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