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                         skb = NULL;
 569                         continue;
 570                 }
 571
 572                 len = qdisc_pkt_len(skb);
 573                 guard = ktime_add_ns(taprio_get_time(q),
 574                                      length_to_duration(q, len));
 575
 576                 /* In the case that there's no gate entry, there's no
 577                  * guard band ...
 578                  */
 579                 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 580                     ktime_after(guard, entry->close_time)) {
 581                         skb = NULL;
 582                         continue;
 583                 }
 584
 585                 /* ... and no budget. */
 586                 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 587                     atomic_sub_return(len, &entry->budget) < 0) {
 588                         skb = NULL;
 589                         continue;
 590                 }
 591
 592                 skb = child->ops->dequeue(child);
 593                 if (unlikely(!skb))
 594                         goto done;
 595
 596 skb_found:
 597                 qdisc_bstats_update(sch, skb);
 598                 qdisc_qstats_backlog_dec(sch, skb);
 599                 sch->q.qlen--;
 600
 601                 goto done;
 602         }
 603
 604 done:
 605         rcu_read_unlock();
 606
 607         return skb;
 608 }
 609
 610 static struct sk_buff *taprio_dequeue_offload(struct Qdisc *sch)
 611 {
 612         struct taprio_sched *q = qdisc_priv(sch);
 613         struct net_device *dev = qdisc_dev(sch);
 614         struct sk_buff *skb;
 615         int i;
 616
 617         for (i = 0; i < dev->num_tx_queues; i++) {
 618                 struct Qdisc *child = q->qdiscs[i];
 619
 620                 if (unlikely(!child))
 621                         continue;
 622
 623                 skb = child->ops->dequeue(child);
 624                 if (unlikely(!skb))
 625                         continue;
 626
 627                 qdisc_bstats_update(sch, skb);
 628                 qdisc_qstats_backlog_dec(sch, skb);
 629                 sch->q.qlen--;
 630
 631                 return skb;
 632         }
 633
 634         return NULL;
 635 }
 636
 637 static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
 638 {
 639         struct taprio_sched *q = qdisc_priv(sch);
 640
 641         return q->dequeue(sch);
 642 }
 643
 644 static bool should_restart_cycle(const struct sched_gate_list *oper,
 645                                  const struct sched_entry *entry)
 646 {
 647         if (list_is_last(&entry->list, &oper->entries))
 648                 return true;
 649
 650         if (ktime_compare(entry->close_time, oper->cycle_close_time) == 0)
 651                 return true;
 652
 653         return false;
 654 }
 655
 656 static bool should_change_schedules(const struct sched_gate_list *admin,
 657                                     const struct sched_gate_list *oper,
 658                                     ktime_t close_time)
 659 {
 660         ktime_t next_base_time, extension_time;
 661
 662         if (!admin)
 663                 return false;
 664
 665         next_base_time = sched_base_time(admin);
 666
 667         /* This is the simple case, the close_time would fall after
 668          * the next schedule base_time.
 669          */
 670         if (ktime_compare(next_base_time, close_time) <= 0)
 671                 return true;
 672
 673         /* This is the cycle_time_extension case, if the close_time
 674          * plus the amount that can be extended would fall after the
 675          * next schedule base_time, we can extend the current schedule
 676          * for that amount.
 677          */
 678         extension_time = ktime_add_ns(close_time, oper->cycle_time_extension);
 679
 680         /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
 681          * how precisely the extension should be made. So after
 682          * conformance testing, this logic may change.
 683          */
 684         if (ktime_compare(next_base_time, extension_time) <= 0)
 685                 return true;
 686
 687         return false;
 688 }
 689
 690 static enum hrtimer_restart advance_sched(struct hrtimer *timer)
 691 {
 692         struct taprio_sched *q = container_of(timer, struct taprio_sched,
 693                                               advance_timer);
 694         struct sched_gate_list *oper, *admin;
 695         struct sched_entry *entry, *next;
 696         struct Qdisc *sch = q->root;
 697         ktime_t close_time;
 698
 699         spin_lock(&q->current_entry_lock);
 700         entry = rcu_dereference_protected(q->current_entry,
 701                                           lockdep_is_held(&q->current_entry_lock));
 702         oper = rcu_dereference_protected(q->oper_sched,
 703                                          lockdep_is_held(&q->current_entry_lock));
 704         admin = rcu_dereference_protected(q->admin_sched,
 705                                           lockdep_is_held(&q->current_entry_lock));
 706
 707         if (!oper)
 708                 switch_schedules(q, &admin, &oper);
 709
 710         /* This can happen in two cases: 1. this is the very first run
 711          * of this function (i.e. we weren't running any schedule
 712          * previously); 2. The previous schedule just ended. The first
 713          * entry of all schedules are pre-calculated during the
 714          * schedule initialization.
 715          */
 716         if (unlikely(!entry || entry->close_time == oper->base_time)) {
 717                 next = list_first_entry(&oper->entries, struct sched_entry,
 718                                         list);
 719                 close_time = next->close_time;
 720                 goto first_run;
 721         }
 722
 723         if (should_restart_cycle(oper, entry)) {
 724                 next = list_first_entry(&oper->entries, struct sched_entry,
 725                                         list);
 726                 oper->cycle_close_time = ktime_add_ns(oper->cycle_close_time,
 727                                                       oper->cycle_time);
 728         } else {
 729                 next = list_next_entry(entry, list);
 730         }
 731
 732         close_time = ktime_add_ns(entry->close_time, next->interval);
 733         close_time = min_t(ktime_t, close_time, oper->cycle_close_time);
 734
 735         if (should_change_schedules(admin, oper, close_time)) {
 736                 /* Set things so the next time this runs, the new
 737                  * schedule runs.
 738                  */
 739                 close_time = sched_base_time(admin);
 740                 switch_schedules(q, &admin, &oper);
 741         }
 742
 743         next->close_time = close_time;
 744         taprio_set_budget(q, next);
 745
 746 first_run:
 747         rcu_assign_pointer(q->current_entry, next);
 748         spin_unlock(&q->current_entry_lock);
 749
 750         hrtimer_set_expires(&q->advance_timer, close_time);
 751
 752         rcu_read_lock();
 753         __netif_schedule(sch);
 754         rcu_read_unlock();
 755
 756         return HRTIMER_RESTART;
 757 }
 758
 759 static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
 760         [TCA_TAPRIO_SCHED_ENTRY_INDEX]     = { .type = NLA_U32 },
 761         [TCA_TAPRIO_SCHED_ENTRY_CMD]       = { .type = NLA_U8 },
 762         [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
 763         [TCA_TAPRIO_SCHED_ENTRY_INTERVAL]  = { .type = NLA_U32 },
 764 };
 765
 766 static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
 767         [TCA_TAPRIO_ATTR_PRIOMAP]              = {
 768                 .len = sizeof(struct tc_mqprio_qopt)
 769         },
 770         [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST]           = { .type = NLA_NESTED },
 771         [TCA_TAPRIO_ATTR_SCHED_BASE_TIME]            = { .type = NLA_S64 },
 772         [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]         = { .type = NLA_NESTED },
 773         [TCA_TAPRIO_ATTR_SCHED_CLOCKID]              = { .type = NLA_S32 },
 774         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]           = { .type = NLA_S64 },
 775         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
 776         [TCA_TAPRIO_ATTR_FLAGS]                      = { .type = NLA_U32 },
 777         [TCA_TAPRIO_ATTR_TXTIME_DELAY]               = { .type = NLA_U32 },
 778 };
 779
 780 static int fill_sched_entry(struct nlattr **tb, struct sched_entry *entry,
 781                             struct netlink_ext_ack *extack)
 782 {
 783         u32 interval = 0;
 784
 785         if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
 786                 entry->command = nla_get_u8(
 787                         tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
 788
 789         if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
 790                 entry->gate_mask = nla_get_u32(
 791                         tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
 792
 793         if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
 794                 interval = nla_get_u32(
 795                         tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
 796
 797         if (interval == 0) {
 798                 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
 799                 return -EINVAL;
 800         }
 801
 802         entry->interval = interval;
 803
 804         return 0;
 805 }
 806
 807 static int parse_sched_entry(struct nlattr *n, struct sched_entry *entry,
 808                              int index, struct netlink_ext_ack *extack)
 809 {
 810         struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
 811         int err;
 812
 813         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
 814                                           entry_policy, NULL);
 815         if (err < 0) {
 816                 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
 817                 return -EINVAL;
 818         }
 819
 820         entry->index = index;
 821
 822         return fill_sched_entry(tb, entry, extack);
 823 }
 824
 825 static int parse_sched_list(struct nlattr *list,
 826                             struct sched_gate_list *sched,
 827                             struct netlink_ext_ack *extack)
 828 {
 829         struct nlattr *n;
 830         int err, rem;
 831         int i = 0;
 832
 833         if (!list)
 834                 return -EINVAL;
 835
 836         nla_for_each_nested(n, list, rem) {
 837                 struct sched_entry *entry;
 838
 839                 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
 840                         NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
 841                         continue;
 842                 }
 843
 844                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 845                 if (!entry) {
 846                         NL_SET_ERR_MSG(extack, "Not enough memory for entry");
 847                         return -ENOMEM;
 848                 }
 849
 850                 err = parse_sched_entry(n, entry, i, extack);
 851                 if (err < 0) {
 852                         kfree(entry);
 853                         return err;
 854                 }
 855
 856                 list_add_tail(&entry->list, &sched->entries);
 857                 i++;
 858         }
 859
 860         sched->num_entries = i;
 861
 862         return i;
 863 }
 864
 865 static int parse_taprio_schedule(struct nlattr **tb,
 866                                  struct sched_gate_list *new,
 867                                  struct netlink_ext_ack *extack)
 868 {
 869         int err = 0;
 870
 871         if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
 872                 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
 873                 return -ENOTSUPP;
 874         }
 875
 876         if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
 877                 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
 878
 879         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
 880                 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
 881
 882         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
 883                 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
 884
 885         if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
 886                 err = parse_sched_list(
 887                         tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST], new, extack);
 888         if (err < 0)
 889                 return err;
 890
 891         if (!new->cycle_time) {
 892                 struct sched_entry *entry;
 893                 ktime_t cycle = 0;
 894
 895                 list_for_each_entry(entry, &new->entries, list)
 896                         cycle = ktime_add_ns(cycle, entry->interval);
 897                 new->cycle_time = cycle;
 898         }
 899
 900         return 0;
 901 }
 902
 903 static int taprio_parse_mqprio_opt(struct net_device *dev,
 904                                    struct tc_mqprio_qopt *qopt,
 905                                    struct netlink_ext_ack *extack,
 906                                    u32 taprio_flags)
 907 {
 908         int i, j;
 909
 910         if (!qopt && !dev->num_tc) {
 911                 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
 912                 return -EINVAL;
 913         }
 914
 915         /* If num_tc is already set, it means that the user already
 916          * configured the mqprio part
 917          */
 918         if (dev->num_tc)
 919                 return 0;
 920
 921         /* Verify num_tc is not out of max range */
 922         if (qopt->num_tc > TC_MAX_QUEUE) {
 923                 NL_SET_ERR_MSG(extack, "Number of traffic classes is outside valid range");
 924                 return -EINVAL;
 925         }
 926
 927         /* taprio imposes that traffic classes map 1:n to tx queues */
 928         if (qopt->num_tc > dev->num_tx_queues) {
 929                 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
 930                 return -EINVAL;
 931         }
 932
 933         /* Verify priority mapping uses valid tcs */
 934         for (i = 0; i <= TC_BITMASK; i++) {
 935                 if (qopt->prio_tc_map[i] >= qopt->num_tc) {
 936                         NL_SET_ERR_MSG(extack, "Invalid traffic class in priority to traffic class mapping");
 937                         return -EINVAL;
 938                 }
 939         }
 940
 941         for (i = 0; i < qopt->num_tc; i++) {
 942                 unsigned int last = qopt->offset[i] + qopt->count[i];
 943
 944                 /* Verify the queue count is in tx range being equal to the
 945                  * real_num_tx_queues indicates the last queue is in use.
 946                  */
 947                 if (qopt->offset[i] >= dev->num_tx_queues ||
 948                     !qopt->count[i] ||
 949                     last > dev->real_num_tx_queues) {
 950                         NL_SET_ERR_MSG(extack, "Invalid queue in traffic class to queue mapping");
 951                         return -EINVAL;
 952                 }
 953
 954                 if (TXTIME_ASSIST_IS_ENABLED(taprio_flags))
 955                         continue;
 956
 957                 /* Verify that the offset and counts do not overlap */
 958                 for (j = i + 1; j < qopt->num_tc; j++) {
 959                         if (last > qopt->offset[j]) {
 960                                 NL_SET_ERR_MSG(extack, "Detected overlap in the traffic class to queue mapping");
 961                                 return -EINVAL;
 962                         }
 963                 }
 964         }
 965
 966         return 0;
 967 }
 968
 969 static int taprio_get_start_time(struct Qdisc *sch,
 970                                  struct sched_gate_list *sched,
 971                                  ktime_t *start)
 972 {
 973         struct taprio_sched *q = qdisc_priv(sch);
 974         ktime_t now, base, cycle;
 975         s64 n;
 976
 977         base = sched_base_time(sched);
 978         now = taprio_get_time(q);
 979
 980         if (ktime_after(base, now)) {
 981                 *start = base;
 982                 return 0;
 983         }
 984
 985         cycle = sched->cycle_time;
 986
 987         /* The qdisc is expected to have at least one sched_entry.  Moreover,
 988          * any entry must have 'interval' > 0. Thus if the cycle time is zero,
 989          * something went really wrong. In that case, we should warn about this
 990          * inconsistent state and return error.
 991          */
 992         if (WARN_ON(!cycle))
 993                 return -EFAULT;
 994
 995         /* Schedule the start time for the beginning of the next
 996          * cycle.
 997          */
 998         n = div64_s64(ktime_sub_ns(now, base), cycle);
 999         *start = ktime_add_ns(base, (n + 1) * cycle);
1000         return 0;
1001 }
1002
1003 static void setup_first_close_time(struct taprio_sched *q,
1004                                    struct sched_gate_list *sched, ktime_t base)
1005 {
1006         struct sched_entry *first;
1007         ktime_t cycle;
1008
1009         first = list_first_entry(&sched->entries,
1010                                  struct sched_entry, list);
1011
1012         cycle = sched->cycle_time;
1013
1014         /* FIXME: find a better place to do this */
1015         sched->cycle_close_time = ktime_add_ns(base, cycle);
1016
1017         first->close_time = ktime_add_ns(base, first->interval);
1018         taprio_set_budget(q, first);
1019         rcu_assign_pointer(q->current_entry, NULL);
1020 }
1021
1022 static void taprio_start_sched(struct Qdisc *sch,
1023                                ktime_t start, struct sched_gate_list *new)
1024 {
1025         struct taprio_sched *q = qdisc_priv(sch);
1026         ktime_t expires;
1027
1028         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1029                 return;
1030
1031         expires = hrtimer_get_expires(&q->advance_timer);
1032         if (expires == 0)
1033                 expires = KTIME_MAX;
1034
1035         /* If the new schedule starts before the next expiration, we
1036          * reprogram it to the earliest one, so we change the admin
1037          * schedule to the operational one at the right time.
1038          */
1039         start = min_t(ktime_t, start, expires);
1040
1041         hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1042 }
1043
1044 static void taprio_set_picos_per_byte(struct net_device *dev,
1045                                       struct taprio_sched *q)
1046 {
1047         struct ethtool_link_ksettings ecmd;
1048         int speed = SPEED_10;
1049         int picos_per_byte;
1050         int err;
1051
1052         err = __ethtool_get_link_ksettings(dev, &ecmd);
1053         if (err < 0)
1054                 goto skip;
1055
1056         if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1057                 speed = ecmd.base.speed;
1058
1059 skip:
1060         picos_per_byte = (USEC_PER_SEC * 8) / speed;
1061
1062         atomic64_set(&q->picos_per_byte, picos_per_byte);
1063         netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1064                    dev->name, (long long)atomic64_read(&q->picos_per_byte),
1065                    ecmd.base.speed);
1066 }
1067
1068 static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1069                                void *ptr)
1070 {
1071         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1072         struct net_device *qdev;
1073         struct taprio_sched *q;
1074         bool found = false;
1075
1076         ASSERT_RTNL();
1077
1078         if (event != NETDEV_UP && event != NETDEV_CHANGE)
1079                 return NOTIFY_DONE;
1080
1081         spin_lock(&taprio_list_lock);
1082         list_for_each_entry(q, &taprio_list, taprio_list) {
1083                 qdev = qdisc_dev(q->root);
1084                 if (qdev == dev) {
1085                         found = true;
1086                         break;
1087                 }
1088         }
1089         spin_unlock(&taprio_list_lock);
1090
1091         if (found)
1092                 taprio_set_picos_per_byte(dev, q);
1093
1094         return NOTIFY_DONE;
1095 }
1096
1097 static void setup_txtime(struct taprio_sched *q,
1098                          struct sched_gate_list *sched, ktime_t base)
1099 {
1100         struct sched_entry *entry;
1101         u32 interval = 0;
1102
1103         list_for_each_entry(entry, &sched->entries, list) {
1104                 entry->next_txtime = ktime_add_ns(base, interval);
1105                 interval += entry->interval;
1106         }
1107 }
1108
1109 static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1110 {
1111         struct __tc_taprio_qopt_offload *__offload;
1112
1113         __offload = kzalloc(struct_size(__offload, offload.entries, num_entries),
1114                             GFP_KERNEL);
1115         if (!__offload)
1116                 return NULL;
1117
1118         refcount_set(&__offload->users, 1);
1119
1120         return &__offload->offload;
1121 }
1122
1123 struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1124                                                   *offload)
1125 {
1126         struct __tc_taprio_qopt_offload *__offload;
1127
1128         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1129                                  offload);
1130
1131         refcount_inc(&__offload->users);
1132
1133         return offload;
1134 }
1135 EXPORT_SYMBOL_GPL(taprio_offload_get);
1136
1137 void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1138 {
1139         struct __tc_taprio_qopt_offload *__offload;
1140
1141         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1142                                  offload);
1143
1144         if (!refcount_dec_and_test(&__offload->users))
1145                 return;
1146
1147         kfree(__offload);
1148 }
1149 EXPORT_SYMBOL_GPL(taprio_offload_free);
1150
1151 /* The function will only serve to keep the pointers to the "oper" and "admin"
1152  * schedules valid in relation to their base times, so when calling dump() the
1153  * users looks at the right schedules.
1154  * When using full offload, the admin configuration is promoted to oper at the
1155  * base_time in the PHC time domain.  But because the system time is not
1156  * necessarily in sync with that, we can't just trigger a hrtimer to call
1157  * switch_schedules at the right hardware time.
1158  * At the moment we call this by hand right away from taprio, but in the future
1159  * it will be useful to create a mechanism for drivers to notify taprio of the
1160  * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1161  * This is left as TODO.
1162  */
1163 static void taprio_offload_config_changed(struct taprio_sched *q)
1164 {
1165         struct sched_gate_list *oper, *admin;
1166
1167         spin_lock(&q->current_entry_lock);
1168
1169         oper = rcu_dereference_protected(q->oper_sched,
1170                                          lockdep_is_held(&q->current_entry_lock));
1171         admin = rcu_dereference_protected(q->admin_sched,
1172                                           lockdep_is_held(&q->current_entry_lock));
1173
1174         switch_schedules(q, &admin, &oper);
1175
1176         spin_unlock(&q->current_entry_lock);
1177 }
1178
1179 static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
1180 {
1181         u32 i, queue_mask = 0;
1182
1183         for (i = 0; i < dev->num_tc; i++) {
1184                 u32 offset, count;
1185
1186                 if (!(tc_mask & BIT(i)))
1187                         continue;
1188
1189                 offset = dev->tc_to_txq[i].offset;
1190                 count = dev->tc_to_txq[i].count;
1191
1192                 queue_mask |= GENMASK(offset + count - 1, offset);
1193         }
1194
1195         return queue_mask;
1196 }
1197
1198 static void taprio_sched_to_offload(struct net_device *dev,
1199                                     struct sched_gate_list *sched,
1200                                     struct tc_taprio_qopt_offload *offload)
1201 {
1202         struct sched_entry *entry;
1203         int i = 0;
1204
1205         offload->base_time = sched->base_time;
1206         offload->cycle_time = sched->cycle_time;
1207         offload->cycle_time_extension = sched->cycle_time_extension;
1208
1209         list_for_each_entry(entry, &sched->entries, list) {
1210                 struct tc_taprio_sched_entry *e = &offload->entries[i];
1211
1212                 e->command = entry->command;
1213                 e->interval = entry->interval;
1214                 e->gate_mask = tc_map_to_queue_mask(dev, entry->gate_mask);
1215
1216                 i++;
1217         }
1218
1219         offload->num_entries = i;
1220 }
1221
1222 static int taprio_enable_offload(struct net_device *dev,
1223                                  struct taprio_sched *q,
1224                                  struct sched_gate_list *sched,
1225                                  struct netlink_ext_ack *extack)
1226 {
1227         const struct net_device_ops *ops = dev->netdev_ops;
1228         struct tc_taprio_qopt_offload *offload;
1229         int err = 0;
1230
1231         if (!ops->ndo_setup_tc) {
1232                 NL_SET_ERR_MSG(extack,
1233                                "Device does not support taprio offload");
1234                 return -EOPNOTSUPP;
1235         }
1236
1237         offload = taprio_offload_alloc(sched->num_entries);
1238         if (!offload) {
1239                 NL_SET_ERR_MSG(extack,
1240                                "Not enough memory for enabling offload mode");
1241                 return -ENOMEM;
1242         }
1243         offload->enable = 1;
1244         taprio_sched_to_offload(dev, sched, offload);
1245
1246         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1247         if (err < 0) {
1248                 NL_SET_ERR_MSG(extack,
1249                                "Device failed to setup taprio offload");
1250                 goto done;
1251         }
1252
1253 done:
1254         taprio_offload_free(offload);
1255
1256         return err;
1257 }
1258
1259 static int taprio_disable_offload(struct net_device *dev,
1260                                   struct taprio_sched *q,
1261                                   struct netlink_ext_ack *extack)
1262 {
1263         const struct net_device_ops *ops = dev->netdev_ops;
1264         struct tc_taprio_qopt_offload *offload;
1265         int err;
1266
1267         if (!FULL_OFFLOAD_IS_ENABLED(q->flags))
1268                 return 0;
1269
1270         if (!ops->ndo_setup_tc)
1271                 return -EOPNOTSUPP;
1272
1273         offload = taprio_offload_alloc(0);
1274         if (!offload) {
1275                 NL_SET_ERR_MSG(extack,
1276                                "Not enough memory to disable offload mode");
1277                 return -ENOMEM;
1278         }
1279         offload->enable = 0;
1280
1281         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1282         if (err < 0) {
1283                 NL_SET_ERR_MSG(extack,
1284                                "Device failed to disable offload");
1285                 goto out;
1286         }
1287
1288 out:
1289         taprio_offload_free(offload);
1290
1291         return err;
1292 }
1293
1294 /* If full offload is enabled, the only possible clockid is the net device's
1295  * PHC. For that reason, specifying a clockid through netlink is incorrect.
1296  * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1297  * in sync with the specified clockid via a user space daemon such as phc2sys.
1298  * For both software taprio and txtime-assist, the clockid is used for the
1299  * hrtimer that advances the schedule and hence mandatory.
1300  */
1301 static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1302                                 struct netlink_ext_ack *extack)
1303 {
1304         struct taprio_sched *q = qdisc_priv(sch);
1305         struct net_device *dev = qdisc_dev(sch);
1306         int err = -EINVAL;
1307
1308         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1309                 const struct ethtool_ops *ops = dev->ethtool_ops;
1310                 struct ethtool_ts_info info = {
1311                         .cmd = ETHTOOL_GET_TS_INFO,
1312                         .phc_index = -1,
1313                 };
1314
1315                 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1316                         NL_SET_ERR_MSG(extack,
1317                                        "The 'clockid' cannot be specified for full offload");
1318                         goto out;
1319                 }
1320
1321                 if (ops && ops->get_ts_info)
1322                         err = ops->get_ts_info(dev, &info);
1323
1324                 if (err || info.phc_index < 0) {
1325                         NL_SET_ERR_MSG(extack,
1326                                        "Device does not have a PTP clock");
1327                         err = -ENOTSUPP;
1328                         goto out;
1329                 }
1330         } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1331                 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1332
1333                 /* We only support static clockids and we don't allow
1334                  * for it to be modified after the first init.
1335                  */
1336                 if (clockid < 0 ||
1337                     (q->clockid != -1 && q->clockid != clockid)) {
1338                         NL_SET_ERR_MSG(extack,
1339                                        "Changing the 'clockid' of a running schedule is not supported");
1340                         err = -ENOTSUPP;
1341                         goto out;
1342                 }
1343
1344                 switch (clockid) {
1345                 case CLOCK_REALTIME:
1346                         q->tk_offset = TK_OFFS_REAL;
1347                         break;
1348                 case CLOCK_MONOTONIC:
1349                         q->tk_offset = TK_OFFS_MAX;
1350                         break;
1351                 case CLOCK_BOOTTIME:
1352                         q->tk_offset = TK_OFFS_BOOT;
1353                         break;
1354                 case CLOCK_TAI:
1355                         q->tk_offset = TK_OFFS_TAI;
1356                         break;
1357                 default:
1358                         NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1359                         err = -EINVAL;
1360                         goto out;
1361                 }
1362
1363                 q->clockid = clockid;
1364         } else {
1365                 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1366                 goto out;
1367         }
1368
1369         /* Everything went ok, return success. */
1370         err = 0;
1371
1372 out:
1373         return err;
1374 }
1375
1376 static int taprio_mqprio_cmp(const struct net_device *dev,
1377                              const struct tc_mqprio_qopt *mqprio)
1378 {
1379         int i;
1380
1381         if (!mqprio || mqprio->num_tc != dev->num_tc)
1382                 return -1;
1383
1384         for (i = 0; i < mqprio->num_tc; i++)
1385                 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1386                     dev->tc_to_txq[i].offset != mqprio->offset[i])
1387                         return -1;
1388
1389         for (i = 0; i <= TC_BITMASK; i++)
1390                 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1391                         return -1;
1392
1393         return 0;
1394 }
1395
1396 /* The semantics of the 'flags' argument in relation to 'change()'
1397  * requests, are interpreted following two rules (which are applied in
1398  * this order): (1) an omitted 'flags' argument is interpreted as
1399  * zero; (2) the 'flags' of a "running" taprio instance cannot be
1400  * changed.
1401  */
1402 static int taprio_new_flags(const struct nlattr *attr, u32 old,
1403                             struct netlink_ext_ack *extack)
1404 {
1405         u32 new = 0;
1406
1407         if (attr)
1408                 new = nla_get_u32(attr);
1409
1410         if (old != TAPRIO_FLAGS_INVALID && old != new) {
1411                 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1412                 return -EOPNOTSUPP;
1413         }
1414
1415         if (!taprio_flags_valid(new)) {
1416                 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1417                 return -EINVAL;
1418         }
1419
1420         return new;
1421 }
1422
1423 static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1424                          struct netlink_ext_ack *extack)
1425 {
1426         struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1427         struct sched_gate_list *oper, *admin, *new_admin;
1428         struct taprio_sched *q = qdisc_priv(sch);
1429         struct net_device *dev = qdisc_dev(sch);
1430         struct tc_mqprio_qopt *mqprio = NULL;
1431         unsigned long flags;
1432         ktime_t start;
1433         int i, err;
1434
1435         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1436                                           taprio_policy, extack);
1437         if (err < 0)
1438                 return err;
1439
1440         if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1441                 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1442
1443         err = taprio_new_flags(tb[TCA_TAPRIO_ATTR_FLAGS],
1444                                q->flags, extack);
1445         if (err < 0)
1446                 return err;
1447
1448         q->flags = err;
1449
1450         err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1451         if (err < 0)
1452                 return err;
1453
1454         new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1455         if (!new_admin) {
1456                 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1457                 return -ENOMEM;
1458         }
1459         INIT_LIST_HEAD(&new_admin->entries);
1460
1461         rcu_read_lock();
1462         oper = rcu_dereference(q->oper_sched);
1463         admin = rcu_dereference(q->admin_sched);
1464         rcu_read_unlock();
1465
1466         /* no changes - no new mqprio settings */
1467         if (!taprio_mqprio_cmp(dev, mqprio))
1468                 mqprio = NULL;
1469
1470         if (mqprio && (oper || admin)) {
1471                 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1472                 err = -ENOTSUPP;
1473                 goto free_sched;
1474         }
1475
1476         err = parse_taprio_schedule(tb, new_admin, extack);
1477         if (err < 0)
1478                 goto free_sched;
1479
1480         if (new_admin->num_entries == 0) {
1481                 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1482                 err = -EINVAL;
1483                 goto free_sched;
1484         }
1485
1486         err = taprio_parse_clockid(sch, tb, extack);
1487         if (err < 0)
1488                 goto free_sched;
1489
1490         taprio_set_picos_per_byte(dev, q);
1491
1492         if (mqprio) {
1493                 netdev_set_num_tc(dev, mqprio->num_tc);
1494                 for (i = 0; i < mqprio->num_tc; i++)
1495                         netdev_set_tc_queue(dev, i,
1496                                             mqprio->count[i],
1497                                             mqprio->offset[i]);
1498
1499                 /* Always use supplied priority mappings */
1500                 for (i = 0; i <= TC_BITMASK; i++)
1501                         netdev_set_prio_tc_map(dev, i,
1502                                                mqprio->prio_tc_map[i]);
1503         }
1504
1505         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1506                 err = taprio_enable_offload(dev, q, new_admin, extack);
1507         else
1508                 err = taprio_disable_offload(dev, q, extack);
1509         if (err)
1510                 goto free_sched;
1511
1512         /* Protects against enqueue()/dequeue() */
1513         spin_lock_bh(qdisc_lock(sch));
1514
1515         if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1516                 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1517                         NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1518                         err = -EINVAL;
1519                         goto unlock;
1520                 }
1521
1522                 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1523         }
1524
1525         if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1526             !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1527             !hrtimer_active(&q->advance_timer)) {
1528                 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1529                 q->advance_timer.function = advance_sched;
1530         }
1531
1532         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1533                 q->dequeue = taprio_dequeue_offload;
1534                 q->peek = taprio_peek_offload;
1535         } else {
1536                 /* Be sure to always keep the function pointers
1537                  * in a consistent state.
1538                  */
1539                 q->dequeue = taprio_dequeue_soft;
1540                 q->peek = taprio_peek_soft;
1541         }
1542
1543         err = taprio_get_start_time(sch, new_admin, &start);
1544         if (err < 0) {
1545                 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1546                 goto unlock;
1547         }
1548
1549         setup_txtime(q, new_admin, start);
1550
1551         if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1552                 if (!oper) {
1553                         rcu_assign_pointer(q->oper_sched, new_admin);
1554                         err = 0;
1555                         new_admin = NULL;
1556                         goto unlock;
1557                 }
1558
1559                 rcu_assign_pointer(q->admin_sched, new_admin);
1560                 if (admin)
1561                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1562         } else {
1563                 setup_first_close_time(q, new_admin, start);
1564
1565                 /* Protects against advance_sched() */
1566                 spin_lock_irqsave(&q->current_entry_lock, flags);
1567
1568                 taprio_start_sched(sch, start, new_admin);
1569
1570                 rcu_assign_pointer(q->admin_sched, new_admin);
1571                 if (admin)
1572                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1573
1574                 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1575
1576                 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1577                         taprio_offload_config_changed(q);
1578         }
1579
1580         new_admin = NULL;
1581         err = 0;
1582
1583 unlock:
1584         spin_unlock_bh(qdisc_lock(sch));
1585
1586 free_sched:
1587         if (new_admin)
1588                 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
1589
1590         return err;
1591 }
1592
1593 static void taprio_destroy(struct Qdisc *sch)
1594 {
1595         struct taprio_sched *q = qdisc_priv(sch);
1596         struct net_device *dev = qdisc_dev(sch);
1597         unsigned int i;
1598
1599         spin_lock(&taprio_list_lock);
1600         list_del(&q->taprio_list);
1601         spin_unlock(&taprio_list_lock);
1602
1603         hrtimer_cancel(&q->advance_timer);
1604
1605         taprio_disable_offload(dev, q, NULL);
1606
1607         if (q->qdiscs) {
1608                 for (i = 0; i < dev->num_tx_queues && q->qdiscs[i]; i++)
1609                         qdisc_put(q->qdiscs[i]);
1610
1611                 kfree(q->qdiscs);
1612         }
1613         q->qdiscs = NULL;
1614
1615         netdev_reset_tc(dev);
1616
1617         if (q->oper_sched)
1618                 call_rcu(&q->oper_sched->rcu, taprio_free_sched_cb);
1619
1620         if (q->admin_sched)
1621                 call_rcu(&q->admin_sched->rcu, taprio_free_sched_cb);
1622 }
1623
1624 static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
1625                        struct netlink_ext_ack *extack)
1626 {
1627         struct taprio_sched *q = qdisc_priv(sch);
1628         struct net_device *dev = qdisc_dev(sch);
1629         int i;
1630
1631         spin_lock_init(&q->current_entry_lock);
1632
1633         hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
1634         q->advance_timer.function = advance_sched;
1635
1636         q->dequeue = taprio_dequeue_soft;
1637         q->peek = taprio_peek_soft;
1638
1639         q->root = sch;
1640
1641         /* We only support static clockids. Use an invalid value as default
1642          * and get the valid one on taprio_change().
1643          */
1644         q->clockid = -1;
1645         q->flags = TAPRIO_FLAGS_INVALID;
1646
1647         spin_lock(&taprio_list_lock);
1648         list_add(&q->taprio_list, &taprio_list);
1649         spin_unlock(&taprio_list_lock);
1650
1651         if (sch->parent != TC_H_ROOT)
1652                 return -EOPNOTSUPP;
1653
1654         if (!netif_is_multiqueue(dev))
1655                 return -EOPNOTSUPP;
1656
1657         /* pre-allocate qdisc, attachment can't fail */
1658         q->qdiscs = kcalloc(dev->num_tx_queues,
1659                             sizeof(q->qdiscs[0]),
1660                             GFP_KERNEL);
1661
1662         if (!q->qdiscs)
1663                 return -ENOMEM;
1664
1665         if (!opt)
1666                 return -EINVAL;
1667
1668         for (i = 0; i < dev->num_tx_queues; i++) {
1669                 struct netdev_queue *dev_queue;
1670                 struct Qdisc *qdisc;
1671
1672                 dev_queue = netdev_get_tx_queue(dev, i);
1673                 qdisc = qdisc_create_dflt(dev_queue,
1674                                           &pfifo_qdisc_ops,
1675                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
1676                                                     TC_H_MIN(i + 1)),
1677                                           extack);
1678                 if (!qdisc)
1679                         return -ENOMEM;
1680
1681                 if (i < dev->real_num_tx_queues)
1682                         qdisc_hash_add(qdisc, false);
1683
1684                 q->qdiscs[i] = qdisc;
1685         }
1686
1687         return taprio_change(sch, opt, extack);
1688 }
1689
1690 static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
1691                                              unsigned long cl)
1692 {
1693         struct net_device *dev = qdisc_dev(sch);
1694         unsigned long ntx = cl - 1;
1695
1696         if (ntx >= dev->num_tx_queues)
1697                 return NULL;
1698
1699         return netdev_get_tx_queue(dev, ntx);
1700 }
1701
1702 static int taprio_graft(struct Qdisc *sch, unsigned long cl,
1703                         struct Qdisc *new, struct Qdisc **old,
1704                         struct netlink_ext_ack *extack)
1705 {
1706         struct taprio_sched *q = qdisc_priv(sch);
1707         struct net_device *dev = qdisc_dev(sch);
1708         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1709
1710         if (!dev_queue)
1711                 return -EINVAL;
1712
1713         if (dev->flags & IFF_UP)
1714                 dev_deactivate(dev);
1715
1716         *old = q->qdiscs[cl - 1];
1717         q->qdiscs[cl - 1] = new;
1718
1719         if (new)
1720                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1721
1722         if (dev->flags & IFF_UP)
1723                 dev_activate(dev);
1724
1725         return 0;
1726 }
1727
1728 static int dump_entry(struct sk_buff *msg,
1729                       const struct sched_entry *entry)
1730 {
1731         struct nlattr *item;
1732
1733         item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
1734         if (!item)
1735                 return -ENOSPC;
1736
1737         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
1738                 goto nla_put_failure;
1739
1740         if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
1741                 goto nla_put_failure;
1742
1743         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
1744                         entry->gate_mask))
1745                 goto nla_put_failure;
1746
1747         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
1748                         entry->interval))
1749                 goto nla_put_failure;
1750
1751         return nla_nest_end(msg, item);
1752
1753 nla_put_failure:
1754         nla_nest_cancel(msg, item);
1755         return -1;
1756 }
1757
1758 static int dump_schedule(struct sk_buff *msg,
1759                          const struct sched_gate_list *root)
1760 {
1761         struct nlattr *entry_list;
1762         struct sched_entry *entry;
1763
1764         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
1765                         root->base_time, TCA_TAPRIO_PAD))
1766                 return -1;
1767
1768         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
1769                         root->cycle_time, TCA_TAPRIO_PAD))
1770                 return -1;
1771
1772         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
1773                         root->cycle_time_extension, TCA_TAPRIO_PAD))
1774                 return -1;
1775
1776         entry_list = nla_nest_start_noflag(msg,
1777                                            TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
1778         if (!entry_list)
1779                 goto error_nest;
1780
1781         list_for_each_entry(entry, &root->entries, list) {
1782                 if (dump_entry(msg, entry) < 0)
1783                         goto error_nest;
1784         }
1785
1786         nla_nest_end(msg, entry_list);
1787         return 0;
1788
1789 error_nest:
1790         nla_nest_cancel(msg, entry_list);
1791         return -1;
1792 }
1793
1794 static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
1795 {
1796         struct taprio_sched *q = qdisc_priv(sch);
1797         struct net_device *dev = qdisc_dev(sch);
1798         struct sched_gate_list *oper, *admin;
1799         struct tc_mqprio_qopt opt = { 0 };
1800         struct nlattr *nest, *sched_nest;
1801         unsigned int i;
1802
1803         rcu_read_lock();
1804         oper = rcu_dereference(q->oper_sched);
1805         admin = rcu_dereference(q->admin_sched);
1806
1807         opt.num_tc = netdev_get_num_tc(dev);
1808         memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
1809
1810         for (i = 0; i < netdev_get_num_tc(dev); i++) {
1811                 opt.count[i] = dev->tc_to_txq[i].count;
1812                 opt.offset[i] = dev->tc_to_txq[i].offset;
1813         }
1814
1815         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1816         if (!nest)
1817                 goto start_error;
1818
1819         if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
1820                 goto options_error;
1821
1822         if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1823             nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
1824                 goto options_error;
1825
1826         if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
1827                 goto options_error;
1828
1829         if (q->txtime_delay &&
1830             nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
1831                 goto options_error;
1832
1833         if (oper && dump_schedule(skb, oper))
1834                 goto options_error;
1835
1836         if (!admin)
1837                 goto done;
1838
1839         sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
1840         if (!sched_nest)
1841                 goto options_error;
1842
1843         if (dump_schedule(skb, admin))
1844                 goto admin_error;
1845
1846         nla_nest_end(skb, sched_nest);
1847
1848 done:
1849         rcu_read_unlock();
1850
1851         return nla_nest_end(skb, nest);
1852
1853 admin_error:
1854         nla_nest_cancel(skb, sched_nest);
1855
1856 options_error:
1857         nla_nest_cancel(skb, nest);
1858
1859 start_error:
1860         rcu_read_unlock();
1861         return -ENOSPC;
1862 }
1863
1864 static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
1865 {
1866         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1867
1868         if (!dev_queue)
1869                 return NULL;
1870
1871         return dev_queue->qdisc_sleeping;
1872 }
1873
1874 static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
1875 {
1876         unsigned int ntx = TC_H_MIN(classid);
1877
1878         if (!taprio_queue_get(sch, ntx))
1879                 return 0;
1880         return ntx;
1881 }
1882
1883 static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
1884                              struct sk_buff *skb, struct tcmsg *tcm)
1885 {
1886         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1887
1888         tcm->tcm_parent = TC_H_ROOT;
1889         tcm->tcm_handle |= TC_H_MIN(cl);
1890         tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
1891
1892         return 0;
1893 }
1894
1895 static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
1896                                    struct gnet_dump *d)
1897         __releases(d->lock)
1898         __acquires(d->lock)
1899 {
1900         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1901
1902         sch = dev_queue->qdisc_sleeping;
1903         if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
1904             qdisc_qstats_copy(d, sch) < 0)
1905                 return -1;
1906         return 0;
1907 }
1908
1909 static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1910 {
1911         struct net_device *dev = qdisc_dev(sch);
1912         unsigned long ntx;
1913
1914         if (arg->stop)
1915                 return;
1916
1917         arg->count = arg->skip;
1918         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
1919                 if (arg->fn(sch, ntx + 1, arg) < 0) {
1920                         arg->stop = 1;
1921                         break;
1922                 }
1923                 arg->count++;
1924         }
1925 }
1926
1927 static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
1928                                                 struct tcmsg *tcm)
1929 {
1930         return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
1931 }
1932
1933 static const struct Qdisc_class_ops taprio_class_ops = {
1934         .graft          = taprio_graft,
1935         .leaf           = taprio_leaf,
1936         .find           = taprio_find,
1937         .walk           = taprio_walk,
1938         .dump           = taprio_dump_class,
1939         .dump_stats     = taprio_dump_class_stats,
1940         .select_queue   = taprio_select_queue,
1941 };
1942
1943 static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
1944         .cl_ops         = &taprio_class_ops,
1945         .id             = "taprio",
1946         .priv_size      = sizeof(struct taprio_sched),
1947         .init           = taprio_init,
1948         .change         = taprio_change,
1949         .destroy        = taprio_destroy,
1950         .peek           = taprio_peek,
1951         .dequeue        = taprio_dequeue,
1952         .enqueue        = taprio_enqueue,
1953         .dump           = taprio_dump,
1954         .owner          = THIS_MODULE,
1955 };
1956
1957 static struct notifier_block taprio_device_notifier = {
1958         .notifier_call = taprio_dev_notifier,
1959 };
1960
1961 static int __init taprio_module_init(void)
1962 {
1963         int err = register_netdevice_notifier(&taprio_device_notifier);
1964
1965         if (err)
1966                 return err;
1967
1968         return register_qdisc(&taprio_qdisc_ops);
1969 }
1970
1971 static void __exit taprio_module_exit(void)
1972 {
1973         unregister_qdisc(&taprio_qdisc_ops);
1974         unregister_netdevice_notifier(&taprio_device_notifier);
1975 }
1976
1977 module_init(taprio_module_init);
1978 module_exit(taprio_module_exit);
1979 MODULE_LICENSE("GPL");