txq_priority(struct Qdisc *sch,
807 struct sched_entry *entry,
808 u32 gate_mask)
809{
810 struct net_device *dev = qdisc_dev(sch);
811 struct sk_buff *skb;
812 int i;
813
814 for (i = 0; i < dev->num_tx_queues; i++) {
815 skb = taprio_dequeue_from_txq(sch, i, entry, gate_mask);
816 if (skb)
817 return skb;
818 }
819
820 return NULL;
821}
822
823/* Will not be called in the full offload case, since the TX queues are
824 * attached to the Qdisc created using qdisc_create_dflt()
825 */
826static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
827{
828 struct taprio_sched *q = qdisc_priv(sch);
829 struct sk_buff *skb = NULL;
830 struct sched_entry *entry;
831 u32 gate_mask;
832
833 rcu_read_lock();
834 entry = rcu_dereference(q->current_entry);
835 /* if there's no entry, it means that the schedule didn't
836 * start yet, so force all gates to be open, this is in
837 * accordance to IEEE 802.1Qbv-2015 Section 8.6.9.4.5
838 * "AdminGateStates"
839 */
840 gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
841 if (!gate_mask)
842 goto done;
843
844 if (static_branch_unlikely(&taprio_have_broken_mqprio) &&
845 !static_branch_likely(&taprio_have_working_mqprio)) {
846 /* Single NIC kind which is broken */
847 skb = taprio_dequeue_txq_priority(sch, entry, gate_mask);
848 } else if (static_branch_likely(&taprio_have_working_mqprio) &&
849 !static_branch_unlikely(&taprio_have_broken_mqprio)) {
850 /* Single NIC kind which prioritizes properly */
851 skb = taprio_dequeue_tc_priority(sch, entry, gate_mask);
852 } else {
853 /* Mixed NIC kinds present in system, need dynamic testing */
854 if (q->broken_mqprio)
855 skb = taprio_dequeue_txq_priority(sch, entry, gate_mask);
856 else
857 skb = taprio_dequeue_tc_priority(sch, entry, gate_mask);
858 }
859
860done:
861 rcu_read_unlock();
862
863 return skb;
864}
865
866static bool should_restart_cycle(const struct sched_gate_list *oper,
867 const struct sched_entry *entry)
868{
869 if (list_is_last(&entry->list, &oper->entries))
870 return true;
871
872 if (ktime_compare(entry->end_time, oper->cycle_end_time) == 0)
873 return true;
874
875 return false;
876}
877
878static bool should_change_schedules(const struct sched_gate_list *admin,
879 const struct sched_gate_list *oper,
880 ktime_t end_time)
881{
882 ktime_t next_base_time, extension_time;
883
884 if (!admin)
885 return false;
886
887 next_base_time = sched_base_time(admin);
888
889 /* This is the simple case, the end_time would fall after
890 * the next schedule base_time.
891 */
892 if (ktime_compare(next_base_time, end_time) <= 0)
893 return true;
894
895 /* This is the cycle_time_extension case, if the end_time
896 * plus the amount that can be extended would fall after the
897 * next schedule base_time, we can extend the current schedule
898 * for that amount.
899 */
900 extension_time = ktime_add_ns(end_time, oper->cycle_time_extension);
901
902 /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
903 * how precisely the extension should be made. So after
904 * conformance testing, this logic may change.
905 */
906 if (ktime_compare(next_base_time, extension_time) <= 0)
907 return true;
908
909 return false;
910}
911
912static enum hrtimer_restart advance_sched(struct hrtimer *timer)
913{
914 struct taprio_sched *q = container_of(timer, struct taprio_sched,
915 advance_timer);
916 struct net_device *dev = qdisc_dev(q->root);
917 struct sched_gate_list *oper, *admin;
918 int num_tc = netdev_get_num_tc(dev);
919 struct sched_entry *entry, *next;
920 struct Qdisc *sch = q->root;
921 ktime_t end_time;
922 int tc;
923
924 spin_lock(&q->current_entry_lock);
925 entry = rcu_dereference_protected(q->current_entry,
926 lockdep_is_held(&q->current_entry_lock));
927 oper = rcu_dereference_protected(q->oper_sched,
928 lockdep_is_held(&q->current_entry_lock));
929 admin = rcu_dereference_protected(q->admin_sched,
930 lockdep_is_held(&q->current_entry_lock));
931
932 if (!oper)
933 switch_schedules(q, &admin, &oper);
934
935 /* This can happen in two cases: 1. this is the very first run
936 * of this function (i.e. we weren't running any schedule
937 * previously); 2. The previous schedule just ended. The first
938 * entry of all schedules are pre-calculated during the
939 * schedule initialization.
940 */
941 if (unlikely(!entry || entry->end_time == oper->base_time)) {
942 next = list_first_entry(&oper->entries, struct sched_entry,
943 list);
944 end_time = next->end_time;
945 goto first_run;
946 }
947
948 if (should_restart_cycle(oper, entry)) {
949 next = list_first_entry(&oper->entries, struct sched_entry,
950 list);
951 oper->cycle_end_time = ktime_add_ns(oper->cycle_end_time,
952 oper->cycle_time);
953 } else {
954 next = list_next_entry(entry, list);
955 }
956
957 end_time = ktime_add_ns(entry->end_time, next->interval);
958 end_time = min_t(ktime_t, end_time, oper->cycle_end_time);
959
960 for (tc = 0; tc < num_tc; tc++) {
961 if (next->gate_duration[tc] == oper->cycle_time)
962 next->gate_close_time[tc] = KTIME_MAX;
963 else
964 next->gate_close_time[tc] = ktime_add_ns(entry->end_time,
965 next->gate_duration[tc]);
966 }
967
968 if (should_change_schedules(admin, oper, end_time)) {
969 /* Set things so the next time this runs, the new
970 * schedule runs.
971 */
972 end_time = sched_base_time(admin);
973 switch_schedules(q, &admin, &oper);
974 }
975
976 next->end_time = end_time;
977 taprio_set_budgets(q, oper, next);
978
979first_run:
980 rcu_assign_pointer(q->current_entry, next);
981 spin_unlock(&q->current_entry_lock);
982
983 hrtimer_set_expires(&q->advance_timer, end_time);
984
985 rcu_read_lock();
986 __netif_schedule(sch);
987 rcu_read_unlock();
988
989 return HRTIMER_RESTART;
990}
991
992static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
993 [TCA_TAPRIO_SCHED_ENTRY_INDEX] = { .type = NLA_U32 },
994 [TCA_TAPRIO_SCHED_ENTRY_CMD] = { .type = NLA_U8 },
995 [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
996 [TCA_TAPRIO_SCHED_ENTRY_INTERVAL] = { .type = NLA_U32 },
997};
998
999static const struct nla_policy taprio_tc_policy[TCA_TAPRIO_TC_ENTRY_MAX + 1] = {
1000 [TCA_TAPRIO_TC_ENTRY_INDEX] = NLA_POLICY_MAX(NLA_U32,
1001 TC_QOPT_MAX_QUEUE),
1002 [TCA_TAPRIO_TC_ENTRY_MAX_SDU] = { .type = NLA_U32 },
1003 [TCA_TAPRIO_TC_ENTRY_FP] = NLA_POLICY_RANGE(NLA_U32,
1004 TC_FP_EXPRESS,
1005 TC_FP_PREEMPTIBLE),
1006};
1007
1008static const struct netlink_range_validation_signed taprio_cycle_time_range = {
1009 .min = 0,
1010 .max = INT_MAX,
1011};
1012
1013static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
1014 [TCA_TAPRIO_ATTR_PRIOMAP] = {
1015 .len = sizeof(struct tc_mqprio_qopt)
1016 },
1017 [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST] = { .type = NLA_NESTED },
1018 [TCA_TAPRIO_ATTR_SCHED_BASE_TIME] = { .type = NLA_S64 },
1019 [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY] = { .type = NLA_NESTED },
1020 [TCA_TAPRIO_ATTR_SCHED_CLOCKID] = { .type = NLA_S32 },
1021 [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME] =
1022 NLA_POLICY_FULL_RANGE_SIGNED(NLA_S64, &taprio_cycle_time_range),
1023 [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
1024 [TCA_TAPRIO_ATTR_FLAGS] =
1025 NLA_POLICY_MASK(NLA_U32, TAPRIO_SUPPORTED_FLAGS),
1026 [TCA_TAPRIO_ATTR_TXTIME_DELAY] = { .type = NLA_U32 },
1027 [TCA_TAPRIO_ATTR_TC_ENTRY] = { .type = NLA_NESTED },
1028};
1029
1030static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
1031 struct sched_entry *entry,
1032 struct netlink_ext_ack *extack)
1033{
1034 int min_duration = length_to_duration(q, ETH_ZLEN);
1035 u32 interval = 0;
1036
1037 if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
1038 entry->command = nla_get_u8(
1039 tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
1040
1041 if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
1042 entry->gate_mask = nla_get_u32(
1043 tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
1044
1045 if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
1046 interval = nla_get_u32(
1047 tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
1048
1049 /* The interval should allow at least the minimum ethernet
1050 * frame to go out.
1051 */
1052 if (interval < min_duration) {
1053 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
1054 return -EINVAL;
1055 }
1056
1057 entry->interval = interval;
1058
1059 return 0;
1060}
1061
1062static int parse_sched_entry(struct taprio_sched *q, struct nlattr *n,
1063 struct sched_entry *entry, int index,
1064 struct netlink_ext_ack *extack)
1065{
1066 struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
1067 int err;
1068
1069 err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
1070 entry_policy, NULL);
1071 if (err < 0) {
1072 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
1073 return -EINVAL;
1074 }
1075
1076 entry->index = index;
1077
1078 return fill_sched_entry(q, tb, entry, extack);
1079}
1080
1081static int parse_sched_list(struct taprio_sched *q, struct nlattr *list,
1082 struct sched_gate_list *sched,
1083 struct netlink_ext_ack *extack)
1084{
1085 struct nlattr *n;
1086 int err, rem;
1087 int i = 0;
1088
1089 if (!list)
1090 return -EINVAL;
1091
1092 nla_for_each_nested(n, list, rem) {
1093 struct sched_entry *entry;
1094
1095 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
1096 NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
1097 continue;
1098 }
1099
1100 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1101 if (!entry) {
1102 NL_SET_ERR_MSG(extack, "Not enough memory for entry");
1103 return -ENOMEM;
1104 }
1105
1106 err = parse_sched_entry(q, n, entry, i, extack);
1107 if (err < 0) {
1108 kfree(entry);
1109 return err;
1110 }
1111
1112 list_add_tail(&entry->list, &sched->entries);
1113 i++;
1114 }
1115
1116 sched->num_entries = i;
1117
1118 return i;
1119}
1120
1121static int parse_taprio_schedule(struct taprio_sched *q, struct nlattr **tb,
1122 struct sched_gate_list *new,
1123 struct netlink_ext_ack *extack)
1124{
1125 int err = 0;
1126
1127 if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
1128 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
1129 return -ENOTSUPP;
1130 }
1131
1132 if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
1133 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
1134
1135 if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
1136 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
1137
1138 if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
1139 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
1140
1141 if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
1142 err = parse_sched_list(q, tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST],
1143 new, extack);
1144 if (err < 0)
1145 return err;
1146
1147 if (!new->cycle_time) {
1148 struct sched_entry *entry;
1149 ktime_t cycle = 0;
1150
1151 list_for_each_entry(entry, &new->entries, list)
1152 cycle = ktime_add_ns(cycle, entry->interval);
1153
1154 if (cycle < 0 || cycle > INT_MAX) {
1155 NL_SET_ERR_MSG(extack, "'cycle_time' is too big");
1156 return -EINVAL;
1157 }
1158
1159 new->cycle_time = cycle;
1160 }
1161
1162 if (new->cycle_time < new->num_entries * length_to_duration(q, ETH_ZLEN)) {
1163 NL_SET_ERR_MSG(extack, "'cycle_time' is too small");
1164 return -EINVAL;
1165 }
1166
1167 taprio_calculate_gate_durations(q, new);
1168
1169 return 0;
1170}
1171
1172static int taprio_parse_mqprio_opt(struct net_device *dev,
1173 struct tc_mqprio_qopt *qopt,
1174 struct netlink_ext_ack *extack,
1175 u32 taprio_flags)
1176{
1177 bool allow_overlapping_txqs = TXTIME_ASSIST_IS_ENABLED(taprio_flags);
1178
1179 if (!qopt) {
1180 if (!dev->num_tc) {
1181 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
1182 return -EINVAL;
1183 }
1184 return 0;
1185 }
1186
1187 /* taprio imposes that traffic classes map 1:n to tx queues */
1188 if (qopt->num_tc > dev->num_tx_queues) {
1189 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
1190 return -EINVAL;
1191 }
1192
1193 /* For some reason, in txtime-assist mode, we allow TXQ ranges for
1194 * different TCs to overlap, and just validate the TXQ ranges.
1195 */
1196 return mqprio_validate_qopt(dev, qopt, true, allow_overlapping_txqs,
1197 extack);
1198}
1199
1200static int taprio_get_start_time(struct Qdisc *sch,
1201 struct sched_gate_list *sched,
1202 ktime_t *start)
1203{
1204 struct taprio_sched *q = qdisc_priv(sch);
1205 ktime_t now, base, cycle;
1206 s64 n;
1207
1208 base = sched_base_time(sched);
1209 now = taprio_get_time(q);
1210
1211 if (ktime_after(base, now)) {
1212 *start = base;
1213 return 0;
1214 }
1215
1216 cycle = sched->cycle_time;
1217
1218 /* The qdisc is expected to have at least one sched_entry. Moreover,
1219 * any entry must have 'interval' > 0. Thus if the cycle time is zero,
1220 * something went really wrong. In that case, we should warn about this
1221 * inconsistent state and return error.
1222 */
1223 if (WARN_ON(!cycle))
1224 return -EFAULT;
1225
1226 /* Schedule the start time for the beginning of the next
1227 * cycle.
1228 */
1229 n = div64_s64(ktime_sub_ns(now, base), cycle);
1230 *start = ktime_add_ns(base, (n + 1) * cycle);
1231 return 0;
1232}
1233
1234static void setup_first_end_time(struct taprio_sched *q,
1235 struct sched_gate_list *sched, ktime_t base)
1236{
1237 struct net_device *dev = qdisc_dev(q->root);
1238 int num_tc = netdev_get_num_tc(dev);
1239 struct sched_entry *first;
1240 ktime_t cycle;
1241 int tc;
1242
1243 first = list_first_entry(&sched->entries,
1244 struct sched_entry, list);
1245
1246 cycle = sched->cycle_time;
1247
1248 /* FIXME: find a better place to do this */
1249 sched->cycle_end_time = ktime_add_ns(base, cycle);
1250
1251 first->end_time = ktime_add_ns(base, first->interval);
1252 taprio_set_budgets(q, sched, first);
1253
1254 for (tc = 0; tc < num_tc; tc++) {
1255 if (first->gate_duration[tc] == sched->cycle_time)
1256 first->gate_close_time[tc] = KTIME_MAX;
1257 else
1258 first->gate_close_time[tc] = ktime_add_ns(base, first->gate_duration[tc]);
1259 }
1260
1261 rcu_assign_pointer(q->current_entry, NULL);
1262}
1263
1264static void taprio_start_sched(struct Qdisc *sch,
1265 ktime_t start, struct sched_gate_list *new)
1266{
1267 struct taprio_sched *q = qdisc_priv(sch);
1268 ktime_t expires;
1269
1270 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1271 return;
1272
1273 expires = hrtimer_get_expires(&q->advance_timer);
1274 if (expires == 0)
1275 expires = KTIME_MAX;
1276
1277 /* If the new schedule starts before the next expiration, we
1278 * reprogram it to the earliest one, so we change the admin
1279 * schedule to the operational one at the right time.
1280 */
1281 start = min_t(ktime_t, start, expires);
1282
1283 hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1284}
1285
1286static void taprio_set_picos_per_byte(struct net_device *dev,
1287 struct taprio_sched *q)
1288{
1289 struct ethtool_link_ksettings ecmd;
1290 int speed = SPEED_10;
1291 int picos_per_byte;
1292 int err;
1293
1294 err = __ethtool_get_link_ksettings(dev, &ecmd);
1295 if (err < 0)
1296 goto skip;
1297
1298 if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1299 speed = ecmd.base.speed;
1300
1301skip:
1302 picos_per_byte = (USEC_PER_SEC * 8) / speed;
1303
1304 atomic64_set(&q->picos_per_byte, picos_per_byte);
1305 netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1306 dev->name, (long long)atomic64_read(&q->picos_per_byte),
1307 ecmd.base.speed);
1308}
1309
1310static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1311 void *ptr)
1312{
1313 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1314 struct sched_gate_list *oper, *admin;
1315 struct qdisc_size_table *stab;
1316 struct taprio_sched *q;
1317
1318 ASSERT_RTNL();
1319
1320 if (event != NETDEV_UP && event != NETDEV_CHANGE)
1321 return NOTIFY_DONE;
1322
1323 list_for_each_entry(q, &taprio_list, taprio_list) {
1324 if (dev != qdisc_dev(q->root))
1325 continue;
1326
1327 taprio_set_picos_per_byte(dev, q);
1328
1329 stab = rtnl_dereference(q->root->stab);
1330
1331 oper = rtnl_dereference(q->oper_sched);
1332 if (oper)
1333 taprio_update_queue_max_sdu(q, oper, stab);
1334
1335 admin = rtnl_dereference(q->admin_sched);
1336 if (admin)
1337 taprio_update_queue_max_sdu(q, admin, stab);
1338
1339 break;
1340 }
1341
1342 return NOTIFY_DONE;
1343}
1344
1345static void setup_txtime(struct taprio_sched *q,
1346 struct sched_gate_list *sched, ktime_t base)
1347{
1348 struct sched_entry *entry;
1349 u64 interval = 0;
1350
1351 list_for_each_entry(entry, &sched->entries, list) {
1352 entry->next_txtime = ktime_add_ns(base, interval);
1353 interval += entry->interval;
1354 }
1355}
1356
1357static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1358{
1359 struct __tc_taprio_qopt_offload *__offload;
1360
1361 __offload = kzalloc(struct_size(__offload, offload.entries, num_entries),
1362 GFP_KERNEL);
1363 if (!__offload)
1364 return NULL;
1365
1366 refcount_set(&__offload->users, 1);
1367
1368 return &__offload->offload;
1369}
1370
1371struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1372 *offload)
1373{
1374 struct __tc_taprio_qopt_offload *__offload;
1375
1376 __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1377 offload);
1378
1379 refcount_inc(&__offload->users);
1380
1381 return offload;
1382}
1383EXPORT_SYMBOL_GPL(taprio_offload_get);
1384
1385void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1386{
1387 struct __tc_taprio_qopt_offload *__offload;
1388
1389 __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1390 offload);
1391
1392 if (!refcount_dec_and_test(&__offload->users))
1393 return;
1394
1395 kfree(__offload);
1396}
1397EXPORT_SYMBOL_GPL(taprio_offload_free);
1398
1399/* The function will only serve to keep the pointers to the "oper" and "admin"
1400 * schedules valid in relation to their base times, so when calling dump() the
1401 * users looks at the right schedules.
1402 * When using full offload, the admin configuration is promoted to oper at the
1403 * base_time in the PHC time domain. But because the system time is not
1404 * necessarily in sync with that, we can't just trigger a hrtimer to call
1405 * switch_schedules at the right hardware time.
1406 * At the moment we call this by hand right away from taprio, but in the future
1407 * it will be useful to create a mechanism for drivers to notify taprio of the
1408 * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1409 * This is left as TODO.
1410 */
1411static void taprio_offload_config_changed(struct taprio_sched *q)
1412{
1413 struct sched_gate_list *oper, *admin;
1414
1415 oper = rtnl_dereference(q->oper_sched);
1416 admin = rtnl_dereference(q->admin_sched);
1417
1418 switch_schedules(q, &admin, &oper);
1419}
1420
1421static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
1422{
1423 u32 i, queue_mask = 0;
1424
1425 for (i = 0; i < dev->num_tc; i++) {
1426 u32 offset, count;
1427
1428 if (!(tc_mask & BIT(i)))
1429 continue;
1430
1431 offset = dev->tc_to_txq[i].offset;
1432 count = dev->tc_to_txq[i].count;
1433
1434 queue_mask |= GENMASK(offset + count - 1, offset);
1435 }
1436
1437 return queue_mask;
1438}
1439
1440static void taprio_sched_to_offload(struct net_device *dev,
1441 struct sched_gate_list *sched,
1442 struct tc_taprio_qopt_offload *offload,
1443 const struct tc_taprio_caps *caps)
1444{
1445 struct sched_entry *entry;
1446 int i = 0;
1447
1448 offload->base_time = sched->base_time;
1449 offload->cycle_time = sched->cycle_time;
1450 offload->cycle_time_extension = sched->cycle_time_extension;
1451
1452 list_for_each_entry(entry, &sched->entries, list) {
1453 struct tc_taprio_sched_entry *e = &offload->entries[i];
1454
1455 e->command = entry->command;
1456 e->interval = entry->interval;
1457 if (caps->gate_mask_per_txq)
1458 e->gate_mask = tc_map_to_queue_mask(dev,
1459 entry->gate_mask);
1460 else
1461 e->gate_mask = entry->gate_mask;
1462
1463 i++;
1464 }
1465
1466 offload->num_entries = i;
1467}
1468
1469static void taprio_detect_broken_mqprio(struct taprio_sched *q)
1470{
1471 struct net_device *dev = qdisc_dev(q->root);
1472 struct tc_taprio_caps caps;
1473
1474 qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1475 &caps, sizeof(caps));
1476
1477 q->broken_mqprio = caps.broken_mqprio;
1478 if (q->broken_mqprio)
1479 static_branch_inc(&taprio_have_broken_mqprio);
1480 else
1481 static_branch_inc(&taprio_have_working_mqprio);
1482
1483 q->detected_mqprio = true;
1484}
1485
1486static void taprio_cleanup_broken_mqprio(struct taprio_sched *q)
1487{
1488 if (!q->detected_mqprio)
1489 return;
1490
1491 if (q->broken_mqprio)
1492 static_branch_dec(&taprio_have_broken_mqprio);
1493 else
1494 static_branch_dec(&taprio_have_working_mqprio);
1495}
1496
1497static int taprio_enable_offload(struct net_device *dev,
1498 struct taprio_sched *q,
1499 struct sched_gate_list *sched,
1500 struct netlink_ext_ack *extack)
1501{
1502 const struct net_device_ops *ops = dev->netdev_ops;
1503 struct tc_taprio_qopt_offload *offload;
1504 struct tc_taprio_caps caps;
1505 int tc, err = 0;
1506
1507 if (!ops->ndo_setup_tc) {
1508 NL_SET_ERR_MSG(extack,
1509 "Device does not support taprio offload");
1510 return -EOPNOTSUPP;
1511 }
1512
1513 qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1514 &caps, sizeof(caps));
1515
1516 if (!caps.supports_queue_max_sdu) {
1517 for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
1518 if (q->max_sdu[tc]) {
1519 NL_SET_ERR_MSG_MOD(extack,
1520 "Device does not handle queueMaxSDU");
1521 return -EOPNOTSUPP;
1522 }
1523 }
1524 }
1525
1526 offload = taprio_offload_alloc(sched->num_entries);
1527 if (!offload) {
1528 NL_SET_ERR_MSG(extack,
1529 "Not enough memory for enabling offload mode");
1530 return -ENOMEM;
1531 }
1532 offload->cmd = TAPRIO_CMD_REPLACE;
1533 offload->extack = extack;
1534 mqprio_qopt_reconstruct(dev, &offload->mqprio.qopt);
1535 offload->mqprio.extack = extack;
1536 taprio_sched_to_offload(dev, sched, offload, &caps);
1537 mqprio_fp_to_offload(q->fp, &offload->mqprio);
1538
1539 for (tc = 0; tc < TC_MAX_QUEUE; tc++)
1540 offload->max_sdu[tc] = q->max_sdu[tc];
1541
1542 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1543 if (err < 0) {
1544 NL_SET_ERR_MSG_WEAK(extack,
1545 "Device failed to setup taprio offload");
1546 goto done;
1547 }
1548
1549 q->offloaded = true;
1550
1551done:
1552 /* The offload structure may linger around via a reference taken by the
1553 * device driver, so clear up the netlink extack pointer so that the
1554 * driver isn't tempted to dereference data which stopped being valid
1555 */
1556 offload->extack = NULL;
1557 offload->mqprio.extack = NULL;
1558 taprio_offload_free(offload);
1559
1560 return err;
1561}
1562
1563static int taprio_disable_offload(struct net_device *dev,
1564 struct taprio_sched *q,
1565 struct netlink_ext_ack *extack)
1566{
1567 const struct net_device_ops *ops = dev->netdev_ops;
1568 struct tc_taprio_qopt_offload *offload;
1569 int err;
1570
1571 if (!q->offloaded)
1572 return 0;
1573
1574 offload = taprio_offload_alloc(0);
1575 if (!offload) {
1576 NL_SET_ERR_MSG(extack,
1577 "Not enough memory to disable offload mode");
1578 return -ENOMEM;
1579 }
1580 offload->cmd = TAPRIO_CMD_DESTROY;
1581
1582 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1583 if (err < 0) {
1584 NL_SET_ERR_MSG(extack,
1585 "Device failed to disable offload");
1586 goto out;
1587 }
1588
1589 q->offloaded = false;
1590
1591out:
1592 taprio_offload_free(offload);
1593
1594 return err;
1595}
1596
1597/* If full offload is enabled, the only possible clockid is the net device's
1598 * PHC. For that reason, specifying a clockid through netlink is incorrect.
1599 * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1600 * in sync with the specified clockid via a user space daemon such as phc2sys.
1601 * For both software taprio and txtime-assist, the clockid is used for the
1602 * hrtimer that advances the schedule and hence mandatory.
1603 */
1604static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1605 struct netlink_ext_ack *extack)
1606{
1607 struct taprio_sched *q = qdisc_priv(sch);
1608 struct net_device *dev = qdisc_dev(sch);
1609 int err = -EINVAL;
1610
1611 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1612 const struct ethtool_ops *ops = dev->ethtool_ops;
1613 struct kernel_ethtool_ts_info info = {
1614 .cmd = ETHTOOL_GET_TS_INFO,
1615 .phc_index = -1,
1616 };
1617
1618 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1619 NL_SET_ERR_MSG(extack,
1620 "The 'clockid' cannot be specified for full offload");
1621 goto out;
1622 }
1623
1624 if (ops && ops->get_ts_info)
1625 err = ops->get_ts_info(dev, &info);
1626
1627 if (err || info.phc_index < 0) {
1628 NL_SET_ERR_MSG(extack,
1629 "Device does not have a PTP clock");
1630 err = -ENOTSUPP;
1631 goto out;
1632 }
1633 } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1634 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1635 enum tk_offsets tk_offset;
1636
1637 /* We only support static clockids and we don't allow
1638 * for it to be modified after the first init.
1639 */
1640 if (clockid < 0 ||
1641 (q->clockid != -1 && q->clockid != clockid)) {
1642 NL_SET_ERR_MSG(extack,
1643 "Changing the 'clockid' of a running schedule is not supported");
1644 err = -ENOTSUPP;
1645 goto out;
1646 }
1647
1648 switch (clockid) {
1649 case CLOCK_REALTIME:
1650 tk_offset = TK_OFFS_REAL;
1651 break;
1652 case CLOCK_MONOTONIC:
1653 tk_offset = TK_OFFS_MAX;
1654 break;
1655 case CLOCK_BOOTTIME:
1656 tk_offset = TK_OFFS_BOOT;
1657 break;
1658 case CLOCK_TAI:
1659 tk_offset = TK_OFFS_TAI;
1660 break;
1661 default:
1662 NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1663 err = -EINVAL;
1664 goto out;
1665 }
1666 /* This pairs with READ_ONCE() in taprio_mono_to_any */
1667 WRITE_ONCE(q->tk_offset, tk_offset);
1668
1669 q->clockid = clockid;
1670 } else {
1671 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1672 goto out;
1673 }
1674
1675 /* Everything went ok, return success. */
1676 err = 0;
1677
1678out:
1679 return err;
1680}
1681
1682static int taprio_parse_tc_entry(struct Qdisc *sch,
1683 struct nlattr *opt,
1684 u32 max_sdu[TC_QOPT_MAX_QUEUE],
1685 u32 fp[TC_QOPT_MAX_QUEUE],
1686 unsigned long *seen_tcs,
1687 struct netlink_ext_ack *extack)
1688{
1689 struct nlattr *tb[TCA_TAPRIO_TC_ENTRY_MAX + 1] = { };
1690 struct net_device *dev = qdisc_dev(sch);
1691 int err, tc;
1692 u32 val;
1693
1694 err = nla_parse_nested(tb, TCA_TAPRIO_TC_ENTRY_MAX, opt,
1695 taprio_tc_policy, extack);
1696 if (err < 0)
1697 return err;
1698
1699 if (!tb[TCA_TAPRIO_TC_ENTRY_INDEX]) {
1700 NL_SET_ERR_MSG_MOD(extack, "TC entry index missing");
1701 return -EINVAL;
1702 }
1703
1704 tc = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_INDEX]);
1705 if (tc >= TC_QOPT_MAX_QUEUE) {
1706 NL_SET_ERR_MSG_MOD(extack, "TC entry index out of range");
1707 return -ERANGE;
1708 }
1709
1710 if (*seen_tcs & BIT(tc)) {
1711 NL_SET_ERR_MSG_MOD(extack, "Duplicate TC entry");
1712 return -EINVAL;
1713 }
1714
1715 *seen_tcs |= BIT(tc);
1716
1717 if (tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]) {
1718 val = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]);
1719 if (val > dev->max_mtu) {
1720 NL_SET_ERR_MSG_MOD(extack, "TC max SDU exceeds device max MTU");
1721 return -ERANGE;
1722 }
1723
1724 max_sdu[tc] = val;
1725 }
1726
1727 if (tb[TCA_TAPRIO_TC_ENTRY_FP])
1728 fp[tc] = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_FP]);
1729
1730 return 0;
1731}
1732
1733static int taprio_parse_tc_entries(struct Qdisc *sch,
1734 struct nlattr *opt,
1735 struct netlink_ext_ack *extack)
1736{
1737 struct taprio_sched *q = qdisc_priv(sch);
1738 struct net_device *dev = qdisc_dev(sch);
1739 u32 max_sdu[TC_QOPT_MAX_QUEUE];
1740 bool have_preemption = false;
1741 unsigned long seen_tcs = 0;
1742 u32 fp[TC_QOPT_MAX_QUEUE];
1743 struct nlattr *n;
1744 int tc, rem;
1745 int err = 0;
1746
1747 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1748 max_sdu[tc] = q->max_sdu[tc];
1749 fp[tc] = q->fp[tc];
1750 }
1751
1752 nla_for_each_nested_type(n, TCA_TAPRIO_ATTR_TC_ENTRY, opt, rem) {
1753 err = taprio_parse_tc_entry(sch, n, max_sdu, fp, &seen_tcs,
1754 extack);
1755 if (err)
1756 return err;
1757 }
1758
1759 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1760 q->max_sdu[tc] = max_sdu[tc];
1761 q->fp[tc] = fp[tc];
1762 if (fp[tc] != TC_FP_EXPRESS)
1763 have_preemption = true;
1764 }
1765
1766 if (have_preemption) {
1767 if (!FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1768 NL_SET_ERR_MSG(extack,
1769 "Preemption only supported with full offload");
1770 return -EOPNOTSUPP;
1771 }
1772
1773 if (!ethtool_dev_mm_supported(dev)) {
1774 NL_SET_ERR_MSG(extack,
1775 "Device does not support preemption");
1776 return -EOPNOTSUPP;
1777 }
1778 }
1779
1780 return err;
1781}
1782
1783static int taprio_mqprio_cmp(const struct net_device *dev,
1784 const struct tc_mqprio_qopt *mqprio)
1785{
1786 int i;
1787
1788 if (!mqprio || mqprio->num_tc != dev->num_tc)
1789 return -1;
1790
1791 for (i = 0; i < mqprio->num_tc; i++)
1792 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1793 dev->tc_to_txq[i].offset != mqprio->offset[i])
1794 return -1;
1795
1796 for (i = 0; i <= TC_BITMASK; i++)
1797 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1798 return -1;
1799
1800 return 0;
1801}
1802
1803static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1804 struct netlink_ext_ack *extack)
1805{
1806 struct qdisc_size_table *stab = rtnl_dereference(sch->stab);
1807 struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1808 struct sched_gate_list *oper, *admin, *new_admin;
1809 struct taprio_sched *q = qdisc_priv(sch);
1810 struct net_device *dev = qdisc_dev(sch);
1811 struct tc_mqprio_qopt *mqprio = NULL;
1812 unsigned long flags;
1813 u32 taprio_flags;
1814 ktime_t start;
1815 int i, err;
1816
1817 err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1818 taprio_policy, extack);
1819 if (err < 0)
1820 return err;
1821
1822 if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1823 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1824
1825 /* The semantics of the 'flags' argument in relation to 'change()'
1826 * requests, are interpreted following two rules (which are applied in
1827 * this order): (1) an omitted 'flags' argument is interpreted as
1828 * zero; (2) the 'flags' of a "running" taprio instance cannot be
1829 * changed.
1830 */
1831 taprio_flags = nla_get_u32_default(tb[TCA_TAPRIO_ATTR_FLAGS], 0);
1832
1833 /* txtime-assist and full offload are mutually exclusive */
1834 if ((taprio_flags & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST) &&
1835 (taprio_flags & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)) {
1836 NL_SET_ERR_MSG_ATTR(extack, tb[TCA_TAPRIO_ATTR_FLAGS],
1837 "TXTIME_ASSIST and FULL_OFFLOAD are mutually exclusive");
1838 return -EINVAL;
1839 }
1840
1841 if (q->flags != TAPRIO_FLAGS_INVALID && q->flags != taprio_flags) {
1842 NL_SET_ERR_MSG_MOD(extack,
1843 "Changing 'flags' of a running schedule is not supported");
1844 return -EOPNOTSUPP;
1845 }
1846 q->flags = taprio_flags;
1847
1848 /* Needed for length_to_duration() during netlink attribute parsing */
1849 taprio_set_picos_per_byte(dev, q);
1850
1851 err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1852 if (err < 0)
1853 return err;
1854
1855 err = taprio_parse_tc_entries(sch, opt, extack);
1856 if (err)
1857 return err;
1858
1859 new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1860 if (!new_admin) {
1861 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1862 return -ENOMEM;
1863 }
1864 INIT_LIST_HEAD(&new_admin->entries);
1865
1866 oper = rtnl_dereference(q->oper_sched);
1867 admin = rtnl_dereference(q->admin_sched);
1868
1869 /* no changes - no new mqprio settings */
1870 if (!taprio_mqprio_cmp(dev, mqprio))
1871 mqprio = NULL;
1872
1873 if (mqprio && (oper || admin)) {
1874 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1875 err = -ENOTSUPP;
1876 goto free_sched;
1877 }
1878
1879 if (mqprio) {
1880 err = netdev_set_num_tc(dev, mqprio->num_tc);
1881 if (err)
1882 goto free_sched;
1883 for (i = 0; i < mqprio->num_tc; i++) {
1884 netdev_set_tc_queue(dev, i,
1885 mqprio->count[i],
1886 mqprio->offset[i]);
1887 q->cur_txq[i] = mqprio->offset[i];
1888 }
1889
1890 /* Always use supplied priority mappings */
1891 for (i = 0; i <= TC_BITMASK; i++)
1892 netdev_set_prio_tc_map(dev, i,
1893 mqprio->prio_tc_map[i]);
1894 }
1895
1896 err = parse_taprio_schedule(q, tb, new_admin, extack);
1897 if (err < 0)
1898 goto free_sched;
1899
1900 if (new_admin->num_entries == 0) {
1901 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1902 err = -EINVAL;
1903 goto free_sched;
1904 }
1905
1906 err = taprio_parse_clockid(sch, tb, extack);
1907 if (err < 0)
1908 goto free_sched;
1909
1910 taprio_update_queue_max_sdu(q, new_admin, stab);
1911
1912 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1913 err = taprio_enable_offload(dev, q, new_admin, extack);
1914 else
1915 err = taprio_disable_offload(dev, q, extack);
1916 if (err)
1917 goto free_sched;
1918
1919 /* Protects against enqueue()/dequeue() */
1920 spin_lock_bh(qdisc_lock(sch));
1921
1922 if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1923 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1924 NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1925 err = -EINVAL;
1926 goto unlock;
1927 }
1928
1929 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1930 }
1931
1932 if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1933 !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1934 !hrtimer_active(&q->advance_timer)) {
1935 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1936 q->advance_timer.function = advance_sched;
1937 }
1938
1939 err = taprio_get_start_time(sch, new_admin, &start);
1940 if (err < 0) {
1941 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1942 goto unlock;
1943 }
1944
1945 setup_txtime(q, new_admin, start);
1946
1947 if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1948 if (!oper) {
1949 rcu_assign_pointer(q->oper_sched, new_admin);
1950 err = 0;
1951 new_admin = NULL;
1952 goto unlock;
1953 }
1954
1955 /* Not going to race against advance_sched(), but still */
1956 admin = rcu_replace_pointer(q->admin_sched, new_admin,
1957 lockdep_rtnl_is_held());
1958 if (admin)
1959 call_rcu(&admin->rcu, taprio_free_sched_cb);
1960 } else {
1961 setup_first_end_time(q, new_admin, start);
1962
1963 /* Protects against advance_sched() */
1964 spin_lock_irqsave(&q->current_entry_lock, flags);
1965
1966 taprio_start_sched(sch, start, new_admin);
1967
1968 admin = rcu_replace_pointer(q->admin_sched, new_admin,
1969 lockdep_rtnl_is_held());
1970 if (admin)
1971 call_rcu(&admin->rcu, taprio_free_sched_cb);
1972
1973 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1974
1975 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1976 taprio_offload_config_changed(q);
1977 }
1978
1979 new_admin = NULL;
1980 err = 0;
1981
1982 if (!stab)
1983 NL_SET_ERR_MSG_MOD(extack,
1984 "Size table not specified, frame length estimations may be inaccurate");
1985
1986unlock:
1987 spin_unlock_bh(qdisc_lock(sch));
1988
1989free_sched:
1990 if (new_admin)
1991 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
1992
1993 return err;
1994}
1995
1996static void taprio_reset(struct Qdisc *sch)
1997{
1998 struct taprio_sched *q = qdisc_priv(sch);
1999 struct net_device *dev = qdisc_dev(sch);
2000 int i;
2001
2002 hrtimer_cancel(&q->advance_timer);
2003
2004 if (q->qdiscs) {
2005 for (i = 0; i < dev->num_tx_queues; i++)
2006 if (q->qdiscs[i])
2007 qdisc_reset(q->qdiscs[i]);
2008 }
2009}
2010
2011static void taprio_destroy(struct Qdisc *sch)
2012{
2013 struct taprio_sched *q = qdisc_priv(sch);
2014 struct net_device *dev = qdisc_dev(sch);
2015 struct sched_gate_list *oper, *admin;
2016 unsigned int i;
2017
2018 list_del(&q->taprio_list);
2019
2020 /* Note that taprio_reset() might not be called if an error
2021 * happens in qdisc_create(), after taprio_init() has been called.
2022 */
2023 hrtimer_cancel(&q->advance_timer);
2024 qdisc_synchronize(sch);
2025
2026 taprio_disable_offload(dev, q, NULL);
2027
2028 if (q->qdiscs) {
2029 for (i = 0; i < dev->num_tx_queues; i++)
2030 qdisc_put(q->qdiscs[i]);
2031
2032 kfree(q->qdiscs);
2033 }
2034 q->qdiscs = NULL;
2035
2036 netdev_reset_tc(dev);
2037
2038 oper = rtnl_dereference(q->oper_sched);
2039 admin = rtnl_dereference(q->admin_sched);
2040
2041 if (oper)
2042 call_rcu(&oper->rcu, taprio_free_sched_cb);
2043
2044 if (admin)
2045 call_rcu(&admin->rcu, taprio_free_sched_cb);
2046
2047 taprio_cleanup_broken_mqprio(q);
2048}
2049
2050static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
2051 struct netlink_ext_ack *extack)
2052{
2053 struct taprio_sched *q = qdisc_priv(sch);
2054 struct net_device *dev = qdisc_dev(sch);
2055 int i, tc;
2056
2057 spin_lock_init(&q->current_entry_lock);
2058
2059 hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
2060 q->advance_timer.function = advance_sched;
2061
2062 q->root = sch;
2063
2064 /* We only support static clockids. Use an invalid value as default
2065 * and get the valid one on taprio_change().
2066 */
2067 q->clockid = -1;
2068 q->flags = TAPRIO_FLAGS_INVALID;
2069
2070 list_add(&q->taprio_list, &taprio_list);
2071
2072 if (sch->parent != TC_H_ROOT) {
2073 NL_SET_ERR_MSG_MOD(extack, "Can only be attached as root qdisc");
2074 return -EOPNOTSUPP;
2075 }
2076
2077 if (!netif_is_multiqueue(dev)) {
2078 NL_SET_ERR_MSG_MOD(extack, "Multi-queue device is required");
2079 return -EOPNOTSUPP;
2080 }
2081
2082 q->qdiscs = kcalloc(dev->num_tx_queues, sizeof(q->qdiscs[0]),
2083 GFP_KERNEL);
2084 if (!q->qdiscs)
2085 return -ENOMEM;
2086
2087 if (!opt)
2088 return -EINVAL;
2089
2090 for (i = 0; i < dev->num_tx_queues; i++) {
2091 struct netdev_queue *dev_queue;
2092 struct Qdisc *qdisc;
2093
2094 dev_queue = netdev_get_tx_queue(dev, i);
2095 qdisc = qdisc_create_dflt(dev_queue,
2096 &pfifo_qdisc_ops,
2097 TC_H_MAKE(TC_H_MAJ(sch->handle),
2098 TC_H_MIN(i + 1)),
2099 extack);
2100 if (!qdisc)
2101 return -ENOMEM;
2102
2103 if (i < dev->real_num_tx_queues)
2104 qdisc_hash_add(qdisc, false);
2105
2106 q->qdiscs[i] = qdisc;
2107 }
2108
2109 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++)
2110 q->fp[tc] = TC_FP_EXPRESS;
2111
2112 taprio_detect_broken_mqprio(q);
2113
2114 return taprio_change(sch, opt, extack);
2115}
2116
2117static void taprio_attach(struct Qdisc *sch)
2118{
2119 struct taprio_sched *q = qdisc_priv(sch);
2120 struct net_device *dev = qdisc_dev(sch);
2121 unsigned int ntx;
2122
2123 /* Attach underlying qdisc */
2124 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
2125 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
2126 struct Qdisc *old, *dev_queue_qdisc;
2127
2128 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2129 struct Qdisc *qdisc = q->qdiscs[ntx];
2130
2131 /* In offload mode, the root taprio qdisc is bypassed
2132 * and the netdev TX queues see the children directly
2133 */
2134 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2135 dev_queue_qdisc = qdisc;
2136 } else {
2137 /* In software mode, attach the root taprio qdisc
2138 * to all netdev TX queues, so that dev_qdisc_enqueue()
2139 * goes through taprio_enqueue().
2140 */
2141 dev_queue_qdisc = sch;
2142 }
2143 old = dev_graft_qdisc(dev_queue, dev_queue_qdisc);
2144 /* The qdisc's refcount requires to be elevated once
2145 * for each netdev TX queue it is grafted onto
2146 */
2147 qdisc_refcount_inc(dev_queue_qdisc);
2148 if (old)
2149 qdisc_put(old);
2150 }
2151}
2152
2153static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
2154 unsigned long cl)
2155{
2156 struct net_device *dev = qdisc_dev(sch);
2157 unsigned long ntx = cl - 1;
2158
2159 if (ntx >= dev->num_tx_queues)
2160 return NULL;
2161
2162 return netdev_get_tx_queue(dev, ntx);
2163}
2164
2165static int taprio_graft(struct Qdisc *sch, unsigned long cl,
2166 struct Qdisc *new, struct Qdisc **old,
2167 struct netlink_ext_ack *extack)
2168{
2169 struct taprio_sched *q = qdisc_priv(sch);
2170 struct net_device *dev = qdisc_dev(sch);
2171 struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2172
2173 if (!dev_queue)
2174 return -EINVAL;
2175
2176 if (dev->flags & IFF_UP)
2177 dev_deactivate(dev);
2178
2179 /* In offload mode, the child Qdisc is directly attached to the netdev
2180 * TX queue, and thus, we need to keep its refcount elevated in order
2181 * to counteract qdisc_graft()'s call to qdisc_put() once per TX queue.
2182 * However, save the reference to the new qdisc in the private array in
2183 * both software and offload cases, to have an up-to-date reference to
2184 * our children.
2185 */
2186 *old = q->qdiscs[cl - 1];
2187 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2188 WARN_ON_ONCE(dev_graft_qdisc(dev_queue, new) != *old);
2189 if (new)
2190 qdisc_refcount_inc(new);
2191 if (*old)
2192 qdisc_put(*old);
2193 }
2194
2195 q->qdiscs[cl - 1] = new;
2196 if (new)
2197 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2198
2199 if (dev->flags & IFF_UP)
2200 dev_activate(dev);
2201
2202 return 0;
2203}
2204
2205static int dump_entry(struct sk_buff *msg,
2206 const struct sched_entry *entry)
2207{
2208 struct nlattr *item;
2209
2210 item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
2211 if (!item)
2212 return -ENOSPC;
2213
2214 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
2215 goto nla_put_failure;
2216
2217 if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
2218 goto nla_put_failure;
2219
2220 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
2221 entry->gate_mask))
2222 goto nla_put_failure;
2223
2224 if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
2225 entry->interval))
2226 goto nla_put_failure;
2227
2228 return nla_nest_end(msg, item);
2229
2230nla_put_failure:
2231 nla_nest_cancel(msg, item);
2232 return -1;
2233}
2234
2235static int dump_schedule(struct sk_buff *msg,
2236 const struct sched_gate_list *root)
2237{
2238 struct nlattr *entry_list;
2239 struct sched_entry *entry;
2240
2241 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
2242 root->base_time, TCA_TAPRIO_PAD))
2243 return -1;
2244
2245 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
2246 root->cycle_time, TCA_TAPRIO_PAD))
2247 return -1;
2248
2249 if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
2250 root->cycle_time_extension, TCA_TAPRIO_PAD))
2251 return -1;
2252
2253 entry_list = nla_nest_start_noflag(msg,
2254 TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
2255 if (!entry_list)
2256 goto error_nest;
2257
2258 list_for_each_entry(entry, &root->entries, list) {
2259 if (dump_entry(msg, entry) < 0)
2260 goto error_nest;
2261 }
2262
2263 nla_nest_end(msg, entry_list);
2264 return 0;
2265
2266error_nest:
2267 nla_nest_cancel(msg, entry_list);
2268 return -1;
2269}
2270
2271static int taprio_dump_tc_entries(struct sk_buff *skb,
2272 struct taprio_sched *q,
2273 struct sched_gate_list *sched)
2274{
2275 struct nlattr *n;
2276 int tc;
2277
2278 for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
2279 n = nla_nest_start(skb, TCA_TAPRIO_ATTR_TC_ENTRY);
2280 if (!n)
2281 return -EMSGSIZE;
2282
2283 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_INDEX, tc))
2284 goto nla_put_failure;
2285
2286 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_MAX_SDU,
2287 sched->max_sdu[tc]))
2288 goto nla_put_failure;
2289
2290 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_FP, q->fp[tc]))
2291 goto nla_put_failure;
2292
2293 nla_nest_end(skb, n);
2294 }
2295
2296 return 0;
2297
2298nla_put_failure:
2299 nla_nest_cancel(skb, n);
2300 return -EMSGSIZE;
2301}
2302
2303static int taprio_put_stat(struct sk_buff *skb, u64 val, u16 attrtype)
2304{
2305 if (val == TAPRIO_STAT_NOT_SET)
2306 return 0;
2307 if (nla_put_u64_64bit(skb, attrtype, val, TCA_TAPRIO_OFFLOAD_STATS_PAD))
2308 return -EMSGSIZE;
2309 return 0;
2310}
2311
2312static int taprio_dump_xstats(struct Qdisc *sch, struct gnet_dump *d,
2313 struct tc_taprio_qopt_offload *offload,
2314 struct tc_taprio_qopt_stats *stats)
2315{
2316 struct net_device *dev = qdisc_dev(sch);
2317 const struct net_device_ops *ops;
2318 struct sk_buff *skb = d->skb;
2319 struct nlattr *xstats;
2320 int err;
2321
2322 ops = qdisc_dev(sch)->netdev_ops;
2323
2324 /* FIXME I could use qdisc_offload_dump_helper(), but that messes
2325 * with sch->flags depending on whether the device reports taprio
2326 * stats, and I'm not sure whether that's a good idea, considering
2327 * that stats are optional to the offload itself
2328 */
2329 if (!ops->ndo_setup_tc)
2330 return 0;
2331
2332 memset(stats, 0xff, sizeof(*stats));
2333
2334 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
2335 if (err == -EOPNOTSUPP)
2336 return 0;
2337 if (err)
2338 return err;
2339
2340 xstats = nla_nest_start(skb, TCA_STATS_APP);
2341 if (!xstats)
2342 goto err;
2343
2344 if (taprio_put_stat(skb, stats->window_drops,
2345 TCA_TAPRIO_OFFLOAD_STATS_WINDOW_DROPS) ||
2346 taprio_put_stat(skb, stats->tx_overruns,
2347 TCA_TAPRIO_OFFLOAD_STATS_TX_OVERRUNS))
2348 goto err_cancel;
2349
2350 nla_nest_end(skb, xstats);
2351
2352 return 0;
2353
2354err_cancel:
2355 nla_nest_cancel(skb, xstats);
2356err:
2357 return -EMSGSIZE;
2358}
2359
2360static int taprio_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
2361{
2362 struct tc_taprio_qopt_offload offload = {
2363 .cmd = TAPRIO_CMD_STATS,
2364 };
2365
2366 return taprio_dump_xstats(sch, d, &offload, &offload.stats);
2367}
2368
2369static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
2370{
2371 struct taprio_sched *q = qdisc_priv(sch);
2372 struct net_device *dev = qdisc_dev(sch);
2373 struct sched_gate_list *oper, *admin;
2374 struct tc_mqprio_qopt opt = { 0 };
2375 struct nlattr *nest, *sched_nest;
2376
2377 mqprio_qopt_reconstruct(dev, &opt);
2378
2379 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2380 if (!nest)
2381 goto start_error;
2382
2383 if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
2384 goto options_error;
2385
2386 if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
2387 nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
2388 goto options_error;
2389
2390 if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
2391 goto options_error;
2392
2393 if (q->txtime_delay &&
2394 nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
2395 goto options_error;
2396
2397 rcu_read_lock();
2398
2399 oper = rtnl_dereference(q->oper_sched);
2400 admin = rtnl_dereference(q->admin_sched);
2401
2402 if (oper && taprio_dump_tc_entries(skb, q, oper))
2403 goto options_error_rcu;
2404
2405 if (oper && dump_schedule(skb, oper))
2406 goto options_error_rcu;
2407
2408 if (!admin)
2409 goto done;
2410
2411 sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
2412 if (!sched_nest)
2413 goto options_error_rcu;
2414
2415 if (dump_schedule(skb, admin))
2416 goto admin_error;
2417
2418 nla_nest_end(skb, sched_nest);
2419
2420done:
2421 rcu_read_unlock();
2422 return nla_nest_end(skb, nest);
2423
2424admin_error:
2425 nla_nest_cancel(skb, sched_nest);
2426
2427options_error_rcu:
2428 rcu_read_unlock();
2429
2430options_error:
2431 nla_nest_cancel(skb, nest);
2432
2433start_error:
2434 return -ENOSPC;
2435}
2436
2437static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
2438{
2439 struct taprio_sched *q = qdisc_priv(sch);
2440 struct net_device *dev = qdisc_dev(sch);
2441 unsigned int ntx = cl - 1;
2442
2443 if (ntx >= dev->num_tx_queues)
2444 return NULL;
2445
2446 return q->qdiscs[ntx];
2447}
2448
2449static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
2450{
2451 unsigned int ntx = TC_H_MIN(classid);
2452
2453 if (!taprio_queue_get(sch, ntx))
2454 return 0;
2455 return ntx;
2456}
2457
2458static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
2459 struct sk_buff *skb, struct tcmsg *tcm)
2460{
2461 struct Qdisc *child = taprio_leaf(sch, cl);
2462
2463 tcm->tcm_parent = TC_H_ROOT;
2464 tcm->tcm_handle |= TC_H_MIN(cl);
2465 tcm->tcm_info = child->handle;
2466
2467 return 0;
2468}
2469
2470static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
2471 struct gnet_dump *d)
2472 __releases(d->lock)
2473 __acquires(d->lock)
2474{
2475 struct Qdisc *child = taprio_leaf(sch, cl);
2476 struct tc_taprio_qopt_offload offload = {
2477 .cmd = TAPRIO_CMD_QUEUE_STATS,
2478 .queue_stats = {
2479 .queue = cl - 1,
2480 },
2481 };
2482
2483 if (gnet_stats_copy_basic(d, NULL, &child->bstats, true) < 0 ||
2484 qdisc_qstats_copy(d, child) < 0)
2485 return -1;
2486
2487 return taprio_dump_xstats(sch, d, &offload, &offload.queue_stats.stats);
2488}
2489
2490static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2491{
2492 struct net_device *dev = qdisc_dev(sch);
2493 unsigned long ntx;
2494
2495 if (arg->stop)
2496 return;
2497
2498 arg->count = arg->skip;
2499 for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
2500 if (!tc_qdisc_stats_dump(sch, ntx + 1, arg))
2501 break;
2502 }
2503}
2504
2505static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
2506 struct tcmsg *tcm)
2507{
2508 return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
2509}
2510
2511static const struct Qdisc_class_ops taprio_class_ops = {
2512 .graft = taprio_graft,
2513 .leaf = taprio_leaf,
2514 .find = taprio_find,
2515 .walk = taprio_walk,
2516 .dump = taprio_dump_class,
2517 .dump_stats = taprio_dump_class_stats,
2518 .select_queue = taprio_select_queue,
2519};
2520
2521static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
2522 .cl_ops = &taprio_class_ops,
2523 .id = "taprio",
2524 .priv_size = sizeof(struct taprio_sched),
2525 .init = taprio_init,
2526 .change = taprio_change,
2527 .destroy = taprio_destroy,
2528 .reset = taprio_reset,
2529 .attach = taprio_attach,
2530 .peek = taprio_peek,
2531 .dequeue = taprio_dequeue,
2532 .enqueue = taprio_enqueue,
2533 .dump = taprio_dump,
2534 .dump_stats = taprio_dump_stats,
2535 .owner = THIS_MODULE,
2536};
2537MODULE_ALIAS_NET_SCH("taprio");
2538
2539static struct notifier_block taprio_device_notifier = {
2540 .notifier_call = taprio_dev_notifier,
2541};
2542
2543static int __init taprio_module_init(void)
2544{
2545 int err = register_netdevice_notifier(&taprio_device_notifier);
2546
2547 if (err)
2548 return err;
2549
2550 return register_qdisc(&taprio_qdisc_ops);
2551}
2552
2553static void __exit taprio_module_exit(void)
2554{
2555 unregister_qdisc(&taprio_qdisc_ops);
2556 unregister_netdevice_notifier(&taprio_device_notifier);
2557}
2558
2559module_init(taprio_module_init);
2560module_exit(taprio_module_exit);
2561MODULE_LICENSE("GPL");
2562MODULE_DESCRIPTION("Time Aware Priority qdisc");