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         int 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 entry_list_policy[TCA_TAPRIO_SCHED_MAX + 1] = {
 681         [TCA_TAPRIO_SCHED_ENTRY] = { .type = NLA_NESTED },
 682 };
 683
 684 static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
 685         [TCA_TAPRIO_ATTR_PRIOMAP]              = {
 686                 .len = sizeof(struct tc_mqprio_qopt)
 687         },
 688         [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST]           = { .type = NLA_NESTED },
 689         [TCA_TAPRIO_ATTR_SCHED_BASE_TIME]            = { .type = NLA_S64 },
 690         [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]         = { .type = NLA_NESTED },
 691         [TCA_TAPRIO_ATTR_SCHED_CLOCKID]              = { .type = NLA_S32 },
 692         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]           = { .type = NLA_S64 },
 693         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
 694 };
 695
 696 static int fill_sched_entry(struct nlattr **tb, struct sched_entry *entry,
 697                             struct netlink_ext_ack *extack)
 698 {
 699         u32 interval = 0;
 700
 701         if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
 702                 entry->command = nla_get_u8(
 703                         tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
 704
 705         if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
 706                 entry->gate_mask = nla_get_u32(
 707                         tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
 708
 709         if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
 710                 interval = nla_get_u32(
 711                         tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
 712
 713         if (interval == 0) {
 714                 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
 715                 return -EINVAL;
 716         }
 717
 718         entry->interval = interval;
 719
 720         return 0;
 721 }
 722
 723 static int parse_sched_entry(struct nlattr *n, struct sched_entry *entry,
 724                              int index, struct netlink_ext_ack *extack)
 725 {
 726         struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
 727         int err;
 728
 729         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
 730                                           entry_policy, NULL);
 731         if (err < 0) {
 732                 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
 733                 return -EINVAL;
 734         }
 735
 736         entry->index = index;
 737
 738         return fill_sched_entry(tb, entry, extack);
 739 }
 740
 741 static int parse_sched_list(struct nlattr *list,
 742                             struct sched_gate_list *sched,
 743                             struct netlink_ext_ack *extack)
 744 {
 745         struct nlattr *n;
 746         int err, rem;
 747         int i = 0;
 748
 749         if (!list)
 750                 return -EINVAL;
 751
 752         nla_for_each_nested(n, list, rem) {
 753                 struct sched_entry *entry;
 754
 755                 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
 756                         NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
 757                         continue;
 758                 }
 759
 760                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 761                 if (!entry) {
 762                         NL_SET_ERR_MSG(extack, "Not enough memory for entry");
 763                         return -ENOMEM;
 764                 }
 765
 766                 err = parse_sched_entry(n, entry, i, extack);
 767                 if (err < 0) {
 768                         kfree(entry);
 769                         return err;
 770                 }
 771
 772                 list_add_tail(&entry->list, &sched->entries);
 773                 i++;
 774         }
 775
 776         sched->num_entries = i;
 777
 778         return i;
 779 }
 780
 781 static int parse_taprio_schedule(struct nlattr **tb,
 782                                  struct sched_gate_list *new,
 783                                  struct netlink_ext_ack *extack)
 784 {
 785         int err = 0;
 786
 787         if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
 788                 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
 789                 return -ENOTSUPP;
 790         }
 791
 792         if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
 793                 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
 794
 795         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
 796                 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
 797
 798         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
 799                 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
 800
 801         if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
 802                 err = parse_sched_list(
 803                         tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST], new, extack);
 804         if (err < 0)
 805                 return err;
 806
 807         if (!new->cycle_time) {
 808                 struct sched_entry *entry;
 809                 ktime_t cycle = 0;
 810
 811                 list_for_each_entry(entry, &new->entries, list)
 812                         cycle = ktime_add_ns(cycle, entry->interval);
 813                 new->cycle_time = cycle;
 814         }
 815
 816         return 0;
 817 }
 818
 819 static int taprio_parse_mqprio_opt(struct net_device *dev,
 820                                    struct tc_mqprio_qopt *qopt,
 821                                    struct netlink_ext_ack *extack,
 822                                    u32 taprio_flags)
 823 {
 824         int i, j;
 825
 826         if (!qopt && !dev->num_tc) {
 827                 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
 828                 return -EINVAL;
 829         }
 830
 831         /* If num_tc is already set, it means that the user already
 832          * configured the mqprio part
 833          */
 834         if (dev->num_tc)
 835                 return 0;
 836
 837         /* Verify num_tc is not out of max range */
 838         if (qopt->num_tc > TC_MAX_QUEUE) {
 839                 NL_SET_ERR_MSG(extack, "Number of traffic classes is outside valid range");
 840                 return -EINVAL;
 841         }
 842
 843         /* taprio imposes that traffic classes map 1:n to tx queues */
 844         if (qopt->num_tc > dev->num_tx_queues) {
 845                 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
 846                 return -EINVAL;
 847         }
 848
 849         /* Verify priority mapping uses valid tcs */
 850         for (i = 0; i < TC_BITMASK + 1; i++) {
 851                 if (qopt->prio_tc_map[i] >= qopt->num_tc) {
 852                         NL_SET_ERR_MSG(extack, "Invalid traffic class in priority to traffic class mapping");
 853                         return -EINVAL;
 854                 }
 855         }
 856
 857         for (i = 0; i < qopt->num_tc; i++) {
 858                 unsigned int last = qopt->offset[i] + qopt->count[i];
 859
 860                 /* Verify the queue count is in tx range being equal to the
 861                  * real_num_tx_queues indicates the last queue is in use.
 862                  */
 863                 if (qopt->offset[i] >= dev->num_tx_queues ||
 864                     !qopt->count[i] ||
 865                     last > dev->real_num_tx_queues) {
 866                         NL_SET_ERR_MSG(extack, "Invalid queue in traffic class to queue mapping");
 867                         return -EINVAL;
 868                 }
 869
 870                 if (TXTIME_ASSIST_IS_ENABLED(taprio_flags))
 871                         continue;
 872
 873                 /* Verify that the offset and counts do not overlap */
 874                 for (j = i + 1; j < qopt->num_tc; j++) {
 875                         if (last > qopt->offset[j]) {
 876                                 NL_SET_ERR_MSG(extack, "Detected overlap in the traffic class to queue mapping");
 877                                 return -EINVAL;
 878                         }
 879                 }
 880         }
 881
 882         return 0;
 883 }
 884
 885 static int taprio_get_start_time(struct Qdisc *sch,
 886                                  struct sched_gate_list *sched,
 887                                  ktime_t *start)
 888 {
 889         struct taprio_sched *q = qdisc_priv(sch);
 890         ktime_t now, base, cycle;
 891         s64 n;
 892
 893         base = sched_base_time(sched);
 894         now = taprio_get_time(q);
 895
 896         if (ktime_after(base, now)) {
 897                 *start = base;
 898                 return 0;
 899         }
 900
 901         cycle = sched->cycle_time;
 902
 903         /* The qdisc is expected to have at least one sched_entry.  Moreover,
 904          * any entry must have 'interval' > 0. Thus if the cycle time is zero,
 905          * something went really wrong. In that case, we should warn about this
 906          * inconsistent state and return error.
 907          */
 908         if (WARN_ON(!cycle))
 909                 return -EFAULT;
 910
 911         /* Schedule the start time for the beginning of the next
 912          * cycle.
 913          */
 914         n = div64_s64(ktime_sub_ns(now, base), cycle);
 915         *start = ktime_add_ns(base, (n + 1) * cycle);
 916         return 0;
 917 }
 918
 919 static void setup_first_close_time(struct taprio_sched *q,
 920                                    struct sched_gate_list *sched, ktime_t base)
 921 {
 922         struct sched_entry *first;
 923         ktime_t cycle;
 924
 925         first = list_first_entry(&sched->entries,
 926                                  struct sched_entry, list);
 927
 928         cycle = sched->cycle_time;
 929
 930         /* FIXME: find a better place to do this */
 931         sched->cycle_close_time = ktime_add_ns(base, cycle);
 932
 933         first->close_time = ktime_add_ns(base, first->interval);
 934         taprio_set_budget(q, first);
 935         rcu_assign_pointer(q->current_entry, NULL);
 936 }
 937
 938 static void taprio_start_sched(struct Qdisc *sch,
 939                                ktime_t start, struct sched_gate_list *new)
 940 {
 941         struct taprio_sched *q = qdisc_priv(sch);
 942         ktime_t expires;
 943
 944         expires = hrtimer_get_expires(&q->advance_timer);
 945         if (expires == 0)
 946                 expires = KTIME_MAX;
 947
 948         /* If the new schedule starts before the next expiration, we
 949          * reprogram it to the earliest one, so we change the admin
 950          * schedule to the operational one at the right time.
 951          */
 952         start = min_t(ktime_t, start, expires);
 953
 954         hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
 955 }
 956
 957 static void taprio_set_picos_per_byte(struct net_device *dev,
 958                                       struct taprio_sched *q)
 959 {
 960         struct ethtool_link_ksettings ecmd;
 961         int picos_per_byte = -1;
 962
 963         if (!__ethtool_get_link_ksettings(dev, &ecmd) &&
 964             ecmd.base.speed != SPEED_UNKNOWN)
 965                 picos_per_byte = div64_s64(NSEC_PER_SEC * 1000LL * 8,
 966                                            ecmd.base.speed * 1000 * 1000);
 967
 968         atomic64_set(&q->picos_per_byte, picos_per_byte);
 969         netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
 970                    dev->name, (long long)atomic64_read(&q->picos_per_byte),
 971                    ecmd.base.speed);
 972 }
 973
 974 static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
 975                                void *ptr)
 976 {
 977         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 978         struct net_device *qdev;
 979         struct taprio_sched *q;
 980         bool found = false;
 981
 982         ASSERT_RTNL();
 983
 984         if (event != NETDEV_UP && event != NETDEV_CHANGE)
 985                 return NOTIFY_DONE;
 986
 987         spin_lock(&taprio_list_lock);
 988         list_for_each_entry(q, &taprio_list, taprio_list) {
 989                 qdev = qdisc_dev(q->root);
 990                 if (qdev == dev) {
 991                         found = true;
 992                         break;
 993                 }
 994         }
 995         spin_unlock(&taprio_list_lock);
 996
 997         if (found)
 998                 taprio_set_picos_per_byte(dev, q);
 999
1000         return NOTIFY_DONE;
1001 }
1002
1003 static void setup_txtime(struct taprio_sched *q,
1004                          struct sched_gate_list *sched, ktime_t base)
1005 {
1006         struct sched_entry *entry;
1007         u32 interval = 0;
1008
1009         list_for_each_entry(entry, &sched->entries, list) {
1010                 entry->next_txtime = ktime_add_ns(base, interval);
1011                 interval += entry->interval;
1012         }
1013 }
1014
1015 static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1016                          struct netlink_ext_ack *extack)
1017 {
1018         struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1019         struct sched_gate_list *oper, *admin, *new_admin;
1020         struct taprio_sched *q = qdisc_priv(sch);
1021         struct net_device *dev = qdisc_dev(sch);
1022         struct tc_mqprio_qopt *mqprio = NULL;
1023         u32 taprio_flags = 0;
1024         int i, err, clockid;
1025         unsigned long flags;
1026         ktime_t start;
1027
1028         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1029                                           taprio_policy, extack);
1030         if (err < 0)
1031                 return err;
1032
1033         if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1034                 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1035
1036         if (tb[TCA_TAPRIO_ATTR_FLAGS]) {
1037                 taprio_flags = nla_get_u32(tb[TCA_TAPRIO_ATTR_FLAGS]);
1038
1039                 if (q->flags != 0 && q->flags != taprio_flags) {
1040                         NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1041                         return -EOPNOTSUPP;
1042                 } else if (!FLAGS_VALID(taprio_flags)) {
1043                         NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1044                         return -EINVAL;
1045                 }
1046
1047                 q->flags = taprio_flags;
1048         }
1049
1050         err = taprio_parse_mqprio_opt(dev, mqprio, extack, taprio_flags);
1051         if (err < 0)
1052                 return err;
1053
1054         new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1055         if (!new_admin) {
1056                 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1057                 return -ENOMEM;
1058         }
1059         INIT_LIST_HEAD(&new_admin->entries);
1060
1061         rcu_read_lock();
1062         oper = rcu_dereference(q->oper_sched);
1063         admin = rcu_dereference(q->admin_sched);
1064         rcu_read_unlock();
1065
1066         if (mqprio && (oper || admin)) {
1067                 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1068                 err = -ENOTSUPP;
1069                 goto free_sched;
1070         }
1071
1072         err = parse_taprio_schedule(tb, new_admin, extack);
1073         if (err < 0)
1074                 goto free_sched;
1075
1076         if (new_admin->num_entries == 0) {
1077                 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1078                 err = -EINVAL;
1079                 goto free_sched;
1080         }
1081
1082         if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1083                 clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1084
1085                 /* We only support static clockids and we don't allow
1086                  * for it to be modified after the first init.
1087                  */
1088                 if (clockid < 0 ||
1089                     (q->clockid != -1 && q->clockid != clockid)) {
1090                         NL_SET_ERR_MSG(extack, "Changing the 'clockid' of a running schedule is not supported");
1091                         err = -ENOTSUPP;
1092                         goto free_sched;
1093                 }
1094
1095                 q->clockid = clockid;
1096         }
1097
1098         if (q->clockid == -1 && !tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1099                 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1100                 err = -EINVAL;
1101                 goto free_sched;
1102         }
1103
1104         taprio_set_picos_per_byte(dev, q);
1105
1106         /* Protects against enqueue()/dequeue() */
1107         spin_lock_bh(qdisc_lock(sch));
1108
1109         if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1110                 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1111                         NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1112                         err = -EINVAL;
1113                         goto unlock;
1114                 }
1115
1116                 q->txtime_delay = nla_get_s32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1117         }
1118
1119         if (!TXTIME_ASSIST_IS_ENABLED(taprio_flags) &&
1120             !hrtimer_active(&q->advance_timer)) {
1121                 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1122                 q->advance_timer.function = advance_sched;
1123         }
1124
1125         if (mqprio) {
1126                 netdev_set_num_tc(dev, mqprio->num_tc);
1127                 for (i = 0; i < mqprio->num_tc; i++)
1128                         netdev_set_tc_queue(dev, i,
1129                                             mqprio->count[i],
1130                                             mqprio->offset[i]);
1131
1132                 /* Always use supplied priority mappings */
1133                 for (i = 0; i < TC_BITMASK + 1; i++)
1134                         netdev_set_prio_tc_map(dev, i,
1135                                                mqprio->prio_tc_map[i]);
1136         }
1137
1138         switch (q->clockid) {
1139         case CLOCK_REALTIME:
1140                 q->tk_offset = TK_OFFS_REAL;
1141                 break;
1142         case CLOCK_MONOTONIC:
1143                 q->tk_offset = TK_OFFS_MAX;
1144                 break;
1145         case CLOCK_BOOTTIME:
1146                 q->tk_offset = TK_OFFS_BOOT;
1147                 break;
1148         case CLOCK_TAI:
1149                 q->tk_offset = TK_OFFS_TAI;
1150                 break;
1151         default:
1152                 NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1153                 err = -EINVAL;
1154                 goto unlock;
1155         }
1156
1157         err = taprio_get_start_time(sch, new_admin, &start);
1158         if (err < 0) {
1159                 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1160                 goto unlock;
1161         }
1162
1163         if (TXTIME_ASSIST_IS_ENABLED(taprio_flags)) {
1164                 setup_txtime(q, new_admin, start);
1165
1166                 if (!oper) {
1167                         rcu_assign_pointer(q->oper_sched, new_admin);
1168                         err = 0;
1169                         new_admin = NULL;
1170                         goto unlock;
1171                 }
1172
1173                 rcu_assign_pointer(q->admin_sched, new_admin);
1174                 if (admin)
1175                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1176         } else {
1177                 setup_first_close_time(q, new_admin, start);
1178
1179                 /* Protects against advance_sched() */
1180                 spin_lock_irqsave(&q->current_entry_lock, flags);
1181
1182                 taprio_start_sched(sch, start, new_admin);
1183
1184                 rcu_assign_pointer(q->admin_sched, new_admin);
1185                 if (admin)
1186                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1187
1188                 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1189         }
1190
1191         new_admin = NULL;
1192         err = 0;
1193
1194 unlock:
1195         spin_unlock_bh(qdisc_lock(sch));
1196
1197 free_sched:
1198         kfree(new_admin);
1199
1200         return err;
1201 }
1202
1203 static void taprio_destroy(struct Qdisc *sch)
1204 {
1205         struct taprio_sched *q = qdisc_priv(sch);
1206         struct net_device *dev = qdisc_dev(sch);
1207         unsigned int i;
1208
1209         spin_lock(&taprio_list_lock);
1210         list_del(&q->taprio_list);
1211         spin_unlock(&taprio_list_lock);
1212
1213         hrtimer_cancel(&q->advance_timer);
1214
1215         if (q->qdiscs) {
1216                 for (i = 0; i < dev->num_tx_queues && q->qdiscs[i]; i++)
1217                         qdisc_put(q->qdiscs[i]);
1218
1219                 kfree(q->qdiscs);
1220         }
1221         q->qdiscs = NULL;
1222
1223         netdev_set_num_tc(dev, 0);
1224
1225         if (q->oper_sched)
1226                 call_rcu(&q->oper_sched->rcu, taprio_free_sched_cb);
1227
1228         if (q->admin_sched)
1229                 call_rcu(&q->admin_sched->rcu, taprio_free_sched_cb);
1230 }
1231
1232 static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
1233                        struct netlink_ext_ack *extack)
1234 {
1235         struct taprio_sched *q = qdisc_priv(sch);
1236         struct net_device *dev = qdisc_dev(sch);
1237         int i;
1238
1239         spin_lock_init(&q->current_entry_lock);
1240
1241         hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
1242         q->advance_timer.function = advance_sched;
1243
1244         q->root = sch;
1245
1246         /* We only support static clockids. Use an invalid value as default
1247          * and get the valid one on taprio_change().
1248          */
1249         q->clockid = -1;
1250
1251         if (sch->parent != TC_H_ROOT)
1252                 return -EOPNOTSUPP;
1253
1254         if (!netif_is_multiqueue(dev))
1255                 return -EOPNOTSUPP;
1256
1257         /* pre-allocate qdisc, attachment can't fail */
1258         q->qdiscs = kcalloc(dev->num_tx_queues,
1259                             sizeof(q->qdiscs[0]),
1260                             GFP_KERNEL);
1261
1262         if (!q->qdiscs)
1263                 return -ENOMEM;
1264
1265         if (!opt)
1266                 return -EINVAL;
1267
1268         spin_lock(&taprio_list_lock);
1269         list_add(&q->taprio_list, &taprio_list);
1270         spin_unlock(&taprio_list_lock);
1271
1272         for (i = 0; i < dev->num_tx_queues; i++) {
1273                 struct netdev_queue *dev_queue;
1274                 struct Qdisc *qdisc;
1275
1276                 dev_queue = netdev_get_tx_queue(dev, i);
1277                 qdisc = qdisc_create_dflt(dev_queue,
1278                                           &pfifo_qdisc_ops,
1279                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
1280                                                     TC_H_MIN(i + 1)),
1281                                           extack);
1282                 if (!qdisc)
1283                         return -ENOMEM;
1284
1285                 if (i < dev->real_num_tx_queues)
1286                         qdisc_hash_add(qdisc, false);
1287
1288                 q->qdiscs[i] = qdisc;
1289         }
1290
1291         return taprio_change(sch, opt, extack);
1292 }
1293
1294 static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
1295                                              unsigned long cl)
1296 {
1297         struct net_device *dev = qdisc_dev(sch);
1298         unsigned long ntx = cl - 1;
1299
1300         if (ntx >= dev->num_tx_queues)
1301                 return NULL;
1302
1303         return netdev_get_tx_queue(dev, ntx);
1304 }
1305
1306 static int taprio_graft(struct Qdisc *sch, unsigned long cl,
1307                         struct Qdisc *new, struct Qdisc **old,
1308                         struct netlink_ext_ack *extack)
1309 {
1310         struct taprio_sched *q = qdisc_priv(sch);
1311         struct net_device *dev = qdisc_dev(sch);
1312         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1313
1314         if (!dev_queue)
1315                 return -EINVAL;
1316
1317         if (dev->flags & IFF_UP)
1318                 dev_deactivate(dev);
1319
1320         *old = q->qdiscs[cl - 1];
1321         q->qdiscs[cl - 1] = new;
1322
1323         if (new)
1324                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1325
1326         if (dev->flags & IFF_UP)
1327                 dev_activate(dev);
1328
1329         return 0;
1330 }
1331
1332 static int dump_entry(struct sk_buff *msg,
1333                       const struct sched_entry *entry)
1334 {
1335         struct nlattr *item;
1336
1337         item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
1338         if (!item)
1339                 return -ENOSPC;
1340
1341         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
1342                 goto nla_put_failure;
1343
1344         if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
1345                 goto nla_put_failure;
1346
1347         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
1348                         entry->gate_mask))
1349                 goto nla_put_failure;
1350
1351         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
1352                         entry->interval))
1353                 goto nla_put_failure;
1354
1355         return nla_nest_end(msg, item);
1356
1357 nla_put_failure:
1358         nla_nest_cancel(msg, item);
1359         return -1;
1360 }
1361
1362 static int dump_schedule(struct sk_buff *msg,
1363                          const struct sched_gate_list *root)
1364 {
1365         struct nlattr *entry_list;
1366         struct sched_entry *entry;
1367
1368         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
1369                         root->base_time, TCA_TAPRIO_PAD))
1370                 return -1;
1371
1372         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
1373                         root->cycle_time, TCA_TAPRIO_PAD))
1374                 return -1;
1375
1376         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
1377                         root->cycle_time_extension, TCA_TAPRIO_PAD))
1378                 return -1;
1379
1380         entry_list = nla_nest_start_noflag(msg,
1381                                            TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
1382         if (!entry_list)
1383                 goto error_nest;
1384
1385         list_for_each_entry(entry, &root->entries, list) {
1386                 if (dump_entry(msg, entry) < 0)
1387                         goto error_nest;
1388         }
1389
1390         nla_nest_end(msg, entry_list);
1391         return 0;
1392
1393 error_nest:
1394         nla_nest_cancel(msg, entry_list);
1395         return -1;
1396 }
1397
1398 static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
1399 {
1400         struct taprio_sched *q = qdisc_priv(sch);
1401         struct net_device *dev = qdisc_dev(sch);
1402         struct sched_gate_list *oper, *admin;
1403         struct tc_mqprio_qopt opt = { 0 };
1404         struct nlattr *nest, *sched_nest;
1405         unsigned int i;
1406
1407         rcu_read_lock();
1408         oper = rcu_dereference(q->oper_sched);
1409         admin = rcu_dereference(q->admin_sched);
1410
1411         opt.num_tc = netdev_get_num_tc(dev);
1412         memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
1413
1414         for (i = 0; i < netdev_get_num_tc(dev); i++) {
1415                 opt.count[i] = dev->tc_to_txq[i].count;
1416                 opt.offset[i] = dev->tc_to_txq[i].offset;
1417         }
1418
1419         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1420         if (!nest)
1421                 goto start_error;
1422
1423         if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
1424                 goto options_error;
1425
1426         if (nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
1427                 goto options_error;
1428
1429         if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
1430                 goto options_error;
1431
1432         if (q->txtime_delay &&
1433             nla_put_s32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
1434                 goto options_error;
1435
1436         if (oper && dump_schedule(skb, oper))
1437                 goto options_error;
1438
1439         if (!admin)
1440                 goto done;
1441
1442         sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
1443         if (!sched_nest)
1444                 goto options_error;
1445
1446         if (dump_schedule(skb, admin))
1447                 goto admin_error;
1448
1449         nla_nest_end(skb, sched_nest);
1450
1451 done:
1452         rcu_read_unlock();
1453
1454         return nla_nest_end(skb, nest);
1455
1456 admin_error:
1457         nla_nest_cancel(skb, sched_nest);
1458
1459 options_error:
1460         nla_nest_cancel(skb, nest);
1461
1462 start_error:
1463         rcu_read_unlock();
1464         return -ENOSPC;
1465 }
1466
1467 static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
1468 {
1469         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1470
1471         if (!dev_queue)
1472                 return NULL;
1473
1474         return dev_queue->qdisc_sleeping;
1475 }
1476
1477 static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
1478 {
1479         unsigned int ntx = TC_H_MIN(classid);
1480
1481         if (!taprio_queue_get(sch, ntx))
1482                 return 0;
1483         return ntx;
1484 }
1485
1486 static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
1487                              struct sk_buff *skb, struct tcmsg *tcm)
1488 {
1489         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1490
1491         tcm->tcm_parent = TC_H_ROOT;
1492         tcm->tcm_handle |= TC_H_MIN(cl);
1493         tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
1494
1495         return 0;
1496 }
1497
1498 static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
1499                                    struct gnet_dump *d)
1500         __releases(d->lock)
1501         __acquires(d->lock)
1502 {
1503         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1504
1505         sch = dev_queue->qdisc_sleeping;
1506         if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
1507             qdisc_qstats_copy(d, sch) < 0)
1508                 return -1;
1509         return 0;
1510 }
1511
1512 static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1513 {
1514         struct net_device *dev = qdisc_dev(sch);
1515         unsigned long ntx;
1516
1517         if (arg->stop)
1518                 return;
1519
1520         arg->count = arg->skip;
1521         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
1522                 if (arg->fn(sch, ntx + 1, arg) < 0) {
1523                         arg->stop = 1;
1524                         break;
1525                 }
1526                 arg->count++;
1527         }
1528 }
1529
1530 static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
1531                                                 struct tcmsg *tcm)
1532 {
1533         return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
1534 }
1535
1536 static const struct Qdisc_class_ops taprio_class_ops = {
1537         .graft          = taprio_graft,
1538         .leaf           = taprio_leaf,
1539         .find           = taprio_find,
1540         .walk           = taprio_walk,
1541         .dump           = taprio_dump_class,
1542         .dump_stats     = taprio_dump_class_stats,
1543         .select_queue   = taprio_select_queue,
1544 };
1545
1546 static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
1547         .cl_ops         = &taprio_class_ops,
1548         .id             = "taprio",
1549         .priv_size      = sizeof(struct taprio_sched),
1550         .init           = taprio_init,
1551         .change         = taprio_change,
1552         .destroy        = taprio_destroy,
1553         .peek           = taprio_peek,
1554         .dequeue        = taprio_dequeue,
1555         .enqueue        = taprio_enqueue,
1556         .dump           = taprio_dump,
1557         .owner          = THIS_MODULE,
1558 };
1559
1560 static struct notifier_block taprio_device_notifier = {
1561         .notifier_call = taprio_dev_notifier,
1562 };
1563
1564 static int __init taprio_module_init(void)
1565 {
1566         int err = register_netdevice_notifier(&taprio_device_notifier);
1567
1568         if (err)
1569                 return err;
1570
1571         return register_qdisc(&taprio_qdisc_ops);
1572 }
1573
1574 static void __exit taprio_module_exit(void)
1575 {
1576         unregister_qdisc(&taprio_qdisc_ops);
1577         unregister_netdevice_notifier(&taprio_device_notifier);
1578 }
1579
1580 module_init(taprio_module_init);
1581 module_exit(taprio_module_exit);
1582 MODULE_LICENSE("GPL");