net/sched/sch_taprio.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* net/sched/sch_taprio.c        Time Aware Priority Scheduler
   4  *
   5  * Authors:     Vinicius Costa Gomes <vinicius.gomes@intel.com>
   6  *
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/kernel.h>
  12 #include <linux/string.h>
  13 #include <linux/list.h>
  14 #include <linux/errno.h>
  15 #include <linux/skbuff.h>
  16 #include <linux/math64.h>
  17 #include <linux/module.h>
  18 #include <linux/spinlock.h>
  19 #include <linux/rcupdate.h>
  20 #include <net/netlink.h>
  21 #include <net/pkt_sched.h>
  22 #include <net/pkt_cls.h>
  23 #include <net/sch_generic.h>
  24 #include <net/sock.h>
  25 #include <net/tcp.h>
  26
  27 static LIST_HEAD(taprio_list);
  28 static DEFINE_SPINLOCK(taprio_list_lock);
  29
  30 #define TAPRIO_ALL_GATES_OPEN -1
  31
  32 #define TXTIME_ASSIST_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST)
  33 #define FULL_OFFLOAD_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)
  34 #define TAPRIO_FLAGS_INVALID U32_MAX
  35
  36 struct sched_entry {
  37         struct list_head list;
  38
  39         /* The instant that this entry "closes" and the next one
  40          * should open, the qdisc will make some effort so that no
  41          * packet leaves after this time.
  42          */
  43         ktime_t close_time;
  44         ktime_t next_txtime;
  45         atomic_t budget;
  46         int index;
  47         u32 gate_mask;
  48         u32 interval;
  49         u8 command;
  50 };
  51
  52 struct sched_gate_list {
  53         struct rcu_head rcu;
  54         struct list_head entries;
  55         size_t num_entries;
  56         ktime_t cycle_close_time;
  57         s64 cycle_time;
  58         s64 cycle_time_extension;
  59         s64 base_time;
  60 };
  61
  62 struct taprio_sched {
  63         struct Qdisc **qdiscs;
  64         struct Qdisc *root;
  65         u32 flags;
  66         enum tk_offsets tk_offset;
  67         int clockid;
  68         atomic64_t picos_per_byte; /* Using picoseconds because for 10Gbps+
  69                                     * speeds it's sub-nanoseconds per byte
  70                                     */
  71
  72         /* Protects the update side of the RCU protected current_entry */
  73         spinlock_t current_entry_lock;
  74         struct sched_entry __rcu *current_entry;
  75         struct sched_gate_list __rcu *oper_sched;
  76         struct sched_gate_list __rcu *admin_sched;
  77         struct hrtimer advance_timer;
  78         struct list_head taprio_list;
  79         struct sk_buff *(*dequeue)(struct Qdisc *sch);
  80         struct sk_buff *(*peek)(struct Qdisc *sch);
  81         u32 txtime_delay;
  82 };
  83
  84 struct __tc_taprio_qopt_offload {
  85         refcount_t users;
  86         struct tc_taprio_qopt_offload offload;
  87 };
  88
  89 static ktime_t sched_base_time(const struct sched_gate_list *sched)
  90 {
  91         if (!sched)
  92                 return KTIME_MAX;
  93
  94         return ns_to_ktime(sched->base_time);
  95 }
  96
  97 static ktime_t taprio_get_time(struct taprio_sched *q)
  98 {
  99         ktime_t mono = ktime_get();
 100
 101         switch (q->tk_offset) {
 102         case TK_OFFS_MAX:
 103                 return mono;
 104         default:
 105                 return ktime_mono_to_any(mono, q->tk_offset);
 106         }
 107
 108         return KTIME_MAX;
 109 }
 110
 111 static void taprio_free_sched_cb(struct rcu_head *head)
 112 {
 113         struct sched_gate_list *sched = container_of(head, struct sched_gate_list, rcu);
 114         struct sched_entry *entry, *n;
 115
 116         if (!sched)
 117                 return;
 118
 119         list_for_each_entry_safe(entry, n, &sched->entries, list) {
 120                 list_del(&entry->list);
 121                 kfree(entry);
 122         }
 123
 124         kfree(sched);
 125 }
 126
 127 static void switch_schedules(struct taprio_sched *q,
 128                              struct sched_gate_list **admin,
 129                              struct sched_gate_list **oper)
 130 {
 131         rcu_assign_pointer(q->oper_sched, *admin);
 132         rcu_assign_pointer(q->admin_sched, NULL);
 133
 134         if (*oper)
 135                 call_rcu(&(*oper)->rcu, taprio_free_sched_cb);
 136
 137         *oper = *admin;
 138         *admin = NULL;
 139 }
 140
 141 /* Get how much time has been already elapsed in the current cycle. */
 142 static s32 get_cycle_time_elapsed(struct sched_gate_list *sched, ktime_t time)
 143 {
 144         ktime_t time_since_sched_start;
 145         s32 time_elapsed;
 146
 147         time_since_sched_start = ktime_sub(time, sched->base_time);
 148         div_s64_rem(time_since_sched_start, sched->cycle_time, &time_elapsed);
 149
 150         return time_elapsed;
 151 }
 152
 153 static ktime_t get_interval_end_time(struct sched_gate_list *sched,
 154                                      struct sched_gate_list *admin,
 155                                      struct sched_entry *entry,
 156                                      ktime_t intv_start)
 157 {
 158         s32 cycle_elapsed = get_cycle_time_elapsed(sched, intv_start);
 159         ktime_t intv_end, cycle_ext_end, cycle_end;
 160
 161         cycle_end = ktime_add_ns(intv_start, sched->cycle_time - cycle_elapsed);
 162         intv_end = ktime_add_ns(intv_start, entry->interval);
 163         cycle_ext_end = ktime_add(cycle_end, sched->cycle_time_extension);
 164
 165         if (ktime_before(intv_end, cycle_end))
 166                 return intv_end;
 167         else if (admin && admin != sched &&
 168                  ktime_after(admin->base_time, cycle_end) &&
 169                  ktime_before(admin->base_time, cycle_ext_end))
 170                 return admin->base_time;
 171         else
 172                 return cycle_end;
 173 }
 174
 175 static int length_to_duration(struct taprio_sched *q, int len)
 176 {
 177         return div_u64(len * atomic64_read(&q->picos_per_byte), 1000);
 178 }
 179
 180 /* Returns the entry corresponding to next available interval. If
 181  * validate_interval is set, it only validates whether the timestamp occurs
 182  * when the gate corresponding to the skb's traffic class is open.
 183  */
 184 static struct sched_entry *find_entry_to_transmit(struct sk_buff *skb,
 185                                                   struct Qdisc *sch,
 186                                                   struct sched_gate_list *sched,
 187                                                   struct sched_gate_list *admin,
 188                                                   ktime_t time,
 189                                                   ktime_t *interval_start,
 190                                                   ktime_t *interval_end,
 191                                                   bool validate_interval)
 192 {
 193         ktime_t curr_intv_start, curr_intv_end, cycle_end, packet_transmit_time;
 194         ktime_t earliest_txtime = KTIME_MAX, txtime, cycle, transmit_end_time;
 195         struct sched_entry *entry = NULL, *entry_found = NULL;
 196         struct taprio_sched *q = qdisc_priv(sch);
 197         struct net_device *dev = qdisc_dev(sch);
 198         bool entry_available = false;
 199         s32 cycle_elapsed;
 200         int tc, n;
 201
 202         tc = netdev_get_prio_tc_map(dev, skb->priority);
 203         packet_transmit_time = length_to_duration(q, qdisc_pkt_len(skb));
 204
 205         *interval_start = 0;
 206         *interval_end = 0;
 207
 208         if (!sched)
 209                 return NULL;
 210
 211         cycle = sched->cycle_time;
 212         cycle_elapsed = get_cycle_time_elapsed(sched, time);
 213         curr_intv_end = ktime_sub_ns(time, cycle_elapsed);
 214         cycle_end = ktime_add_ns(curr_intv_end, cycle);
 215
 216         list_for_each_entry(entry, &sched->entries, list) {
 217                 curr_intv_start = curr_intv_end;
 218                 curr_intv_end = get_interval_end_time(sched, admin, entry,
 219                                                       curr_intv_start);
 220
 221                 if (ktime_after(curr_intv_start, cycle_end))
 222                         break;
 223
 224                 if (!(entry->gate_mask & BIT(tc)) ||
 225                     packet_transmit_time > entry->interval)
 226                         continue;
 227
 228                 txtime = entry->next_txtime;
 229
 230                 if (ktime_before(txtime, time) || validate_interval) {
 231                         transmit_end_time = ktime_add_ns(time, packet_transmit_time);
 232                         if ((ktime_before(curr_intv_start, time) &&
 233                              ktime_before(transmit_end_time, curr_intv_end)) ||
 234                             (ktime_after(curr_intv_start, time) && !validate_interval)) {
 235                                 entry_found = entry;
 236                                 *interval_start = curr_intv_start;
 237                                 *interval_end = curr_intv_end;
 238                                 break;
 239                         } else if (!entry_available && !validate_interval) {
 240                                 /* Here, we are just trying to find out the
 241                                  * first available interval in the next cycle.
 242                                  */
 243                                 entry_available = 1;
 244                                 entry_found = entry;
 245                                 *interval_start = ktime_add_ns(curr_intv_start, cycle);
 246                                 *interval_end = ktime_add_ns(curr_intv_end, cycle);
 247                         }
 248                 } else if (ktime_before(txtime, earliest_txtime) &&
 249                            !entry_available) {
 250                         earliest_txtime = txtime;
 251                         entry_found = entry;
 252                         n = div_s64(ktime_sub(txtime, curr_intv_start), cycle);
 253                         *interval_start = ktime_add(curr_intv_start, n * cycle);
 254                         *interval_end = ktime_add(curr_intv_end, n * cycle);
 255                 }
 256         }
 257
 258         return entry_found;
 259 }
 260
 261 static bool is_valid_interval(struct sk_buff *skb, struct Qdisc *sch)
 262 {
 263         struct taprio_sched *q = qdisc_priv(sch);
 264         struct sched_gate_list *sched, *admin;
 265         ktime_t interval_start, interval_end;
 266         struct sched_entry *entry;
 267
 268         rcu_read_lock();
 269         sched = rcu_dereference(q->oper_sched);
 270         admin = rcu_dereference(q->admin_sched);
 271
 272         entry = find_entry_to_transmit(skb, sch, sched, admin, skb->tstamp,
 273                                        &interval_start, &interval_end, true);
 274         rcu_read_unlock();
 275
 276         return entry;
 277 }
 278
 279 static bool taprio_flags_valid(u32 flags)
 280 {
 281         /* Make sure no other flag bits are set. */
 282         if (flags & ~(TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST |
 283                       TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 284                 return false;
 285         /* txtime-assist and full offload are mutually exclusive */
 286         if ((flags & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST) &&
 287             (flags & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 288                 return false;
 289         return true;
 290 }
 291
 292 /* This returns the tstamp value set by TCP in terms of the set clock. */
 293 static ktime_t get_tcp_tstamp(struct taprio_sched *q, struct sk_buff *skb)
 294 {
 295         unsigned int offset = skb_network_offset(skb);
 296         const struct ipv6hdr *ipv6h;
 297         const struct iphdr *iph;
 298         struct ipv6hdr _ipv6h;
 299
 300         ipv6h = skb_header_pointer(skb, offset, sizeof(_ipv6h), &_ipv6h);
 301         if (!ipv6h)
 302                 return 0;
 303
 304         if (ipv6h->version == 4) {
 305                 iph = (struct iphdr *)ipv6h;
 306                 offset += iph->ihl * 4;
 307
 308                 /* special-case 6in4 tunnelling, as that is a common way to get
 309                  * v6 connectivity in the home
 310                  */
 311                 if (iph->protocol == IPPROTO_IPV6) {
 312                         ipv6h = skb_header_pointer(skb, offset,
 313                                                    sizeof(_ipv6h), &_ipv6h);
 314
 315                         if (!ipv6h || ipv6h->nexthdr != IPPROTO_TCP)
 316                                 return 0;
 317                 } else if (iph->protocol != IPPROTO_TCP) {
 318                         return 0;
 319                 }
 320         } else if (ipv6h->version == 6 && ipv6h->nexthdr != IPPROTO_TCP) {
 321                 return 0;
 322         }
 323
 324         return ktime_mono_to_any(skb->skb_mstamp_ns, q->tk_offset);
 325 }
 326
 327 /* There are a few scenarios where we will have to modify the txtime from
 328  * what is read from next_txtime in sched_entry. They are:
 329  * 1. If txtime is in the past,
 330  *    a. The gate for the traffic class is currently open and packet can be
 331  *       transmitted before it closes, schedule the packet right away.
 332  *    b. If the gate corresponding to the traffic class is going to open later
 333  *       in the cycle, set the txtime of packet to the interval start.
 334  * 2. If txtime is in the future, there are packets corresponding to the
 335  *    current traffic class waiting to be transmitted. So, the following
 336  *    possibilities exist:
 337  *    a. We can transmit the packet before the window containing the txtime
 338  *       closes.
 339  *    b. The window might close before the transmission can be completed
 340  *       successfully. So, schedule the packet in the next open window.
 341  */
 342 static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)
 343 {
 344         ktime_t transmit_end_time, interval_end, interval_start, tcp_tstamp;
 345         struct taprio_sched *q = qdisc_priv(sch);
 346         struct sched_gate_list *sched, *admin;
 347         ktime_t minimum_time, now, txtime;
 348         int len, packet_transmit_time;
 349         struct sched_entry *entry;
 350         bool sched_changed;
 351
 352         now = taprio_get_time(q);
 353         minimum_time = ktime_add_ns(now, q->txtime_delay);
 354
 355         tcp_tstamp = get_tcp_tstamp(q, skb);
 356         minimum_time = max_t(ktime_t, minimum_time, tcp_tstamp);
 357
 358         rcu_read_lock();
 359         admin = rcu_dereference(q->admin_sched);
 360         sched = rcu_dereference(q->oper_sched);
 361         if (admin && ktime_after(minimum_time, admin->base_time))
 362                 switch_schedules(q, &admin, &sched);
 363
 364         /* Until the schedule starts, all the queues are open */
 365         if (!sched || ktime_before(minimum_time, sched->base_time)) {
 366                 txtime = minimum_time;
 367                 goto done;
 368         }
 369
 370         len = qdisc_pkt_len(skb);
 371         packet_transmit_time = length_to_duration(q, len);
 372
 373         do {
 374                 sched_changed = 0;
 375
 376                 entry = find_entry_to_transmit(skb, sch, sched, admin,
 377                                                minimum_time,
 378                                                &interval_start, &interval_end,
 379                                                false);
 380                 if (!entry) {
 381                         txtime = 0;
 382                         goto done;
 383                 }
 384
 385                 txtime = entry->next_txtime;
 386                 txtime = max_t(ktime_t, txtime, minimum_time);
 387                 txtime = max_t(ktime_t, txtime, interval_start);
 388
 389                 if (admin && admin != sched &&
 390                     ktime_after(txtime, admin->base_time)) {
 391                         sched = admin;
 392                         sched_changed = 1;
 393                         continue;
 394                 }
 395
 396                 transmit_end_time = ktime_add(txtime, packet_transmit_time);
 397                 minimum_time = transmit_end_time;
 398
 399                 /* Update the txtime of current entry to the next time it's
 400                  * interval starts.
 401                  */
 402                 if (ktime_after(transmit_end_time, interval_end))
 403                         entry->next_txtime = ktime_add(interval_start, sched->cycle_time);
 404         } while (sched_changed || ktime_after(transmit_end_time, interval_end));
 405
 406         entry->next_txtime = transmit_end_time;
 407
 408 done:
 409         rcu_read_unlock();
 410         return txtime;
 411 }
 412
 413 static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 414                           struct sk_buff **to_free)
 415 {
 416         struct taprio_sched *q = qdisc_priv(sch);
 417         struct Qdisc *child;
 418         int queue;
 419
 420         queue = skb_get_queue_mapping(skb);
 421
 422         child = q->qdiscs[queue];
 423         if (unlikely(!child))
 424                 return qdisc_drop(skb, sch, to_free);
 425
 426         if (skb->sk && sock_flag(skb->sk, SOCK_TXTIME)) {
 427                 if (!is_valid_interval(skb, sch))
 428                         return qdisc_drop(skb, sch, to_free);
 429         } else if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
 430                 skb->tstamp = get_packet_txtime(skb, sch);
 431                 if (!skb->tstamp)
 432                         return qdisc_drop(skb, sch, to_free);
 433         }
 434
 435         qdisc_qstats_backlog_inc(sch, skb);
 436         sch->q.qlen++;
 437
 438         return qdisc_enqueue(skb, child, to_free);
 439 }
 440
 441 static struct sk_buff *taprio_peek_soft(struct Qdisc *sch)
 442 {
 443         struct taprio_sched *q = qdisc_priv(sch);
 444         struct net_device *dev = qdisc_dev(sch);
 445         struct sched_entry *entry;
 446         struct sk_buff *skb;
 447         u32 gate_mask;
 448         int i;
 449
 450         rcu_read_lock();
 451         entry = rcu_dereference(q->current_entry);
 452         gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
 453         rcu_read_unlock();
 454
 455         if (!gate_mask)
 456                 return NULL;
 457
 458         for (i = 0; i < dev->num_tx_queues; i++) {
 459                 struct Qdisc *child = q->qdiscs[i];
 460                 int prio;
 461                 u8 tc;
 462
 463                 if (unlikely(!child))
 464                         continue;
 465
 466                 skb = child->ops->peek(child);
 467                 if (!skb)
 468                         continue;
 469
 470                 if (TXTIME_ASSIST_IS_ENABLED(q->flags))
 471                         return skb;
 472
 473                 prio = skb->priority;
 474                 tc = netdev_get_prio_tc_map(dev, prio);
 475
 476                 if (!(gate_mask & BIT(tc)))
 477                         continue;
 478
 479                 return skb;
 480         }
 481
 482         return NULL;
 483 }
 484
 485 static struct sk_buff *taprio_peek_offload(struct Qdisc *sch)
 486 {
 487         struct taprio_sched *q = qdisc_priv(sch);
 488         struct net_device *dev = qdisc_dev(sch);
 489         struct sk_buff *skb;
 490         int i;
 491
 492         for (i = 0; i < dev->num_tx_queues; i++) {
 493                 struct Qdisc *child = q->qdiscs[i];
 494
 495                 if (unlikely(!child))
 496                         continue;
 497
 498                 skb = child->ops->peek(child);
 499                 if (!skb)
 500                         continue;
 501
 502                 return skb;
 503         }
 504
 505         return NULL;
 506 }
 507
 508 static struct sk_buff *taprio_peek(struct Qdisc *sch)
 509 {
 510         struct taprio_sched *q = qdisc_priv(sch);
 511
 512         return q->peek(sch);
 513 }
 514
 515 static void taprio_set_budget(struct taprio_sched *q, struct sched_entry *entry)
 516 {
 517         atomic_set(&entry->budget,
 518                    div64_u64((u64)entry->interval * 1000,
 519                              atomic64_read(&q->picos_per_byte)));
 520 }
 521
 522 static struct sk_buff *taprio_dequeue_soft(struct Qdisc *sch)
 523 {
 524         struct taprio_sched *q = qdisc_priv(sch);
 525         struct net_device *dev = qdisc_dev(sch);
 526         struct sk_buff *skb = NULL;
 527         struct sched_entry *entry;
 528         u32 gate_mask;
 529         int i;
 530
 531         rcu_read_lock();
 532         entry = rcu_dereference(q->current_entry);
 533         /* if there's no entry, it means that the schedule didn't
 534          * start yet, so force all gates to be open, this is in
 535          * accordance to IEEE 802.1Qbv-2015 Section 8.6.9.4.5
 536          * "AdminGateSates"
 537          */
 538         gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
 539
 540         if (!gate_mask)
 541                 goto done;
 542
 543         for (i = 0; i < dev->num_tx_queues; i++) {
 544                 struct Qdisc *child = q->qdiscs[i];
 545                 ktime_t guard;
 546                 int prio;
 547                 int len;
 548                 u8 tc;
 549
 550                 if (unlikely(!child))
 551                         continue;
 552
 553                 if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
 554                         skb = child->ops->dequeue(child);
 555                         if (!skb)
 556                                 continue;
 557                         goto skb_found;
 558                 }
 559
 560                 skb = child->ops->peek(child);
 561                 if (!skb)
 562                         continue;
 563
 564                 prio = skb->priority;
 565                 tc = netdev_get_prio_tc_map(dev, prio);
 566
 567                 if (!(gate_mask & BIT(tc)))
 568                         continue;
 569
 570                 len = qdisc_pkt_len(skb);
 571                 guard = ktime_add_ns(taprio_get_time(q),
 572                                      length_to_duration(q, len));
 573
 574                 /* In the case that there's no gate entry, there's no
 575                  * guard band ...
 576                  */
 577                 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 578                     ktime_after(guard, entry->close_time))
 579                         continue;
 580
 581                 /* ... and no budget. */
 582                 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 583                     atomic_sub_return(len, &entry->budget) < 0)
 584                         continue;
 585
 586                 skb = child->ops->dequeue(child);
 587                 if (unlikely(!skb))
 588                         goto done;
 589
 590 skb_found:
 591                 qdisc_bstats_update(sch, skb);
 592                 qdisc_qstats_backlog_dec(sch, skb);
 593                 sch->q.qlen--;
 594
 595                 goto done;
 596         }
 597
 598 done:
 599         rcu_read_unlock();
 600
 601         return skb;
 602 }
 603
 604 static struct sk_buff *taprio_dequeue_offload(struct Qdisc *sch)
 605 {
 606         struct taprio_sched *q = qdisc_priv(sch);
 607         struct net_device *dev = qdisc_dev(sch);
 608         struct sk_buff *skb;
 609         int i;
 610
 611         for (i = 0; i < dev->num_tx_queues; i++) {
 612                 struct Qdisc *child = q->qdiscs[i];
 613
 614                 if (unlikely(!child))
 615                         continue;
 616
 617                 skb = child->ops->dequeue(child);
 618                 if (unlikely(!skb))
 619                         continue;
 620
 621                 qdisc_bstats_update(sch, skb);
 622                 qdisc_qstats_backlog_dec(sch, skb);
 623                 sch->q.qlen--;
 624
 625                 return skb;
 626         }
 627
 628         return NULL;
 629 }
 630
 631 static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
 632 {
 633         struct taprio_sched *q = qdisc_priv(sch);
 634
 635         return q->dequeue(sch);
 636 }
 637
 638 static bool should_restart_cycle(const struct sched_gate_list *oper,
 639                                  const struct sched_entry *entry)
 640 {
 641         if (list_is_last(&entry->list, &oper->entries))
 642                 return true;
 643
 644         if (ktime_compare(entry->close_time, oper->cycle_close_time) == 0)
 645                 return true;
 646
 647         return false;
 648 }
 649
 650 static bool should_change_schedules(const struct sched_gate_list *admin,
 651                                     const struct sched_gate_list *oper,
 652                                     ktime_t close_time)
 653 {
 654         ktime_t next_base_time, extension_time;
 655
 656         if (!admin)
 657                 return false;
 658
 659         next_base_time = sched_base_time(admin);
 660
 661         /* This is the simple case, the close_time would fall after
 662          * the next schedule base_time.
 663          */
 664         if (ktime_compare(next_base_time, close_time) <= 0)
 665                 return true;
 666
 667         /* This is the cycle_time_extension case, if the close_time
 668          * plus the amount that can be extended would fall after the
 669          * next schedule base_time, we can extend the current schedule
 670          * for that amount.
 671          */
 672         extension_time = ktime_add_ns(close_time, oper->cycle_time_extension);
 673
 674         /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
 675          * how precisely the extension should be made. So after
 676          * conformance testing, this logic may change.
 677          */
 678         if (ktime_compare(next_base_time, extension_time) <= 0)
 679                 return true;
 680
 681         return false;
 682 }
 683
 684 static enum hrtimer_restart advance_sched(struct hrtimer *timer)
 685 {
 686         struct taprio_sched *q = container_of(timer, struct taprio_sched,
 687                                               advance_timer);
 688         struct sched_gate_list *oper, *admin;
 689         struct sched_entry *entry, *next;
 690         struct Qdisc *sch = q->root;
 691         ktime_t close_time;
 692
 693         spin_lock(&q->current_entry_lock);
 694         entry = rcu_dereference_protected(q->current_entry,
 695                                           lockdep_is_held(&q->current_entry_lock));
 696         oper = rcu_dereference_protected(q->oper_sched,
 697                                          lockdep_is_held(&q->current_entry_lock));
 698         admin = rcu_dereference_protected(q->admin_sched,
 699                                           lockdep_is_held(&q->current_entry_lock));
 700
 701         if (!oper)
 702                 switch_schedules(q, &admin, &oper);
 703
 704         /* This can happen in two cases: 1. this is the very first run
 705          * of this function (i.e. we weren't running any schedule
 706          * previously); 2. The previous schedule just ended. The first
 707          * entry of all schedules are pre-calculated during the
 708          * schedule initialization.
 709          */
 710         if (unlikely(!entry || entry->close_time == oper->base_time)) {
 711                 next = list_first_entry(&oper->entries, struct sched_entry,
 712                                         list);
 713                 close_time = next->close_time;
 714                 goto first_run;
 715         }
 716
 717         if (should_restart_cycle(oper, entry)) {
 718                 next = list_first_entry(&oper->entries, struct sched_entry,
 719                                         list);
 720                 oper->cycle_close_time = ktime_add_ns(oper->cycle_close_time,
 721                                                       oper->cycle_time);
 722         } else {
 723                 next = list_next_entry(entry, list);
 724         }
 725
 726         close_time = ktime_add_ns(entry->close_time, next->interval);
 727         close_time = min_t(ktime_t, close_time, oper->cycle_close_time);
 728
 729         if (should_change_schedules(admin, oper, close_time)) {
 730                 /* Set things so the next time this runs, the new
 731                  * schedule runs.
 732                  */
 733                 close_time = sched_base_time(admin);
 734                 switch_schedules(q, &admin, &oper);
 735         }
 736
 737         next->close_time = close_time;
 738         taprio_set_budget(q, next);
 739
 740 first_run:
 741         rcu_assign_pointer(q->current_entry, next);
 742         spin_unlock(&q->current_entry_lock);
 743
 744         hrtimer_set_expires(&q->advance_timer, close_time);
 745
 746         rcu_read_lock();
 747         __netif_schedule(sch);
 748         rcu_read_unlock();
 749
 750         return HRTIMER_RESTART;
 751 }
 752
 753 static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
 754         [TCA_TAPRIO_SCHED_ENTRY_INDEX]     = { .type = NLA_U32 },
 755         [TCA_TAPRIO_SCHED_ENTRY_CMD]       = { .type = NLA_U8 },
 756         [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
 757         [TCA_TAPRIO_SCHED_ENTRY_INTERVAL]  = { .type = NLA_U32 },
 758 };
 759
 760 static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
 761         [TCA_TAPRIO_ATTR_PRIOMAP]              = {
 762                 .len = sizeof(struct tc_mqprio_qopt)
 763         },
 764         [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST]           = { .type = NLA_NESTED },
 765         [TCA_TAPRIO_ATTR_SCHED_BASE_TIME]            = { .type = NLA_S64 },
 766         [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]         = { .type = NLA_NESTED },
 767         [TCA_TAPRIO_ATTR_SCHED_CLOCKID]              = { .type = NLA_S32 },
 768         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]           = { .type = NLA_S64 },
 769         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
 770         [TCA_TAPRIO_ATTR_FLAGS]                      = { .type = NLA_U32 },
 771 };
 772
 773 static int fill_sched_entry(struct nlattr **tb, struct sched_entry *entry,
 774                             struct netlink_ext_ack *extack)
 775 {
 776         u32 interval = 0;
 777
 778         if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
 779                 entry->command = nla_get_u8(
 780                         tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
 781
 782         if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
 783                 entry->gate_mask = nla_get_u32(
 784                         tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
 785
 786         if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
 787                 interval = nla_get_u32(
 788                         tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
 789
 790         if (interval == 0) {
 791                 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
 792                 return -EINVAL;
 793         }
 794
 795         entry->interval = interval;
 796
 797         return 0;
 798 }
 799
 800 static int parse_sched_entry(struct nlattr *n, struct sched_entry *entry,
 801                              int index, struct netlink_ext_ack *extack)
 802 {
 803         struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
 804         int err;
 805
 806         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
 807                                           entry_policy, NULL);
 808         if (err < 0) {
 809                 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
 810                 return -EINVAL;
 811         }
 812
 813         entry->index = index;
 814
 815         return fill_sched_entry(tb, entry, extack);
 816 }
 817
 818 static int parse_sched_list(struct nlattr *list,
 819                             struct sched_gate_list *sched,
 820                             struct netlink_ext_ack *extack)
 821 {
 822         struct nlattr *n;
 823         int err, rem;
 824         int i = 0;
 825
 826         if (!list)
 827                 return -EINVAL;
 828
 829         nla_for_each_nested(n, list, rem) {
 830                 struct sched_entry *entry;
 831
 832                 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
 833                         NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
 834                         continue;
 835                 }
 836
 837                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 838                 if (!entry) {
 839                         NL_SET_ERR_MSG(extack, "Not enough memory for entry");
 840                         return -ENOMEM;
 841                 }
 842
 843                 err = parse_sched_entry(n, entry, i, extack);
 844                 if (err < 0) {
 845                         kfree(entry);
 846                         return err;
 847                 }
 848
 849                 list_add_tail(&entry->list, &sched->entries);
 850                 i++;
 851         }
 852
 853         sched->num_entries = i;
 854
 855         return i;
 856 }
 857
 858 static int parse_taprio_schedule(struct nlattr **tb,
 859                                  struct sched_gate_list *new,
 860                                  struct netlink_ext_ack *extack)
 861 {
 862         int err = 0;
 863
 864         if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
 865                 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
 866                 return -ENOTSUPP;
 867         }
 868
 869         if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
 870                 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
 871
 872         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
 873                 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
 874
 875         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
 876                 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
 877
 878         if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
 879                 err = parse_sched_list(
 880                         tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST], new, extack);
 881         if (err < 0)
 882                 return err;
 883
 884         if (!new->cycle_time) {
 885                 struct sched_entry *entry;
 886                 ktime_t cycle = 0;
 887
 888                 list_for_each_entry(entry, &new->entries, list)
 889                         cycle = ktime_add_ns(cycle, entry->interval);
 890                 new->cycle_time = cycle;
 891         }
 892
 893         return 0;
 894 }
 895
 896 static int taprio_parse_mqprio_opt(struct net_device *dev,
 897                                    struct tc_mqprio_qopt *qopt,
 898                                    struct netlink_ext_ack *extack,
 899                                    u32 taprio_flags)
 900 {
 901         int i, j;
 902
 903         if (!qopt && !dev->num_tc) {
 904                 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
 905                 return -EINVAL;
 906         }
 907
 908         /* If num_tc is already set, it means that the user already
 909          * configured the mqprio part
 910          */
 911         if (dev->num_tc)
 912                 return 0;
 913
 914         /* Verify num_tc is not out of max range */
 915         if (qopt->num_tc > TC_MAX_QUEUE) {
 916                 NL_SET_ERR_MSG(extack, "Number of traffic classes is outside valid range");
 917                 return -EINVAL;
 918         }
 919
 920         /* taprio imposes that traffic classes map 1:n to tx queues */
 921         if (qopt->num_tc > dev->num_tx_queues) {
 922                 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
 923                 return -EINVAL;
 924         }
 925
 926         /* Verify priority mapping uses valid tcs */
 927         for (i = 0; i <= TC_BITMASK; i++) {
 928                 if (qopt->prio_tc_map[i] >= qopt->num_tc) {
 929                         NL_SET_ERR_MSG(extack, "Invalid traffic class in priority to traffic class mapping");
 930                         return -EINVAL;
 931                 }
 932         }
 933
 934         for (i = 0; i < qopt->num_tc; i++) {
 935                 unsigned int last = qopt->offset[i] + qopt->count[i];
 936
 937                 /* Verify the queue count is in tx range being equal to the
 938                  * real_num_tx_queues indicates the last queue is in use.
 939                  */
 940                 if (qopt->offset[i] >= dev->num_tx_queues ||
 941                     !qopt->count[i] ||
 942                     last > dev->real_num_tx_queues) {
 943                         NL_SET_ERR_MSG(extack, "Invalid queue in traffic class to queue mapping");
 944                         return -EINVAL;
 945                 }
 946
 947                 if (TXTIME_ASSIST_IS_ENABLED(taprio_flags))
 948                         continue;
 949
 950                 /* Verify that the offset and counts do not overlap */
 951                 for (j = i + 1; j < qopt->num_tc; j++) {
 952                         if (last > qopt->offset[j]) {
 953                                 NL_SET_ERR_MSG(extack, "Detected overlap in the traffic class to queue mapping");
 954                                 return -EINVAL;
 955                         }
 956                 }
 957         }
 958
 959         return 0;
 960 }
 961
 962 static int taprio_get_start_time(struct Qdisc *sch,
 963                                  struct sched_gate_list *sched,
 964                                  ktime_t *start)
 965 {
 966         struct taprio_sched *q = qdisc_priv(sch);
 967         ktime_t now, base, cycle;
 968         s64 n;
 969
 970         base = sched_base_time(sched);
 971         now = taprio_get_time(q);
 972
 973         if (ktime_after(base, now)) {
 974                 *start = base;
 975                 return 0;
 976         }
 977
 978         cycle = sched->cycle_time;
 979
 980         /* The qdisc is expected to have at least one sched_entry.  Moreover,
 981          * any entry must have 'interval' > 0. Thus if the cycle time is zero,
 982          * something went really wrong. In that case, we should warn about this
 983          * inconsistent state and return error.
 984          */
 985         if (WARN_ON(!cycle))
 986                 return -EFAULT;
 987
 988         /* Schedule the start time for the beginning of the next
 989          * cycle.
 990          */
 991         n = div64_s64(ktime_sub_ns(now, base), cycle);
 992         *start = ktime_add_ns(base, (n + 1) * cycle);
 993         return 0;
 994 }
 995
 996 static void setup_first_close_time(struct taprio_sched *q,
 997                                    struct sched_gate_list *sched, ktime_t base)
 998 {
 999         struct sched_entry *first;
1000         ktime_t cycle;
1001
1002         first = list_first_entry(&sched->entries,
1003                                  struct sched_entry, list);
1004
1005         cycle = sched->cycle_time;
1006
1007         /* FIXME: find a better place to do this */
1008         sched->cycle_close_time = ktime_add_ns(base, cycle);
1009
1010         first->close_time = ktime_add_ns(base, first->interval);
1011         taprio_set_budget(q, first);
1012         rcu_assign_pointer(q->current_entry, NULL);
1013 }
1014
1015 static void taprio_start_sched(struct Qdisc *sch,
1016                                ktime_t start, struct sched_gate_list *new)
1017 {
1018         struct taprio_sched *q = qdisc_priv(sch);
1019         ktime_t expires;
1020
1021         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1022                 return;
1023
1024         expires = hrtimer_get_expires(&q->advance_timer);
1025         if (expires == 0)
1026                 expires = KTIME_MAX;
1027
1028         /* If the new schedule starts before the next expiration, we
1029          * reprogram it to the earliest one, so we change the admin
1030          * schedule to the operational one at the right time.
1031          */
1032         start = min_t(ktime_t, start, expires);
1033
1034         hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1035 }
1036
1037 static void taprio_set_picos_per_byte(struct net_device *dev,
1038                                       struct taprio_sched *q)
1039 {
1040         struct ethtool_link_ksettings ecmd;
1041         int speed = SPEED_10;
1042         int picos_per_byte;
1043         int err;
1044
1045         err = __ethtool_get_link_ksettings(dev, &ecmd);
1046         if (err < 0)
1047                 goto skip;
1048
1049         if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1050                 speed = ecmd.base.speed;
1051
1052 skip:
1053         picos_per_byte = (USEC_PER_SEC * 8) / speed;
1054
1055         atomic64_set(&q->picos_per_byte, picos_per_byte);
1056         netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1057                    dev->name, (long long)atomic64_read(&q->picos_per_byte),
1058                    ecmd.base.speed);
1059 }
1060
1061 static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1062                                void *ptr)
1063 {
1064         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1065         struct net_device *qdev;
1066         struct taprio_sched *q;
1067         bool found = false;
1068
1069         ASSERT_RTNL();
1070
1071         if (event != NETDEV_UP && event != NETDEV_CHANGE)
1072                 return NOTIFY_DONE;
1073
1074         spin_lock(&taprio_list_lock);
1075         list_for_each_entry(q, &taprio_list, taprio_list) {
1076                 qdev = qdisc_dev(q->root);
1077                 if (qdev == dev) {
1078                         found = true;
1079                         break;
1080                 }
1081         }
1082         spin_unlock(&taprio_list_lock);
1083
1084         if (found)
1085                 taprio_set_picos_per_byte(dev, q);
1086
1087         return NOTIFY_DONE;
1088 }
1089
1090 static void setup_txtime(struct taprio_sched *q,
1091                          struct sched_gate_list *sched, ktime_t base)
1092 {
1093         struct sched_entry *entry;
1094         u32 interval = 0;
1095
1096         list_for_each_entry(entry, &sched->entries, list) {
1097                 entry->next_txtime = ktime_add_ns(base, interval);
1098                 interval += entry->interval;
1099         }
1100 }
1101
1102 static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1103 {
1104         size_t size = sizeof(struct tc_taprio_sched_entry) * num_entries +
1105                       sizeof(struct __tc_taprio_qopt_offload);
1106         struct __tc_taprio_qopt_offload *__offload;
1107
1108         __offload = kzalloc(size, GFP_KERNEL);
1109         if (!__offload)
1110                 return NULL;
1111
1112         refcount_set(&__offload->users, 1);
1113
1114         return &__offload->offload;
1115 }
1116
1117 struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1118                                                   *offload)
1119 {
1120         struct __tc_taprio_qopt_offload *__offload;
1121
1122         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1123                                  offload);
1124
1125         refcount_inc(&__offload->users);
1126
1127         return offload;
1128 }
1129 EXPORT_SYMBOL_GPL(taprio_offload_get);
1130
1131 void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1132 {
1133         struct __tc_taprio_qopt_offload *__offload;
1134
1135         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1136                                  offload);
1137
1138         if (!refcount_dec_and_test(&__offload->users))
1139                 return;
1140
1141         kfree(__offload);
1142 }
1143 EXPORT_SYMBOL_GPL(taprio_offload_free);
1144
1145 /* The function will only serve to keep the pointers to the "oper" and "admin"
1146  * schedules valid in relation to their base times, so when calling dump() the
1147  * users looks at the right schedules.
1148  * When using full offload, the admin configuration is promoted to oper at the
1149  * base_time in the PHC time domain.  But because the system time is not
1150  * necessarily in sync with that, we can't just trigger a hrtimer to call
1151  * switch_schedules at the right hardware time.
1152  * At the moment we call this by hand right away from taprio, but in the future
1153  * it will be useful to create a mechanism for drivers to notify taprio of the
1154  * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1155  * This is left as TODO.
1156  */
1157 static void taprio_offload_config_changed(struct taprio_sched *q)
1158 {
1159         struct sched_gate_list *oper, *admin;
1160
1161         spin_lock(&q->current_entry_lock);
1162
1163         oper = rcu_dereference_protected(q->oper_sched,
1164                                          lockdep_is_held(&q->current_entry_lock));
1165         admin = rcu_dereference_protected(q->admin_sched,
1166                                           lockdep_is_held(&q->current_entry_lock));
1167
1168         switch_schedules(q, &admin, &oper);
1169
1170         spin_unlock(&q->current_entry_lock);
1171 }
1172
1173 static void taprio_sched_to_offload(struct taprio_sched *q,
1174                                     struct sched_gate_list *sched,
1175                                     const struct tc_mqprio_qopt *mqprio,
1176                                     struct tc_taprio_qopt_offload *offload)
1177 {
1178         struct sched_entry *entry;
1179         int i = 0;
1180
1181         offload->base_time = sched->base_time;
1182         offload->cycle_time = sched->cycle_time;
1183         offload->cycle_time_extension = sched->cycle_time_extension;
1184
1185         list_for_each_entry(entry, &sched->entries, list) {
1186                 struct tc_taprio_sched_entry *e = &offload->entries[i];
1187
1188                 e->command = entry->command;
1189                 e->interval = entry->interval;
1190                 e->gate_mask = entry->gate_mask;
1191                 i++;
1192         }
1193
1194         offload->num_entries = i;
1195 }
1196
1197 static int taprio_enable_offload(struct net_device *dev,
1198                                  struct tc_mqprio_qopt *mqprio,
1199                                  struct taprio_sched *q,
1200                                  struct sched_gate_list *sched,
1201                                  struct netlink_ext_ack *extack)
1202 {
1203         const struct net_device_ops *ops = dev->netdev_ops;
1204         struct tc_taprio_qopt_offload *offload;
1205         int err = 0;
1206
1207         if (!ops->ndo_setup_tc) {
1208                 NL_SET_ERR_MSG(extack,
1209                                "Device does not support taprio offload");
1210                 return -EOPNOTSUPP;
1211         }
1212
1213         offload = taprio_offload_alloc(sched->num_entries);
1214         if (!offload) {
1215                 NL_SET_ERR_MSG(extack,
1216                                "Not enough memory for enabling offload mode");
1217                 return -ENOMEM;
1218         }
1219         offload->enable = 1;
1220         taprio_sched_to_offload(q, sched, mqprio, offload);
1221
1222         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1223         if (err < 0) {
1224                 NL_SET_ERR_MSG(extack,
1225                                "Device failed to setup taprio offload");
1226                 goto done;
1227         }
1228
1229 done:
1230         taprio_offload_free(offload);
1231
1232         return err;
1233 }
1234
1235 static int taprio_disable_offload(struct net_device *dev,
1236                                   struct taprio_sched *q,
1237                                   struct netlink_ext_ack *extack)
1238 {
1239         const struct net_device_ops *ops = dev->netdev_ops;
1240         struct tc_taprio_qopt_offload *offload;
1241         int err;
1242
1243         if (!FULL_OFFLOAD_IS_ENABLED(q->flags))
1244                 return 0;
1245
1246         if (!ops->ndo_setup_tc)
1247                 return -EOPNOTSUPP;
1248
1249         offload = taprio_offload_alloc(0);
1250         if (!offload) {
1251                 NL_SET_ERR_MSG(extack,
1252                                "Not enough memory to disable offload mode");
1253                 return -ENOMEM;
1254         }
1255         offload->enable = 0;
1256
1257         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1258         if (err < 0) {
1259                 NL_SET_ERR_MSG(extack,
1260                                "Device failed to disable offload");
1261                 goto out;
1262         }
1263
1264 out:
1265         taprio_offload_free(offload);
1266
1267         return err;
1268 }
1269
1270 /* If full offload is enabled, the only possible clockid is the net device's
1271  * PHC. For that reason, specifying a clockid through netlink is incorrect.
1272  * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1273  * in sync with the specified clockid via a user space daemon such as phc2sys.
1274  * For both software taprio and txtime-assist, the clockid is used for the
1275  * hrtimer that advances the schedule and hence mandatory.
1276  */
1277 static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1278                                 struct netlink_ext_ack *extack)
1279 {
1280         struct taprio_sched *q = qdisc_priv(sch);
1281         struct net_device *dev = qdisc_dev(sch);
1282         int err = -EINVAL;
1283
1284         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1285                 const struct ethtool_ops *ops = dev->ethtool_ops;
1286                 struct ethtool_ts_info info = {
1287                         .cmd = ETHTOOL_GET_TS_INFO,
1288                         .phc_index = -1,
1289                 };
1290
1291                 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1292                         NL_SET_ERR_MSG(extack,
1293                                        "The 'clockid' cannot be specified for full offload");
1294                         goto out;
1295                 }
1296
1297                 if (ops && ops->get_ts_info)
1298                         err = ops->get_ts_info(dev, &info);
1299
1300                 if (err || info.phc_index < 0) {
1301                         NL_SET_ERR_MSG(extack,
1302                                        "Device does not have a PTP clock");
1303                         err = -ENOTSUPP;
1304                         goto out;
1305                 }
1306         } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1307                 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1308
1309                 /* We only support static clockids and we don't allow
1310                  * for it to be modified after the first init.
1311                  */
1312                 if (clockid < 0 ||
1313                     (q->clockid != -1 && q->clockid != clockid)) {
1314                         NL_SET_ERR_MSG(extack,
1315                                        "Changing the 'clockid' of a running schedule is not supported");
1316                         err = -ENOTSUPP;
1317                         goto out;
1318                 }
1319
1320                 switch (clockid) {
1321                 case CLOCK_REALTIME:
1322                         q->tk_offset = TK_OFFS_REAL;
1323                         break;
1324                 case CLOCK_MONOTONIC:
1325                         q->tk_offset = TK_OFFS_MAX;
1326                         break;
1327                 case CLOCK_BOOTTIME:
1328                         q->tk_offset = TK_OFFS_BOOT;
1329                         break;
1330                 case CLOCK_TAI:
1331                         q->tk_offset = TK_OFFS_TAI;
1332                         break;
1333                 default:
1334                         NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1335                         err = -EINVAL;
1336                         goto out;
1337                 }
1338
1339                 q->clockid = clockid;
1340         } else {
1341                 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1342                 goto out;
1343         }
1344
1345         /* Everything went ok, return success. */
1346         err = 0;
1347
1348 out:
1349         return err;
1350 }
1351
1352 static int taprio_mqprio_cmp(const struct net_device *dev,
1353                              const struct tc_mqprio_qopt *mqprio)
1354 {
1355         int i;
1356
1357         if (!mqprio || mqprio->num_tc != dev->num_tc)
1358                 return -1;
1359
1360         for (i = 0; i < mqprio->num_tc; i++)
1361                 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1362                     dev->tc_to_txq[i].offset != mqprio->offset[i])
1363                         return -1;
1364
1365         for (i = 0; i <= TC_BITMASK; i++)
1366                 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1367                         return -1;
1368
1369         return 0;
1370 }
1371
1372 /* The semantics of the 'flags' argument in relation to 'change()'
1373  * requests, are interpreted following two rules (which are applied in
1374  * this order): (1) an omitted 'flags' argument is interpreted as
1375  * zero; (2) the 'flags' of a "running" taprio instance cannot be
1376  * changed.
1377  */
1378 static int taprio_new_flags(const struct nlattr *attr, u32 old,
1379                             struct netlink_ext_ack *extack)
1380 {
1381         u32 new = 0;
1382
1383         if (attr)
1384                 new = nla_get_u32(attr);
1385
1386         if (old != TAPRIO_FLAGS_INVALID && old != new) {
1387                 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1388                 return -EOPNOTSUPP;
1389         }
1390
1391         if (!taprio_flags_valid(new)) {
1392                 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1393                 return -EINVAL;
1394         }
1395
1396         return new;
1397 }
1398
1399 static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1400                          struct netlink_ext_ack *extack)
1401 {
1402         struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1403         struct sched_gate_list *oper, *admin, *new_admin;
1404         struct taprio_sched *q = qdisc_priv(sch);
1405         struct net_device *dev = qdisc_dev(sch);
1406         struct tc_mqprio_qopt *mqprio = NULL;
1407         unsigned long flags;
1408         ktime_t start;
1409         int i, err;
1410
1411         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1412                                           taprio_policy, extack);
1413         if (err < 0)
1414                 return err;
1415
1416         if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1417                 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1418
1419         err = taprio_new_flags(tb[TCA_TAPRIO_ATTR_FLAGS],
1420                                q->flags, extack);
1421         if (err < 0)
1422                 return err;
1423
1424         q->flags = err;
1425
1426         err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1427         if (err < 0)
1428                 return err;
1429
1430         new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1431         if (!new_admin) {
1432                 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1433                 return -ENOMEM;
1434         }
1435         INIT_LIST_HEAD(&new_admin->entries);
1436
1437         rcu_read_lock();
1438         oper = rcu_dereference(q->oper_sched);
1439         admin = rcu_dereference(q->admin_sched);
1440         rcu_read_unlock();
1441
1442         /* no changes - no new mqprio settings */
1443         if (!taprio_mqprio_cmp(dev, mqprio))
1444                 mqprio = NULL;
1445
1446         if (mqprio && (oper || admin)) {
1447                 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1448                 err = -ENOTSUPP;
1449                 goto free_sched;
1450         }
1451
1452         err = parse_taprio_schedule(tb, new_admin, extack);
1453         if (err < 0)
1454                 goto free_sched;
1455
1456         if (new_admin->num_entries == 0) {
1457                 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1458                 err = -EINVAL;
1459                 goto free_sched;
1460         }
1461
1462         err = taprio_parse_clockid(sch, tb, extack);
1463         if (err < 0)
1464                 goto free_sched;
1465
1466         taprio_set_picos_per_byte(dev, q);
1467
1468         if (mqprio) {
1469                 netdev_set_num_tc(dev, mqprio->num_tc);
1470                 for (i = 0; i < mqprio->num_tc; i++)
1471                         netdev_set_tc_queue(dev, i,
1472                                             mqprio->count[i],
1473                                             mqprio->offset[i]);
1474
1475                 /* Always use supplied priority mappings */
1476                 for (i = 0; i <= TC_BITMASK; i++)
1477                         netdev_set_prio_tc_map(dev, i,
1478                                                mqprio->prio_tc_map[i]);
1479         }
1480
1481         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1482                 err = taprio_enable_offload(dev, mqprio, q, new_admin, extack);
1483         else
1484                 err = taprio_disable_offload(dev, q, extack);
1485         if (err)
1486                 goto free_sched;
1487
1488         /* Protects against enqueue()/dequeue() */
1489         spin_lock_bh(qdisc_lock(sch));
1490
1491         if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1492                 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1493                         NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1494                         err = -EINVAL;
1495                         goto unlock;
1496                 }
1497
1498                 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1499         }
1500
1501         if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1502             !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1503             !hrtimer_active(&q->advance_timer)) {
1504                 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1505                 q->advance_timer.function = advance_sched;
1506         }
1507
1508         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1509                 q->dequeue = taprio_dequeue_offload;
1510                 q->peek = taprio_peek_offload;
1511         } else {
1512                 /* Be sure to always keep the function pointers
1513                  * in a consistent state.
1514                  */
1515                 q->dequeue = taprio_dequeue_soft;
1516                 q->peek = taprio_peek_soft;
1517         }
1518
1519         err = taprio_get_start_time(sch, new_admin, &start);
1520         if (err < 0) {
1521                 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1522                 goto unlock;
1523         }
1524
1525         if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1526                 setup_txtime(q, new_admin, start);
1527
1528                 if (!oper) {
1529                         rcu_assign_pointer(q->oper_sched, new_admin);
1530                         err = 0;
1531                         new_admin = NULL;
1532                         goto unlock;
1533                 }
1534
1535                 rcu_assign_pointer(q->admin_sched, new_admin);
1536                 if (admin)
1537                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1538         } else {
1539                 setup_first_close_time(q, new_admin, start);
1540
1541                 /* Protects against advance_sched() */
1542                 spin_lock_irqsave(&q->current_entry_lock, flags);
1543
1544                 taprio_start_sched(sch, start, new_admin);
1545
1546                 rcu_assign_pointer(q->admin_sched, new_admin);
1547                 if (admin)
1548                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1549
1550                 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1551
1552                 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1553                         taprio_offload_config_changed(q);
1554         }
1555
1556         new_admin = NULL;
1557         err = 0;
1558
1559 unlock:
1560         spin_unlock_bh(qdisc_lock(sch));
1561
1562 free_sched:
1563         if (new_admin)
1564                 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
1565
1566         return err;
1567 }
1568
1569 static void taprio_destroy(struct Qdisc *sch)
1570 {
1571         struct taprio_sched *q = qdisc_priv(sch);
1572         struct net_device *dev = qdisc_dev(sch);
1573         unsigned int i;
1574
1575         spin_lock(&taprio_list_lock);
1576         list_del(&q->taprio_list);
1577         spin_unlock(&taprio_list_lock);
1578
1579         hrtimer_cancel(&q->advance_timer);
1580
1581         taprio_disable_offload(dev, q, NULL);
1582
1583         if (q->qdiscs) {
1584                 for (i = 0; i < dev->num_tx_queues && q->qdiscs[i]; i++)
1585                         qdisc_put(q->qdiscs[i]);
1586
1587                 kfree(q->qdiscs);
1588         }
1589         q->qdiscs = NULL;
1590
1591         netdev_reset_tc(dev);
1592
1593         if (q->oper_sched)
1594                 call_rcu(&q->oper_sched->rcu, taprio_free_sched_cb);
1595
1596         if (q->admin_sched)
1597                 call_rcu(&q->admin_sched->rcu, taprio_free_sched_cb);
1598 }
1599
1600 static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
1601                        struct netlink_ext_ack *extack)
1602 {
1603         struct taprio_sched *q = qdisc_priv(sch);
1604         struct net_device *dev = qdisc_dev(sch);
1605         int i;
1606
1607         spin_lock_init(&q->current_entry_lock);
1608
1609         hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
1610         q->advance_timer.function = advance_sched;
1611
1612         q->dequeue = taprio_dequeue_soft;
1613         q->peek = taprio_peek_soft;
1614
1615         q->root = sch;
1616
1617         /* We only support static clockids. Use an invalid value as default
1618          * and get the valid one on taprio_change().
1619          */
1620         q->clockid = -1;
1621         q->flags = TAPRIO_FLAGS_INVALID;
1622
1623         spin_lock(&taprio_list_lock);
1624         list_add(&q->taprio_list, &taprio_list);
1625         spin_unlock(&taprio_list_lock);
1626
1627         if (sch->parent != TC_H_ROOT)
1628                 return -EOPNOTSUPP;
1629
1630         if (!netif_is_multiqueue(dev))
1631                 return -EOPNOTSUPP;
1632
1633         /* pre-allocate qdisc, attachment can't fail */
1634         q->qdiscs = kcalloc(dev->num_tx_queues,
1635                             sizeof(q->qdiscs[0]),
1636                             GFP_KERNEL);
1637
1638         if (!q->qdiscs)
1639                 return -ENOMEM;
1640
1641         if (!opt)
1642                 return -EINVAL;
1643
1644         for (i = 0; i < dev->num_tx_queues; i++) {
1645                 struct netdev_queue *dev_queue;
1646                 struct Qdisc *qdisc;
1647
1648                 dev_queue = netdev_get_tx_queue(dev, i);
1649                 qdisc = qdisc_create_dflt(dev_queue,
1650                                           &pfifo_qdisc_ops,
1651                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
1652                                                     TC_H_MIN(i + 1)),
1653                                           extack);
1654                 if (!qdisc)
1655                         return -ENOMEM;
1656
1657                 if (i < dev->real_num_tx_queues)
1658                         qdisc_hash_add(qdisc, false);
1659
1660                 q->qdiscs[i] = qdisc;
1661         }
1662
1663         return taprio_change(sch, opt, extack);
1664 }
1665
1666 static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
1667                                              unsigned long cl)
1668 {
1669         struct net_device *dev = qdisc_dev(sch);
1670         unsigned long ntx = cl - 1;
1671
1672         if (ntx >= dev->num_tx_queues)
1673                 return NULL;
1674
1675         return netdev_get_tx_queue(dev, ntx);
1676 }
1677
1678 static int taprio_graft(struct Qdisc *sch, unsigned long cl,
1679                         struct Qdisc *new, struct Qdisc **old,
1680                         struct netlink_ext_ack *extack)
1681 {
1682         struct taprio_sched *q = qdisc_priv(sch);
1683         struct net_device *dev = qdisc_dev(sch);
1684         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1685
1686         if (!dev_queue)
1687                 return -EINVAL;
1688
1689         if (dev->flags & IFF_UP)
1690                 dev_deactivate(dev);
1691
1692         *old = q->qdiscs[cl - 1];
1693         q->qdiscs[cl - 1] = new;
1694
1695         if (new)
1696                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1697
1698         if (dev->flags & IFF_UP)
1699                 dev_activate(dev);
1700
1701         return 0;
1702 }
1703
1704 static int dump_entry(struct sk_buff *msg,
1705                       const struct sched_entry *entry)
1706 {
1707         struct nlattr *item;
1708
1709         item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
1710         if (!item)
1711                 return -ENOSPC;
1712
1713         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
1714                 goto nla_put_failure;
1715
1716         if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
1717                 goto nla_put_failure;
1718
1719         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
1720                         entry->gate_mask))
1721                 goto nla_put_failure;
1722
1723         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
1724                         entry->interval))
1725                 goto nla_put_failure;
1726
1727         return nla_nest_end(msg, item);
1728
1729 nla_put_failure:
1730         nla_nest_cancel(msg, item);
1731         return -1;
1732 }
1733
1734 static int dump_schedule(struct sk_buff *msg,
1735                          const struct sched_gate_list *root)
1736 {
1737         struct nlattr *entry_list;
1738         struct sched_entry *entry;
1739
1740         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
1741                         root->base_time, TCA_TAPRIO_PAD))
1742                 return -1;
1743
1744         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
1745                         root->cycle_time, TCA_TAPRIO_PAD))
1746                 return -1;
1747
1748         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
1749                         root->cycle_time_extension, TCA_TAPRIO_PAD))
1750                 return -1;
1751
1752         entry_list = nla_nest_start_noflag(msg,
1753                                            TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
1754         if (!entry_list)
1755                 goto error_nest;
1756
1757         list_for_each_entry(entry, &root->entries, list) {
1758                 if (dump_entry(msg, entry) < 0)
1759                         goto error_nest;
1760         }
1761
1762         nla_nest_end(msg, entry_list);
1763         return 0;
1764
1765 error_nest:
1766         nla_nest_cancel(msg, entry_list);
1767         return -1;
1768 }
1769
1770 static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
1771 {
1772         struct taprio_sched *q = qdisc_priv(sch);
1773         struct net_device *dev = qdisc_dev(sch);
1774         struct sched_gate_list *oper, *admin;
1775         struct tc_mqprio_qopt opt = { 0 };
1776         struct nlattr *nest, *sched_nest;
1777         unsigned int i;
1778
1779         rcu_read_lock();
1780         oper = rcu_dereference(q->oper_sched);
1781         admin = rcu_dereference(q->admin_sched);
1782
1783         opt.num_tc = netdev_get_num_tc(dev);
1784         memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
1785
1786         for (i = 0; i < netdev_get_num_tc(dev); i++) {
1787                 opt.count[i] = dev->tc_to_txq[i].count;
1788                 opt.offset[i] = dev->tc_to_txq[i].offset;
1789         }
1790
1791         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1792         if (!nest)
1793                 goto start_error;
1794
1795         if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
1796                 goto options_error;
1797
1798         if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1799             nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
1800                 goto options_error;
1801
1802         if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
1803                 goto options_error;
1804
1805         if (q->txtime_delay &&
1806             nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
1807                 goto options_error;
1808
1809         if (oper && dump_schedule(skb, oper))
1810                 goto options_error;
1811
1812         if (!admin)
1813                 goto done;
1814
1815         sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
1816         if (!sched_nest)
1817                 goto options_error;
1818
1819         if (dump_schedule(skb, admin))
1820                 goto admin_error;
1821
1822         nla_nest_end(skb, sched_nest);
1823
1824 done:
1825         rcu_read_unlock();
1826
1827         return nla_nest_end(skb, nest);
1828
1829 admin_error:
1830         nla_nest_cancel(skb, sched_nest);
1831
1832 options_error:
1833         nla_nest_cancel(skb, nest);
1834
1835 start_error:
1836         rcu_read_unlock();
1837         return -ENOSPC;
1838 }
1839
1840 static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
1841 {
1842         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1843
1844         if (!dev_queue)
1845                 return NULL;
1846
1847         return dev_queue->qdisc_sleeping;
1848 }
1849
1850 static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
1851 {
1852         unsigned int ntx = TC_H_MIN(classid);
1853
1854         if (!taprio_queue_get(sch, ntx))
1855                 return 0;
1856         return ntx;
1857 }
1858
1859 static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
1860                              struct sk_buff *skb, struct tcmsg *tcm)
1861 {
1862         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1863
1864         tcm->tcm_parent = TC_H_ROOT;
1865         tcm->tcm_handle |= TC_H_MIN(cl);
1866         tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
1867
1868         return 0;
1869 }
1870
1871 static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
1872                                    struct gnet_dump *d)
1873         __releases(d->lock)
1874         __acquires(d->lock)
1875 {
1876         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1877
1878         sch = dev_queue->qdisc_sleeping;
1879         if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
1880             qdisc_qstats_copy(d, sch) < 0)
1881                 return -1;
1882         return 0;
1883 }
1884
1885 static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1886 {
1887         struct net_device *dev = qdisc_dev(sch);
1888         unsigned long ntx;
1889
1890         if (arg->stop)
1891                 return;
1892
1893         arg->count = arg->skip;
1894         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
1895                 if (arg->fn(sch, ntx + 1, arg) < 0) {
1896                         arg->stop = 1;
1897                         break;
1898                 }
1899                 arg->count++;
1900         }
1901 }
1902
1903 static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
1904                                                 struct tcmsg *tcm)
1905 {
1906         return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
1907 }
1908
1909 static const struct Qdisc_class_ops taprio_class_ops = {
1910         .graft          = taprio_graft,
1911         .leaf           = taprio_leaf,
1912         .find           = taprio_find,
1913         .walk           = taprio_walk,
1914         .dump           = taprio_dump_class,
1915         .dump_stats     = taprio_dump_class_stats,
1916         .select_queue   = taprio_select_queue,
1917 };
1918
1919 static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
1920         .cl_ops         = &taprio_class_ops,
1921         .id             = "taprio",
1922         .priv_size      = sizeof(struct taprio_sched),
1923         .init           = taprio_init,
1924         .change         = taprio_change,
1925         .destroy        = taprio_destroy,
1926         .peek           = taprio_peek,
1927         .dequeue        = taprio_dequeue,
1928         .enqueue        = taprio_enqueue,
1929         .dump           = taprio_dump,
1930         .owner          = THIS_MODULE,
1931 };
1932
1933 static struct notifier_block taprio_device_notifier = {
1934         .notifier_call = taprio_dev_notifier,
1935 };
1936
1937 static int __init taprio_module_init(void)
1938 {
1939         int err = register_netdevice_notifier(&taprio_device_notifier);
1940
1941         if (err)
1942                 return err;
1943
1944         return register_qdisc(&taprio_qdisc_ops);
1945 }
1946
1947 static void __exit taprio_module_exit(void)
1948 {
1949         unregister_qdisc(&taprio_qdisc_ops);
1950         unregister_netdevice_notifier(&taprio_device_notifier);
1951 }
1952
1953 module_init(taprio_module_init);
1954 module_exit(taprio_module_exit);
1955 MODULE_LICENSE("GPL");