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]        = NLA_POLICY_MAX(NLA_U32,
1012                                                             TC_QOPT_MAX_QUEUE),
1013         [TCA_TAPRIO_TC_ENTRY_MAX_SDU]      = { .type = NLA_U32 },
1014         [TCA_TAPRIO_TC_ENTRY_FP]           = NLA_POLICY_RANGE(NLA_U32,
1015                                                               TC_FP_EXPRESS,
1016                                                               TC_FP_PREEMPTIBLE),
1017 };
1018
1019 static const struct netlink_range_validation_signed taprio_cycle_time_range = {
1020         .min = 0,
1021         .max = INT_MAX,
1022 };
1023
1024 static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
1025         [TCA_TAPRIO_ATTR_PRIOMAP]              = {
1026                 .len = sizeof(struct tc_mqprio_qopt)
1027         },
1028         [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST]           = { .type = NLA_NESTED },
1029         [TCA_TAPRIO_ATTR_SCHED_BASE_TIME]            = { .type = NLA_S64 },
1030         [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]         = { .type = NLA_NESTED },
1031         [TCA_TAPRIO_ATTR_SCHED_CLOCKID]              = { .type = NLA_S32 },
1032         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]           =
1033                 NLA_POLICY_FULL_RANGE_SIGNED(NLA_S64, &taprio_cycle_time_range),
1034         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
1035         [TCA_TAPRIO_ATTR_FLAGS]                      = { .type = NLA_U32 },
1036         [TCA_TAPRIO_ATTR_TXTIME_DELAY]               = { .type = NLA_U32 },
1037         [TCA_TAPRIO_ATTR_TC_ENTRY]                   = { .type = NLA_NESTED },
1038 };
1039
1040 static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
1041                             struct sched_entry *entry,
1042                             struct netlink_ext_ack *extack)
1043 {
1044         int min_duration = length_to_duration(q, ETH_ZLEN);
1045         u32 interval = 0;
1046
1047         if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
1048                 entry->command = nla_get_u8(
1049                         tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
1050
1051         if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
1052                 entry->gate_mask = nla_get_u32(
1053                         tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
1054
1055         if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
1056                 interval = nla_get_u32(
1057                         tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
1058
1059         /* The interval should allow at least the minimum ethernet
1060          * frame to go out.
1061          */
1062         if (interval < min_duration) {
1063                 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
1064                 return -EINVAL;
1065         }
1066
1067         entry->interval = interval;
1068
1069         return 0;
1070 }
1071
1072 static int parse_sched_entry(struct taprio_sched *q, struct nlattr *n,
1073                              struct sched_entry *entry, int index,
1074                              struct netlink_ext_ack *extack)
1075 {
1076         struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
1077         int err;
1078
1079         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
1080                                           entry_policy, NULL);
1081         if (err < 0) {
1082                 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
1083                 return -EINVAL;
1084         }
1085
1086         entry->index = index;
1087
1088         return fill_sched_entry(q, tb, entry, extack);
1089 }
1090
1091 static int parse_sched_list(struct taprio_sched *q, struct nlattr *list,
1092                             struct sched_gate_list *sched,
1093                             struct netlink_ext_ack *extack)
1094 {
1095         struct nlattr *n;
1096         int err, rem;
1097         int i = 0;
1098
1099         if (!list)
1100                 return -EINVAL;
1101
1102         nla_for_each_nested(n, list, rem) {
1103                 struct sched_entry *entry;
1104
1105                 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
1106                         NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
1107                         continue;
1108                 }
1109
1110                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1111                 if (!entry) {
1112                         NL_SET_ERR_MSG(extack, "Not enough memory for entry");
1113                         return -ENOMEM;
1114                 }
1115
1116                 err = parse_sched_entry(q, n, entry, i, extack);
1117                 if (err < 0) {
1118                         kfree(entry);
1119                         return err;
1120                 }
1121
1122                 list_add_tail(&entry->list, &sched->entries);
1123                 i++;
1124         }
1125
1126         sched->num_entries = i;
1127
1128         return i;
1129 }
1130
1131 static int parse_taprio_schedule(struct taprio_sched *q, struct nlattr **tb,
1132                                  struct sched_gate_list *new,
1133                                  struct netlink_ext_ack *extack)
1134 {
1135         int err = 0;
1136
1137         if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
1138                 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
1139                 return -ENOTSUPP;
1140         }
1141
1142         if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
1143                 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
1144
1145         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
1146                 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
1147
1148         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
1149                 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
1150
1151         if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
1152                 err = parse_sched_list(q, tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST],
1153                                        new, extack);
1154         if (err < 0)
1155                 return err;
1156
1157         if (!new->cycle_time) {
1158                 struct sched_entry *entry;
1159                 ktime_t cycle = 0;
1160
1161                 list_for_each_entry(entry, &new->entries, list)
1162                         cycle = ktime_add_ns(cycle, entry->interval);
1163
1164                 if (!cycle) {
1165                         NL_SET_ERR_MSG(extack, "'cycle_time' can never be 0");
1166                         return -EINVAL;
1167                 }
1168
1169                 if (cycle < 0 || cycle > INT_MAX) {
1170                         NL_SET_ERR_MSG(extack, "'cycle_time' is too big");
1171                         return -EINVAL;
1172                 }
1173
1174                 new->cycle_time = cycle;
1175         }
1176
1177         taprio_calculate_gate_durations(q, new);
1178
1179         return 0;
1180 }
1181
1182 static int taprio_parse_mqprio_opt(struct net_device *dev,
1183                                    struct tc_mqprio_qopt *qopt,
1184                                    struct netlink_ext_ack *extack,
1185                                    u32 taprio_flags)
1186 {
1187         bool allow_overlapping_txqs = TXTIME_ASSIST_IS_ENABLED(taprio_flags);
1188
1189         if (!qopt && !dev->num_tc) {
1190                 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
1191                 return -EINVAL;
1192         }
1193
1194         /* If num_tc is already set, it means that the user already
1195          * configured the mqprio part
1196          */
1197         if (dev->num_tc)
1198                 return 0;
1199
1200         /* taprio imposes that traffic classes map 1:n to tx queues */
1201         if (qopt->num_tc > dev->num_tx_queues) {
1202                 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
1203                 return -EINVAL;
1204         }
1205
1206         /* For some reason, in txtime-assist mode, we allow TXQ ranges for
1207          * different TCs to overlap, and just validate the TXQ ranges.
1208          */
1209         return mqprio_validate_qopt(dev, qopt, true, allow_overlapping_txqs,
1210                                     extack);
1211 }
1212
1213 static int taprio_get_start_time(struct Qdisc *sch,
1214                                  struct sched_gate_list *sched,
1215                                  ktime_t *start)
1216 {
1217         struct taprio_sched *q = qdisc_priv(sch);
1218         ktime_t now, base, cycle;
1219         s64 n;
1220
1221         base = sched_base_time(sched);
1222         now = taprio_get_time(q);
1223
1224         if (ktime_after(base, now)) {
1225                 *start = base;
1226                 return 0;
1227         }
1228
1229         cycle = sched->cycle_time;
1230
1231         /* The qdisc is expected to have at least one sched_entry.  Moreover,
1232          * any entry must have 'interval' > 0. Thus if the cycle time is zero,
1233          * something went really wrong. In that case, we should warn about this
1234          * inconsistent state and return error.
1235          */
1236         if (WARN_ON(!cycle))
1237                 return -EFAULT;
1238
1239         /* Schedule the start time for the beginning of the next
1240          * cycle.
1241          */
1242         n = div64_s64(ktime_sub_ns(now, base), cycle);
1243         *start = ktime_add_ns(base, (n + 1) * cycle);
1244         return 0;
1245 }
1246
1247 static void setup_first_end_time(struct taprio_sched *q,
1248                                  struct sched_gate_list *sched, ktime_t base)
1249 {
1250         struct net_device *dev = qdisc_dev(q->root);
1251         int num_tc = netdev_get_num_tc(dev);
1252         struct sched_entry *first;
1253         ktime_t cycle;
1254         int tc;
1255
1256         first = list_first_entry(&sched->entries,
1257                                  struct sched_entry, list);
1258
1259         cycle = sched->cycle_time;
1260
1261         /* FIXME: find a better place to do this */
1262         sched->cycle_end_time = ktime_add_ns(base, cycle);
1263
1264         first->end_time = ktime_add_ns(base, first->interval);
1265         taprio_set_budgets(q, sched, first);
1266
1267         for (tc = 0; tc < num_tc; tc++) {
1268                 if (first->gate_duration[tc] == sched->cycle_time)
1269                         first->gate_close_time[tc] = KTIME_MAX;
1270                 else
1271                         first->gate_close_time[tc] = ktime_add_ns(base, first->gate_duration[tc]);
1272         }
1273
1274         rcu_assign_pointer(q->current_entry, NULL);
1275 }
1276
1277 static void taprio_start_sched(struct Qdisc *sch,
1278                                ktime_t start, struct sched_gate_list *new)
1279 {
1280         struct taprio_sched *q = qdisc_priv(sch);
1281         ktime_t expires;
1282
1283         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1284                 return;
1285
1286         expires = hrtimer_get_expires(&q->advance_timer);
1287         if (expires == 0)
1288                 expires = KTIME_MAX;
1289
1290         /* If the new schedule starts before the next expiration, we
1291          * reprogram it to the earliest one, so we change the admin
1292          * schedule to the operational one at the right time.
1293          */
1294         start = min_t(ktime_t, start, expires);
1295
1296         hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1297 }
1298
1299 static void taprio_set_picos_per_byte(struct net_device *dev,
1300                                       struct taprio_sched *q)
1301 {
1302         struct ethtool_link_ksettings ecmd;
1303         int speed = SPEED_10;
1304         int picos_per_byte;
1305         int err;
1306
1307         err = __ethtool_get_link_ksettings(dev, &ecmd);
1308         if (err < 0)
1309                 goto skip;
1310
1311         if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1312                 speed = ecmd.base.speed;
1313
1314 skip:
1315         picos_per_byte = (USEC_PER_SEC * 8) / speed;
1316
1317         atomic64_set(&q->picos_per_byte, picos_per_byte);
1318         netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1319                    dev->name, (long long)atomic64_read(&q->picos_per_byte),
1320                    ecmd.base.speed);
1321 }
1322
1323 static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1324                                void *ptr)
1325 {
1326         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1327         struct sched_gate_list *oper, *admin;
1328         struct qdisc_size_table *stab;
1329         struct taprio_sched *q;
1330
1331         ASSERT_RTNL();
1332
1333         if (event != NETDEV_UP && event != NETDEV_CHANGE)
1334                 return NOTIFY_DONE;
1335
1336         list_for_each_entry(q, &taprio_list, taprio_list) {
1337                 if (dev != qdisc_dev(q->root))
1338                         continue;
1339
1340                 taprio_set_picos_per_byte(dev, q);
1341
1342                 stab = rtnl_dereference(q->root->stab);
1343
1344                 oper = rtnl_dereference(q->oper_sched);
1345                 if (oper)
1346                         taprio_update_queue_max_sdu(q, oper, stab);
1347
1348                 admin = rtnl_dereference(q->admin_sched);
1349                 if (admin)
1350                         taprio_update_queue_max_sdu(q, admin, stab);
1351
1352                 break;
1353         }
1354
1355         return NOTIFY_DONE;
1356 }
1357
1358 static void setup_txtime(struct taprio_sched *q,
1359                          struct sched_gate_list *sched, ktime_t base)
1360 {
1361         struct sched_entry *entry;
1362         u64 interval = 0;
1363
1364         list_for_each_entry(entry, &sched->entries, list) {
1365                 entry->next_txtime = ktime_add_ns(base, interval);
1366                 interval += entry->interval;
1367         }
1368 }
1369
1370 static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1371 {
1372         struct __tc_taprio_qopt_offload *__offload;
1373
1374         __offload = kzalloc(struct_size(__offload, offload.entries, num_entries),
1375                             GFP_KERNEL);
1376         if (!__offload)
1377                 return NULL;
1378
1379         refcount_set(&__offload->users, 1);
1380
1381         return &__offload->offload;
1382 }
1383
1384 struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1385                                                   *offload)
1386 {
1387         struct __tc_taprio_qopt_offload *__offload;
1388
1389         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1390                                  offload);
1391
1392         refcount_inc(&__offload->users);
1393
1394         return offload;
1395 }
1396 EXPORT_SYMBOL_GPL(taprio_offload_get);
1397
1398 void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1399 {
1400         struct __tc_taprio_qopt_offload *__offload;
1401
1402         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1403                                  offload);
1404
1405         if (!refcount_dec_and_test(&__offload->users))
1406                 return;
1407
1408         kfree(__offload);
1409 }
1410 EXPORT_SYMBOL_GPL(taprio_offload_free);
1411
1412 /* The function will only serve to keep the pointers to the "oper" and "admin"
1413  * schedules valid in relation to their base times, so when calling dump() the
1414  * users looks at the right schedules.
1415  * When using full offload, the admin configuration is promoted to oper at the
1416  * base_time in the PHC time domain.  But because the system time is not
1417  * necessarily in sync with that, we can't just trigger a hrtimer to call
1418  * switch_schedules at the right hardware time.
1419  * At the moment we call this by hand right away from taprio, but in the future
1420  * it will be useful to create a mechanism for drivers to notify taprio of the
1421  * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1422  * This is left as TODO.
1423  */
1424 static void taprio_offload_config_changed(struct taprio_sched *q)
1425 {
1426         struct sched_gate_list *oper, *admin;
1427
1428         oper = rtnl_dereference(q->oper_sched);
1429         admin = rtnl_dereference(q->admin_sched);
1430
1431         switch_schedules(q, &admin, &oper);
1432 }
1433
1434 static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
1435 {
1436         u32 i, queue_mask = 0;
1437
1438         for (i = 0; i < dev->num_tc; i++) {
1439                 u32 offset, count;
1440
1441                 if (!(tc_mask & BIT(i)))
1442                         continue;
1443
1444                 offset = dev->tc_to_txq[i].offset;
1445                 count = dev->tc_to_txq[i].count;
1446
1447                 queue_mask |= GENMASK(offset + count - 1, offset);
1448         }
1449
1450         return queue_mask;
1451 }
1452
1453 static void taprio_sched_to_offload(struct net_device *dev,
1454                                     struct sched_gate_list *sched,
1455                                     struct tc_taprio_qopt_offload *offload,
1456                                     const struct tc_taprio_caps *caps)
1457 {
1458         struct sched_entry *entry;
1459         int i = 0;
1460
1461         offload->base_time = sched->base_time;
1462         offload->cycle_time = sched->cycle_time;
1463         offload->cycle_time_extension = sched->cycle_time_extension;
1464
1465         list_for_each_entry(entry, &sched->entries, list) {
1466                 struct tc_taprio_sched_entry *e = &offload->entries[i];
1467
1468                 e->command = entry->command;
1469                 e->interval = entry->interval;
1470                 if (caps->gate_mask_per_txq)
1471                         e->gate_mask = tc_map_to_queue_mask(dev,
1472                                                             entry->gate_mask);
1473                 else
1474                         e->gate_mask = entry->gate_mask;
1475
1476                 i++;
1477         }
1478
1479         offload->num_entries = i;
1480 }
1481
1482 static void taprio_detect_broken_mqprio(struct taprio_sched *q)
1483 {
1484         struct net_device *dev = qdisc_dev(q->root);
1485         struct tc_taprio_caps caps;
1486
1487         qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1488                                  &caps, sizeof(caps));
1489
1490         q->broken_mqprio = caps.broken_mqprio;
1491         if (q->broken_mqprio)
1492                 static_branch_inc(&taprio_have_broken_mqprio);
1493         else
1494                 static_branch_inc(&taprio_have_working_mqprio);
1495
1496         q->detected_mqprio = true;
1497 }
1498
1499 static void taprio_cleanup_broken_mqprio(struct taprio_sched *q)
1500 {
1501         if (!q->detected_mqprio)
1502                 return;
1503
1504         if (q->broken_mqprio)
1505                 static_branch_dec(&taprio_have_broken_mqprio);
1506         else
1507                 static_branch_dec(&taprio_have_working_mqprio);
1508 }
1509
1510 static int taprio_enable_offload(struct net_device *dev,
1511                                  struct taprio_sched *q,
1512                                  struct sched_gate_list *sched,
1513                                  struct netlink_ext_ack *extack)
1514 {
1515         const struct net_device_ops *ops = dev->netdev_ops;
1516         struct tc_taprio_qopt_offload *offload;
1517         struct tc_taprio_caps caps;
1518         int tc, err = 0;
1519
1520         if (!ops->ndo_setup_tc) {
1521                 NL_SET_ERR_MSG(extack,
1522                                "Device does not support taprio offload");
1523                 return -EOPNOTSUPP;
1524         }
1525
1526         qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1527                                  &caps, sizeof(caps));
1528
1529         if (!caps.supports_queue_max_sdu) {
1530                 for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
1531                         if (q->max_sdu[tc]) {
1532                                 NL_SET_ERR_MSG_MOD(extack,
1533                                                    "Device does not handle queueMaxSDU");
1534                                 return -EOPNOTSUPP;
1535                         }
1536                 }
1537         }
1538
1539         offload = taprio_offload_alloc(sched->num_entries);
1540         if (!offload) {
1541                 NL_SET_ERR_MSG(extack,
1542                                "Not enough memory for enabling offload mode");
1543                 return -ENOMEM;
1544         }
1545         offload->cmd = TAPRIO_CMD_REPLACE;
1546         offload->extack = extack;
1547         mqprio_qopt_reconstruct(dev, &offload->mqprio.qopt);
1548         offload->mqprio.extack = extack;
1549         taprio_sched_to_offload(dev, sched, offload, &caps);
1550         mqprio_fp_to_offload(q->fp, &offload->mqprio);
1551
1552         for (tc = 0; tc < TC_MAX_QUEUE; tc++)
1553                 offload->max_sdu[tc] = q->max_sdu[tc];
1554
1555         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1556         if (err < 0) {
1557                 NL_SET_ERR_MSG_WEAK(extack,
1558                                     "Device failed to setup taprio offload");
1559                 goto done;
1560         }
1561
1562         q->offloaded = true;
1563
1564 done:
1565         /* The offload structure may linger around via a reference taken by the
1566          * device driver, so clear up the netlink extack pointer so that the
1567          * driver isn't tempted to dereference data which stopped being valid
1568          */
1569         offload->extack = NULL;
1570         offload->mqprio.extack = NULL;
1571         taprio_offload_free(offload);
1572
1573         return err;
1574 }
1575
1576 static int taprio_disable_offload(struct net_device *dev,
1577                                   struct taprio_sched *q,
1578                                   struct netlink_ext_ack *extack)
1579 {
1580         const struct net_device_ops *ops = dev->netdev_ops;
1581         struct tc_taprio_qopt_offload *offload;
1582         int err;
1583
1584         if (!q->offloaded)
1585                 return 0;
1586
1587         offload = taprio_offload_alloc(0);
1588         if (!offload) {
1589                 NL_SET_ERR_MSG(extack,
1590                                "Not enough memory to disable offload mode");
1591                 return -ENOMEM;
1592         }
1593         offload->cmd = TAPRIO_CMD_DESTROY;
1594
1595         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1596         if (err < 0) {
1597                 NL_SET_ERR_MSG(extack,
1598                                "Device failed to disable offload");
1599                 goto out;
1600         }
1601
1602         q->offloaded = false;
1603
1604 out:
1605         taprio_offload_free(offload);
1606
1607         return err;
1608 }
1609
1610 /* If full offload is enabled, the only possible clockid is the net device's
1611  * PHC. For that reason, specifying a clockid through netlink is incorrect.
1612  * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1613  * in sync with the specified clockid via a user space daemon such as phc2sys.
1614  * For both software taprio and txtime-assist, the clockid is used for the
1615  * hrtimer that advances the schedule and hence mandatory.
1616  */
1617 static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1618                                 struct netlink_ext_ack *extack)
1619 {
1620         struct taprio_sched *q = qdisc_priv(sch);
1621         struct net_device *dev = qdisc_dev(sch);
1622         int err = -EINVAL;
1623
1624         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1625                 const struct ethtool_ops *ops = dev->ethtool_ops;
1626                 struct ethtool_ts_info info = {
1627                         .cmd = ETHTOOL_GET_TS_INFO,
1628                         .phc_index = -1,
1629                 };
1630
1631                 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1632                         NL_SET_ERR_MSG(extack,
1633                                        "The 'clockid' cannot be specified for full offload");
1634                         goto out;
1635                 }
1636
1637                 if (ops && ops->get_ts_info)
1638                         err = ops->get_ts_info(dev, &info);
1639
1640                 if (err || info.phc_index < 0) {
1641                         NL_SET_ERR_MSG(extack,
1642                                        "Device does not have a PTP clock");
1643                         err = -ENOTSUPP;
1644                         goto out;
1645                 }
1646         } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1647                 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1648                 enum tk_offsets tk_offset;
1649
1650                 /* We only support static clockids and we don't allow
1651                  * for it to be modified after the first init.
1652                  */
1653                 if (clockid < 0 ||
1654                     (q->clockid != -1 && q->clockid != clockid)) {
1655                         NL_SET_ERR_MSG(extack,
1656                                        "Changing the 'clockid' of a running schedule is not supported");
1657                         err = -ENOTSUPP;
1658                         goto out;
1659                 }
1660
1661                 switch (clockid) {
1662                 case CLOCK_REALTIME:
1663                         tk_offset = TK_OFFS_REAL;
1664                         break;
1665                 case CLOCK_MONOTONIC:
1666                         tk_offset = TK_OFFS_MAX;
1667                         break;
1668                 case CLOCK_BOOTTIME:
1669                         tk_offset = TK_OFFS_BOOT;
1670                         break;
1671                 case CLOCK_TAI:
1672                         tk_offset = TK_OFFS_TAI;
1673                         break;
1674                 default:
1675                         NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1676                         err = -EINVAL;
1677                         goto out;
1678                 }
1679                 /* This pairs with READ_ONCE() in taprio_mono_to_any */
1680                 WRITE_ONCE(q->tk_offset, tk_offset);
1681
1682                 q->clockid = clockid;
1683         } else {
1684                 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1685                 goto out;
1686         }
1687
1688         /* Everything went ok, return success. */
1689         err = 0;
1690
1691 out:
1692         return err;
1693 }
1694
1695 static int taprio_parse_tc_entry(struct Qdisc *sch,
1696                                  struct nlattr *opt,
1697                                  u32 max_sdu[TC_QOPT_MAX_QUEUE],
1698                                  u32 fp[TC_QOPT_MAX_QUEUE],
1699                                  unsigned long *seen_tcs,
1700                                  struct netlink_ext_ack *extack)
1701 {
1702         struct nlattr *tb[TCA_TAPRIO_TC_ENTRY_MAX + 1] = { };
1703         struct net_device *dev = qdisc_dev(sch);
1704         int err, tc;
1705         u32 val;
1706
1707         err = nla_parse_nested(tb, TCA_TAPRIO_TC_ENTRY_MAX, opt,
1708                                taprio_tc_policy, extack);
1709         if (err < 0)
1710                 return err;
1711
1712         if (!tb[TCA_TAPRIO_TC_ENTRY_INDEX]) {
1713                 NL_SET_ERR_MSG_MOD(extack, "TC entry index missing");
1714                 return -EINVAL;
1715         }
1716
1717         tc = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_INDEX]);
1718         if (tc >= TC_QOPT_MAX_QUEUE) {
1719                 NL_SET_ERR_MSG_MOD(extack, "TC entry index out of range");
1720                 return -ERANGE;
1721         }
1722
1723         if (*seen_tcs & BIT(tc)) {
1724                 NL_SET_ERR_MSG_MOD(extack, "Duplicate TC entry");
1725                 return -EINVAL;
1726         }
1727
1728         *seen_tcs |= BIT(tc);
1729
1730         if (tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]) {
1731                 val = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]);
1732                 if (val > dev->max_mtu) {
1733                         NL_SET_ERR_MSG_MOD(extack, "TC max SDU exceeds device max MTU");
1734                         return -ERANGE;
1735                 }
1736
1737                 max_sdu[tc] = val;
1738         }
1739
1740         if (tb[TCA_TAPRIO_TC_ENTRY_FP])
1741                 fp[tc] = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_FP]);
1742
1743         return 0;
1744 }
1745
1746 static int taprio_parse_tc_entries(struct Qdisc *sch,
1747                                    struct nlattr *opt,
1748                                    struct netlink_ext_ack *extack)
1749 {
1750         struct taprio_sched *q = qdisc_priv(sch);
1751         struct net_device *dev = qdisc_dev(sch);
1752         u32 max_sdu[TC_QOPT_MAX_QUEUE];
1753         bool have_preemption = false;
1754         unsigned long seen_tcs = 0;
1755         u32 fp[TC_QOPT_MAX_QUEUE];
1756         struct nlattr *n;
1757         int tc, rem;
1758         int err = 0;
1759
1760         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1761                 max_sdu[tc] = q->max_sdu[tc];
1762                 fp[tc] = q->fp[tc];
1763         }
1764
1765         nla_for_each_nested(n, opt, rem) {
1766                 if (nla_type(n) != TCA_TAPRIO_ATTR_TC_ENTRY)
1767                         continue;
1768
1769                 err = taprio_parse_tc_entry(sch, n, max_sdu, fp, &seen_tcs,
1770                                             extack);
1771                 if (err)
1772                         return err;
1773         }
1774
1775         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1776                 q->max_sdu[tc] = max_sdu[tc];
1777                 q->fp[tc] = fp[tc];
1778                 if (fp[tc] != TC_FP_EXPRESS)
1779                         have_preemption = true;
1780         }
1781
1782         if (have_preemption) {
1783                 if (!FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1784                         NL_SET_ERR_MSG(extack,
1785                                        "Preemption only supported with full offload");
1786                         return -EOPNOTSUPP;
1787                 }
1788
1789                 if (!ethtool_dev_mm_supported(dev)) {
1790                         NL_SET_ERR_MSG(extack,
1791                                        "Device does not support preemption");
1792                         return -EOPNOTSUPP;
1793                 }
1794         }
1795
1796         return err;
1797 }
1798
1799 static int taprio_mqprio_cmp(const struct net_device *dev,
1800                              const struct tc_mqprio_qopt *mqprio)
1801 {
1802         int i;
1803
1804         if (!mqprio || mqprio->num_tc != dev->num_tc)
1805                 return -1;
1806
1807         for (i = 0; i < mqprio->num_tc; i++)
1808                 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1809                     dev->tc_to_txq[i].offset != mqprio->offset[i])
1810                         return -1;
1811
1812         for (i = 0; i <= TC_BITMASK; i++)
1813                 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1814                         return -1;
1815
1816         return 0;
1817 }
1818
1819 /* The semantics of the 'flags' argument in relation to 'change()'
1820  * requests, are interpreted following two rules (which are applied in
1821  * this order): (1) an omitted 'flags' argument is interpreted as
1822  * zero; (2) the 'flags' of a "running" taprio instance cannot be
1823  * changed.
1824  */
1825 static int taprio_new_flags(const struct nlattr *attr, u32 old,
1826                             struct netlink_ext_ack *extack)
1827 {
1828         u32 new = 0;
1829
1830         if (attr)
1831                 new = nla_get_u32(attr);
1832
1833         if (old != TAPRIO_FLAGS_INVALID && old != new) {
1834                 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1835                 return -EOPNOTSUPP;
1836         }
1837
1838         if (!taprio_flags_valid(new)) {
1839                 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1840                 return -EINVAL;
1841         }
1842
1843         return new;
1844 }
1845
1846 static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1847                          struct netlink_ext_ack *extack)
1848 {
1849         struct qdisc_size_table *stab = rtnl_dereference(sch->stab);
1850         struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1851         struct sched_gate_list *oper, *admin, *new_admin;
1852         struct taprio_sched *q = qdisc_priv(sch);
1853         struct net_device *dev = qdisc_dev(sch);
1854         struct tc_mqprio_qopt *mqprio = NULL;
1855         unsigned long flags;
1856         ktime_t start;
1857         int i, err;
1858
1859         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1860                                           taprio_policy, extack);
1861         if (err < 0)
1862                 return err;
1863
1864         if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1865                 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1866
1867         err = taprio_new_flags(tb[TCA_TAPRIO_ATTR_FLAGS],
1868                                q->flags, extack);
1869         if (err < 0)
1870                 return err;
1871
1872         q->flags = err;
1873
1874         err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1875         if (err < 0)
1876                 return err;
1877
1878         err = taprio_parse_tc_entries(sch, opt, extack);
1879         if (err)
1880                 return err;
1881
1882         new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1883         if (!new_admin) {
1884                 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1885                 return -ENOMEM;
1886         }
1887         INIT_LIST_HEAD(&new_admin->entries);
1888
1889         oper = rtnl_dereference(q->oper_sched);
1890         admin = rtnl_dereference(q->admin_sched);
1891
1892         /* no changes - no new mqprio settings */
1893         if (!taprio_mqprio_cmp(dev, mqprio))
1894                 mqprio = NULL;
1895
1896         if (mqprio && (oper || admin)) {
1897                 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1898                 err = -ENOTSUPP;
1899                 goto free_sched;
1900         }
1901
1902         if (mqprio) {
1903                 err = netdev_set_num_tc(dev, mqprio->num_tc);
1904                 if (err)
1905                         goto free_sched;
1906                 for (i = 0; i < mqprio->num_tc; i++) {
1907                         netdev_set_tc_queue(dev, i,
1908                                             mqprio->count[i],
1909                                             mqprio->offset[i]);
1910                         q->cur_txq[i] = mqprio->offset[i];
1911                 }
1912
1913                 /* Always use supplied priority mappings */
1914                 for (i = 0; i <= TC_BITMASK; i++)
1915                         netdev_set_prio_tc_map(dev, i,
1916                                                mqprio->prio_tc_map[i]);
1917         }
1918
1919         err = parse_taprio_schedule(q, tb, new_admin, extack);
1920         if (err < 0)
1921                 goto free_sched;
1922
1923         if (new_admin->num_entries == 0) {
1924                 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1925                 err = -EINVAL;
1926                 goto free_sched;
1927         }
1928
1929         err = taprio_parse_clockid(sch, tb, extack);
1930         if (err < 0)
1931                 goto free_sched;
1932
1933         taprio_set_picos_per_byte(dev, q);
1934         taprio_update_queue_max_sdu(q, new_admin, stab);
1935
1936         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1937                 err = taprio_enable_offload(dev, q, new_admin, extack);
1938         else
1939                 err = taprio_disable_offload(dev, q, extack);
1940         if (err)
1941                 goto free_sched;
1942
1943         /* Protects against enqueue()/dequeue() */
1944         spin_lock_bh(qdisc_lock(sch));
1945
1946         if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1947                 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1948                         NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1949                         err = -EINVAL;
1950                         goto unlock;
1951                 }
1952
1953                 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1954         }
1955
1956         if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1957             !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1958             !hrtimer_active(&q->advance_timer)) {
1959                 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1960                 q->advance_timer.function = advance_sched;
1961         }
1962
1963         err = taprio_get_start_time(sch, new_admin, &start);
1964         if (err < 0) {
1965                 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1966                 goto unlock;
1967         }
1968
1969         setup_txtime(q, new_admin, start);
1970
1971         if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1972                 if (!oper) {
1973                         rcu_assign_pointer(q->oper_sched, new_admin);
1974                         err = 0;
1975                         new_admin = NULL;
1976                         goto unlock;
1977                 }
1978
1979                 rcu_assign_pointer(q->admin_sched, new_admin);
1980                 if (admin)
1981                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1982         } else {
1983                 setup_first_end_time(q, new_admin, start);
1984
1985                 /* Protects against advance_sched() */
1986                 spin_lock_irqsave(&q->current_entry_lock, flags);
1987
1988                 taprio_start_sched(sch, start, new_admin);
1989
1990                 rcu_assign_pointer(q->admin_sched, new_admin);
1991                 if (admin)
1992                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1993
1994                 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1995
1996                 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1997                         taprio_offload_config_changed(q);
1998         }
1999
2000         new_admin = NULL;
2001         err = 0;
2002
2003         if (!stab)
2004                 NL_SET_ERR_MSG_MOD(extack,
2005                                    "Size table not specified, frame length estimations may be inaccurate");
2006
2007 unlock:
2008         spin_unlock_bh(qdisc_lock(sch));
2009
2010 free_sched:
2011         if (new_admin)
2012                 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
2013
2014         return err;
2015 }
2016
2017 static void taprio_reset(struct Qdisc *sch)
2018 {
2019         struct taprio_sched *q = qdisc_priv(sch);
2020         struct net_device *dev = qdisc_dev(sch);
2021         int i;
2022
2023         hrtimer_cancel(&q->advance_timer);
2024
2025         if (q->qdiscs) {
2026                 for (i = 0; i < dev->num_tx_queues; i++)
2027                         if (q->qdiscs[i])
2028                                 qdisc_reset(q->qdiscs[i]);
2029         }
2030 }
2031
2032 static void taprio_destroy(struct Qdisc *sch)
2033 {
2034         struct taprio_sched *q = qdisc_priv(sch);
2035         struct net_device *dev = qdisc_dev(sch);
2036         struct sched_gate_list *oper, *admin;
2037         unsigned int i;
2038
2039         list_del(&q->taprio_list);
2040
2041         /* Note that taprio_reset() might not be called if an error
2042          * happens in qdisc_create(), after taprio_init() has been called.
2043          */
2044         hrtimer_cancel(&q->advance_timer);
2045         qdisc_synchronize(sch);
2046
2047         taprio_disable_offload(dev, q, NULL);
2048
2049         if (q->qdiscs) {
2050                 for (i = 0; i < dev->num_tx_queues; i++)
2051                         qdisc_put(q->qdiscs[i]);
2052
2053                 kfree(q->qdiscs);
2054         }
2055         q->qdiscs = NULL;
2056
2057         netdev_reset_tc(dev);
2058
2059         oper = rtnl_dereference(q->oper_sched);
2060         admin = rtnl_dereference(q->admin_sched);
2061
2062         if (oper)
2063                 call_rcu(&oper->rcu, taprio_free_sched_cb);
2064
2065         if (admin)
2066                 call_rcu(&admin->rcu, taprio_free_sched_cb);
2067
2068         taprio_cleanup_broken_mqprio(q);
2069 }
2070
2071 static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
2072                        struct netlink_ext_ack *extack)
2073 {
2074         struct taprio_sched *q = qdisc_priv(sch);
2075         struct net_device *dev = qdisc_dev(sch);
2076         int i, tc;
2077
2078         spin_lock_init(&q->current_entry_lock);
2079
2080         hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
2081         q->advance_timer.function = advance_sched;
2082
2083         q->root = sch;
2084
2085         /* We only support static clockids. Use an invalid value as default
2086          * and get the valid one on taprio_change().
2087          */
2088         q->clockid = -1;
2089         q->flags = TAPRIO_FLAGS_INVALID;
2090
2091         list_add(&q->taprio_list, &taprio_list);
2092
2093         if (sch->parent != TC_H_ROOT) {
2094                 NL_SET_ERR_MSG_MOD(extack, "Can only be attached as root qdisc");
2095                 return -EOPNOTSUPP;
2096         }
2097
2098         if (!netif_is_multiqueue(dev)) {
2099                 NL_SET_ERR_MSG_MOD(extack, "Multi-queue device is required");
2100                 return -EOPNOTSUPP;
2101         }
2102
2103         q->qdiscs = kcalloc(dev->num_tx_queues, sizeof(q->qdiscs[0]),
2104                             GFP_KERNEL);
2105         if (!q->qdiscs)
2106                 return -ENOMEM;
2107
2108         if (!opt)
2109                 return -EINVAL;
2110
2111         for (i = 0; i < dev->num_tx_queues; i++) {
2112                 struct netdev_queue *dev_queue;
2113                 struct Qdisc *qdisc;
2114
2115                 dev_queue = netdev_get_tx_queue(dev, i);
2116                 qdisc = qdisc_create_dflt(dev_queue,
2117                                           &pfifo_qdisc_ops,
2118                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
2119                                                     TC_H_MIN(i + 1)),
2120                                           extack);
2121                 if (!qdisc)
2122                         return -ENOMEM;
2123
2124                 if (i < dev->real_num_tx_queues)
2125                         qdisc_hash_add(qdisc, false);
2126
2127                 q->qdiscs[i] = qdisc;
2128         }
2129
2130         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++)
2131                 q->fp[tc] = TC_FP_EXPRESS;
2132
2133         taprio_detect_broken_mqprio(q);
2134
2135         return taprio_change(sch, opt, extack);
2136 }
2137
2138 static void taprio_attach(struct Qdisc *sch)
2139 {
2140         struct taprio_sched *q = qdisc_priv(sch);
2141         struct net_device *dev = qdisc_dev(sch);
2142         unsigned int ntx;
2143
2144         /* Attach underlying qdisc */
2145         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
2146                 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
2147                 struct Qdisc *old, *dev_queue_qdisc;
2148
2149                 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2150                         struct Qdisc *qdisc = q->qdiscs[ntx];
2151
2152                         /* In offload mode, the root taprio qdisc is bypassed
2153                          * and the netdev TX queues see the children directly
2154                          */
2155                         qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2156                         dev_queue_qdisc = qdisc;
2157                 } else {
2158                         /* In software mode, attach the root taprio qdisc
2159                          * to all netdev TX queues, so that dev_qdisc_enqueue()
2160                          * goes through taprio_enqueue().
2161                          */
2162                         dev_queue_qdisc = sch;
2163                 }
2164                 old = dev_graft_qdisc(dev_queue, dev_queue_qdisc);
2165                 /* The qdisc's refcount requires to be elevated once
2166                  * for each netdev TX queue it is grafted onto
2167                  */
2168                 qdisc_refcount_inc(dev_queue_qdisc);
2169                 if (old)
2170                         qdisc_put(old);
2171         }
2172 }
2173
2174 static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
2175                                              unsigned long cl)
2176 {
2177         struct net_device *dev = qdisc_dev(sch);
2178         unsigned long ntx = cl - 1;
2179
2180         if (ntx >= dev->num_tx_queues)
2181                 return NULL;
2182
2183         return netdev_get_tx_queue(dev, ntx);
2184 }
2185
2186 static int taprio_graft(struct Qdisc *sch, unsigned long cl,
2187                         struct Qdisc *new, struct Qdisc **old,
2188                         struct netlink_ext_ack *extack)
2189 {
2190         struct taprio_sched *q = qdisc_priv(sch);
2191         struct net_device *dev = qdisc_dev(sch);
2192         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2193
2194         if (!dev_queue)
2195                 return -EINVAL;
2196
2197         if (dev->flags & IFF_UP)
2198                 dev_deactivate(dev);
2199
2200         /* In offload mode, the child Qdisc is directly attached to the netdev
2201          * TX queue, and thus, we need to keep its refcount elevated in order
2202          * to counteract qdisc_graft()'s call to qdisc_put() once per TX queue.
2203          * However, save the reference to the new qdisc in the private array in
2204          * both software and offload cases, to have an up-to-date reference to
2205          * our children.
2206          */
2207         *old = q->qdiscs[cl - 1];
2208         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2209                 WARN_ON_ONCE(dev_graft_qdisc(dev_queue, new) != *old);
2210                 if (new)
2211                         qdisc_refcount_inc(new);
2212                 if (*old)
2213                         qdisc_put(*old);
2214         }
2215
2216         q->qdiscs[cl - 1] = new;
2217         if (new)
2218                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2219
2220         if (dev->flags & IFF_UP)
2221                 dev_activate(dev);
2222
2223         return 0;
2224 }
2225
2226 static int dump_entry(struct sk_buff *msg,
2227                       const struct sched_entry *entry)
2228 {
2229         struct nlattr *item;
2230
2231         item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
2232         if (!item)
2233                 return -ENOSPC;
2234
2235         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
2236                 goto nla_put_failure;
2237
2238         if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
2239                 goto nla_put_failure;
2240
2241         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
2242                         entry->gate_mask))
2243                 goto nla_put_failure;
2244
2245         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
2246                         entry->interval))
2247                 goto nla_put_failure;
2248
2249         return nla_nest_end(msg, item);
2250
2251 nla_put_failure:
2252         nla_nest_cancel(msg, item);
2253         return -1;
2254 }
2255
2256 static int dump_schedule(struct sk_buff *msg,
2257                          const struct sched_gate_list *root)
2258 {
2259         struct nlattr *entry_list;
2260         struct sched_entry *entry;
2261
2262         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
2263                         root->base_time, TCA_TAPRIO_PAD))
2264                 return -1;
2265
2266         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
2267                         root->cycle_time, TCA_TAPRIO_PAD))
2268                 return -1;
2269
2270         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
2271                         root->cycle_time_extension, TCA_TAPRIO_PAD))
2272                 return -1;
2273
2274         entry_list = nla_nest_start_noflag(msg,
2275                                            TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
2276         if (!entry_list)
2277                 goto error_nest;
2278
2279         list_for_each_entry(entry, &root->entries, list) {
2280                 if (dump_entry(msg, entry) < 0)
2281                         goto error_nest;
2282         }
2283
2284         nla_nest_end(msg, entry_list);
2285         return 0;
2286
2287 error_nest:
2288         nla_nest_cancel(msg, entry_list);
2289         return -1;
2290 }
2291
2292 static int taprio_dump_tc_entries(struct sk_buff *skb,
2293                                   struct taprio_sched *q,
2294                                   struct sched_gate_list *sched)
2295 {
2296         struct nlattr *n;
2297         int tc;
2298
2299         for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
2300                 n = nla_nest_start(skb, TCA_TAPRIO_ATTR_TC_ENTRY);
2301                 if (!n)
2302                         return -EMSGSIZE;
2303
2304                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_INDEX, tc))
2305                         goto nla_put_failure;
2306
2307                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_MAX_SDU,
2308                                 sched->max_sdu[tc]))
2309                         goto nla_put_failure;
2310
2311                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_FP, q->fp[tc]))
2312                         goto nla_put_failure;
2313
2314                 nla_nest_end(skb, n);
2315         }
2316
2317         return 0;
2318
2319 nla_put_failure:
2320         nla_nest_cancel(skb, n);
2321         return -EMSGSIZE;
2322 }
2323
2324 static int taprio_put_stat(struct sk_buff *skb, u64 val, u16 attrtype)
2325 {
2326         if (val == TAPRIO_STAT_NOT_SET)
2327                 return 0;
2328         if (nla_put_u64_64bit(skb, attrtype, val, TCA_TAPRIO_OFFLOAD_STATS_PAD))
2329                 return -EMSGSIZE;
2330         return 0;
2331 }
2332
2333 static int taprio_dump_xstats(struct Qdisc *sch, struct gnet_dump *d,
2334                               struct tc_taprio_qopt_offload *offload,
2335                               struct tc_taprio_qopt_stats *stats)
2336 {
2337         struct net_device *dev = qdisc_dev(sch);
2338         const struct net_device_ops *ops;
2339         struct sk_buff *skb = d->skb;
2340         struct nlattr *xstats;
2341         int err;
2342
2343         ops = qdisc_dev(sch)->netdev_ops;
2344
2345         /* FIXME I could use qdisc_offload_dump_helper(), but that messes
2346          * with sch->flags depending on whether the device reports taprio
2347          * stats, and I'm not sure whether that's a good idea, considering
2348          * that stats are optional to the offload itself
2349          */
2350         if (!ops->ndo_setup_tc)
2351                 return 0;
2352
2353         memset(stats, 0xff, sizeof(*stats));
2354
2355         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
2356         if (err == -EOPNOTSUPP)
2357                 return 0;
2358         if (err)
2359                 return err;
2360
2361         xstats = nla_nest_start(skb, TCA_STATS_APP);
2362         if (!xstats)
2363                 goto err;
2364
2365         if (taprio_put_stat(skb, stats->window_drops,
2366                             TCA_TAPRIO_OFFLOAD_STATS_WINDOW_DROPS) ||
2367             taprio_put_stat(skb, stats->tx_overruns,
2368                             TCA_TAPRIO_OFFLOAD_STATS_TX_OVERRUNS))
2369                 goto err_cancel;
2370
2371         nla_nest_end(skb, xstats);
2372
2373         return 0;
2374
2375 err_cancel:
2376         nla_nest_cancel(skb, xstats);
2377 err:
2378         return -EMSGSIZE;
2379 }
2380
2381 static int taprio_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
2382 {
2383         struct tc_taprio_qopt_offload offload = {
2384                 .cmd = TAPRIO_CMD_STATS,
2385         };
2386
2387         return taprio_dump_xstats(sch, d, &offload, &offload.stats);
2388 }
2389
2390 static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
2391 {
2392         struct taprio_sched *q = qdisc_priv(sch);
2393         struct net_device *dev = qdisc_dev(sch);
2394         struct sched_gate_list *oper, *admin;
2395         struct tc_mqprio_qopt opt = { 0 };
2396         struct nlattr *nest, *sched_nest;
2397
2398         oper = rtnl_dereference(q->oper_sched);
2399         admin = rtnl_dereference(q->admin_sched);
2400
2401         mqprio_qopt_reconstruct(dev, &opt);
2402
2403         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2404         if (!nest)
2405                 goto start_error;
2406
2407         if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
2408                 goto options_error;
2409
2410         if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
2411             nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
2412                 goto options_error;
2413
2414         if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
2415                 goto options_error;
2416
2417         if (q->txtime_delay &&
2418             nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
2419                 goto options_error;
2420
2421         if (oper && taprio_dump_tc_entries(skb, q, oper))
2422                 goto options_error;
2423
2424         if (oper && dump_schedule(skb, oper))
2425                 goto options_error;
2426
2427         if (!admin)
2428                 goto done;
2429
2430         sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
2431         if (!sched_nest)
2432                 goto options_error;
2433
2434         if (dump_schedule(skb, admin))
2435                 goto admin_error;
2436
2437         nla_nest_end(skb, sched_nest);
2438
2439 done:
2440         return nla_nest_end(skb, nest);
2441
2442 admin_error:
2443         nla_nest_cancel(skb, sched_nest);
2444
2445 options_error:
2446         nla_nest_cancel(skb, nest);
2447
2448 start_error:
2449         return -ENOSPC;
2450 }
2451
2452 static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
2453 {
2454         struct taprio_sched *q = qdisc_priv(sch);
2455         struct net_device *dev = qdisc_dev(sch);
2456         unsigned int ntx = cl - 1;
2457
2458         if (ntx >= dev->num_tx_queues)
2459                 return NULL;
2460
2461         return q->qdiscs[ntx];
2462 }
2463
2464 static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
2465 {
2466         unsigned int ntx = TC_H_MIN(classid);
2467
2468         if (!taprio_queue_get(sch, ntx))
2469                 return 0;
2470         return ntx;
2471 }
2472
2473 static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
2474                              struct sk_buff *skb, struct tcmsg *tcm)
2475 {
2476         struct Qdisc *child = taprio_leaf(sch, cl);
2477
2478         tcm->tcm_parent = TC_H_ROOT;
2479         tcm->tcm_handle |= TC_H_MIN(cl);
2480         tcm->tcm_info = child->handle;
2481
2482         return 0;
2483 }
2484
2485 static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
2486                                    struct gnet_dump *d)
2487         __releases(d->lock)
2488         __acquires(d->lock)
2489 {
2490         struct Qdisc *child = taprio_leaf(sch, cl);
2491         struct tc_taprio_qopt_offload offload = {
2492                 .cmd = TAPRIO_CMD_QUEUE_STATS,
2493                 .queue_stats = {
2494                         .queue = cl - 1,
2495                 },
2496         };
2497
2498         if (gnet_stats_copy_basic(d, NULL, &child->bstats, true) < 0 ||
2499             qdisc_qstats_copy(d, child) < 0)
2500                 return -1;
2501
2502         return taprio_dump_xstats(sch, d, &offload, &offload.queue_stats.stats);
2503 }
2504
2505 static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2506 {
2507         struct net_device *dev = qdisc_dev(sch);
2508         unsigned long ntx;
2509
2510         if (arg->stop)
2511                 return;
2512
2513         arg->count = arg->skip;
2514         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
2515                 if (!tc_qdisc_stats_dump(sch, ntx + 1, arg))
2516                         break;
2517         }
2518 }
2519
2520 static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
2521                                                 struct tcmsg *tcm)
2522 {
2523         return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
2524 }
2525
2526 static const struct Qdisc_class_ops taprio_class_ops = {
2527         .graft          = taprio_graft,
2528         .leaf           = taprio_leaf,
2529         .find           = taprio_find,
2530         .walk           = taprio_walk,
2531         .dump           = taprio_dump_class,
2532         .dump_stats     = taprio_dump_class_stats,
2533         .select_queue   = taprio_select_queue,
2534 };
2535
2536 static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
2537         .cl_ops         = &taprio_class_ops,
2538         .id             = "taprio",
2539         .priv_size      = sizeof(struct taprio_sched),
2540         .init           = taprio_init,
2541         .change         = taprio_change,
2542         .destroy        = taprio_destroy,
2543         .reset          = taprio_reset,
2544         .attach         = taprio_attach,
2545         .peek           = taprio_peek,
2546         .dequeue        = taprio_dequeue,
2547         .enqueue        = taprio_enqueue,
2548         .dump           = taprio_dump,
2549         .dump_stats     = taprio_dump_stats,
2550         .owner          = THIS_MODULE,
2551 };
2552
2553 static struct notifier_block taprio_device_notifier = {
2554         .notifier_call = taprio_dev_notifier,
2555 };
2556
2557 static int __init taprio_module_init(void)
2558 {
2559         int err = register_netdevice_notifier(&taprio_device_notifier);
2560
2561         if (err)
2562                 return err;
2563
2564         return register_qdisc(&taprio_qdisc_ops);
2565 }
2566
2567 static void __exit taprio_module_exit(void)
2568 {
2569         unregister_qdisc(&taprio_qdisc_ops);
2570         unregister_netdevice_notifier(&taprio_device_notifier);
2571 }
2572
2573 module_init(taprio_module_init);
2574 module_exit(taprio_module_exit);
2575 MODULE_LICENSE("GPL");
2576 MODULE_DESCRIPTION("Time Aware Priority qdisc");