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/ethtool_netlink.h>
  11 #include <linux/types.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/string.h>
  15 #include <linux/list.h>
  16 #include <linux/errno.h>
  17 #include <linux/skbuff.h>
  18 #include <linux/math64.h>
  19 #include <linux/module.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/time.h>
  23 #include <net/gso.h>
  24 #include <net/netlink.h>
  25 #include <net/pkt_sched.h>
  26 #include <net/pkt_cls.h>
  27 #include <net/sch_generic.h>
  28 #include <net/sock.h>
  29 #include <net/tcp.h>
  30
  31 #define TAPRIO_STAT_NOT_SET     (~0ULL)
  32
  33 #include "sch_mqprio_lib.h"
  34
  35 static LIST_HEAD(taprio_list);
  36 static struct static_key_false taprio_have_broken_mqprio;
  37 static struct static_key_false taprio_have_working_mqprio;
  38
  39 #define TAPRIO_ALL_GATES_OPEN -1
  40
  41 #define TXTIME_ASSIST_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST)
  42 #define FULL_OFFLOAD_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)
  43 #define TAPRIO_FLAGS_INVALID U32_MAX
  44
  45 struct sched_entry {
  46         /* Durations between this GCL entry and the GCL entry where the
  47          * respective traffic class gate closes
  48          */
  49         u64 gate_duration[TC_MAX_QUEUE];
  50         atomic_t budget[TC_MAX_QUEUE];
  51         /* The qdisc makes some effort so that no packet leaves
  52          * after this time
  53          */
  54         ktime_t gate_close_time[TC_MAX_QUEUE];
  55         struct list_head list;
  56         /* Used to calculate when to advance the schedule */
  57         ktime_t end_time;
  58         ktime_t next_txtime;
  59         int index;
  60         u32 gate_mask;
  61         u32 interval;
  62         u8 command;
  63 };
  64
  65 struct sched_gate_list {
  66         /* Longest non-zero contiguous gate durations per traffic class,
  67          * or 0 if a traffic class gate never opens during the schedule.
  68          */
  69         u64 max_open_gate_duration[TC_MAX_QUEUE];
  70         u32 max_frm_len[TC_MAX_QUEUE]; /* for the fast path */
  71         u32 max_sdu[TC_MAX_QUEUE]; /* for dump */
  72         struct rcu_head rcu;
  73         struct list_head entries;
  74         size_t num_entries;
  75         ktime_t cycle_end_time;
  76         s64 cycle_time;
  77         s64 cycle_time_extension;
  78         s64 base_time;
  79 };
  80
  81 struct taprio_sched {
  82         struct Qdisc **qdiscs;
  83         struct Qdisc *root;
  84         u32 flags;
  85         enum tk_offsets tk_offset;
  86         int clockid;
  87         bool offloaded;
  88         bool detected_mqprio;
  89         bool broken_mqprio;
  90         atomic64_t picos_per_byte; /* Using picoseconds because for 10Gbps+
  91                                     * speeds it's sub-nanoseconds per byte
  92                                     */
  93
  94         /* Protects the update side of the RCU protected current_entry */
  95         spinlock_t current_entry_lock;
  96         struct sched_entry __rcu *current_entry;
  97         struct sched_gate_list __rcu *oper_sched;
  98         struct sched_gate_list __rcu *admin_sched;
  99         struct hrtimer advance_timer;
 100         struct list_head taprio_list;
 101         int cur_txq[TC_MAX_QUEUE];
 102         u32 max_sdu[TC_MAX_QUEUE]; /* save info from the user */
 103         u32 fp[TC_QOPT_MAX_QUEUE]; /* only for dump and offloading */
 104         u32 txtime_delay;
 105 };
 106
 107 struct __tc_taprio_qopt_offload {
 108         refcount_t users;
 109         struct tc_taprio_qopt_offload offload;
 110 };
 111
 112 static void taprio_calculate_gate_durations(struct taprio_sched *q,
 113                                             struct sched_gate_list *sched)
 114 {
 115         struct net_device *dev = qdisc_dev(q->root);
 116         int num_tc = netdev_get_num_tc(dev);
 117         struct sched_entry *entry, *cur;
 118         int tc;
 119
 120         list_for_each_entry(entry, &sched->entries, list) {
 121                 u32 gates_still_open = entry->gate_mask;
 122
 123                 /* For each traffic class, calculate each open gate duration,
 124                  * starting at this schedule entry and ending at the schedule
 125                  * entry containing a gate close event for that TC.
 126                  */
 127                 cur = entry;
 128
 129                 do {
 130                         if (!gates_still_open)
 131                                 break;
 132
 133                         for (tc = 0; tc < num_tc; tc++) {
 134                                 if (!(gates_still_open & BIT(tc)))
 135                                         continue;
 136
 137                                 if (cur->gate_mask & BIT(tc))
 138                                         entry->gate_duration[tc] += cur->interval;
 139                                 else
 140                                         gates_still_open &= ~BIT(tc);
 141                         }
 142
 143                         cur = list_next_entry_circular(cur, &sched->entries, list);
 144                 } while (cur != entry);
 145
 146                 /* Keep track of the maximum gate duration for each traffic
 147                  * class, taking care to not confuse a traffic class which is
 148                  * temporarily closed with one that is always closed.
 149                  */
 150                 for (tc = 0; tc < num_tc; tc++)
 151                         if (entry->gate_duration[tc] &&
 152                             sched->max_open_gate_duration[tc] < entry->gate_duration[tc])
 153                                 sched->max_open_gate_duration[tc] = entry->gate_duration[tc];
 154         }
 155 }
 156
 157 static bool taprio_entry_allows_tx(ktime_t skb_end_time,
 158                                    struct sched_entry *entry, int tc)
 159 {
 160         return ktime_before(skb_end_time, entry->gate_close_time[tc]);
 161 }
 162
 163 static ktime_t sched_base_time(const struct sched_gate_list *sched)
 164 {
 165         if (!sched)
 166                 return KTIME_MAX;
 167
 168         return ns_to_ktime(sched->base_time);
 169 }
 170
 171 static ktime_t taprio_mono_to_any(const struct taprio_sched *q, ktime_t mono)
 172 {
 173         /* This pairs with WRITE_ONCE() in taprio_parse_clockid() */
 174         enum tk_offsets tk_offset = READ_ONCE(q->tk_offset);
 175
 176         switch (tk_offset) {
 177         case TK_OFFS_MAX:
 178                 return mono;
 179         default:
 180                 return ktime_mono_to_any(mono, tk_offset);
 181         }
 182 }
 183
 184 static ktime_t taprio_get_time(const struct taprio_sched *q)
 185 {
 186         return taprio_mono_to_any(q, ktime_get());
 187 }
 188
 189 static void taprio_free_sched_cb(struct rcu_head *head)
 190 {
 191         struct sched_gate_list *sched = container_of(head, struct sched_gate_list, rcu);
 192         struct sched_entry *entry, *n;
 193
 194         list_for_each_entry_safe(entry, n, &sched->entries, list) {
 195                 list_del(&entry->list);
 196                 kfree(entry);
 197         }
 198
 199         kfree(sched);
 200 }
 201
 202 static void switch_schedules(struct taprio_sched *q,
 203                              struct sched_gate_list **admin,
 204                              struct sched_gate_list **oper)
 205 {
 206         rcu_assign_pointer(q->oper_sched, *admin);
 207         rcu_assign_pointer(q->admin_sched, NULL);
 208
 209         if (*oper)
 210                 call_rcu(&(*oper)->rcu, taprio_free_sched_cb);
 211
 212         *oper = *admin;
 213         *admin = NULL;
 214 }
 215
 216 /* Get how much time has been already elapsed in the current cycle. */
 217 static s32 get_cycle_time_elapsed(struct sched_gate_list *sched, ktime_t time)
 218 {
 219         ktime_t time_since_sched_start;
 220         s32 time_elapsed;
 221
 222         time_since_sched_start = ktime_sub(time, sched->base_time);
 223         div_s64_rem(time_since_sched_start, sched->cycle_time, &time_elapsed);
 224
 225         return time_elapsed;
 226 }
 227
 228 static ktime_t get_interval_end_time(struct sched_gate_list *sched,
 229                                      struct sched_gate_list *admin,
 230                                      struct sched_entry *entry,
 231                                      ktime_t intv_start)
 232 {
 233         s32 cycle_elapsed = get_cycle_time_elapsed(sched, intv_start);
 234         ktime_t intv_end, cycle_ext_end, cycle_end;
 235
 236         cycle_end = ktime_add_ns(intv_start, sched->cycle_time - cycle_elapsed);
 237         intv_end = ktime_add_ns(intv_start, entry->interval);
 238         cycle_ext_end = ktime_add(cycle_end, sched->cycle_time_extension);
 239
 240         if (ktime_before(intv_end, cycle_end))
 241                 return intv_end;
 242         else if (admin && admin != sched &&
 243                  ktime_after(admin->base_time, cycle_end) &&
 244                  ktime_before(admin->base_time, cycle_ext_end))
 245                 return admin->base_time;
 246         else
 247                 return cycle_end;
 248 }
 249
 250 static int length_to_duration(struct taprio_sched *q, int len)
 251 {
 252         return div_u64(len * atomic64_read(&q->picos_per_byte), PSEC_PER_NSEC);
 253 }
 254
 255 static int duration_to_length(struct taprio_sched *q, u64 duration)
 256 {
 257         return div_u64(duration * PSEC_PER_NSEC, atomic64_read(&q->picos_per_byte));
 258 }
 259
 260 /* Sets sched->max_sdu[] and sched->max_frm_len[] to the minimum between the
 261  * q->max_sdu[] requested by the user and the max_sdu dynamically determined by
 262  * the maximum open gate durations at the given link speed.
 263  */
 264 static void taprio_update_queue_max_sdu(struct taprio_sched *q,
 265                                         struct sched_gate_list *sched,
 266                                         struct qdisc_size_table *stab)
 267 {
 268         struct net_device *dev = qdisc_dev(q->root);
 269         int num_tc = netdev_get_num_tc(dev);
 270         u32 max_sdu_from_user;
 271         u32 max_sdu_dynamic;
 272         u32 max_sdu;
 273         int tc;
 274
 275         for (tc = 0; tc < num_tc; tc++) {
 276                 max_sdu_from_user = q->max_sdu[tc] ?: U32_MAX;
 277
 278                 /* TC gate never closes => keep the queueMaxSDU
 279                  * selected by the user
 280                  */
 281                 if (sched->max_open_gate_duration[tc] == sched->cycle_time) {
 282                         max_sdu_dynamic = U32_MAX;
 283                 } else {
 284                         u32 max_frm_len;
 285
 286                         max_frm_len = duration_to_length(q, sched->max_open_gate_duration[tc]);
 287                         /* Compensate for L1 overhead from size table,
 288                          * but don't let the frame size go negative
 289                          */
 290                         if (stab) {
 291                                 max_frm_len -= stab->szopts.overhead;
 292                                 max_frm_len = max_t(int, max_frm_len,
 293                                                     dev->hard_header_len + 1);
 294                         }
 295                         max_sdu_dynamic = max_frm_len - dev->hard_header_len;
 296                         if (max_sdu_dynamic > dev->max_mtu)
 297                                 max_sdu_dynamic = U32_MAX;
 298                 }
 299
 300                 max_sdu = min(max_sdu_dynamic, max_sdu_from_user);
 301
 302                 if (max_sdu != U32_MAX) {
 303                         sched->max_frm_len[tc] = max_sdu + dev->hard_header_len;
 304                         sched->max_sdu[tc] = max_sdu;
 305                 } else {
 306                         sched->max_frm_len[tc] = U32_MAX; /* never oversized */
 307                         sched->max_sdu[tc] = 0;
 308                 }
 309         }
 310 }
 311
 312 /* Returns the entry corresponding to next available interval. If
 313  * validate_interval is set, it only validates whether the timestamp occurs
 314  * when the gate corresponding to the skb's traffic class is open.
 315  */
 316 static struct sched_entry *find_entry_to_transmit(struct sk_buff *skb,
 317                                                   struct Qdisc *sch,
 318                                                   struct sched_gate_list *sched,
 319                                                   struct sched_gate_list *admin,
 320                                                   ktime_t time,
 321                                                   ktime_t *interval_start,
 322                                                   ktime_t *interval_end,
 323                                                   bool validate_interval)
 324 {
 325         ktime_t curr_intv_start, curr_intv_end, cycle_end, packet_transmit_time;
 326         ktime_t earliest_txtime = KTIME_MAX, txtime, cycle, transmit_end_time;
 327         struct sched_entry *entry = NULL, *entry_found = NULL;
 328         struct taprio_sched *q = qdisc_priv(sch);
 329         struct net_device *dev = qdisc_dev(sch);
 330         bool entry_available = false;
 331         s32 cycle_elapsed;
 332         int tc, n;
 333
 334         tc = netdev_get_prio_tc_map(dev, skb->priority);
 335         packet_transmit_time = length_to_duration(q, qdisc_pkt_len(skb));
 336
 337         *interval_start = 0;
 338         *interval_end = 0;
 339
 340         if (!sched)
 341                 return NULL;
 342
 343         cycle = sched->cycle_time;
 344         cycle_elapsed = get_cycle_time_elapsed(sched, time);
 345         curr_intv_end = ktime_sub_ns(time, cycle_elapsed);
 346         cycle_end = ktime_add_ns(curr_intv_end, cycle);
 347
 348         list_for_each_entry(entry, &sched->entries, list) {
 349                 curr_intv_start = curr_intv_end;
 350                 curr_intv_end = get_interval_end_time(sched, admin, entry,
 351                                                       curr_intv_start);
 352
 353                 if (ktime_after(curr_intv_start, cycle_end))
 354                         break;
 355
 356                 if (!(entry->gate_mask & BIT(tc)) ||
 357                     packet_transmit_time > entry->interval)
 358                         continue;
 359
 360                 txtime = entry->next_txtime;
 361
 362                 if (ktime_before(txtime, time) || validate_interval) {
 363                         transmit_end_time = ktime_add_ns(time, packet_transmit_time);
 364                         if ((ktime_before(curr_intv_start, time) &&
 365                              ktime_before(transmit_end_time, curr_intv_end)) ||
 366                             (ktime_after(curr_intv_start, time) && !validate_interval)) {
 367                                 entry_found = entry;
 368                                 *interval_start = curr_intv_start;
 369                                 *interval_end = curr_intv_end;
 370                                 break;
 371                         } else if (!entry_available && !validate_interval) {
 372                                 /* Here, we are just trying to find out the
 373                                  * first available interval in the next cycle.
 374                                  */
 375                                 entry_available = true;
 376                                 entry_found = entry;
 377                                 *interval_start = ktime_add_ns(curr_intv_start, cycle);
 378                                 *interval_end = ktime_add_ns(curr_intv_end, cycle);
 379                         }
 380                 } else if (ktime_before(txtime, earliest_txtime) &&
 381                            !entry_available) {
 382                         earliest_txtime = txtime;
 383                         entry_found = entry;
 384                         n = div_s64(ktime_sub(txtime, curr_intv_start), cycle);
 385                         *interval_start = ktime_add(curr_intv_start, n * cycle);
 386                         *interval_end = ktime_add(curr_intv_end, n * cycle);
 387                 }
 388         }
 389
 390         return entry_found;
 391 }
 392
 393 static bool is_valid_interval(struct sk_buff *skb, struct Qdisc *sch)
 394 {
 395         struct taprio_sched *q = qdisc_priv(sch);
 396         struct sched_gate_list *sched, *admin;
 397         ktime_t interval_start, interval_end;
 398         struct sched_entry *entry;
 399
 400         rcu_read_lock();
 401         sched = rcu_dereference(q->oper_sched);
 402         admin = rcu_dereference(q->admin_sched);
 403
 404         entry = find_entry_to_transmit(skb, sch, sched, admin, skb->tstamp,
 405                                        &interval_start, &interval_end, true);
 406         rcu_read_unlock();
 407
 408         return entry;
 409 }
 410
 411 static bool taprio_flags_valid(u32 flags)
 412 {
 413         /* Make sure no other flag bits are set. */
 414         if (flags & ~(TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST |
 415                       TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 416                 return false;
 417         /* txtime-assist and full offload are mutually exclusive */
 418         if ((flags & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST) &&
 419             (flags & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 420                 return false;
 421         return true;
 422 }
 423
 424 /* This returns the tstamp value set by TCP in terms of the set clock. */
 425 static ktime_t get_tcp_tstamp(struct taprio_sched *q, struct sk_buff *skb)
 426 {
 427         unsigned int offset = skb_network_offset(skb);
 428         const struct ipv6hdr *ipv6h;
 429         const struct iphdr *iph;
 430         struct ipv6hdr _ipv6h;
 431
 432         ipv6h = skb_header_pointer(skb, offset, sizeof(_ipv6h), &_ipv6h);
 433         if (!ipv6h)
 434                 return 0;
 435
 436         if (ipv6h->version == 4) {
 437                 iph = (struct iphdr *)ipv6h;
 438                 offset += iph->ihl * 4;
 439
 440                 /* special-case 6in4 tunnelling, as that is a common way to get
 441                  * v6 connectivity in the home
 442                  */
 443                 if (iph->protocol == IPPROTO_IPV6) {
 444                         ipv6h = skb_header_pointer(skb, offset,
 445                                                    sizeof(_ipv6h), &_ipv6h);
 446
 447                         if (!ipv6h || ipv6h->nexthdr != IPPROTO_TCP)
 448                                 return 0;
 449                 } else if (iph->protocol != IPPROTO_TCP) {
 450                         return 0;
 451                 }
 452         } else if (ipv6h->version == 6 && ipv6h->nexthdr != IPPROTO_TCP) {
 453                 return 0;
 454         }
 455
 456         return taprio_mono_to_any(q, skb->skb_mstamp_ns);
 457 }
 458
 459 /* There are a few scenarios where we will have to modify the txtime from
 460  * what is read from next_txtime in sched_entry. They are:
 461  * 1. If txtime is in the past,
 462  *    a. The gate for the traffic class is currently open and packet can be
 463  *       transmitted before it closes, schedule the packet right away.
 464  *    b. If the gate corresponding to the traffic class is going to open later
 465  *       in the cycle, set the txtime of packet to the interval start.
 466  * 2. If txtime is in the future, there are packets corresponding to the
 467  *    current traffic class waiting to be transmitted. So, the following
 468  *    possibilities exist:
 469  *    a. We can transmit the packet before the window containing the txtime
 470  *       closes.
 471  *    b. The window might close before the transmission can be completed
 472  *       successfully. So, schedule the packet in the next open window.
 473  */
 474 static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)
 475 {
 476         ktime_t transmit_end_time, interval_end, interval_start, tcp_tstamp;
 477         struct taprio_sched *q = qdisc_priv(sch);
 478         struct sched_gate_list *sched, *admin;
 479         ktime_t minimum_time, now, txtime;
 480         int len, packet_transmit_time;
 481         struct sched_entry *entry;
 482         bool sched_changed;
 483
 484         now = taprio_get_time(q);
 485         minimum_time = ktime_add_ns(now, q->txtime_delay);
 486
 487         tcp_tstamp = get_tcp_tstamp(q, skb);
 488         minimum_time = max_t(ktime_t, minimum_time, tcp_tstamp);
 489
 490         rcu_read_lock();
 491         admin = rcu_dereference(q->admin_sched);
 492         sched = rcu_dereference(q->oper_sched);
 493         if (admin && ktime_after(minimum_time, admin->base_time))
 494                 switch_schedules(q, &admin, &sched);
 495
 496         /* Until the schedule starts, all the queues are open */
 497         if (!sched || ktime_before(minimum_time, sched->base_time)) {
 498                 txtime = minimum_time;
 499                 goto done;
 500         }
 501
 502         len = qdisc_pkt_len(skb);
 503         packet_transmit_time = length_to_duration(q, len);
 504
 505         do {
 506                 sched_changed = false;
 507
 508                 entry = find_entry_to_transmit(skb, sch, sched, admin,
 509                                                minimum_time,
 510                                                &interval_start, &interval_end,
 511                                                false);
 512                 if (!entry) {
 513                         txtime = 0;
 514                         goto done;
 515                 }
 516
 517                 txtime = entry->next_txtime;
 518                 txtime = max_t(ktime_t, txtime, minimum_time);
 519                 txtime = max_t(ktime_t, txtime, interval_start);
 520
 521                 if (admin && admin != sched &&
 522                     ktime_after(txtime, admin->base_time)) {
 523                         sched = admin;
 524                         sched_changed = true;
 525                         continue;
 526                 }
 527
 528                 transmit_end_time = ktime_add(txtime, packet_transmit_time);
 529                 minimum_time = transmit_end_time;
 530
 531                 /* Update the txtime of current entry to the next time it's
 532                  * interval starts.
 533                  */
 534                 if (ktime_after(transmit_end_time, interval_end))
 535                         entry->next_txtime = ktime_add(interval_start, sched->cycle_time);
 536         } while (sched_changed || ktime_after(transmit_end_time, interval_end));
 537
 538         entry->next_txtime = transmit_end_time;
 539
 540 done:
 541         rcu_read_unlock();
 542         return txtime;
 543 }
 544
 545 /* Devices with full offload are expected to honor this in hardware */
 546 static bool taprio_skb_exceeds_queue_max_sdu(struct Qdisc *sch,
 547                                              struct sk_buff *skb)
 548 {
 549         struct taprio_sched *q = qdisc_priv(sch);
 550         struct net_device *dev = qdisc_dev(sch);
 551         struct sched_gate_list *sched;
 552         int prio = skb->priority;
 553         bool exceeds = false;
 554         u8 tc;
 555
 556         tc = netdev_get_prio_tc_map(dev, prio);
 557
 558         rcu_read_lock();
 559         sched = rcu_dereference(q->oper_sched);
 560         if (sched && skb->len > sched->max_frm_len[tc])
 561                 exceeds = true;
 562         rcu_read_unlock();
 563
 564         return exceeds;
 565 }
 566
 567 static int taprio_enqueue_one(struct sk_buff *skb, struct Qdisc *sch,
 568                               struct Qdisc *child, struct sk_buff **to_free)
 569 {
 570         struct taprio_sched *q = qdisc_priv(sch);
 571
 572         /* sk_flags are only safe to use on full sockets. */
 573         if (skb->sk && sk_fullsock(skb->sk) && sock_flag(skb->sk, SOCK_TXTIME)) {
 574                 if (!is_valid_interval(skb, sch))
 575                         return qdisc_drop(skb, sch, to_free);
 576         } else if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
 577                 skb->tstamp = get_packet_txtime(skb, sch);
 578                 if (!skb->tstamp)
 579                         return qdisc_drop(skb, sch, to_free);
 580         }
 581
 582         qdisc_qstats_backlog_inc(sch, skb);
 583         sch->q.qlen++;
 584
 585         return qdisc_enqueue(skb, child, to_free);
 586 }
 587
 588 static int taprio_enqueue_segmented(struct sk_buff *skb, struct Qdisc *sch,
 589                                     struct Qdisc *child,
 590                                     struct sk_buff **to_free)
 591 {
 592         unsigned int slen = 0, numsegs = 0, len = qdisc_pkt_len(skb);
 593         netdev_features_t features = netif_skb_features(skb);
 594         struct sk_buff *segs, *nskb;
 595         int ret;
 596
 597         segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 598         if (IS_ERR_OR_NULL(segs))
 599                 return qdisc_drop(skb, sch, to_free);
 600
 601         skb_list_walk_safe(segs, segs, nskb) {
 602                 skb_mark_not_on_list(segs);
 603                 qdisc_skb_cb(segs)->pkt_len = segs->len;
 604                 slen += segs->len;
 605
 606                 /* FIXME: we should be segmenting to a smaller size
 607                  * rather than dropping these
 608                  */
 609                 if (taprio_skb_exceeds_queue_max_sdu(sch, segs))
 610                         ret = qdisc_drop(segs, sch, to_free);
 611                 else
 612                         ret = taprio_enqueue_one(segs, sch, child, to_free);
 613
 614                 if (ret != NET_XMIT_SUCCESS) {
 615                         if (net_xmit_drop_count(ret))
 616                                 qdisc_qstats_drop(sch);
 617                 } else {
 618                         numsegs++;
 619                 }
 620         }
 621
 622         if (numsegs > 1)
 623                 qdisc_tree_reduce_backlog(sch, 1 - numsegs, len - slen);
 624         consume_skb(skb);
 625
 626         return numsegs > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
 627 }
 628
 629 /* Will not be called in the full offload case, since the TX queues are
 630  * attached to the Qdisc created using qdisc_create_dflt()
 631  */
 632 static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 633                           struct sk_buff **to_free)
 634 {
 635         struct taprio_sched *q = qdisc_priv(sch);
 636         struct Qdisc *child;
 637         int queue;
 638
 639         queue = skb_get_queue_mapping(skb);
 640
 641         child = q->qdiscs[queue];
 642         if (unlikely(!child))
 643                 return qdisc_drop(skb, sch, to_free);
 644
 645         if (taprio_skb_exceeds_queue_max_sdu(sch, skb)) {
 646                 /* Large packets might not be transmitted when the transmission
 647                  * duration exceeds any configured interval. Therefore, segment
 648                  * the skb into smaller chunks. Drivers with full offload are
 649                  * expected to handle this in hardware.
 650                  */
 651                 if (skb_is_gso(skb))
 652                         return taprio_enqueue_segmented(skb, sch, child,
 653                                                         to_free);
 654
 655                 return qdisc_drop(skb, sch, to_free);
 656         }
 657
 658         return taprio_enqueue_one(skb, sch, child, to_free);
 659 }
 660
 661 static struct sk_buff *taprio_peek(struct Qdisc *sch)
 662 {
 663         WARN_ONCE(1, "taprio only supports operating as root qdisc, peek() not implemented");
 664         return NULL;
 665 }
 666
 667 static void taprio_set_budgets(struct taprio_sched *q,
 668                                struct sched_gate_list *sched,
 669                                struct sched_entry *entry)
 670 {
 671         struct net_device *dev = qdisc_dev(q->root);
 672         int num_tc = netdev_get_num_tc(dev);
 673         int tc, budget;
 674
 675         for (tc = 0; tc < num_tc; tc++) {
 676                 /* Traffic classes which never close have infinite budget */
 677                 if (entry->gate_duration[tc] == sched->cycle_time)
 678                         budget = INT_MAX;
 679                 else
 680                         budget = div64_u64((u64)entry->gate_duration[tc] * PSEC_PER_NSEC,
 681                                            atomic64_read(&q->picos_per_byte));
 682
 683                 atomic_set(&entry->budget[tc], budget);
 684         }
 685 }
 686
 687 /* When an skb is sent, it consumes from the budget of all traffic classes */
 688 static int taprio_update_budgets(struct sched_entry *entry, size_t len,
 689                                  int tc_consumed, int num_tc)
 690 {
 691         int tc, budget, new_budget = 0;
 692
 693         for (tc = 0; tc < num_tc; tc++) {
 694                 budget = atomic_read(&entry->budget[tc]);
 695                 /* Don't consume from infinite budget */
 696                 if (budget == INT_MAX) {
 697                         if (tc == tc_consumed)
 698                                 new_budget = budget;
 699                         continue;
 700                 }
 701
 702                 if (tc == tc_consumed)
 703                         new_budget = atomic_sub_return(len, &entry->budget[tc]);
 704                 else
 705                         atomic_sub(len, &entry->budget[tc]);
 706         }
 707
 708         return new_budget;
 709 }
 710
 711 static struct sk_buff *taprio_dequeue_from_txq(struct Qdisc *sch, int txq,
 712                                                struct sched_entry *entry,
 713                                                u32 gate_mask)
 714 {
 715         struct taprio_sched *q = qdisc_priv(sch);
 716         struct net_device *dev = qdisc_dev(sch);
 717         struct Qdisc *child = q->qdiscs[txq];
 718         int num_tc = netdev_get_num_tc(dev);
 719         struct sk_buff *skb;
 720         ktime_t guard;
 721         int prio;
 722         int len;
 723         u8 tc;
 724
 725         if (unlikely(!child))
 726                 return NULL;
 727
 728         if (TXTIME_ASSIST_IS_ENABLED(q->flags))
 729                 goto skip_peek_checks;
 730
 731         skb = child->ops->peek(child);
 732         if (!skb)
 733                 return NULL;
 734
 735         prio = skb->priority;
 736         tc = netdev_get_prio_tc_map(dev, prio);
 737
 738         if (!(gate_mask & BIT(tc)))
 739                 return NULL;
 740
 741         len = qdisc_pkt_len(skb);
 742         guard = ktime_add_ns(taprio_get_time(q), length_to_duration(q, len));
 743
 744         /* In the case that there's no gate entry, there's no
 745          * guard band ...
 746          */
 747         if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 748             !taprio_entry_allows_tx(guard, entry, tc))
 749                 return NULL;
 750
 751         /* ... and no budget. */
 752         if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 753             taprio_update_budgets(entry, len, tc, num_tc) < 0)
 754                 return NULL;
 755
 756 skip_peek_checks:
 757         skb = child->ops->dequeue(child);
 758         if (unlikely(!skb))
 759                 return NULL;
 760
 761         qdisc_bstats_update(sch, skb);
 762         qdisc_qstats_backlog_dec(sch, skb);
 763         sch->q.qlen--;
 764
 765         return skb;
 766 }
 767
 768 static void taprio_next_tc_txq(struct net_device *dev, int tc, int *txq)
 769 {
 770         int offset = dev->tc_to_txq[tc].offset;
 771         int count = dev->tc_to_txq[tc].count;
 772
 773         (*txq)++;
 774         if (*txq == offset + count)
 775                 *txq = offset;
 776 }
 777
 778 /* Prioritize higher traffic classes, and select among TXQs belonging to the
 779  * same TC using round robin
 780  */
 781 static struct sk_buff *taprio_dequeue_tc_priority(struct Qdisc *sch,
 782                                                   struct sched_entry *entry,
 783                                                   u32 gate_mask)
 784 {
 785         struct taprio_sched *q = qdisc_priv(sch);
 786         struct net_device *dev = qdisc_dev(sch);
 787         int num_tc = netdev_get_num_tc(dev);
 788         struct sk_buff *skb;
 789         int tc;
 790
 791         for (tc = num_tc - 1; tc >= 0; tc--) {
 792                 int first_txq = q->cur_txq[tc];
 793
 794                 if (!(gate_mask & BIT(tc)))
 795                         continue;
 796
 797                 do {
 798                         skb = taprio_dequeue_from_txq(sch, q->cur_txq[tc],
 799                                                       entry, gate_mask);
 800
 801                         taprio_next_tc_txq(dev, tc, &q->cur_txq[tc]);
 802
 803                         if (q->cur_txq[tc] >= dev->num_tx_queues)
 804                                 q->cur_txq[tc] = first_txq;
 805
 806                         if (skb)
 807                                 return skb;
 808                 } while (q->cur_txq[tc] != first_txq);
 809         }
 810
 811         return NULL;
 812 }
 813
 814 /* Broken way of prioritizing smaller TXQ indices and ignoring the traffic
 815  * class other than to determine whether the gate is open or not
 816  */
 817 static struct sk_buff *taprio_dequeue_txq_priority(struct Qdisc *sch,
 818                                                    struct sched_entry *entry,
 819                                                    u32 gate_mask)
 820 {
 821         struct net_device *dev = qdisc_dev(sch);
 822         struct sk_buff *skb;
 823         int i;
 824
 825         for (i = 0; i < dev->num_tx_queues; i++) {
 826                 skb = taprio_dequeue_from_txq(sch, i, entry, gate_mask);
 827                 if (skb)
 828                         return skb;
 829         }
 830
 831         return NULL;
 832 }
 833
 834 /* Will not be called in the full offload case, since the TX queues are
 835  * attached to the Qdisc created using qdisc_create_dflt()
 836  */
 837 static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
 838 {
 839         struct taprio_sched *q = qdisc_priv(sch);
 840         struct sk_buff *skb = NULL;
 841         struct sched_entry *entry;
 842         u32 gate_mask;
 843
 844         rcu_read_lock();
 845         entry = rcu_dereference(q->current_entry);
 846         /* if there's no entry, it means that the schedule didn't
 847          * start yet, so force all gates to be open, this is in
 848          * accordance to IEEE 802.1Qbv-2015 Section 8.6.9.4.5
 849          * "AdminGateStates"
 850          */
 851         gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
 852         if (!gate_mask)
 853                 goto done;
 854
 855         if (static_branch_unlikely(&taprio_have_broken_mqprio) &&
 856             !static_branch_likely(&taprio_have_working_mqprio)) {
 857                 /* Single NIC kind which is broken */
 858                 skb = taprio_dequeue_txq_priority(sch, entry, gate_mask);
 859         } else if (static_branch_likely(&taprio_have_working_mqprio) &&
 860                    !static_branch_unlikely(&taprio_have_broken_mqprio)) {
 861                 /* Single NIC kind which prioritizes properly */
 862                 skb = taprio_dequeue_tc_priority(sch, entry, gate_mask);
 863         } else {
 864                 /* Mixed NIC kinds present in system, need dynamic testing */
 865                 if (q->broken_mqprio)
 866                         skb = taprio_dequeue_txq_priority(sch, entry, gate_mask);
 867                 else
 868                         skb = taprio_dequeue_tc_priority(sch, entry, gate_mask);
 869         }
 870
 871 done:
 872         rcu_read_unlock();
 873
 874         return skb;
 875 }
 876
 877 static bool should_restart_cycle(const struct sched_gate_list *oper,
 878                                  const struct sched_entry *entry)
 879 {
 880         if (list_is_last(&entry->list, &oper->entries))
 881                 return true;
 882
 883         if (ktime_compare(entry->end_time, oper->cycle_end_time) == 0)
 884                 return true;
 885
 886         return false;
 887 }
 888
 889 static bool should_change_schedules(const struct sched_gate_list *admin,
 890                                     const struct sched_gate_list *oper,
 891                                     ktime_t end_time)
 892 {
 893         ktime_t next_base_time, extension_time;
 894
 895         if (!admin)
 896                 return false;
 897
 898         next_base_time = sched_base_time(admin);
 899
 900         /* This is the simple case, the end_time would fall after
 901          * the next schedule base_time.
 902          */
 903         if (ktime_compare(next_base_time, end_time) <= 0)
 904                 return true;
 905
 906         /* This is the cycle_time_extension case, if the end_time
 907          * plus the amount that can be extended would fall after the
 908          * next schedule base_time, we can extend the current schedule
 909          * for that amount.
 910          */
 911         extension_time = ktime_add_ns(end_time, oper->cycle_time_extension);
 912
 913         /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
 914          * how precisely the extension should be made. So after
 915          * conformance testing, this logic may change.
 916          */
 917         if (ktime_compare(next_base_time, extension_time) <= 0)
 918                 return true;
 919
 920         return false;
 921 }
 922
 923 static enum hrtimer_restart advance_sched(struct hrtimer *timer)
 924 {
 925         struct taprio_sched *q = container_of(timer, struct taprio_sched,
 926                                               advance_timer);
 927         struct net_device *dev = qdisc_dev(q->root);
 928         struct sched_gate_list *oper, *admin;
 929         int num_tc = netdev_get_num_tc(dev);
 930         struct sched_entry *entry, *next;
 931         struct Qdisc *sch = q->root;
 932         ktime_t end_time;
 933         int tc;
 934
 935         spin_lock(&q->current_entry_lock);
 936         entry = rcu_dereference_protected(q->current_entry,
 937                                           lockdep_is_held(&q->current_entry_lock));
 938         oper = rcu_dereference_protected(q->oper_sched,
 939                                          lockdep_is_held(&q->current_entry_lock));
 940         admin = rcu_dereference_protected(q->admin_sched,
 941                                           lockdep_is_held(&q->current_entry_lock));
 942
 943         if (!oper)
 944                 switch_schedules(q, &admin, &oper);
 945
 946         /* This can happen in two cases: 1. this is the very first run
 947          * of this function (i.e. we weren't running any schedule
 948          * previously); 2. The previous schedule just ended. The first
 949          * entry of all schedules are pre-calculated during the
 950          * schedule initialization.
 951          */
 952         if (unlikely(!entry || entry->end_time == oper->base_time)) {
 953                 next = list_first_entry(&oper->entries, struct sched_entry,
 954                                         list);
 955                 end_time = next->end_time;
 956                 goto first_run;
 957         }
 958
 959         if (should_restart_cycle(oper, entry)) {
 960                 next = list_first_entry(&oper->entries, struct sched_entry,
 961                                         list);
 962                 oper->cycle_end_time = ktime_add_ns(oper->cycle_end_time,
 963                                                     oper->cycle_time);
 964         } else {
 965                 next = list_next_entry(entry, list);
 966         }
 967
 968         end_time = ktime_add_ns(entry->end_time, next->interval);
 969         end_time = min_t(ktime_t, end_time, oper->cycle_end_time);
 970
 971         for (tc = 0; tc < num_tc; tc++) {
 972                 if (next->gate_duration[tc] == oper->cycle_time)
 973                         next->gate_close_time[tc] = KTIME_MAX;
 974                 else
 975                         next->gate_close_time[tc] = ktime_add_ns(entry->end_time,
 976                                                                  next->gate_duration[tc]);
 977         }
 978
 979         if (should_change_schedules(admin, oper, end_time)) {
 980                 /* Set things so the next time this runs, the new
 981                  * schedule runs.
 982                  */
 983                 end_time = sched_base_time(admin);
 984                 switch_schedules(q, &admin, &oper);
 985         }
 986
 987         next->end_time = end_time;
 988         taprio_set_budgets(q, oper, next);
 989
 990 first_run:
 991         rcu_assign_pointer(q->current_entry, next);
 992         spin_unlock(&q->current_entry_lock);
 993
 994         hrtimer_set_expires(&q->advance_timer, end_time);
 995
 996         rcu_read_lock();
 997         __netif_schedule(sch);
 998         rcu_read_unlock();
 999
1000         return HRTIMER_RESTART;
1001 }
1002
1003 static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
1004         [TCA_TAPRIO_SCHED_ENTRY_INDEX]     = { .type = NLA_U32 },
1005         [TCA_TAPRIO_SCHED_ENTRY_CMD]       = { .type = NLA_U8 },
1006         [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
1007         [TCA_TAPRIO_SCHED_ENTRY_INTERVAL]  = { .type = NLA_U32 },
1008 };
1009
1010 static const struct nla_policy taprio_tc_policy[TCA_TAPRIO_TC_ENTRY_MAX + 1] = {
1011         [TCA_TAPRIO_TC_ENTRY_INDEX]        = { .type = NLA_U32 },
1012         [TCA_TAPRIO_TC_ENTRY_MAX_SDU]      = { .type = NLA_U32 },
1013         [TCA_TAPRIO_TC_ENTRY_FP]           = NLA_POLICY_RANGE(NLA_U32,
1014                                                               TC_FP_EXPRESS,
1015                                                               TC_FP_PREEMPTIBLE),
1016 };
1017
1018 static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
1019         [TCA_TAPRIO_ATTR_PRIOMAP]              = {
1020                 .len = sizeof(struct tc_mqprio_qopt)
1021         },
1022         [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST]           = { .type = NLA_NESTED },
1023         [TCA_TAPRIO_ATTR_SCHED_BASE_TIME]            = { .type = NLA_S64 },
1024         [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]         = { .type = NLA_NESTED },
1025         [TCA_TAPRIO_ATTR_SCHED_CLOCKID]              = { .type = NLA_S32 },
1026         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]           = { .type = NLA_S64 },
1027         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
1028         [TCA_TAPRIO_ATTR_FLAGS]                      = { .type = NLA_U32 },
1029         [TCA_TAPRIO_ATTR_TXTIME_DELAY]               = { .type = NLA_U32 },
1030         [TCA_TAPRIO_ATTR_TC_ENTRY]                   = { .type = NLA_NESTED },
1031 };
1032
1033 static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
1034                             struct sched_entry *entry,
1035                             struct netlink_ext_ack *extack)
1036 {
1037         int min_duration = length_to_duration(q, ETH_ZLEN);
1038         u32 interval = 0;
1039
1040         if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
1041                 entry->command = nla_get_u8(
1042                         tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
1043
1044         if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
1045                 entry->gate_mask = nla_get_u32(
1046                         tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
1047
1048         if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
1049                 interval = nla_get_u32(
1050                         tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
1051
1052         /* The interval should allow at least the minimum ethernet
1053          * frame to go out.
1054          */
1055         if (interval < min_duration) {
1056                 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
1057                 return -EINVAL;
1058         }
1059
1060         entry->interval = interval;
1061
1062         return 0;
1063 }
1064
1065 static int parse_sched_entry(struct taprio_sched *q, struct nlattr *n,
1066                              struct sched_entry *entry, int index,
1067                              struct netlink_ext_ack *extack)
1068 {
1069         struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
1070         int err;
1071
1072         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
1073                                           entry_policy, NULL);
1074         if (err < 0) {
1075                 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
1076                 return -EINVAL;
1077         }
1078
1079         entry->index = index;
1080
1081         return fill_sched_entry(q, tb, entry, extack);
1082 }
1083
1084 static int parse_sched_list(struct taprio_sched *q, struct nlattr *list,
1085                             struct sched_gate_list *sched,
1086                             struct netlink_ext_ack *extack)
1087 {
1088         struct nlattr *n;
1089         int err, rem;
1090         int i = 0;
1091
1092         if (!list)
1093                 return -EINVAL;
1094
1095         nla_for_each_nested(n, list, rem) {
1096                 struct sched_entry *entry;
1097
1098                 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
1099                         NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
1100                         continue;
1101                 }
1102
1103                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1104                 if (!entry) {
1105                         NL_SET_ERR_MSG(extack, "Not enough memory for entry");
1106                         return -ENOMEM;
1107                 }
1108
1109                 err = parse_sched_entry(q, n, entry, i, extack);
1110                 if (err < 0) {
1111                         kfree(entry);
1112                         return err;
1113                 }
1114
1115                 list_add_tail(&entry->list, &sched->entries);
1116                 i++;
1117         }
1118
1119         sched->num_entries = i;
1120
1121         return i;
1122 }
1123
1124 static int parse_taprio_schedule(struct taprio_sched *q, struct nlattr **tb,
1125                                  struct sched_gate_list *new,
1126                                  struct netlink_ext_ack *extack)
1127 {
1128         int err = 0;
1129
1130         if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
1131                 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
1132                 return -ENOTSUPP;
1133         }
1134
1135         if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
1136                 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
1137
1138         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
1139                 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
1140
1141         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
1142                 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
1143
1144         if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
1145                 err = parse_sched_list(q, tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST],
1146                                        new, extack);
1147         if (err < 0)
1148                 return err;
1149
1150         if (!new->cycle_time) {
1151                 struct sched_entry *entry;
1152                 ktime_t cycle = 0;
1153
1154                 list_for_each_entry(entry, &new->entries, list)
1155                         cycle = ktime_add_ns(cycle, entry->interval);
1156
1157                 if (!cycle) {
1158                         NL_SET_ERR_MSG(extack, "'cycle_time' can never be 0");
1159                         return -EINVAL;
1160                 }
1161
1162                 new->cycle_time = cycle;
1163         }
1164
1165         taprio_calculate_gate_durations(q, new);
1166
1167         return 0;
1168 }
1169
1170 static int taprio_parse_mqprio_opt(struct net_device *dev,
1171                                    struct tc_mqprio_qopt *qopt,
1172                                    struct netlink_ext_ack *extack,
1173                                    u32 taprio_flags)
1174 {
1175         bool allow_overlapping_txqs = TXTIME_ASSIST_IS_ENABLED(taprio_flags);
1176
1177         if (!qopt && !dev->num_tc) {
1178                 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
1179                 return -EINVAL;
1180         }
1181
1182         /* If num_tc is already set, it means that the user already
1183          * configured the mqprio part
1184          */
1185         if (dev->num_tc)
1186                 return 0;
1187
1188         /* taprio imposes that traffic classes map 1:n to tx queues */
1189         if (qopt->num_tc > dev->num_tx_queues) {
1190                 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
1191                 return -EINVAL;
1192         }
1193
1194         /* For some reason, in txtime-assist mode, we allow TXQ ranges for
1195          * different TCs to overlap, and just validate the TXQ ranges.
1196          */
1197         return mqprio_validate_qopt(dev, qopt, true, allow_overlapping_txqs,
1198                                     extack);
1199 }
1200
1201 static int taprio_get_start_time(struct Qdisc *sch,
1202                                  struct sched_gate_list *sched,
1203                                  ktime_t *start)
1204 {
1205         struct taprio_sched *q = qdisc_priv(sch);
1206         ktime_t now, base, cycle;
1207         s64 n;
1208
1209         base = sched_base_time(sched);
1210         now = taprio_get_time(q);
1211
1212         if (ktime_after(base, now)) {
1213                 *start = base;
1214                 return 0;
1215         }
1216
1217         cycle = sched->cycle_time;
1218
1219         /* The qdisc is expected to have at least one sched_entry.  Moreover,
1220          * any entry must have 'interval' > 0. Thus if the cycle time is zero,
1221          * something went really wrong. In that case, we should warn about this
1222          * inconsistent state and return error.
1223          */
1224         if (WARN_ON(!cycle))
1225                 return -EFAULT;
1226
1227         /* Schedule the start time for the beginning of the next
1228          * cycle.
1229          */
1230         n = div64_s64(ktime_sub_ns(now, base), cycle);
1231         *start = ktime_add_ns(base, (n + 1) * cycle);
1232         return 0;
1233 }
1234
1235 static void setup_first_end_time(struct taprio_sched *q,
1236                                  struct sched_gate_list *sched, ktime_t base)
1237 {
1238         struct net_device *dev = qdisc_dev(q->root);
1239         int num_tc = netdev_get_num_tc(dev);
1240         struct sched_entry *first;
1241         ktime_t cycle;
1242         int tc;
1243
1244         first = list_first_entry(&sched->entries,
1245                                  struct sched_entry, list);
1246
1247         cycle = sched->cycle_time;
1248
1249         /* FIXME: find a better place to do this */
1250         sched->cycle_end_time = ktime_add_ns(base, cycle);
1251
1252         first->end_time = ktime_add_ns(base, first->interval);
1253         taprio_set_budgets(q, sched, first);
1254
1255         for (tc = 0; tc < num_tc; tc++) {
1256                 if (first->gate_duration[tc] == sched->cycle_time)
1257                         first->gate_close_time[tc] = KTIME_MAX;
1258                 else
1259                         first->gate_close_time[tc] = ktime_add_ns(base, first->gate_duration[tc]);
1260         }
1261
1262         rcu_assign_pointer(q->current_entry, NULL);
1263 }
1264
1265 static void taprio_start_sched(struct Qdisc *sch,
1266                                ktime_t start, struct sched_gate_list *new)
1267 {
1268         struct taprio_sched *q = qdisc_priv(sch);
1269         ktime_t expires;
1270
1271         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1272                 return;
1273
1274         expires = hrtimer_get_expires(&q->advance_timer);
1275         if (expires == 0)
1276                 expires = KTIME_MAX;
1277
1278         /* If the new schedule starts before the next expiration, we
1279          * reprogram it to the earliest one, so we change the admin
1280          * schedule to the operational one at the right time.
1281          */
1282         start = min_t(ktime_t, start, expires);
1283
1284         hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1285 }
1286
1287 static void taprio_set_picos_per_byte(struct net_device *dev,
1288                                       struct taprio_sched *q)
1289 {
1290         struct ethtool_link_ksettings ecmd;
1291         int speed = SPEED_10;
1292         int picos_per_byte;
1293         int err;
1294
1295         err = __ethtool_get_link_ksettings(dev, &ecmd);
1296         if (err < 0)
1297                 goto skip;
1298
1299         if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1300                 speed = ecmd.base.speed;
1301
1302 skip:
1303         picos_per_byte = (USEC_PER_SEC * 8) / speed;
1304
1305         atomic64_set(&q->picos_per_byte, picos_per_byte);
1306         netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1307                    dev->name, (long long)atomic64_read(&q->picos_per_byte),
1308                    ecmd.base.speed);
1309 }
1310
1311 static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1312                                void *ptr)
1313 {
1314         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1315         struct sched_gate_list *oper, *admin;
1316         struct qdisc_size_table *stab;
1317         struct taprio_sched *q;
1318
1319         ASSERT_RTNL();
1320
1321         if (event != NETDEV_UP && event != NETDEV_CHANGE)
1322                 return NOTIFY_DONE;
1323
1324         list_for_each_entry(q, &taprio_list, taprio_list) {
1325                 if (dev != qdisc_dev(q->root))
1326                         continue;
1327
1328                 taprio_set_picos_per_byte(dev, q);
1329
1330                 stab = rtnl_dereference(q->root->stab);
1331
1332                 oper = rtnl_dereference(q->oper_sched);
1333                 if (oper)
1334                         taprio_update_queue_max_sdu(q, oper, stab);
1335
1336                 admin = rtnl_dereference(q->admin_sched);
1337                 if (admin)
1338                         taprio_update_queue_max_sdu(q, admin, stab);
1339
1340                 break;
1341         }
1342
1343         return NOTIFY_DONE;
1344 }
1345
1346 static void setup_txtime(struct taprio_sched *q,
1347                          struct sched_gate_list *sched, ktime_t base)
1348 {
1349         struct sched_entry *entry;
1350         u32 interval = 0;
1351
1352         list_for_each_entry(entry, &sched->entries, list) {
1353                 entry->next_txtime = ktime_add_ns(base, interval);
1354                 interval += entry->interval;
1355         }
1356 }
1357
1358 static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1359 {
1360         struct __tc_taprio_qopt_offload *__offload;
1361
1362         __offload = kzalloc(struct_size(__offload, offload.entries, num_entries),
1363                             GFP_KERNEL);
1364         if (!__offload)
1365                 return NULL;
1366
1367         refcount_set(&__offload->users, 1);
1368
1369         return &__offload->offload;
1370 }
1371
1372 struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1373                                                   *offload)
1374 {
1375         struct __tc_taprio_qopt_offload *__offload;
1376
1377         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1378                                  offload);
1379
1380         refcount_inc(&__offload->users);
1381
1382         return offload;
1383 }
1384 EXPORT_SYMBOL_GPL(taprio_offload_get);
1385
1386 void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1387 {
1388         struct __tc_taprio_qopt_offload *__offload;
1389
1390         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1391                                  offload);
1392
1393         if (!refcount_dec_and_test(&__offload->users))
1394                 return;
1395
1396         kfree(__offload);
1397 }
1398 EXPORT_SYMBOL_GPL(taprio_offload_free);
1399
1400 /* The function will only serve to keep the pointers to the "oper" and "admin"
1401  * schedules valid in relation to their base times, so when calling dump() the
1402  * users looks at the right schedules.
1403  * When using full offload, the admin configuration is promoted to oper at the
1404  * base_time in the PHC time domain.  But because the system time is not
1405  * necessarily in sync with that, we can't just trigger a hrtimer to call
1406  * switch_schedules at the right hardware time.
1407  * At the moment we call this by hand right away from taprio, but in the future
1408  * it will be useful to create a mechanism for drivers to notify taprio of the
1409  * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1410  * This is left as TODO.
1411  */
1412 static void taprio_offload_config_changed(struct taprio_sched *q)
1413 {
1414         struct sched_gate_list *oper, *admin;
1415
1416         oper = rtnl_dereference(q->oper_sched);
1417         admin = rtnl_dereference(q->admin_sched);
1418
1419         switch_schedules(q, &admin, &oper);
1420 }
1421
1422 static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
1423 {
1424         u32 i, queue_mask = 0;
1425
1426         for (i = 0; i < dev->num_tc; i++) {
1427                 u32 offset, count;
1428
1429                 if (!(tc_mask & BIT(i)))
1430                         continue;
1431
1432                 offset = dev->tc_to_txq[i].offset;
1433                 count = dev->tc_to_txq[i].count;
1434
1435                 queue_mask |= GENMASK(offset + count - 1, offset);
1436         }
1437
1438         return queue_mask;
1439 }
1440
1441 static void taprio_sched_to_offload(struct net_device *dev,
1442                                     struct sched_gate_list *sched,
1443                                     struct tc_taprio_qopt_offload *offload,
1444                                     const struct tc_taprio_caps *caps)
1445 {
1446         struct sched_entry *entry;
1447         int i = 0;
1448
1449         offload->base_time = sched->base_time;
1450         offload->cycle_time = sched->cycle_time;
1451         offload->cycle_time_extension = sched->cycle_time_extension;
1452
1453         list_for_each_entry(entry, &sched->entries, list) {
1454                 struct tc_taprio_sched_entry *e = &offload->entries[i];
1455
1456                 e->command = entry->command;
1457                 e->interval = entry->interval;
1458                 if (caps->gate_mask_per_txq)
1459                         e->gate_mask = tc_map_to_queue_mask(dev,
1460                                                             entry->gate_mask);
1461                 else
1462                         e->gate_mask = entry->gate_mask;
1463
1464                 i++;
1465         }
1466
1467         offload->num_entries = i;
1468 }
1469
1470 static void taprio_detect_broken_mqprio(struct taprio_sched *q)
1471 {
1472         struct net_device *dev = qdisc_dev(q->root);
1473         struct tc_taprio_caps caps;
1474
1475         qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1476                                  &caps, sizeof(caps));
1477
1478         q->broken_mqprio = caps.broken_mqprio;
1479         if (q->broken_mqprio)
1480                 static_branch_inc(&taprio_have_broken_mqprio);
1481         else
1482                 static_branch_inc(&taprio_have_working_mqprio);
1483
1484         q->detected_mqprio = true;
1485 }
1486
1487 static void taprio_cleanup_broken_mqprio(struct taprio_sched *q)
1488 {
1489         if (!q->detected_mqprio)
1490                 return;
1491
1492         if (q->broken_mqprio)
1493                 static_branch_dec(&taprio_have_broken_mqprio);
1494         else
1495                 static_branch_dec(&taprio_have_working_mqprio);
1496 }
1497
1498 static int taprio_enable_offload(struct net_device *dev,
1499                                  struct taprio_sched *q,
1500                                  struct sched_gate_list *sched,
1501                                  struct netlink_ext_ack *extack)
1502 {
1503         const struct net_device_ops *ops = dev->netdev_ops;
1504         struct tc_taprio_qopt_offload *offload;
1505         struct tc_taprio_caps caps;
1506         int tc, err = 0;
1507
1508         if (!ops->ndo_setup_tc) {
1509                 NL_SET_ERR_MSG(extack,
1510                                "Device does not support taprio offload");
1511                 return -EOPNOTSUPP;
1512         }
1513
1514         qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1515                                  &caps, sizeof(caps));
1516
1517         if (!caps.supports_queue_max_sdu) {
1518                 for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
1519                         if (q->max_sdu[tc]) {
1520                                 NL_SET_ERR_MSG_MOD(extack,
1521                                                    "Device does not handle queueMaxSDU");
1522                                 return -EOPNOTSUPP;
1523                         }
1524                 }
1525         }
1526
1527         offload = taprio_offload_alloc(sched->num_entries);
1528         if (!offload) {
1529                 NL_SET_ERR_MSG(extack,
1530                                "Not enough memory for enabling offload mode");
1531                 return -ENOMEM;
1532         }
1533         offload->cmd = TAPRIO_CMD_REPLACE;
1534         offload->extack = extack;
1535         mqprio_qopt_reconstruct(dev, &offload->mqprio.qopt);
1536         offload->mqprio.extack = extack;
1537         taprio_sched_to_offload(dev, sched, offload, &caps);
1538         mqprio_fp_to_offload(q->fp, &offload->mqprio);
1539
1540         for (tc = 0; tc < TC_MAX_QUEUE; tc++)
1541                 offload->max_sdu[tc] = q->max_sdu[tc];
1542
1543         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1544         if (err < 0) {
1545                 NL_SET_ERR_MSG_WEAK(extack,
1546                                     "Device failed to setup taprio offload");
1547                 goto done;
1548         }
1549
1550         q->offloaded = true;
1551
1552 done:
1553         /* The offload structure may linger around via a reference taken by the
1554          * device driver, so clear up the netlink extack pointer so that the
1555          * driver isn't tempted to dereference data which stopped being valid
1556          */
1557         offload->extack = NULL;
1558         offload->mqprio.extack = NULL;
1559         taprio_offload_free(offload);
1560
1561         return err;
1562 }
1563
1564 static int taprio_disable_offload(struct net_device *dev,
1565                                   struct taprio_sched *q,
1566                                   struct netlink_ext_ack *extack)
1567 {
1568         const struct net_device_ops *ops = dev->netdev_ops;
1569         struct tc_taprio_qopt_offload *offload;
1570         int err;
1571
1572         if (!q->offloaded)
1573                 return 0;
1574
1575         offload = taprio_offload_alloc(0);
1576         if (!offload) {
1577                 NL_SET_ERR_MSG(extack,
1578                                "Not enough memory to disable offload mode");
1579                 return -ENOMEM;
1580         }
1581         offload->cmd = TAPRIO_CMD_DESTROY;
1582
1583         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1584         if (err < 0) {
1585                 NL_SET_ERR_MSG(extack,
1586                                "Device failed to disable offload");
1587                 goto out;
1588         }
1589
1590         q->offloaded = false;
1591
1592 out:
1593         taprio_offload_free(offload);
1594
1595         return err;
1596 }
1597
1598 /* If full offload is enabled, the only possible clockid is the net device's
1599  * PHC. For that reason, specifying a clockid through netlink is incorrect.
1600  * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1601  * in sync with the specified clockid via a user space daemon such as phc2sys.
1602  * For both software taprio and txtime-assist, the clockid is used for the
1603  * hrtimer that advances the schedule and hence mandatory.
1604  */
1605 static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1606                                 struct netlink_ext_ack *extack)
1607 {
1608         struct taprio_sched *q = qdisc_priv(sch);
1609         struct net_device *dev = qdisc_dev(sch);
1610         int err = -EINVAL;
1611
1612         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1613                 const struct ethtool_ops *ops = dev->ethtool_ops;
1614                 struct ethtool_ts_info info = {
1615                         .cmd = ETHTOOL_GET_TS_INFO,
1616                         .phc_index = -1,
1617                 };
1618
1619                 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1620                         NL_SET_ERR_MSG(extack,
1621                                        "The 'clockid' cannot be specified for full offload");
1622                         goto out;
1623                 }
1624
1625                 if (ops && ops->get_ts_info)
1626                         err = ops->get_ts_info(dev, &info);
1627
1628                 if (err || info.phc_index < 0) {
1629                         NL_SET_ERR_MSG(extack,
1630                                        "Device does not have a PTP clock");
1631                         err = -ENOTSUPP;
1632                         goto out;
1633                 }
1634         } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1635                 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1636                 enum tk_offsets tk_offset;
1637
1638                 /* We only support static clockids and we don't allow
1639                  * for it to be modified after the first init.
1640                  */
1641                 if (clockid < 0 ||
1642                     (q->clockid != -1 && q->clockid != clockid)) {
1643                         NL_SET_ERR_MSG(extack,
1644                                        "Changing the 'clockid' of a running schedule is not supported");
1645                         err = -ENOTSUPP;
1646                         goto out;
1647                 }
1648
1649                 switch (clockid) {
1650                 case CLOCK_REALTIME:
1651                         tk_offset = TK_OFFS_REAL;
1652                         break;
1653                 case CLOCK_MONOTONIC:
1654                         tk_offset = TK_OFFS_MAX;
1655                         break;
1656                 case CLOCK_BOOTTIME:
1657                         tk_offset = TK_OFFS_BOOT;
1658                         break;
1659                 case CLOCK_TAI:
1660                         tk_offset = TK_OFFS_TAI;
1661                         break;
1662                 default:
1663                         NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1664                         err = -EINVAL;
1665                         goto out;
1666                 }
1667                 /* This pairs with READ_ONCE() in taprio_mono_to_any */
1668                 WRITE_ONCE(q->tk_offset, tk_offset);
1669
1670                 q->clockid = clockid;
1671         } else {
1672                 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1673                 goto out;
1674         }
1675
1676         /* Everything went ok, return success. */
1677         err = 0;
1678
1679 out:
1680         return err;
1681 }
1682
1683 static int taprio_parse_tc_entry(struct Qdisc *sch,
1684                                  struct nlattr *opt,
1685                                  u32 max_sdu[TC_QOPT_MAX_QUEUE],
1686                                  u32 fp[TC_QOPT_MAX_QUEUE],
1687                                  unsigned long *seen_tcs,
1688                                  struct netlink_ext_ack *extack)
1689 {
1690         struct nlattr *tb[TCA_TAPRIO_TC_ENTRY_MAX + 1] = { };
1691         struct net_device *dev = qdisc_dev(sch);
1692         int err, tc;
1693         u32 val;
1694
1695         err = nla_parse_nested(tb, TCA_TAPRIO_TC_ENTRY_MAX, opt,
1696                                taprio_tc_policy, extack);
1697         if (err < 0)
1698                 return err;
1699
1700         if (!tb[TCA_TAPRIO_TC_ENTRY_INDEX]) {
1701                 NL_SET_ERR_MSG_MOD(extack, "TC entry index missing");
1702                 return -EINVAL;
1703         }
1704
1705         tc = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_INDEX]);
1706         if (tc >= TC_QOPT_MAX_QUEUE) {
1707                 NL_SET_ERR_MSG_MOD(extack, "TC entry index out of range");
1708                 return -ERANGE;
1709         }
1710
1711         if (*seen_tcs & BIT(tc)) {
1712                 NL_SET_ERR_MSG_MOD(extack, "Duplicate TC entry");
1713                 return -EINVAL;
1714         }
1715
1716         *seen_tcs |= BIT(tc);
1717
1718         if (tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]) {
1719                 val = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]);
1720                 if (val > dev->max_mtu) {
1721                         NL_SET_ERR_MSG_MOD(extack, "TC max SDU exceeds device max MTU");
1722                         return -ERANGE;
1723                 }
1724
1725                 max_sdu[tc] = val;
1726         }
1727
1728         if (tb[TCA_TAPRIO_TC_ENTRY_FP])
1729                 fp[tc] = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_FP]);
1730
1731         return 0;
1732 }
1733
1734 static int taprio_parse_tc_entries(struct Qdisc *sch,
1735                                    struct nlattr *opt,
1736                                    struct netlink_ext_ack *extack)
1737 {
1738         struct taprio_sched *q = qdisc_priv(sch);
1739         struct net_device *dev = qdisc_dev(sch);
1740         u32 max_sdu[TC_QOPT_MAX_QUEUE];
1741         bool have_preemption = false;
1742         unsigned long seen_tcs = 0;
1743         u32 fp[TC_QOPT_MAX_QUEUE];
1744         struct nlattr *n;
1745         int tc, rem;
1746         int err = 0;
1747
1748         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1749                 max_sdu[tc] = q->max_sdu[tc];
1750                 fp[tc] = q->fp[tc];
1751         }
1752
1753         nla_for_each_nested(n, opt, rem) {
1754                 if (nla_type(n) != TCA_TAPRIO_ATTR_TC_ENTRY)
1755                         continue;
1756
1757                 err = taprio_parse_tc_entry(sch, n, max_sdu, fp, &seen_tcs,
1758                                             extack);
1759                 if (err)
1760                         return err;
1761         }
1762
1763         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1764                 q->max_sdu[tc] = max_sdu[tc];
1765                 q->fp[tc] = fp[tc];
1766                 if (fp[tc] != TC_FP_EXPRESS)
1767                         have_preemption = true;
1768         }
1769
1770         if (have_preemption) {
1771                 if (!FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1772                         NL_SET_ERR_MSG(extack,
1773                                        "Preemption only supported with full offload");
1774                         return -EOPNOTSUPP;
1775                 }
1776
1777                 if (!ethtool_dev_mm_supported(dev)) {
1778                         NL_SET_ERR_MSG(extack,
1779                                        "Device does not support preemption");
1780                         return -EOPNOTSUPP;
1781                 }
1782         }
1783
1784         return err;
1785 }
1786
1787 static int taprio_mqprio_cmp(const struct net_device *dev,
1788                              const struct tc_mqprio_qopt *mqprio)
1789 {
1790         int i;
1791
1792         if (!mqprio || mqprio->num_tc != dev->num_tc)
1793                 return -1;
1794
1795         for (i = 0; i < mqprio->num_tc; i++)
1796                 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1797                     dev->tc_to_txq[i].offset != mqprio->offset[i])
1798                         return -1;
1799
1800         for (i = 0; i <= TC_BITMASK; i++)
1801                 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1802                         return -1;
1803
1804         return 0;
1805 }
1806
1807 /* The semantics of the 'flags' argument in relation to 'change()'
1808  * requests, are interpreted following two rules (which are applied in
1809  * this order): (1) an omitted 'flags' argument is interpreted as
1810  * zero; (2) the 'flags' of a "running" taprio instance cannot be
1811  * changed.
1812  */
1813 static int taprio_new_flags(const struct nlattr *attr, u32 old,
1814                             struct netlink_ext_ack *extack)
1815 {
1816         u32 new = 0;
1817
1818         if (attr)
1819                 new = nla_get_u32(attr);
1820
1821         if (old != TAPRIO_FLAGS_INVALID && old != new) {
1822                 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1823                 return -EOPNOTSUPP;
1824         }
1825
1826         if (!taprio_flags_valid(new)) {
1827                 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1828                 return -EINVAL;
1829         }
1830
1831         return new;
1832 }
1833
1834 static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1835                          struct netlink_ext_ack *extack)
1836 {
1837         struct qdisc_size_table *stab = rtnl_dereference(sch->stab);
1838         struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1839         struct sched_gate_list *oper, *admin, *new_admin;
1840         struct taprio_sched *q = qdisc_priv(sch);
1841         struct net_device *dev = qdisc_dev(sch);
1842         struct tc_mqprio_qopt *mqprio = NULL;
1843         unsigned long flags;
1844         ktime_t start;
1845         int i, err;
1846
1847         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1848                                           taprio_policy, extack);
1849         if (err < 0)
1850                 return err;
1851
1852         if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1853                 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1854
1855         err = taprio_new_flags(tb[TCA_TAPRIO_ATTR_FLAGS],
1856                                q->flags, extack);
1857         if (err < 0)
1858                 return err;
1859
1860         q->flags = err;
1861
1862         err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1863         if (err < 0)
1864                 return err;
1865
1866         err = taprio_parse_tc_entries(sch, opt, extack);
1867         if (err)
1868                 return err;
1869
1870         new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1871         if (!new_admin) {
1872                 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1873                 return -ENOMEM;
1874         }
1875         INIT_LIST_HEAD(&new_admin->entries);
1876
1877         oper = rtnl_dereference(q->oper_sched);
1878         admin = rtnl_dereference(q->admin_sched);
1879
1880         /* no changes - no new mqprio settings */
1881         if (!taprio_mqprio_cmp(dev, mqprio))
1882                 mqprio = NULL;
1883
1884         if (mqprio && (oper || admin)) {
1885                 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1886                 err = -ENOTSUPP;
1887                 goto free_sched;
1888         }
1889
1890         if (mqprio) {
1891                 err = netdev_set_num_tc(dev, mqprio->num_tc);
1892                 if (err)
1893                         goto free_sched;
1894                 for (i = 0; i < mqprio->num_tc; i++) {
1895                         netdev_set_tc_queue(dev, i,
1896                                             mqprio->count[i],
1897                                             mqprio->offset[i]);
1898                         q->cur_txq[i] = mqprio->offset[i];
1899                 }
1900
1901                 /* Always use supplied priority mappings */
1902                 for (i = 0; i <= TC_BITMASK; i++)
1903                         netdev_set_prio_tc_map(dev, i,
1904                                                mqprio->prio_tc_map[i]);
1905         }
1906
1907         err = parse_taprio_schedule(q, tb, new_admin, extack);
1908         if (err < 0)
1909                 goto free_sched;
1910
1911         if (new_admin->num_entries == 0) {
1912                 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1913                 err = -EINVAL;
1914                 goto free_sched;
1915         }
1916
1917         err = taprio_parse_clockid(sch, tb, extack);
1918         if (err < 0)
1919                 goto free_sched;
1920
1921         taprio_set_picos_per_byte(dev, q);
1922         taprio_update_queue_max_sdu(q, new_admin, stab);
1923
1924         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1925                 err = taprio_enable_offload(dev, q, new_admin, extack);
1926         else
1927                 err = taprio_disable_offload(dev, q, extack);
1928         if (err)
1929                 goto free_sched;
1930
1931         /* Protects against enqueue()/dequeue() */
1932         spin_lock_bh(qdisc_lock(sch));
1933
1934         if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1935                 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1936                         NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1937                         err = -EINVAL;
1938                         goto unlock;
1939                 }
1940
1941                 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1942         }
1943
1944         if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1945             !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1946             !hrtimer_active(&q->advance_timer)) {
1947                 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1948                 q->advance_timer.function = advance_sched;
1949         }
1950
1951         err = taprio_get_start_time(sch, new_admin, &start);
1952         if (err < 0) {
1953                 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1954                 goto unlock;
1955         }
1956
1957         setup_txtime(q, new_admin, start);
1958
1959         if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1960                 if (!oper) {
1961                         rcu_assign_pointer(q->oper_sched, new_admin);
1962                         err = 0;
1963                         new_admin = NULL;
1964                         goto unlock;
1965                 }
1966
1967                 rcu_assign_pointer(q->admin_sched, new_admin);
1968                 if (admin)
1969                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1970         } else {
1971                 setup_first_end_time(q, new_admin, start);
1972
1973                 /* Protects against advance_sched() */
1974                 spin_lock_irqsave(&q->current_entry_lock, flags);
1975
1976                 taprio_start_sched(sch, start, new_admin);
1977
1978                 rcu_assign_pointer(q->admin_sched, new_admin);
1979                 if (admin)
1980                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1981
1982                 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1983
1984                 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1985                         taprio_offload_config_changed(q);
1986         }
1987
1988         new_admin = NULL;
1989         err = 0;
1990
1991         if (!stab)
1992                 NL_SET_ERR_MSG_MOD(extack,
1993                                    "Size table not specified, frame length estimations may be inaccurate");
1994
1995 unlock:
1996         spin_unlock_bh(qdisc_lock(sch));
1997
1998 free_sched:
1999         if (new_admin)
2000                 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
2001
2002         return err;
2003 }
2004
2005 static void taprio_reset(struct Qdisc *sch)
2006 {
2007         struct taprio_sched *q = qdisc_priv(sch);
2008         struct net_device *dev = qdisc_dev(sch);
2009         int i;
2010
2011         hrtimer_cancel(&q->advance_timer);
2012
2013         if (q->qdiscs) {
2014                 for (i = 0; i < dev->num_tx_queues; i++)
2015                         if (q->qdiscs[i])
2016                                 qdisc_reset(q->qdiscs[i]);
2017         }
2018 }
2019
2020 static void taprio_destroy(struct Qdisc *sch)
2021 {
2022         struct taprio_sched *q = qdisc_priv(sch);
2023         struct net_device *dev = qdisc_dev(sch);
2024         struct sched_gate_list *oper, *admin;
2025         unsigned int i;
2026
2027         list_del(&q->taprio_list);
2028
2029         /* Note that taprio_reset() might not be called if an error
2030          * happens in qdisc_create(), after taprio_init() has been called.
2031          */
2032         hrtimer_cancel(&q->advance_timer);
2033         qdisc_synchronize(sch);
2034
2035         taprio_disable_offload(dev, q, NULL);
2036
2037         if (q->qdiscs) {
2038                 for (i = 0; i < dev->num_tx_queues; i++)
2039                         qdisc_put(q->qdiscs[i]);
2040
2041                 kfree(q->qdiscs);
2042         }
2043         q->qdiscs = NULL;
2044
2045         netdev_reset_tc(dev);
2046
2047         oper = rtnl_dereference(q->oper_sched);
2048         admin = rtnl_dereference(q->admin_sched);
2049
2050         if (oper)
2051                 call_rcu(&oper->rcu, taprio_free_sched_cb);
2052
2053         if (admin)
2054                 call_rcu(&admin->rcu, taprio_free_sched_cb);
2055
2056         taprio_cleanup_broken_mqprio(q);
2057 }
2058
2059 static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
2060                        struct netlink_ext_ack *extack)
2061 {
2062         struct taprio_sched *q = qdisc_priv(sch);
2063         struct net_device *dev = qdisc_dev(sch);
2064         int i, tc;
2065
2066         spin_lock_init(&q->current_entry_lock);
2067
2068         hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
2069         q->advance_timer.function = advance_sched;
2070
2071         q->root = sch;
2072
2073         /* We only support static clockids. Use an invalid value as default
2074          * and get the valid one on taprio_change().
2075          */
2076         q->clockid = -1;
2077         q->flags = TAPRIO_FLAGS_INVALID;
2078
2079         list_add(&q->taprio_list, &taprio_list);
2080
2081         if (sch->parent != TC_H_ROOT) {
2082                 NL_SET_ERR_MSG_MOD(extack, "Can only be attached as root qdisc");
2083                 return -EOPNOTSUPP;
2084         }
2085
2086         if (!netif_is_multiqueue(dev)) {
2087                 NL_SET_ERR_MSG_MOD(extack, "Multi-queue device is required");
2088                 return -EOPNOTSUPP;
2089         }
2090
2091         /* pre-allocate qdisc, attachment can't fail */
2092         q->qdiscs = kcalloc(dev->num_tx_queues,
2093                             sizeof(q->qdiscs[0]),
2094                             GFP_KERNEL);
2095
2096         if (!q->qdiscs)
2097                 return -ENOMEM;
2098
2099         if (!opt)
2100                 return -EINVAL;
2101
2102         for (i = 0; i < dev->num_tx_queues; i++) {
2103                 struct netdev_queue *dev_queue;
2104                 struct Qdisc *qdisc;
2105
2106                 dev_queue = netdev_get_tx_queue(dev, i);
2107                 qdisc = qdisc_create_dflt(dev_queue,
2108                                           &pfifo_qdisc_ops,
2109                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
2110                                                     TC_H_MIN(i + 1)),
2111                                           extack);
2112                 if (!qdisc)
2113                         return -ENOMEM;
2114
2115                 if (i < dev->real_num_tx_queues)
2116                         qdisc_hash_add(qdisc, false);
2117
2118                 q->qdiscs[i] = qdisc;
2119         }
2120
2121         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++)
2122                 q->fp[tc] = TC_FP_EXPRESS;
2123
2124         taprio_detect_broken_mqprio(q);
2125
2126         return taprio_change(sch, opt, extack);
2127 }
2128
2129 static void taprio_attach(struct Qdisc *sch)
2130 {
2131         struct taprio_sched *q = qdisc_priv(sch);
2132         struct net_device *dev = qdisc_dev(sch);
2133         unsigned int ntx;
2134
2135         /* Attach underlying qdisc */
2136         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
2137                 struct Qdisc *qdisc = q->qdiscs[ntx];
2138                 struct Qdisc *old;
2139
2140                 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2141                         qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2142                         old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
2143                 } else {
2144                         old = dev_graft_qdisc(qdisc->dev_queue, sch);
2145                         qdisc_refcount_inc(sch);
2146                 }
2147                 if (old)
2148                         qdisc_put(old);
2149         }
2150
2151         /* access to the child qdiscs is not needed in offload mode */
2152         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2153                 kfree(q->qdiscs);
2154                 q->qdiscs = NULL;
2155         }
2156 }
2157
2158 static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
2159                                              unsigned long cl)
2160 {
2161         struct net_device *dev = qdisc_dev(sch);
2162         unsigned long ntx = cl - 1;
2163
2164         if (ntx >= dev->num_tx_queues)
2165                 return NULL;
2166
2167         return netdev_get_tx_queue(dev, ntx);
2168 }
2169
2170 static int taprio_graft(struct Qdisc *sch, unsigned long cl,
2171                         struct Qdisc *new, struct Qdisc **old,
2172                         struct netlink_ext_ack *extack)
2173 {
2174         struct taprio_sched *q = qdisc_priv(sch);
2175         struct net_device *dev = qdisc_dev(sch);
2176         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2177
2178         if (!dev_queue)
2179                 return -EINVAL;
2180
2181         if (dev->flags & IFF_UP)
2182                 dev_deactivate(dev);
2183
2184         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2185                 *old = dev_graft_qdisc(dev_queue, new);
2186         } else {
2187                 *old = q->qdiscs[cl - 1];
2188                 q->qdiscs[cl - 1] = new;
2189         }
2190
2191         if (new)
2192                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2193
2194         if (dev->flags & IFF_UP)
2195                 dev_activate(dev);
2196
2197         return 0;
2198 }
2199
2200 static int dump_entry(struct sk_buff *msg,
2201                       const struct sched_entry *entry)
2202 {
2203         struct nlattr *item;
2204
2205         item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
2206         if (!item)
2207                 return -ENOSPC;
2208
2209         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
2210                 goto nla_put_failure;
2211
2212         if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
2213                 goto nla_put_failure;
2214
2215         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
2216                         entry->gate_mask))
2217                 goto nla_put_failure;
2218
2219         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
2220                         entry->interval))
2221                 goto nla_put_failure;
2222
2223         return nla_nest_end(msg, item);
2224
2225 nla_put_failure:
2226         nla_nest_cancel(msg, item);
2227         return -1;
2228 }
2229
2230 static int dump_schedule(struct sk_buff *msg,
2231                          const struct sched_gate_list *root)
2232 {
2233         struct nlattr *entry_list;
2234         struct sched_entry *entry;
2235
2236         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
2237                         root->base_time, TCA_TAPRIO_PAD))
2238                 return -1;
2239
2240         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
2241                         root->cycle_time, TCA_TAPRIO_PAD))
2242                 return -1;
2243
2244         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
2245                         root->cycle_time_extension, TCA_TAPRIO_PAD))
2246                 return -1;
2247
2248         entry_list = nla_nest_start_noflag(msg,
2249                                            TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
2250         if (!entry_list)
2251                 goto error_nest;
2252
2253         list_for_each_entry(entry, &root->entries, list) {
2254                 if (dump_entry(msg, entry) < 0)
2255                         goto error_nest;
2256         }
2257
2258         nla_nest_end(msg, entry_list);
2259         return 0;
2260
2261 error_nest:
2262         nla_nest_cancel(msg, entry_list);
2263         return -1;
2264 }
2265
2266 static int taprio_dump_tc_entries(struct sk_buff *skb,
2267                                   struct taprio_sched *q,
2268                                   struct sched_gate_list *sched)
2269 {
2270         struct nlattr *n;
2271         int tc;
2272
2273         for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
2274                 n = nla_nest_start(skb, TCA_TAPRIO_ATTR_TC_ENTRY);
2275                 if (!n)
2276                         return -EMSGSIZE;
2277
2278                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_INDEX, tc))
2279                         goto nla_put_failure;
2280
2281                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_MAX_SDU,
2282                                 sched->max_sdu[tc]))
2283                         goto nla_put_failure;
2284
2285                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_FP, q->fp[tc]))
2286                         goto nla_put_failure;
2287
2288                 nla_nest_end(skb, n);
2289         }
2290
2291         return 0;
2292
2293 nla_put_failure:
2294         nla_nest_cancel(skb, n);
2295         return -EMSGSIZE;
2296 }
2297
2298 static int taprio_put_stat(struct sk_buff *skb, u64 val, u16 attrtype)
2299 {
2300         if (val == TAPRIO_STAT_NOT_SET)
2301                 return 0;
2302         if (nla_put_u64_64bit(skb, attrtype, val, TCA_TAPRIO_OFFLOAD_STATS_PAD))
2303                 return -EMSGSIZE;
2304         return 0;
2305 }
2306
2307 static int taprio_dump_xstats(struct Qdisc *sch, struct gnet_dump *d,
2308                               struct tc_taprio_qopt_offload *offload,
2309                               struct tc_taprio_qopt_stats *stats)
2310 {
2311         struct net_device *dev = qdisc_dev(sch);
2312         const struct net_device_ops *ops;
2313         struct sk_buff *skb = d->skb;
2314         struct nlattr *xstats;
2315         int err;
2316
2317         ops = qdisc_dev(sch)->netdev_ops;
2318
2319         /* FIXME I could use qdisc_offload_dump_helper(), but that messes
2320          * with sch->flags depending on whether the device reports taprio
2321          * stats, and I'm not sure whether that's a good idea, considering
2322          * that stats are optional to the offload itself
2323          */
2324         if (!ops->ndo_setup_tc)
2325                 return 0;
2326
2327         memset(stats, 0xff, sizeof(*stats));
2328
2329         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
2330         if (err == -EOPNOTSUPP)
2331                 return 0;
2332         if (err)
2333                 return err;
2334
2335         xstats = nla_nest_start(skb, TCA_STATS_APP);
2336         if (!xstats)
2337                 goto err;
2338
2339         if (taprio_put_stat(skb, stats->window_drops,
2340                             TCA_TAPRIO_OFFLOAD_STATS_WINDOW_DROPS) ||
2341             taprio_put_stat(skb, stats->tx_overruns,
2342                             TCA_TAPRIO_OFFLOAD_STATS_TX_OVERRUNS))
2343                 goto err_cancel;
2344
2345         nla_nest_end(skb, xstats);
2346
2347         return 0;
2348
2349 err_cancel:
2350         nla_nest_cancel(skb, xstats);
2351 err:
2352         return -EMSGSIZE;
2353 }
2354
2355 static int taprio_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
2356 {
2357         struct tc_taprio_qopt_offload offload = {
2358                 .cmd = TAPRIO_CMD_STATS,
2359         };
2360
2361         return taprio_dump_xstats(sch, d, &offload, &offload.stats);
2362 }
2363
2364 static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
2365 {
2366         struct taprio_sched *q = qdisc_priv(sch);
2367         struct net_device *dev = qdisc_dev(sch);
2368         struct sched_gate_list *oper, *admin;
2369         struct tc_mqprio_qopt opt = { 0 };
2370         struct nlattr *nest, *sched_nest;
2371
2372         oper = rtnl_dereference(q->oper_sched);
2373         admin = rtnl_dereference(q->admin_sched);
2374
2375         mqprio_qopt_reconstruct(dev, &opt);
2376
2377         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2378         if (!nest)
2379                 goto start_error;
2380
2381         if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
2382                 goto options_error;
2383
2384         if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
2385             nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
2386                 goto options_error;
2387
2388         if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
2389                 goto options_error;
2390
2391         if (q->txtime_delay &&
2392             nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
2393                 goto options_error;
2394
2395         if (oper && taprio_dump_tc_entries(skb, q, oper))
2396                 goto options_error;
2397
2398         if (oper && dump_schedule(skb, oper))
2399                 goto options_error;
2400
2401         if (!admin)
2402                 goto done;
2403
2404         sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
2405         if (!sched_nest)
2406                 goto options_error;
2407
2408         if (dump_schedule(skb, admin))
2409                 goto admin_error;
2410
2411         nla_nest_end(skb, sched_nest);
2412
2413 done:
2414         return nla_nest_end(skb, nest);
2415
2416 admin_error:
2417         nla_nest_cancel(skb, sched_nest);
2418
2419 options_error:
2420         nla_nest_cancel(skb, nest);
2421
2422 start_error:
2423         return -ENOSPC;
2424 }
2425
2426 static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
2427 {
2428         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2429
2430         if (!dev_queue)
2431                 return NULL;
2432
2433         return rtnl_dereference(dev_queue->qdisc_sleeping);
2434 }
2435
2436 static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
2437 {
2438         unsigned int ntx = TC_H_MIN(classid);
2439
2440         if (!taprio_queue_get(sch, ntx))
2441                 return 0;
2442         return ntx;
2443 }
2444
2445 static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
2446                              struct sk_buff *skb, struct tcmsg *tcm)
2447 {
2448         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2449
2450         tcm->tcm_parent = TC_H_ROOT;
2451         tcm->tcm_handle |= TC_H_MIN(cl);
2452         tcm->tcm_info = rtnl_dereference(dev_queue->qdisc_sleeping)->handle;
2453
2454         return 0;
2455 }
2456
2457 static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
2458                                    struct gnet_dump *d)
2459         __releases(d->lock)
2460         __acquires(d->lock)
2461 {
2462         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2463         struct tc_taprio_qopt_offload offload = {
2464                 .cmd = TAPRIO_CMD_QUEUE_STATS,
2465                 .queue_stats = {
2466                         .queue = cl - 1,
2467                 },
2468         };
2469         struct Qdisc *child;
2470
2471         child = rtnl_dereference(dev_queue->qdisc_sleeping);
2472         if (gnet_stats_copy_basic(d, NULL, &child->bstats, true) < 0 ||
2473             qdisc_qstats_copy(d, child) < 0)
2474                 return -1;
2475
2476         return taprio_dump_xstats(sch, d, &offload, &offload.queue_stats.stats);
2477 }
2478
2479 static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2480 {
2481         struct net_device *dev = qdisc_dev(sch);
2482         unsigned long ntx;
2483
2484         if (arg->stop)
2485                 return;
2486
2487         arg->count = arg->skip;
2488         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
2489                 if (!tc_qdisc_stats_dump(sch, ntx + 1, arg))
2490                         break;
2491         }
2492 }
2493
2494 static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
2495                                                 struct tcmsg *tcm)
2496 {
2497         return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
2498 }
2499
2500 static const struct Qdisc_class_ops taprio_class_ops = {
2501         .graft          = taprio_graft,
2502         .leaf           = taprio_leaf,
2503         .find           = taprio_find,
2504         .walk           = taprio_walk,
2505         .dump           = taprio_dump_class,
2506         .dump_stats     = taprio_dump_class_stats,
2507         .select_queue   = taprio_select_queue,
2508 };
2509
2510 static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
2511         .cl_ops         = &taprio_class_ops,
2512         .id             = "taprio",
2513         .priv_size      = sizeof(struct taprio_sched),
2514         .init           = taprio_init,
2515         .change         = taprio_change,
2516         .destroy        = taprio_destroy,
2517         .reset          = taprio_reset,
2518         .attach         = taprio_attach,
2519         .peek           = taprio_peek,
2520         .dequeue        = taprio_dequeue,
2521         .enqueue        = taprio_enqueue,
2522         .dump           = taprio_dump,
2523         .dump_stats     = taprio_dump_stats,
2524         .owner          = THIS_MODULE,
2525 };
2526
2527 static struct notifier_block taprio_device_notifier = {
2528         .notifier_call = taprio_dev_notifier,
2529 };
2530
2531 static int __init taprio_module_init(void)
2532 {
2533         int err = register_netdevice_notifier(&taprio_device_notifier);
2534
2535         if (err)
2536                 return err;
2537
2538         return register_qdisc(&taprio_qdisc_ops);
2539 }
2540
2541 static void __exit taprio_module_exit(void)
2542 {
2543         unregister_qdisc(&taprio_qdisc_ops);
2544         unregister_netdevice_notifier(&taprio_device_notifier);
2545 }
2546
2547 module_init(taprio_module_init);
2548 module_exit(taprio_module_exit);
2549 MODULE_LICENSE("GPL");