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/ethtool.h>
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/kernel.h>
  13 #include <linux/string.h>
  14 #include <linux/list.h>
  15 #include <linux/errno.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/math64.h>
  18 #include <linux/module.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/rcupdate.h>
  21 #include <net/netlink.h>
  22 #include <net/pkt_sched.h>
  23 #include <net/pkt_cls.h>
  24 #include <net/sch_generic.h>
  25 #include <net/sock.h>
  26 #include <net/tcp.h>
  27
  28 static LIST_HEAD(taprio_list);
  29 static DEFINE_SPINLOCK(taprio_list_lock);
  30
  31 #define TAPRIO_ALL_GATES_OPEN -1
  32
  33 #define TXTIME_ASSIST_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST)
  34 #define FULL_OFFLOAD_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)
  35 #define TAPRIO_FLAGS_INVALID U32_MAX
  36
  37 struct sched_entry {
  38         struct list_head list;
  39
  40         /* The instant that this entry "closes" and the next one
  41          * should open, the qdisc will make some effort so that no
  42          * packet leaves after this time.
  43          */
  44         ktime_t close_time;
  45         ktime_t next_txtime;
  46         atomic_t budget;
  47         int index;
  48         u32 gate_mask;
  49         u32 interval;
  50         u8 command;
  51 };
  52
  53 struct sched_gate_list {
  54         struct rcu_head rcu;
  55         struct list_head entries;
  56         size_t num_entries;
  57         ktime_t cycle_close_time;
  58         s64 cycle_time;
  59         s64 cycle_time_extension;
  60         s64 base_time;
  61 };
  62
  63 struct taprio_sched {
  64         struct Qdisc **qdiscs;
  65         struct Qdisc *root;
  66         u32 flags;
  67         enum tk_offsets tk_offset;
  68         int clockid;
  69         atomic64_t picos_per_byte; /* Using picoseconds because for 10Gbps+
  70                                     * speeds it's sub-nanoseconds per byte
  71                                     */
  72
  73         /* Protects the update side of the RCU protected current_entry */
  74         spinlock_t current_entry_lock;
  75         struct sched_entry __rcu *current_entry;
  76         struct sched_gate_list __rcu *oper_sched;
  77         struct sched_gate_list __rcu *admin_sched;
  78         struct hrtimer advance_timer;
  79         struct list_head taprio_list;
  80         struct sk_buff *(*dequeue)(struct Qdisc *sch);
  81         struct sk_buff *(*peek)(struct Qdisc *sch);
  82         u32 txtime_delay;
  83 };
  84
  85 struct __tc_taprio_qopt_offload {
  86         refcount_t users;
  87         struct tc_taprio_qopt_offload offload;
  88 };
  89
  90 static ktime_t sched_base_time(const struct sched_gate_list *sched)
  91 {
  92         if (!sched)
  93                 return KTIME_MAX;
  94
  95         return ns_to_ktime(sched->base_time);
  96 }
  97
  98 static ktime_t taprio_mono_to_any(const struct taprio_sched *q, ktime_t mono)
  99 {
 100         /* This pairs with WRITE_ONCE() in taprio_parse_clockid() */
 101         enum tk_offsets tk_offset = READ_ONCE(q->tk_offset);
 102
 103         switch (tk_offset) {
 104         case TK_OFFS_MAX:
 105                 return mono;
 106         default:
 107                 return ktime_mono_to_any(mono, tk_offset);
 108         }
 109 }
 110
 111 static ktime_t taprio_get_time(const struct taprio_sched *q)
 112 {
 113         return taprio_mono_to_any(q, ktime_get());
 114 }
 115
 116 static void taprio_free_sched_cb(struct rcu_head *head)
 117 {
 118         struct sched_gate_list *sched = container_of(head, struct sched_gate_list, rcu);
 119         struct sched_entry *entry, *n;
 120
 121         list_for_each_entry_safe(entry, n, &sched->entries, list) {
 122                 list_del(&entry->list);
 123                 kfree(entry);
 124         }
 125
 126         kfree(sched);
 127 }
 128
 129 static void switch_schedules(struct taprio_sched *q,
 130                              struct sched_gate_list **admin,
 131                              struct sched_gate_list **oper)
 132 {
 133         rcu_assign_pointer(q->oper_sched, *admin);
 134         rcu_assign_pointer(q->admin_sched, NULL);
 135
 136         if (*oper)
 137                 call_rcu(&(*oper)->rcu, taprio_free_sched_cb);
 138
 139         *oper = *admin;
 140         *admin = NULL;
 141 }
 142
 143 /* Get how much time has been already elapsed in the current cycle. */
 144 static s32 get_cycle_time_elapsed(struct sched_gate_list *sched, ktime_t time)
 145 {
 146         ktime_t time_since_sched_start;
 147         s32 time_elapsed;
 148
 149         time_since_sched_start = ktime_sub(time, sched->base_time);
 150         div_s64_rem(time_since_sched_start, sched->cycle_time, &time_elapsed);
 151
 152         return time_elapsed;
 153 }
 154
 155 static ktime_t get_interval_end_time(struct sched_gate_list *sched,
 156                                      struct sched_gate_list *admin,
 157                                      struct sched_entry *entry,
 158                                      ktime_t intv_start)
 159 {
 160         s32 cycle_elapsed = get_cycle_time_elapsed(sched, intv_start);
 161         ktime_t intv_end, cycle_ext_end, cycle_end;
 162
 163         cycle_end = ktime_add_ns(intv_start, sched->cycle_time - cycle_elapsed);
 164         intv_end = ktime_add_ns(intv_start, entry->interval);
 165         cycle_ext_end = ktime_add(cycle_end, sched->cycle_time_extension);
 166
 167         if (ktime_before(intv_end, cycle_end))
 168                 return intv_end;
 169         else if (admin && admin != sched &&
 170                  ktime_after(admin->base_time, cycle_end) &&
 171                  ktime_before(admin->base_time, cycle_ext_end))
 172                 return admin->base_time;
 173         else
 174                 return cycle_end;
 175 }
 176
 177 static int length_to_duration(struct taprio_sched *q, int len)
 178 {
 179         return div_u64(len * atomic64_read(&q->picos_per_byte), 1000);
 180 }
 181
 182 /* Returns the entry corresponding to next available interval. If
 183  * validate_interval is set, it only validates whether the timestamp occurs
 184  * when the gate corresponding to the skb's traffic class is open.
 185  */
 186 static struct sched_entry *find_entry_to_transmit(struct sk_buff *skb,
 187                                                   struct Qdisc *sch,
 188                                                   struct sched_gate_list *sched,
 189                                                   struct sched_gate_list *admin,
 190                                                   ktime_t time,
 191                                                   ktime_t *interval_start,
 192                                                   ktime_t *interval_end,
 193                                                   bool validate_interval)
 194 {
 195         ktime_t curr_intv_start, curr_intv_end, cycle_end, packet_transmit_time;
 196         ktime_t earliest_txtime = KTIME_MAX, txtime, cycle, transmit_end_time;
 197         struct sched_entry *entry = NULL, *entry_found = NULL;
 198         struct taprio_sched *q = qdisc_priv(sch);
 199         struct net_device *dev = qdisc_dev(sch);
 200         bool entry_available = false;
 201         s32 cycle_elapsed;
 202         int tc, n;
 203
 204         tc = netdev_get_prio_tc_map(dev, skb->priority);
 205         packet_transmit_time = length_to_duration(q, qdisc_pkt_len(skb));
 206
 207         *interval_start = 0;
 208         *interval_end = 0;
 209
 210         if (!sched)
 211                 return NULL;
 212
 213         cycle = sched->cycle_time;
 214         cycle_elapsed = get_cycle_time_elapsed(sched, time);
 215         curr_intv_end = ktime_sub_ns(time, cycle_elapsed);
 216         cycle_end = ktime_add_ns(curr_intv_end, cycle);
 217
 218         list_for_each_entry(entry, &sched->entries, list) {
 219                 curr_intv_start = curr_intv_end;
 220                 curr_intv_end = get_interval_end_time(sched, admin, entry,
 221                                                       curr_intv_start);
 222
 223                 if (ktime_after(curr_intv_start, cycle_end))
 224                         break;
 225
 226                 if (!(entry->gate_mask & BIT(tc)) ||
 227                     packet_transmit_time > entry->interval)
 228                         continue;
 229
 230                 txtime = entry->next_txtime;
 231
 232                 if (ktime_before(txtime, time) || validate_interval) {
 233                         transmit_end_time = ktime_add_ns(time, packet_transmit_time);
 234                         if ((ktime_before(curr_intv_start, time) &&
 235                              ktime_before(transmit_end_time, curr_intv_end)) ||
 236                             (ktime_after(curr_intv_start, time) && !validate_interval)) {
 237                                 entry_found = entry;
 238                                 *interval_start = curr_intv_start;
 239                                 *interval_end = curr_intv_end;
 240                                 break;
 241                         } else if (!entry_available && !validate_interval) {
 242                                 /* Here, we are just trying to find out the
 243                                  * first available interval in the next cycle.
 244                                  */
 245                                 entry_available = true;
 246                                 entry_found = entry;
 247                                 *interval_start = ktime_add_ns(curr_intv_start, cycle);
 248                                 *interval_end = ktime_add_ns(curr_intv_end, cycle);
 249                         }
 250                 } else if (ktime_before(txtime, earliest_txtime) &&
 251                            !entry_available) {
 252                         earliest_txtime = txtime;
 253                         entry_found = entry;
 254                         n = div_s64(ktime_sub(txtime, curr_intv_start), cycle);
 255                         *interval_start = ktime_add(curr_intv_start, n * cycle);
 256                         *interval_end = ktime_add(curr_intv_end, n * cycle);
 257                 }
 258         }
 259
 260         return entry_found;
 261 }
 262
 263 static bool is_valid_interval(struct sk_buff *skb, struct Qdisc *sch)
 264 {
 265         struct taprio_sched *q = qdisc_priv(sch);
 266         struct sched_gate_list *sched, *admin;
 267         ktime_t interval_start, interval_end;
 268         struct sched_entry *entry;
 269
 270         rcu_read_lock();
 271         sched = rcu_dereference(q->oper_sched);
 272         admin = rcu_dereference(q->admin_sched);
 273
 274         entry = find_entry_to_transmit(skb, sch, sched, admin, skb->tstamp,
 275                                        &interval_start, &interval_end, true);
 276         rcu_read_unlock();
 277
 278         return entry;
 279 }
 280
 281 static bool taprio_flags_valid(u32 flags)
 282 {
 283         /* Make sure no other flag bits are set. */
 284         if (flags & ~(TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST |
 285                       TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 286                 return false;
 287         /* txtime-assist and full offload are mutually exclusive */
 288         if ((flags & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST) &&
 289             (flags & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 290                 return false;
 291         return true;
 292 }
 293
 294 /* This returns the tstamp value set by TCP in terms of the set clock. */
 295 static ktime_t get_tcp_tstamp(struct taprio_sched *q, struct sk_buff *skb)
 296 {
 297         unsigned int offset = skb_network_offset(skb);
 298         const struct ipv6hdr *ipv6h;
 299         const struct iphdr *iph;
 300         struct ipv6hdr _ipv6h;
 301
 302         ipv6h = skb_header_pointer(skb, offset, sizeof(_ipv6h), &_ipv6h);
 303         if (!ipv6h)
 304                 return 0;
 305
 306         if (ipv6h->version == 4) {
 307                 iph = (struct iphdr *)ipv6h;
 308                 offset += iph->ihl * 4;
 309
 310                 /* special-case 6in4 tunnelling, as that is a common way to get
 311                  * v6 connectivity in the home
 312                  */
 313                 if (iph->protocol == IPPROTO_IPV6) {
 314                         ipv6h = skb_header_pointer(skb, offset,
 315                                                    sizeof(_ipv6h), &_ipv6h);
 316
 317                         if (!ipv6h || ipv6h->nexthdr != IPPROTO_TCP)
 318                                 return 0;
 319                 } else if (iph->protocol != IPPROTO_TCP) {
 320                         return 0;
 321                 }
 322         } else if (ipv6h->version == 6 && ipv6h->nexthdr != IPPROTO_TCP) {
 323                 return 0;
 324         }
 325
 326         return taprio_mono_to_any(q, skb->skb_mstamp_ns);
 327 }
 328
 329 /* There are a few scenarios where we will have to modify the txtime from
 330  * what is read from next_txtime in sched_entry. They are:
 331  * 1. If txtime is in the past,
 332  *    a. The gate for the traffic class is currently open and packet can be
 333  *       transmitted before it closes, schedule the packet right away.
 334  *    b. If the gate corresponding to the traffic class is going to open later
 335  *       in the cycle, set the txtime of packet to the interval start.
 336  * 2. If txtime is in the future, there are packets corresponding to the
 337  *    current traffic class waiting to be transmitted. So, the following
 338  *    possibilities exist:
 339  *    a. We can transmit the packet before the window containing the txtime
 340  *       closes.
 341  *    b. The window might close before the transmission can be completed
 342  *       successfully. So, schedule the packet in the next open window.
 343  */
 344 static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)
 345 {
 346         ktime_t transmit_end_time, interval_end, interval_start, tcp_tstamp;
 347         struct taprio_sched *q = qdisc_priv(sch);
 348         struct sched_gate_list *sched, *admin;
 349         ktime_t minimum_time, now, txtime;
 350         int len, packet_transmit_time;
 351         struct sched_entry *entry;
 352         bool sched_changed;
 353
 354         now = taprio_get_time(q);
 355         minimum_time = ktime_add_ns(now, q->txtime_delay);
 356
 357         tcp_tstamp = get_tcp_tstamp(q, skb);
 358         minimum_time = max_t(ktime_t, minimum_time, tcp_tstamp);
 359
 360         rcu_read_lock();
 361         admin = rcu_dereference(q->admin_sched);
 362         sched = rcu_dereference(q->oper_sched);
 363         if (admin && ktime_after(minimum_time, admin->base_time))
 364                 switch_schedules(q, &admin, &sched);
 365
 366         /* Until the schedule starts, all the queues are open */
 367         if (!sched || ktime_before(minimum_time, sched->base_time)) {
 368                 txtime = minimum_time;
 369                 goto done;
 370         }
 371
 372         len = qdisc_pkt_len(skb);
 373         packet_transmit_time = length_to_duration(q, len);
 374
 375         do {
 376                 sched_changed = false;
 377
 378                 entry = find_entry_to_transmit(skb, sch, sched, admin,
 379                                                minimum_time,
 380                                                &interval_start, &interval_end,
 381                                                false);
 382                 if (!entry) {
 383                         txtime = 0;
 384                         goto done;
 385                 }
 386
 387                 txtime = entry->next_txtime;
 388                 txtime = max_t(ktime_t, txtime, minimum_time);
 389                 txtime = max_t(ktime_t, txtime, interval_start);
 390
 391                 if (admin && admin != sched &&
 392                     ktime_after(txtime, admin->base_time)) {
 393                         sched = admin;
 394                         sched_changed = true;
 395                         continue;
 396                 }
 397
 398                 transmit_end_time = ktime_add(txtime, packet_transmit_time);
 399                 minimum_time = transmit_end_time;
 400
 401                 /* Update the txtime of current entry to the next time it's
 402                  * interval starts.
 403                  */
 404                 if (ktime_after(transmit_end_time, interval_end))
 405                         entry->next_txtime = ktime_add(interval_start, sched->cycle_time);
 406         } while (sched_changed || ktime_after(transmit_end_time, interval_end));
 407
 408         entry->next_txtime = transmit_end_time;
 409
 410 done:
 411         rcu_read_unlock();
 412         return txtime;
 413 }
 414
 415 static int taprio_enqueue_one(struct sk_buff *skb, struct Qdisc *sch,
 416                               struct Qdisc *child, struct sk_buff **to_free)
 417 {
 418         struct taprio_sched *q = qdisc_priv(sch);
 419
 420         if (skb->sk && sock_flag(skb->sk, SOCK_TXTIME)) {
 421                 if (!is_valid_interval(skb, sch))
 422                         return qdisc_drop(skb, sch, to_free);
 423         } else if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
 424                 skb->tstamp = get_packet_txtime(skb, sch);
 425                 if (!skb->tstamp)
 426                         return qdisc_drop(skb, sch, to_free);
 427         }
 428
 429         qdisc_qstats_backlog_inc(sch, skb);
 430         sch->q.qlen++;
 431
 432         return qdisc_enqueue(skb, child, to_free);
 433 }
 434
 435 static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 436                           struct sk_buff **to_free)
 437 {
 438         struct taprio_sched *q = qdisc_priv(sch);
 439         struct Qdisc *child;
 440         int queue;
 441
 442         if (unlikely(FULL_OFFLOAD_IS_ENABLED(q->flags))) {
 443                 WARN_ONCE(1, "Trying to enqueue skb into the root of a taprio qdisc configured with full offload\n");
 444                 return qdisc_drop(skb, sch, to_free);
 445         }
 446
 447         queue = skb_get_queue_mapping(skb);
 448
 449         child = q->qdiscs[queue];
 450         if (unlikely(!child))
 451                 return qdisc_drop(skb, sch, to_free);
 452
 453         /* Large packets might not be transmitted when the transmission duration
 454          * exceeds any configured interval. Therefore, segment the skb into
 455          * smaller chunks. Skip it for the full offload case, as the driver
 456          * and/or the hardware is expected to handle this.
 457          */
 458         if (skb_is_gso(skb) && !FULL_OFFLOAD_IS_ENABLED(q->flags)) {
 459                 unsigned int slen = 0, numsegs = 0, len = qdisc_pkt_len(skb);
 460                 netdev_features_t features = netif_skb_features(skb);
 461                 struct sk_buff *segs, *nskb;
 462                 int ret;
 463
 464                 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 465                 if (IS_ERR_OR_NULL(segs))
 466                         return qdisc_drop(skb, sch, to_free);
 467
 468                 skb_list_walk_safe(segs, segs, nskb) {
 469                         skb_mark_not_on_list(segs);
 470                         qdisc_skb_cb(segs)->pkt_len = segs->len;
 471                         slen += segs->len;
 472
 473                         ret = taprio_enqueue_one(segs, sch, child, to_free);
 474                         if (ret != NET_XMIT_SUCCESS) {
 475                                 if (net_xmit_drop_count(ret))
 476                                         qdisc_qstats_drop(sch);
 477                         } else {
 478                                 numsegs++;
 479                         }
 480                 }
 481
 482                 if (numsegs > 1)
 483                         qdisc_tree_reduce_backlog(sch, 1 - numsegs, len - slen);
 484                 consume_skb(skb);
 485
 486                 return numsegs > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
 487         }
 488
 489         return taprio_enqueue_one(skb, sch, child, to_free);
 490 }
 491
 492 static struct sk_buff *taprio_peek_soft(struct Qdisc *sch)
 493 {
 494         struct taprio_sched *q = qdisc_priv(sch);
 495         struct net_device *dev = qdisc_dev(sch);
 496         struct sched_entry *entry;
 497         struct sk_buff *skb;
 498         u32 gate_mask;
 499         int i;
 500
 501         rcu_read_lock();
 502         entry = rcu_dereference(q->current_entry);
 503         gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
 504         rcu_read_unlock();
 505
 506         if (!gate_mask)
 507                 return NULL;
 508
 509         for (i = 0; i < dev->num_tx_queues; i++) {
 510                 struct Qdisc *child = q->qdiscs[i];
 511                 int prio;
 512                 u8 tc;
 513
 514                 if (unlikely(!child))
 515                         continue;
 516
 517                 skb = child->ops->peek(child);
 518                 if (!skb)
 519                         continue;
 520
 521                 if (TXTIME_ASSIST_IS_ENABLED(q->flags))
 522                         return skb;
 523
 524                 prio = skb->priority;
 525                 tc = netdev_get_prio_tc_map(dev, prio);
 526
 527                 if (!(gate_mask & BIT(tc)))
 528                         continue;
 529
 530                 return skb;
 531         }
 532
 533         return NULL;
 534 }
 535
 536 static struct sk_buff *taprio_peek_offload(struct Qdisc *sch)
 537 {
 538         WARN_ONCE(1, "Trying to peek into the root of a taprio qdisc configured with full offload\n");
 539
 540         return NULL;
 541 }
 542
 543 static struct sk_buff *taprio_peek(struct Qdisc *sch)
 544 {
 545         struct taprio_sched *q = qdisc_priv(sch);
 546
 547         return q->peek(sch);
 548 }
 549
 550 static void taprio_set_budget(struct taprio_sched *q, struct sched_entry *entry)
 551 {
 552         atomic_set(&entry->budget,
 553                    div64_u64((u64)entry->interval * 1000,
 554                              atomic64_read(&q->picos_per_byte)));
 555 }
 556
 557 static struct sk_buff *taprio_dequeue_soft(struct Qdisc *sch)
 558 {
 559         struct taprio_sched *q = qdisc_priv(sch);
 560         struct net_device *dev = qdisc_dev(sch);
 561         struct sk_buff *skb = NULL;
 562         struct sched_entry *entry;
 563         u32 gate_mask;
 564         int i;
 565
 566         rcu_read_lock();
 567         entry = rcu_dereference(q->current_entry);
 568         /* if there's no entry, it means that the schedule didn't
 569          * start yet, so force all gates to be open, this is in
 570          * accordance to IEEE 802.1Qbv-2015 Section 8.6.9.4.5
 571          * "AdminGateStates"
 572          */
 573         gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
 574
 575         if (!gate_mask)
 576                 goto done;
 577
 578         for (i = 0; i < dev->num_tx_queues; i++) {
 579                 struct Qdisc *child = q->qdiscs[i];
 580                 ktime_t guard;
 581                 int prio;
 582                 int len;
 583                 u8 tc;
 584
 585                 if (unlikely(!child))
 586                         continue;
 587
 588                 if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
 589                         skb = child->ops->dequeue(child);
 590                         if (!skb)
 591                                 continue;
 592                         goto skb_found;
 593                 }
 594
 595                 skb = child->ops->peek(child);
 596                 if (!skb)
 597                         continue;
 598
 599                 prio = skb->priority;
 600                 tc = netdev_get_prio_tc_map(dev, prio);
 601
 602                 if (!(gate_mask & BIT(tc))) {
 603                         skb = NULL;
 604                         continue;
 605                 }
 606
 607                 len = qdisc_pkt_len(skb);
 608                 guard = ktime_add_ns(taprio_get_time(q),
 609                                      length_to_duration(q, len));
 610
 611                 /* In the case that there's no gate entry, there's no
 612                  * guard band ...
 613                  */
 614                 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 615                     ktime_after(guard, entry->close_time)) {
 616                         skb = NULL;
 617                         continue;
 618                 }
 619
 620                 /* ... and no budget. */
 621                 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 622                     atomic_sub_return(len, &entry->budget) < 0) {
 623                         skb = NULL;
 624                         continue;
 625                 }
 626
 627                 skb = child->ops->dequeue(child);
 628                 if (unlikely(!skb))
 629                         goto done;
 630
 631 skb_found:
 632                 qdisc_bstats_update(sch, skb);
 633                 qdisc_qstats_backlog_dec(sch, skb);
 634                 sch->q.qlen--;
 635
 636                 goto done;
 637         }
 638
 639 done:
 640         rcu_read_unlock();
 641
 642         return skb;
 643 }
 644
 645 static struct sk_buff *taprio_dequeue_offload(struct Qdisc *sch)
 646 {
 647         WARN_ONCE(1, "Trying to dequeue from the root of a taprio qdisc configured with full offload\n");
 648
 649         return NULL;
 650 }
 651
 652 static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
 653 {
 654         struct taprio_sched *q = qdisc_priv(sch);
 655
 656         return q->dequeue(sch);
 657 }
 658
 659 static bool should_restart_cycle(const struct sched_gate_list *oper,
 660                                  const struct sched_entry *entry)
 661 {
 662         if (list_is_last(&entry->list, &oper->entries))
 663                 return true;
 664
 665         if (ktime_compare(entry->close_time, oper->cycle_close_time) == 0)
 666                 return true;
 667
 668         return false;
 669 }
 670
 671 static bool should_change_schedules(const struct sched_gate_list *admin,
 672                                     const struct sched_gate_list *oper,
 673                                     ktime_t close_time)
 674 {
 675         ktime_t next_base_time, extension_time;
 676
 677         if (!admin)
 678                 return false;
 679
 680         next_base_time = sched_base_time(admin);
 681
 682         /* This is the simple case, the close_time would fall after
 683          * the next schedule base_time.
 684          */
 685         if (ktime_compare(next_base_time, close_time) <= 0)
 686                 return true;
 687
 688         /* This is the cycle_time_extension case, if the close_time
 689          * plus the amount that can be extended would fall after the
 690          * next schedule base_time, we can extend the current schedule
 691          * for that amount.
 692          */
 693         extension_time = ktime_add_ns(close_time, oper->cycle_time_extension);
 694
 695         /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
 696          * how precisely the extension should be made. So after
 697          * conformance testing, this logic may change.
 698          */
 699         if (ktime_compare(next_base_time, extension_time) <= 0)
 700                 return true;
 701
 702         return false;
 703 }
 704
 705 static enum hrtimer_restart advance_sched(struct hrtimer *timer)
 706 {
 707         struct taprio_sched *q = container_of(timer, struct taprio_sched,
 708                                               advance_timer);
 709         struct sched_gate_list *oper, *admin;
 710         struct sched_entry *entry, *next;
 711         struct Qdisc *sch = q->root;
 712         ktime_t close_time;
 713
 714         spin_lock(&q->current_entry_lock);
 715         entry = rcu_dereference_protected(q->current_entry,
 716                                           lockdep_is_held(&q->current_entry_lock));
 717         oper = rcu_dereference_protected(q->oper_sched,
 718                                          lockdep_is_held(&q->current_entry_lock));
 719         admin = rcu_dereference_protected(q->admin_sched,
 720                                           lockdep_is_held(&q->current_entry_lock));
 721
 722         if (!oper)
 723                 switch_schedules(q, &admin, &oper);
 724
 725         /* This can happen in two cases: 1. this is the very first run
 726          * of this function (i.e. we weren't running any schedule
 727          * previously); 2. The previous schedule just ended. The first
 728          * entry of all schedules are pre-calculated during the
 729          * schedule initialization.
 730          */
 731         if (unlikely(!entry || entry->close_time == oper->base_time)) {
 732                 next = list_first_entry(&oper->entries, struct sched_entry,
 733                                         list);
 734                 close_time = next->close_time;
 735                 goto first_run;
 736         }
 737
 738         if (should_restart_cycle(oper, entry)) {
 739                 next = list_first_entry(&oper->entries, struct sched_entry,
 740                                         list);
 741                 oper->cycle_close_time = ktime_add_ns(oper->cycle_close_time,
 742                                                       oper->cycle_time);
 743         } else {
 744                 next = list_next_entry(entry, list);
 745         }
 746
 747         close_time = ktime_add_ns(entry->close_time, next->interval);
 748         close_time = min_t(ktime_t, close_time, oper->cycle_close_time);
 749
 750         if (should_change_schedules(admin, oper, close_time)) {
 751                 /* Set things so the next time this runs, the new
 752                  * schedule runs.
 753                  */
 754                 close_time = sched_base_time(admin);
 755                 switch_schedules(q, &admin, &oper);
 756         }
 757
 758         next->close_time = close_time;
 759         taprio_set_budget(q, next);
 760
 761 first_run:
 762         rcu_assign_pointer(q->current_entry, next);
 763         spin_unlock(&q->current_entry_lock);
 764
 765         hrtimer_set_expires(&q->advance_timer, close_time);
 766
 767         rcu_read_lock();
 768         __netif_schedule(sch);
 769         rcu_read_unlock();
 770
 771         return HRTIMER_RESTART;
 772 }
 773
 774 static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
 775         [TCA_TAPRIO_SCHED_ENTRY_INDEX]     = { .type = NLA_U32 },
 776         [TCA_TAPRIO_SCHED_ENTRY_CMD]       = { .type = NLA_U8 },
 777         [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
 778         [TCA_TAPRIO_SCHED_ENTRY_INTERVAL]  = { .type = NLA_U32 },
 779 };
 780
 781 static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
 782         [TCA_TAPRIO_ATTR_PRIOMAP]              = {
 783                 .len = sizeof(struct tc_mqprio_qopt)
 784         },
 785         [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST]           = { .type = NLA_NESTED },
 786         [TCA_TAPRIO_ATTR_SCHED_BASE_TIME]            = { .type = NLA_S64 },
 787         [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]         = { .type = NLA_NESTED },
 788         [TCA_TAPRIO_ATTR_SCHED_CLOCKID]              = { .type = NLA_S32 },
 789         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]           = { .type = NLA_S64 },
 790         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
 791         [TCA_TAPRIO_ATTR_FLAGS]                      = { .type = NLA_U32 },
 792         [TCA_TAPRIO_ATTR_TXTIME_DELAY]               = { .type = NLA_U32 },
 793 };
 794
 795 static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
 796                             struct sched_entry *entry,
 797                             struct netlink_ext_ack *extack)
 798 {
 799         int min_duration = length_to_duration(q, ETH_ZLEN);
 800         u32 interval = 0;
 801
 802         if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
 803                 entry->command = nla_get_u8(
 804                         tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
 805
 806         if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
 807                 entry->gate_mask = nla_get_u32(
 808                         tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
 809
 810         if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
 811                 interval = nla_get_u32(
 812                         tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
 813
 814         /* The interval should allow at least the minimum ethernet
 815          * frame to go out.
 816          */
 817         if (interval < min_duration) {
 818                 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
 819                 return -EINVAL;
 820         }
 821
 822         entry->interval = interval;
 823
 824         return 0;
 825 }
 826
 827 static int parse_sched_entry(struct taprio_sched *q, struct nlattr *n,
 828                              struct sched_entry *entry, int index,
 829                              struct netlink_ext_ack *extack)
 830 {
 831         struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
 832         int err;
 833
 834         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
 835                                           entry_policy, NULL);
 836         if (err < 0) {
 837                 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
 838                 return -EINVAL;
 839         }
 840
 841         entry->index = index;
 842
 843         return fill_sched_entry(q, tb, entry, extack);
 844 }
 845
 846 static int parse_sched_list(struct taprio_sched *q, struct nlattr *list,
 847                             struct sched_gate_list *sched,
 848                             struct netlink_ext_ack *extack)
 849 {
 850         struct nlattr *n;
 851         int err, rem;
 852         int i = 0;
 853
 854         if (!list)
 855                 return -EINVAL;
 856
 857         nla_for_each_nested(n, list, rem) {
 858                 struct sched_entry *entry;
 859
 860                 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
 861                         NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
 862                         continue;
 863                 }
 864
 865                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 866                 if (!entry) {
 867                         NL_SET_ERR_MSG(extack, "Not enough memory for entry");
 868                         return -ENOMEM;
 869                 }
 870
 871                 err = parse_sched_entry(q, n, entry, i, extack);
 872                 if (err < 0) {
 873                         kfree(entry);
 874                         return err;
 875                 }
 876
 877                 list_add_tail(&entry->list, &sched->entries);
 878                 i++;
 879         }
 880
 881         sched->num_entries = i;
 882
 883         return i;
 884 }
 885
 886 static int parse_taprio_schedule(struct taprio_sched *q, struct nlattr **tb,
 887                                  struct sched_gate_list *new,
 888                                  struct netlink_ext_ack *extack)
 889 {
 890         int err = 0;
 891
 892         if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
 893                 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
 894                 return -ENOTSUPP;
 895         }
 896
 897         if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
 898                 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
 899
 900         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
 901                 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
 902
 903         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
 904                 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
 905
 906         if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
 907                 err = parse_sched_list(q, tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST],
 908                                        new, extack);
 909         if (err < 0)
 910                 return err;
 911
 912         if (!new->cycle_time) {
 913                 struct sched_entry *entry;
 914                 ktime_t cycle = 0;
 915
 916                 list_for_each_entry(entry, &new->entries, list)
 917                         cycle = ktime_add_ns(cycle, entry->interval);
 918
 919                 if (!cycle) {
 920                         NL_SET_ERR_MSG(extack, "'cycle_time' can never be 0");
 921                         return -EINVAL;
 922                 }
 923
 924                 new->cycle_time = cycle;
 925         }
 926
 927         return 0;
 928 }
 929
 930 static int taprio_parse_mqprio_opt(struct net_device *dev,
 931                                    struct tc_mqprio_qopt *qopt,
 932                                    struct netlink_ext_ack *extack,
 933                                    u32 taprio_flags)
 934 {
 935         int i, j;
 936
 937         if (!qopt && !dev->num_tc) {
 938                 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
 939                 return -EINVAL;
 940         }
 941
 942         /* If num_tc is already set, it means that the user already
 943          * configured the mqprio part
 944          */
 945         if (dev->num_tc)
 946                 return 0;
 947
 948         /* Verify num_tc is not out of max range */
 949         if (qopt->num_tc > TC_MAX_QUEUE) {
 950                 NL_SET_ERR_MSG(extack, "Number of traffic classes is outside valid range");
 951                 return -EINVAL;
 952         }
 953
 954         /* taprio imposes that traffic classes map 1:n to tx queues */
 955         if (qopt->num_tc > dev->num_tx_queues) {
 956                 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
 957                 return -EINVAL;
 958         }
 959
 960         /* Verify priority mapping uses valid tcs */
 961         for (i = 0; i <= TC_BITMASK; i++) {
 962                 if (qopt->prio_tc_map[i] >= qopt->num_tc) {
 963                         NL_SET_ERR_MSG(extack, "Invalid traffic class in priority to traffic class mapping");
 964                         return -EINVAL;
 965                 }
 966         }
 967
 968         for (i = 0; i < qopt->num_tc; i++) {
 969                 unsigned int last = qopt->offset[i] + qopt->count[i];
 970
 971                 /* Verify the queue count is in tx range being equal to the
 972                  * real_num_tx_queues indicates the last queue is in use.
 973                  */
 974                 if (qopt->offset[i] >= dev->num_tx_queues ||
 975                     !qopt->count[i] ||
 976                     last > dev->real_num_tx_queues) {
 977                         NL_SET_ERR_MSG(extack, "Invalid queue in traffic class to queue mapping");
 978                         return -EINVAL;
 979                 }
 980
 981                 if (TXTIME_ASSIST_IS_ENABLED(taprio_flags))
 982                         continue;
 983
 984                 /* Verify that the offset and counts do not overlap */
 985                 for (j = i + 1; j < qopt->num_tc; j++) {
 986                         if (last > qopt->offset[j]) {
 987                                 NL_SET_ERR_MSG(extack, "Detected overlap in the traffic class to queue mapping");
 988                                 return -EINVAL;
 989                         }
 990                 }
 991         }
 992
 993         return 0;
 994 }
 995
 996 static int taprio_get_start_time(struct Qdisc *sch,
 997                                  struct sched_gate_list *sched,
 998                                  ktime_t *start)
 999 {
1000         struct taprio_sched *q = qdisc_priv(sch);
1001         ktime_t now, base, cycle;
1002         s64 n;
1003
1004         base = sched_base_time(sched);
1005         now = taprio_get_time(q);
1006
1007         if (ktime_after(base, now)) {
1008                 *start = base;
1009                 return 0;
1010         }
1011
1012         cycle = sched->cycle_time;
1013
1014         /* The qdisc is expected to have at least one sched_entry.  Moreover,
1015          * any entry must have 'interval' > 0. Thus if the cycle time is zero,
1016          * something went really wrong. In that case, we should warn about this
1017          * inconsistent state and return error.
1018          */
1019         if (WARN_ON(!cycle))
1020                 return -EFAULT;
1021
1022         /* Schedule the start time for the beginning of the next
1023          * cycle.
1024          */
1025         n = div64_s64(ktime_sub_ns(now, base), cycle);
1026         *start = ktime_add_ns(base, (n + 1) * cycle);
1027         return 0;
1028 }
1029
1030 static void setup_first_close_time(struct taprio_sched *q,
1031                                    struct sched_gate_list *sched, ktime_t base)
1032 {
1033         struct sched_entry *first;
1034         ktime_t cycle;
1035
1036         first = list_first_entry(&sched->entries,
1037                                  struct sched_entry, list);
1038
1039         cycle = sched->cycle_time;
1040
1041         /* FIXME: find a better place to do this */
1042         sched->cycle_close_time = ktime_add_ns(base, cycle);
1043
1044         first->close_time = ktime_add_ns(base, first->interval);
1045         taprio_set_budget(q, first);
1046         rcu_assign_pointer(q->current_entry, NULL);
1047 }
1048
1049 static void taprio_start_sched(struct Qdisc *sch,
1050                                ktime_t start, struct sched_gate_list *new)
1051 {
1052         struct taprio_sched *q = qdisc_priv(sch);
1053         ktime_t expires;
1054
1055         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1056                 return;
1057
1058         expires = hrtimer_get_expires(&q->advance_timer);
1059         if (expires == 0)
1060                 expires = KTIME_MAX;
1061
1062         /* If the new schedule starts before the next expiration, we
1063          * reprogram it to the earliest one, so we change the admin
1064          * schedule to the operational one at the right time.
1065          */
1066         start = min_t(ktime_t, start, expires);
1067
1068         hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1069 }
1070
1071 static void taprio_set_picos_per_byte(struct net_device *dev,
1072                                       struct taprio_sched *q)
1073 {
1074         struct ethtool_link_ksettings ecmd;
1075         int speed = SPEED_10;
1076         int picos_per_byte;
1077         int err;
1078
1079         err = __ethtool_get_link_ksettings(dev, &ecmd);
1080         if (err < 0)
1081                 goto skip;
1082
1083         if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1084                 speed = ecmd.base.speed;
1085
1086 skip:
1087         picos_per_byte = (USEC_PER_SEC * 8) / speed;
1088
1089         atomic64_set(&q->picos_per_byte, picos_per_byte);
1090         netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1091                    dev->name, (long long)atomic64_read(&q->picos_per_byte),
1092                    ecmd.base.speed);
1093 }
1094
1095 static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1096                                void *ptr)
1097 {
1098         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1099         struct net_device *qdev;
1100         struct taprio_sched *q;
1101         bool found = false;
1102
1103         ASSERT_RTNL();
1104
1105         if (event != NETDEV_UP && event != NETDEV_CHANGE)
1106                 return NOTIFY_DONE;
1107
1108         spin_lock(&taprio_list_lock);
1109         list_for_each_entry(q, &taprio_list, taprio_list) {
1110                 qdev = qdisc_dev(q->root);
1111                 if (qdev == dev) {
1112                         found = true;
1113                         break;
1114                 }
1115         }
1116         spin_unlock(&taprio_list_lock);
1117
1118         if (found)
1119                 taprio_set_picos_per_byte(dev, q);
1120
1121         return NOTIFY_DONE;
1122 }
1123
1124 static void setup_txtime(struct taprio_sched *q,
1125                          struct sched_gate_list *sched, ktime_t base)
1126 {
1127         struct sched_entry *entry;
1128         u32 interval = 0;
1129
1130         list_for_each_entry(entry, &sched->entries, list) {
1131                 entry->next_txtime = ktime_add_ns(base, interval);
1132                 interval += entry->interval;
1133         }
1134 }
1135
1136 static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1137 {
1138         struct __tc_taprio_qopt_offload *__offload;
1139
1140         __offload = kzalloc(struct_size(__offload, offload.entries, num_entries),
1141                             GFP_KERNEL);
1142         if (!__offload)
1143                 return NULL;
1144
1145         refcount_set(&__offload->users, 1);
1146
1147         return &__offload->offload;
1148 }
1149
1150 struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1151                                                   *offload)
1152 {
1153         struct __tc_taprio_qopt_offload *__offload;
1154
1155         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1156                                  offload);
1157
1158         refcount_inc(&__offload->users);
1159
1160         return offload;
1161 }
1162 EXPORT_SYMBOL_GPL(taprio_offload_get);
1163
1164 void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1165 {
1166         struct __tc_taprio_qopt_offload *__offload;
1167
1168         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1169                                  offload);
1170
1171         if (!refcount_dec_and_test(&__offload->users))
1172                 return;
1173
1174         kfree(__offload);
1175 }
1176 EXPORT_SYMBOL_GPL(taprio_offload_free);
1177
1178 /* The function will only serve to keep the pointers to the "oper" and "admin"
1179  * schedules valid in relation to their base times, so when calling dump() the
1180  * users looks at the right schedules.
1181  * When using full offload, the admin configuration is promoted to oper at the
1182  * base_time in the PHC time domain.  But because the system time is not
1183  * necessarily in sync with that, we can't just trigger a hrtimer to call
1184  * switch_schedules at the right hardware time.
1185  * At the moment we call this by hand right away from taprio, but in the future
1186  * it will be useful to create a mechanism for drivers to notify taprio of the
1187  * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1188  * This is left as TODO.
1189  */
1190 static void taprio_offload_config_changed(struct taprio_sched *q)
1191 {
1192         struct sched_gate_list *oper, *admin;
1193
1194         spin_lock(&q->current_entry_lock);
1195
1196         oper = rcu_dereference_protected(q->oper_sched,
1197                                          lockdep_is_held(&q->current_entry_lock));
1198         admin = rcu_dereference_protected(q->admin_sched,
1199                                           lockdep_is_held(&q->current_entry_lock));
1200
1201         switch_schedules(q, &admin, &oper);
1202
1203         spin_unlock(&q->current_entry_lock);
1204 }
1205
1206 static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
1207 {
1208         u32 i, queue_mask = 0;
1209
1210         for (i = 0; i < dev->num_tc; i++) {
1211                 u32 offset, count;
1212
1213                 if (!(tc_mask & BIT(i)))
1214                         continue;
1215
1216                 offset = dev->tc_to_txq[i].offset;
1217                 count = dev->tc_to_txq[i].count;
1218
1219                 queue_mask |= GENMASK(offset + count - 1, offset);
1220         }
1221
1222         return queue_mask;
1223 }
1224
1225 static void taprio_sched_to_offload(struct net_device *dev,
1226                                     struct sched_gate_list *sched,
1227                                     struct tc_taprio_qopt_offload *offload)
1228 {
1229         struct sched_entry *entry;
1230         int i = 0;
1231
1232         offload->base_time = sched->base_time;
1233         offload->cycle_time = sched->cycle_time;
1234         offload->cycle_time_extension = sched->cycle_time_extension;
1235
1236         list_for_each_entry(entry, &sched->entries, list) {
1237                 struct tc_taprio_sched_entry *e = &offload->entries[i];
1238
1239                 e->command = entry->command;
1240                 e->interval = entry->interval;
1241                 e->gate_mask = tc_map_to_queue_mask(dev, entry->gate_mask);
1242
1243                 i++;
1244         }
1245
1246         offload->num_entries = i;
1247 }
1248
1249 static int taprio_enable_offload(struct net_device *dev,
1250                                  struct taprio_sched *q,
1251                                  struct sched_gate_list *sched,
1252                                  struct netlink_ext_ack *extack)
1253 {
1254         const struct net_device_ops *ops = dev->netdev_ops;
1255         struct tc_taprio_qopt_offload *offload;
1256         int err = 0;
1257
1258         if (!ops->ndo_setup_tc) {
1259                 NL_SET_ERR_MSG(extack,
1260                                "Device does not support taprio offload");
1261                 return -EOPNOTSUPP;
1262         }
1263
1264         offload = taprio_offload_alloc(sched->num_entries);
1265         if (!offload) {
1266                 NL_SET_ERR_MSG(extack,
1267                                "Not enough memory for enabling offload mode");
1268                 return -ENOMEM;
1269         }
1270         offload->enable = 1;
1271         taprio_sched_to_offload(dev, sched, offload);
1272
1273         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1274         if (err < 0) {
1275                 NL_SET_ERR_MSG(extack,
1276                                "Device failed to setup taprio offload");
1277                 goto done;
1278         }
1279
1280 done:
1281         taprio_offload_free(offload);
1282
1283         return err;
1284 }
1285
1286 static int taprio_disable_offload(struct net_device *dev,
1287                                   struct taprio_sched *q,
1288                                   struct netlink_ext_ack *extack)
1289 {
1290         const struct net_device_ops *ops = dev->netdev_ops;
1291         struct tc_taprio_qopt_offload *offload;
1292         int err;
1293
1294         if (!FULL_OFFLOAD_IS_ENABLED(q->flags))
1295                 return 0;
1296
1297         if (!ops->ndo_setup_tc)
1298                 return -EOPNOTSUPP;
1299
1300         offload = taprio_offload_alloc(0);
1301         if (!offload) {
1302                 NL_SET_ERR_MSG(extack,
1303                                "Not enough memory to disable offload mode");
1304                 return -ENOMEM;
1305         }
1306         offload->enable = 0;
1307
1308         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1309         if (err < 0) {
1310                 NL_SET_ERR_MSG(extack,
1311                                "Device failed to disable offload");
1312                 goto out;
1313         }
1314
1315 out:
1316         taprio_offload_free(offload);
1317
1318         return err;
1319 }
1320
1321 /* If full offload is enabled, the only possible clockid is the net device's
1322  * PHC. For that reason, specifying a clockid through netlink is incorrect.
1323  * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1324  * in sync with the specified clockid via a user space daemon such as phc2sys.
1325  * For both software taprio and txtime-assist, the clockid is used for the
1326  * hrtimer that advances the schedule and hence mandatory.
1327  */
1328 static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1329                                 struct netlink_ext_ack *extack)
1330 {
1331         struct taprio_sched *q = qdisc_priv(sch);
1332         struct net_device *dev = qdisc_dev(sch);
1333         int err = -EINVAL;
1334
1335         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1336                 const struct ethtool_ops *ops = dev->ethtool_ops;
1337                 struct ethtool_ts_info info = {
1338                         .cmd = ETHTOOL_GET_TS_INFO,
1339                         .phc_index = -1,
1340                 };
1341
1342                 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1343                         NL_SET_ERR_MSG(extack,
1344                                        "The 'clockid' cannot be specified for full offload");
1345                         goto out;
1346                 }
1347
1348                 if (ops && ops->get_ts_info)
1349                         err = ops->get_ts_info(dev, &info);
1350
1351                 if (err || info.phc_index < 0) {
1352                         NL_SET_ERR_MSG(extack,
1353                                        "Device does not have a PTP clock");
1354                         err = -ENOTSUPP;
1355                         goto out;
1356                 }
1357         } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1358                 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1359                 enum tk_offsets tk_offset;
1360
1361                 /* We only support static clockids and we don't allow
1362                  * for it to be modified after the first init.
1363                  */
1364                 if (clockid < 0 ||
1365                     (q->clockid != -1 && q->clockid != clockid)) {
1366                         NL_SET_ERR_MSG(extack,
1367                                        "Changing the 'clockid' of a running schedule is not supported");
1368                         err = -ENOTSUPP;
1369                         goto out;
1370                 }
1371
1372                 switch (clockid) {
1373                 case CLOCK_REALTIME:
1374                         tk_offset = TK_OFFS_REAL;
1375                         break;
1376                 case CLOCK_MONOTONIC:
1377                         tk_offset = TK_OFFS_MAX;
1378                         break;
1379                 case CLOCK_BOOTTIME:
1380                         tk_offset = TK_OFFS_BOOT;
1381                         break;
1382                 case CLOCK_TAI:
1383                         tk_offset = TK_OFFS_TAI;
1384                         break;
1385                 default:
1386                         NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1387                         err = -EINVAL;
1388                         goto out;
1389                 }
1390                 /* This pairs with READ_ONCE() in taprio_mono_to_any */
1391                 WRITE_ONCE(q->tk_offset, tk_offset);
1392
1393                 q->clockid = clockid;
1394         } else {
1395                 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1396                 goto out;
1397         }
1398
1399         /* Everything went ok, return success. */
1400         err = 0;
1401
1402 out:
1403         return err;
1404 }
1405
1406 static int taprio_mqprio_cmp(const struct net_device *dev,
1407                              const struct tc_mqprio_qopt *mqprio)
1408 {
1409         int i;
1410
1411         if (!mqprio || mqprio->num_tc != dev->num_tc)
1412                 return -1;
1413
1414         for (i = 0; i < mqprio->num_tc; i++)
1415                 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1416                     dev->tc_to_txq[i].offset != mqprio->offset[i])
1417                         return -1;
1418
1419         for (i = 0; i <= TC_BITMASK; i++)
1420                 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1421                         return -1;
1422
1423         return 0;
1424 }
1425
1426 /* The semantics of the 'flags' argument in relation to 'change()'
1427  * requests, are interpreted following two rules (which are applied in
1428  * this order): (1) an omitted 'flags' argument is interpreted as
1429  * zero; (2) the 'flags' of a "running" taprio instance cannot be
1430  * changed.
1431  */
1432 static int taprio_new_flags(const struct nlattr *attr, u32 old,
1433                             struct netlink_ext_ack *extack)
1434 {
1435         u32 new = 0;
1436
1437         if (attr)
1438                 new = nla_get_u32(attr);
1439
1440         if (old != TAPRIO_FLAGS_INVALID && old != new) {
1441                 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1442                 return -EOPNOTSUPP;
1443         }
1444
1445         if (!taprio_flags_valid(new)) {
1446                 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1447                 return -EINVAL;
1448         }
1449
1450         return new;
1451 }
1452
1453 static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1454                          struct netlink_ext_ack *extack)
1455 {
1456         struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1457         struct sched_gate_list *oper, *admin, *new_admin;
1458         struct taprio_sched *q = qdisc_priv(sch);
1459         struct net_device *dev = qdisc_dev(sch);
1460         struct tc_mqprio_qopt *mqprio = NULL;
1461         unsigned long flags;
1462         ktime_t start;
1463         int i, err;
1464
1465         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1466                                           taprio_policy, extack);
1467         if (err < 0)
1468                 return err;
1469
1470         if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1471                 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1472
1473         err = taprio_new_flags(tb[TCA_TAPRIO_ATTR_FLAGS],
1474                                q->flags, extack);
1475         if (err < 0)
1476                 return err;
1477
1478         q->flags = err;
1479
1480         err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1481         if (err < 0)
1482                 return err;
1483
1484         new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1485         if (!new_admin) {
1486                 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1487                 return -ENOMEM;
1488         }
1489         INIT_LIST_HEAD(&new_admin->entries);
1490
1491         rcu_read_lock();
1492         oper = rcu_dereference(q->oper_sched);
1493         admin = rcu_dereference(q->admin_sched);
1494         rcu_read_unlock();
1495
1496         /* no changes - no new mqprio settings */
1497         if (!taprio_mqprio_cmp(dev, mqprio))
1498                 mqprio = NULL;
1499
1500         if (mqprio && (oper || admin)) {
1501                 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1502                 err = -ENOTSUPP;
1503                 goto free_sched;
1504         }
1505
1506         err = parse_taprio_schedule(q, tb, new_admin, extack);
1507         if (err < 0)
1508                 goto free_sched;
1509
1510         if (new_admin->num_entries == 0) {
1511                 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1512                 err = -EINVAL;
1513                 goto free_sched;
1514         }
1515
1516         err = taprio_parse_clockid(sch, tb, extack);
1517         if (err < 0)
1518                 goto free_sched;
1519
1520         taprio_set_picos_per_byte(dev, q);
1521
1522         if (mqprio) {
1523                 err = netdev_set_num_tc(dev, mqprio->num_tc);
1524                 if (err)
1525                         goto free_sched;
1526                 for (i = 0; i < mqprio->num_tc; i++)
1527                         netdev_set_tc_queue(dev, i,
1528                                             mqprio->count[i],
1529                                             mqprio->offset[i]);
1530
1531                 /* Always use supplied priority mappings */
1532                 for (i = 0; i <= TC_BITMASK; i++)
1533                         netdev_set_prio_tc_map(dev, i,
1534                                                mqprio->prio_tc_map[i]);
1535         }
1536
1537         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1538                 err = taprio_enable_offload(dev, q, new_admin, extack);
1539         else
1540                 err = taprio_disable_offload(dev, q, extack);
1541         if (err)
1542                 goto free_sched;
1543
1544         /* Protects against enqueue()/dequeue() */
1545         spin_lock_bh(qdisc_lock(sch));
1546
1547         if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1548                 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1549                         NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1550                         err = -EINVAL;
1551                         goto unlock;
1552                 }
1553
1554                 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1555         }
1556
1557         if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1558             !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1559             !hrtimer_active(&q->advance_timer)) {
1560                 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1561                 q->advance_timer.function = advance_sched;
1562         }
1563
1564         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1565                 q->dequeue = taprio_dequeue_offload;
1566                 q->peek = taprio_peek_offload;
1567         } else {
1568                 /* Be sure to always keep the function pointers
1569                  * in a consistent state.
1570                  */
1571                 q->dequeue = taprio_dequeue_soft;
1572                 q->peek = taprio_peek_soft;
1573         }
1574
1575         err = taprio_get_start_time(sch, new_admin, &start);
1576         if (err < 0) {
1577                 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1578                 goto unlock;
1579         }
1580
1581         setup_txtime(q, new_admin, start);
1582
1583         if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1584                 if (!oper) {
1585                         rcu_assign_pointer(q->oper_sched, new_admin);
1586                         err = 0;
1587                         new_admin = NULL;
1588                         goto unlock;
1589                 }
1590
1591                 rcu_assign_pointer(q->admin_sched, new_admin);
1592                 if (admin)
1593                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1594         } else {
1595                 setup_first_close_time(q, new_admin, start);
1596
1597                 /* Protects against advance_sched() */
1598                 spin_lock_irqsave(&q->current_entry_lock, flags);
1599
1600                 taprio_start_sched(sch, start, new_admin);
1601
1602                 rcu_assign_pointer(q->admin_sched, new_admin);
1603                 if (admin)
1604                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1605
1606                 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1607
1608                 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1609                         taprio_offload_config_changed(q);
1610         }
1611
1612         new_admin = NULL;
1613         err = 0;
1614
1615 unlock:
1616         spin_unlock_bh(qdisc_lock(sch));
1617
1618 free_sched:
1619         if (new_admin)
1620                 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
1621
1622         return err;
1623 }
1624
1625 static void taprio_reset(struct Qdisc *sch)
1626 {
1627         struct taprio_sched *q = qdisc_priv(sch);
1628         struct net_device *dev = qdisc_dev(sch);
1629         int i;
1630
1631         hrtimer_cancel(&q->advance_timer);
1632         if (q->qdiscs) {
1633                 for (i = 0; i < dev->num_tx_queues; i++)
1634                         if (q->qdiscs[i])
1635                                 qdisc_reset(q->qdiscs[i]);
1636         }
1637         sch->qstats.backlog = 0;
1638         sch->q.qlen = 0;
1639 }
1640
1641 static void taprio_destroy(struct Qdisc *sch)
1642 {
1643         struct taprio_sched *q = qdisc_priv(sch);
1644         struct net_device *dev = qdisc_dev(sch);
1645         unsigned int i;
1646
1647         spin_lock(&taprio_list_lock);
1648         list_del(&q->taprio_list);
1649         spin_unlock(&taprio_list_lock);
1650
1651         /* Note that taprio_reset() might not be called if an error
1652          * happens in qdisc_create(), after taprio_init() has been called.
1653          */
1654         hrtimer_cancel(&q->advance_timer);
1655
1656         taprio_disable_offload(dev, q, NULL);
1657
1658         if (q->qdiscs) {
1659                 for (i = 0; i < dev->num_tx_queues; i++)
1660                         qdisc_put(q->qdiscs[i]);
1661
1662                 kfree(q->qdiscs);
1663         }
1664         q->qdiscs = NULL;
1665
1666         netdev_reset_tc(dev);
1667
1668         if (q->oper_sched)
1669                 call_rcu(&q->oper_sched->rcu, taprio_free_sched_cb);
1670
1671         if (q->admin_sched)
1672                 call_rcu(&q->admin_sched->rcu, taprio_free_sched_cb);
1673 }
1674
1675 static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
1676                        struct netlink_ext_ack *extack)
1677 {
1678         struct taprio_sched *q = qdisc_priv(sch);
1679         struct net_device *dev = qdisc_dev(sch);
1680         int i;
1681
1682         spin_lock_init(&q->current_entry_lock);
1683
1684         hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
1685         q->advance_timer.function = advance_sched;
1686
1687         q->dequeue = taprio_dequeue_soft;
1688         q->peek = taprio_peek_soft;
1689
1690         q->root = sch;
1691
1692         /* We only support static clockids. Use an invalid value as default
1693          * and get the valid one on taprio_change().
1694          */
1695         q->clockid = -1;
1696         q->flags = TAPRIO_FLAGS_INVALID;
1697
1698         spin_lock(&taprio_list_lock);
1699         list_add(&q->taprio_list, &taprio_list);
1700         spin_unlock(&taprio_list_lock);
1701
1702         if (sch->parent != TC_H_ROOT)
1703                 return -EOPNOTSUPP;
1704
1705         if (!netif_is_multiqueue(dev))
1706                 return -EOPNOTSUPP;
1707
1708         /* pre-allocate qdisc, attachment can't fail */
1709         q->qdiscs = kcalloc(dev->num_tx_queues,
1710                             sizeof(q->qdiscs[0]),
1711                             GFP_KERNEL);
1712
1713         if (!q->qdiscs)
1714                 return -ENOMEM;
1715
1716         if (!opt)
1717                 return -EINVAL;
1718
1719         for (i = 0; i < dev->num_tx_queues; i++) {
1720                 struct netdev_queue *dev_queue;
1721                 struct Qdisc *qdisc;
1722
1723                 dev_queue = netdev_get_tx_queue(dev, i);
1724                 qdisc = qdisc_create_dflt(dev_queue,
1725                                           &pfifo_qdisc_ops,
1726                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
1727                                                     TC_H_MIN(i + 1)),
1728                                           extack);
1729                 if (!qdisc)
1730                         return -ENOMEM;
1731
1732                 if (i < dev->real_num_tx_queues)
1733                         qdisc_hash_add(qdisc, false);
1734
1735                 q->qdiscs[i] = qdisc;
1736         }
1737
1738         return taprio_change(sch, opt, extack);
1739 }
1740
1741 static void taprio_attach(struct Qdisc *sch)
1742 {
1743         struct taprio_sched *q = qdisc_priv(sch);
1744         struct net_device *dev = qdisc_dev(sch);
1745         unsigned int ntx;
1746
1747         /* Attach underlying qdisc */
1748         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
1749                 struct Qdisc *qdisc = q->qdiscs[ntx];
1750                 struct Qdisc *old;
1751
1752                 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1753                         qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1754                         old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
1755                 } else {
1756                         old = dev_graft_qdisc(qdisc->dev_queue, sch);
1757                         qdisc_refcount_inc(sch);
1758                 }
1759                 if (old)
1760                         qdisc_put(old);
1761         }
1762
1763         /* access to the child qdiscs is not needed in offload mode */
1764         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1765                 kfree(q->qdiscs);
1766                 q->qdiscs = NULL;
1767         }
1768 }
1769
1770 static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
1771                                              unsigned long cl)
1772 {
1773         struct net_device *dev = qdisc_dev(sch);
1774         unsigned long ntx = cl - 1;
1775
1776         if (ntx >= dev->num_tx_queues)
1777                 return NULL;
1778
1779         return netdev_get_tx_queue(dev, ntx);
1780 }
1781
1782 static int taprio_graft(struct Qdisc *sch, unsigned long cl,
1783                         struct Qdisc *new, struct Qdisc **old,
1784                         struct netlink_ext_ack *extack)
1785 {
1786         struct taprio_sched *q = qdisc_priv(sch);
1787         struct net_device *dev = qdisc_dev(sch);
1788         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1789
1790         if (!dev_queue)
1791                 return -EINVAL;
1792
1793         if (dev->flags & IFF_UP)
1794                 dev_deactivate(dev);
1795
1796         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1797                 *old = dev_graft_qdisc(dev_queue, new);
1798         } else {
1799                 *old = q->qdiscs[cl - 1];
1800                 q->qdiscs[cl - 1] = new;
1801         }
1802
1803         if (new)
1804                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1805
1806         if (dev->flags & IFF_UP)
1807                 dev_activate(dev);
1808
1809         return 0;
1810 }
1811
1812 static int dump_entry(struct sk_buff *msg,
1813                       const struct sched_entry *entry)
1814 {
1815         struct nlattr *item;
1816
1817         item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
1818         if (!item)
1819                 return -ENOSPC;
1820
1821         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
1822                 goto nla_put_failure;
1823
1824         if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
1825                 goto nla_put_failure;
1826
1827         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
1828                         entry->gate_mask))
1829                 goto nla_put_failure;
1830
1831         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
1832                         entry->interval))
1833                 goto nla_put_failure;
1834
1835         return nla_nest_end(msg, item);
1836
1837 nla_put_failure:
1838         nla_nest_cancel(msg, item);
1839         return -1;
1840 }
1841
1842 static int dump_schedule(struct sk_buff *msg,
1843                          const struct sched_gate_list *root)
1844 {
1845         struct nlattr *entry_list;
1846         struct sched_entry *entry;
1847
1848         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
1849                         root->base_time, TCA_TAPRIO_PAD))
1850                 return -1;
1851
1852         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
1853                         root->cycle_time, TCA_TAPRIO_PAD))
1854                 return -1;
1855
1856         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
1857                         root->cycle_time_extension, TCA_TAPRIO_PAD))
1858                 return -1;
1859
1860         entry_list = nla_nest_start_noflag(msg,
1861                                            TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
1862         if (!entry_list)
1863                 goto error_nest;
1864
1865         list_for_each_entry(entry, &root->entries, list) {
1866                 if (dump_entry(msg, entry) < 0)
1867                         goto error_nest;
1868         }
1869
1870         nla_nest_end(msg, entry_list);
1871         return 0;
1872
1873 error_nest:
1874         nla_nest_cancel(msg, entry_list);
1875         return -1;
1876 }
1877
1878 static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
1879 {
1880         struct taprio_sched *q = qdisc_priv(sch);
1881         struct net_device *dev = qdisc_dev(sch);
1882         struct sched_gate_list *oper, *admin;
1883         struct tc_mqprio_qopt opt = { 0 };
1884         struct nlattr *nest, *sched_nest;
1885         unsigned int i;
1886
1887         rcu_read_lock();
1888         oper = rcu_dereference(q->oper_sched);
1889         admin = rcu_dereference(q->admin_sched);
1890
1891         opt.num_tc = netdev_get_num_tc(dev);
1892         memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
1893
1894         for (i = 0; i < netdev_get_num_tc(dev); i++) {
1895                 opt.count[i] = dev->tc_to_txq[i].count;
1896                 opt.offset[i] = dev->tc_to_txq[i].offset;
1897         }
1898
1899         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1900         if (!nest)
1901                 goto start_error;
1902
1903         if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
1904                 goto options_error;
1905
1906         if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1907             nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
1908                 goto options_error;
1909
1910         if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
1911                 goto options_error;
1912
1913         if (q->txtime_delay &&
1914             nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
1915                 goto options_error;
1916
1917         if (oper && dump_schedule(skb, oper))
1918                 goto options_error;
1919
1920         if (!admin)
1921                 goto done;
1922
1923         sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
1924         if (!sched_nest)
1925                 goto options_error;
1926
1927         if (dump_schedule(skb, admin))
1928                 goto admin_error;
1929
1930         nla_nest_end(skb, sched_nest);
1931
1932 done:
1933         rcu_read_unlock();
1934
1935         return nla_nest_end(skb, nest);
1936
1937 admin_error:
1938         nla_nest_cancel(skb, sched_nest);
1939
1940 options_error:
1941         nla_nest_cancel(skb, nest);
1942
1943 start_error:
1944         rcu_read_unlock();
1945         return -ENOSPC;
1946 }
1947
1948 static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
1949 {
1950         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1951
1952         if (!dev_queue)
1953                 return NULL;
1954
1955         return dev_queue->qdisc_sleeping;
1956 }
1957
1958 static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
1959 {
1960         unsigned int ntx = TC_H_MIN(classid);
1961
1962         if (!taprio_queue_get(sch, ntx))
1963                 return 0;
1964         return ntx;
1965 }
1966
1967 static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
1968                              struct sk_buff *skb, struct tcmsg *tcm)
1969 {
1970         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1971
1972         tcm->tcm_parent = TC_H_ROOT;
1973         tcm->tcm_handle |= TC_H_MIN(cl);
1974         tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
1975
1976         return 0;
1977 }
1978
1979 static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
1980                                    struct gnet_dump *d)
1981         __releases(d->lock)
1982         __acquires(d->lock)
1983 {
1984         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1985
1986         sch = dev_queue->qdisc_sleeping;
1987         if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
1988             qdisc_qstats_copy(d, sch) < 0)
1989                 return -1;
1990         return 0;
1991 }
1992
1993 static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1994 {
1995         struct net_device *dev = qdisc_dev(sch);
1996         unsigned long ntx;
1997
1998         if (arg->stop)
1999                 return;
2000
2001         arg->count = arg->skip;
2002         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
2003                 if (arg->fn(sch, ntx + 1, arg) < 0) {
2004                         arg->stop = 1;
2005                         break;
2006                 }
2007                 arg->count++;
2008         }
2009 }
2010
2011 static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
2012                                                 struct tcmsg *tcm)
2013 {
2014         return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
2015 }
2016
2017 static const struct Qdisc_class_ops taprio_class_ops = {
2018         .graft          = taprio_graft,
2019         .leaf           = taprio_leaf,
2020         .find           = taprio_find,
2021         .walk           = taprio_walk,
2022         .dump           = taprio_dump_class,
2023         .dump_stats     = taprio_dump_class_stats,
2024         .select_queue   = taprio_select_queue,
2025 };
2026
2027 static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
2028         .cl_ops         = &taprio_class_ops,
2029         .id             = "taprio",
2030         .priv_size      = sizeof(struct taprio_sched),
2031         .init           = taprio_init,
2032         .change         = taprio_change,
2033         .destroy        = taprio_destroy,
2034         .reset          = taprio_reset,
2035         .attach         = taprio_attach,
2036         .peek           = taprio_peek,
2037         .dequeue        = taprio_dequeue,
2038         .enqueue        = taprio_enqueue,
2039         .dump           = taprio_dump,
2040         .owner          = THIS_MODULE,
2041 };
2042
2043 static struct notifier_block taprio_device_notifier = {
2044         .notifier_call = taprio_dev_notifier,
2045 };
2046
2047 static int __init taprio_module_init(void)
2048 {
2049         int err = register_netdevice_notifier(&taprio_device_notifier);
2050
2051         if (err)
2052                 return err;
2053
2054         return register_qdisc(&taprio_qdisc_ops);
2055 }
2056
2057 static void __exit taprio_module_exit(void)
2058 {
2059         unregister_qdisc(&taprio_qdisc_ops);
2060         unregister_netdevice_notifier(&taprio_device_notifier);
2061 }
2062
2063 module_init(taprio_module_init);
2064 module_exit(taprio_module_exit);
2065 MODULE_LICENSE("GPL");