net/sched/sch_taprio.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* net/sched/sch_taprio.c        Time Aware Priority Scheduler
   4  *
   5  * Authors:     Vinicius Costa Gomes <vinicius.gomes@intel.com>
   6  *
   7  */
   8
   9 #include <linux/ethtool.h>
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/kernel.h>
  13 #include <linux/string.h>
  14 #include <linux/list.h>
  15 #include <linux/errno.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/math64.h>
  18 #include <linux/module.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/rcupdate.h>
  21 #include <linux/time.h>
  22 #include <net/netlink.h>
  23 #include <net/pkt_sched.h>
  24 #include <net/pkt_cls.h>
  25 #include <net/sch_generic.h>
  26 #include <net/sock.h>
  27 #include <net/tcp.h>
  28
  29 static LIST_HEAD(taprio_list);
  30
  31 #define TAPRIO_ALL_GATES_OPEN -1
  32
  33 #define TXTIME_ASSIST_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST)
  34 #define FULL_OFFLOAD_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)
  35 #define TAPRIO_FLAGS_INVALID U32_MAX
  36
  37 struct sched_entry {
  38         struct list_head list;
  39
  40         /* The instant that this entry "closes" and the next one
  41          * should open, the qdisc will make some effort so that no
  42          * packet leaves after this time.
  43          */
  44         ktime_t close_time;
  45         ktime_t next_txtime;
  46         atomic_t budget;
  47         int index;
  48         u32 gate_mask;
  49         u32 interval;
  50         u8 command;
  51 };
  52
  53 struct sched_gate_list {
  54         struct rcu_head rcu;
  55         struct list_head entries;
  56         size_t num_entries;
  57         ktime_t cycle_close_time;
  58         s64 cycle_time;
  59         s64 cycle_time_extension;
  60         s64 base_time;
  61 };
  62
  63 struct taprio_sched {
  64         struct Qdisc **qdiscs;
  65         struct Qdisc *root;
  66         u32 flags;
  67         enum tk_offsets tk_offset;
  68         int clockid;
  69         bool offloaded;
  70         atomic64_t picos_per_byte; /* Using picoseconds because for 10Gbps+
  71                                     * speeds it's sub-nanoseconds per byte
  72                                     */
  73
  74         /* Protects the update side of the RCU protected current_entry */
  75         spinlock_t current_entry_lock;
  76         struct sched_entry __rcu *current_entry;
  77         struct sched_gate_list __rcu *oper_sched;
  78         struct sched_gate_list __rcu *admin_sched;
  79         struct hrtimer advance_timer;
  80         struct list_head taprio_list;
  81         u32 max_frm_len[TC_MAX_QUEUE]; /* for the fast path */
  82         u32 max_sdu[TC_MAX_QUEUE]; /* for dump and offloading */
  83         u32 txtime_delay;
  84 };
  85
  86 struct __tc_taprio_qopt_offload {
  87         refcount_t users;
  88         struct tc_taprio_qopt_offload offload;
  89 };
  90
  91 static ktime_t sched_base_time(const struct sched_gate_list *sched)
  92 {
  93         if (!sched)
  94                 return KTIME_MAX;
  95
  96         return ns_to_ktime(sched->base_time);
  97 }
  98
  99 static ktime_t taprio_mono_to_any(const struct taprio_sched *q, ktime_t mono)
 100 {
 101         /* This pairs with WRITE_ONCE() in taprio_parse_clockid() */
 102         enum tk_offsets tk_offset = READ_ONCE(q->tk_offset);
 103
 104         switch (tk_offset) {
 105         case TK_OFFS_MAX:
 106                 return mono;
 107         default:
 108                 return ktime_mono_to_any(mono, tk_offset);
 109         }
 110 }
 111
 112 static ktime_t taprio_get_time(const struct taprio_sched *q)
 113 {
 114         return taprio_mono_to_any(q, ktime_get());
 115 }
 116
 117 static void taprio_free_sched_cb(struct rcu_head *head)
 118 {
 119         struct sched_gate_list *sched = container_of(head, struct sched_gate_list, rcu);
 120         struct sched_entry *entry, *n;
 121
 122         list_for_each_entry_safe(entry, n, &sched->entries, list) {
 123                 list_del(&entry->list);
 124                 kfree(entry);
 125         }
 126
 127         kfree(sched);
 128 }
 129
 130 static void switch_schedules(struct taprio_sched *q,
 131                              struct sched_gate_list **admin,
 132                              struct sched_gate_list **oper)
 133 {
 134         rcu_assign_pointer(q->oper_sched, *admin);
 135         rcu_assign_pointer(q->admin_sched, NULL);
 136
 137         if (*oper)
 138                 call_rcu(&(*oper)->rcu, taprio_free_sched_cb);
 139
 140         *oper = *admin;
 141         *admin = NULL;
 142 }
 143
 144 /* Get how much time has been already elapsed in the current cycle. */
 145 static s32 get_cycle_time_elapsed(struct sched_gate_list *sched, ktime_t time)
 146 {
 147         ktime_t time_since_sched_start;
 148         s32 time_elapsed;
 149
 150         time_since_sched_start = ktime_sub(time, sched->base_time);
 151         div_s64_rem(time_since_sched_start, sched->cycle_time, &time_elapsed);
 152
 153         return time_elapsed;
 154 }
 155
 156 static ktime_t get_interval_end_time(struct sched_gate_list *sched,
 157                                      struct sched_gate_list *admin,
 158                                      struct sched_entry *entry,
 159                                      ktime_t intv_start)
 160 {
 161         s32 cycle_elapsed = get_cycle_time_elapsed(sched, intv_start);
 162         ktime_t intv_end, cycle_ext_end, cycle_end;
 163
 164         cycle_end = ktime_add_ns(intv_start, sched->cycle_time - cycle_elapsed);
 165         intv_end = ktime_add_ns(intv_start, entry->interval);
 166         cycle_ext_end = ktime_add(cycle_end, sched->cycle_time_extension);
 167
 168         if (ktime_before(intv_end, cycle_end))
 169                 return intv_end;
 170         else if (admin && admin != sched &&
 171                  ktime_after(admin->base_time, cycle_end) &&
 172                  ktime_before(admin->base_time, cycle_ext_end))
 173                 return admin->base_time;
 174         else
 175                 return cycle_end;
 176 }
 177
 178 static int length_to_duration(struct taprio_sched *q, int len)
 179 {
 180         return div_u64(len * atomic64_read(&q->picos_per_byte), PSEC_PER_NSEC);
 181 }
 182
 183 /* Returns the entry corresponding to next available interval. If
 184  * validate_interval is set, it only validates whether the timestamp occurs
 185  * when the gate corresponding to the skb's traffic class is open.
 186  */
 187 static struct sched_entry *find_entry_to_transmit(struct sk_buff *skb,
 188                                                   struct Qdisc *sch,
 189                                                   struct sched_gate_list *sched,
 190                                                   struct sched_gate_list *admin,
 191                                                   ktime_t time,
 192                                                   ktime_t *interval_start,
 193                                                   ktime_t *interval_end,
 194                                                   bool validate_interval)
 195 {
 196         ktime_t curr_intv_start, curr_intv_end, cycle_end, packet_transmit_time;
 197         ktime_t earliest_txtime = KTIME_MAX, txtime, cycle, transmit_end_time;
 198         struct sched_entry *entry = NULL, *entry_found = NULL;
 199         struct taprio_sched *q = qdisc_priv(sch);
 200         struct net_device *dev = qdisc_dev(sch);
 201         bool entry_available = false;
 202         s32 cycle_elapsed;
 203         int tc, n;
 204
 205         tc = netdev_get_prio_tc_map(dev, skb->priority);
 206         packet_transmit_time = length_to_duration(q, qdisc_pkt_len(skb));
 207
 208         *interval_start = 0;
 209         *interval_end = 0;
 210
 211         if (!sched)
 212                 return NULL;
 213
 214         cycle = sched->cycle_time;
 215         cycle_elapsed = get_cycle_time_elapsed(sched, time);
 216         curr_intv_end = ktime_sub_ns(time, cycle_elapsed);
 217         cycle_end = ktime_add_ns(curr_intv_end, cycle);
 218
 219         list_for_each_entry(entry, &sched->entries, list) {
 220                 curr_intv_start = curr_intv_end;
 221                 curr_intv_end = get_interval_end_time(sched, admin, entry,
 222                                                       curr_intv_start);
 223
 224                 if (ktime_after(curr_intv_start, cycle_end))
 225                         break;
 226
 227                 if (!(entry->gate_mask & BIT(tc)) ||
 228                     packet_transmit_time > entry->interval)
 229                         continue;
 230
 231                 txtime = entry->next_txtime;
 232
 233                 if (ktime_before(txtime, time) || validate_interval) {
 234                         transmit_end_time = ktime_add_ns(time, packet_transmit_time);
 235                         if ((ktime_before(curr_intv_start, time) &&
 236                              ktime_before(transmit_end_time, curr_intv_end)) ||
 237                             (ktime_after(curr_intv_start, time) && !validate_interval)) {
 238                                 entry_found = entry;
 239                                 *interval_start = curr_intv_start;
 240                                 *interval_end = curr_intv_end;
 241                                 break;
 242                         } else if (!entry_available && !validate_interval) {
 243                                 /* Here, we are just trying to find out the
 244                                  * first available interval in the next cycle.
 245                                  */
 246                                 entry_available = true;
 247                                 entry_found = entry;
 248                                 *interval_start = ktime_add_ns(curr_intv_start, cycle);
 249                                 *interval_end = ktime_add_ns(curr_intv_end, cycle);
 250                         }
 251                 } else if (ktime_before(txtime, earliest_txtime) &&
 252                            !entry_available) {
 253                         earliest_txtime = txtime;
 254                         entry_found = entry;
 255                         n = div_s64(ktime_sub(txtime, curr_intv_start), cycle);
 256                         *interval_start = ktime_add(curr_intv_start, n * cycle);
 257                         *interval_end = ktime_add(curr_intv_end, n * cycle);
 258                 }
 259         }
 260
 261         return entry_found;
 262 }
 263
 264 static bool is_valid_interval(struct sk_buff *skb, struct Qdisc *sch)
 265 {
 266         struct taprio_sched *q = qdisc_priv(sch);
 267         struct sched_gate_list *sched, *admin;
 268         ktime_t interval_start, interval_end;
 269         struct sched_entry *entry;
 270
 271         rcu_read_lock();
 272         sched = rcu_dereference(q->oper_sched);
 273         admin = rcu_dereference(q->admin_sched);
 274
 275         entry = find_entry_to_transmit(skb, sch, sched, admin, skb->tstamp,
 276                                        &interval_start, &interval_end, true);
 277         rcu_read_unlock();
 278
 279         return entry;
 280 }
 281
 282 static bool taprio_flags_valid(u32 flags)
 283 {
 284         /* Make sure no other flag bits are set. */
 285         if (flags & ~(TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST |
 286                       TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 287                 return false;
 288         /* txtime-assist and full offload are mutually exclusive */
 289         if ((flags & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST) &&
 290             (flags & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 291                 return false;
 292         return true;
 293 }
 294
 295 /* This returns the tstamp value set by TCP in terms of the set clock. */
 296 static ktime_t get_tcp_tstamp(struct taprio_sched *q, struct sk_buff *skb)
 297 {
 298         unsigned int offset = skb_network_offset(skb);
 299         const struct ipv6hdr *ipv6h;
 300         const struct iphdr *iph;
 301         struct ipv6hdr _ipv6h;
 302
 303         ipv6h = skb_header_pointer(skb, offset, sizeof(_ipv6h), &_ipv6h);
 304         if (!ipv6h)
 305                 return 0;
 306
 307         if (ipv6h->version == 4) {
 308                 iph = (struct iphdr *)ipv6h;
 309                 offset += iph->ihl * 4;
 310
 311                 /* special-case 6in4 tunnelling, as that is a common way to get
 312                  * v6 connectivity in the home
 313                  */
 314                 if (iph->protocol == IPPROTO_IPV6) {
 315                         ipv6h = skb_header_pointer(skb, offset,
 316                                                    sizeof(_ipv6h), &_ipv6h);
 317
 318                         if (!ipv6h || ipv6h->nexthdr != IPPROTO_TCP)
 319                                 return 0;
 320                 } else if (iph->protocol != IPPROTO_TCP) {
 321                         return 0;
 322                 }
 323         } else if (ipv6h->version == 6 && ipv6h->nexthdr != IPPROTO_TCP) {
 324                 return 0;
 325         }
 326
 327         return taprio_mono_to_any(q, skb->skb_mstamp_ns);
 328 }
 329
 330 /* There are a few scenarios where we will have to modify the txtime from
 331  * what is read from next_txtime in sched_entry. They are:
 332  * 1. If txtime is in the past,
 333  *    a. The gate for the traffic class is currently open and packet can be
 334  *       transmitted before it closes, schedule the packet right away.
 335  *    b. If the gate corresponding to the traffic class is going to open later
 336  *       in the cycle, set the txtime of packet to the interval start.
 337  * 2. If txtime is in the future, there are packets corresponding to the
 338  *    current traffic class waiting to be transmitted. So, the following
 339  *    possibilities exist:
 340  *    a. We can transmit the packet before the window containing the txtime
 341  *       closes.
 342  *    b. The window might close before the transmission can be completed
 343  *       successfully. So, schedule the packet in the next open window.
 344  */
 345 static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)
 346 {
 347         ktime_t transmit_end_time, interval_end, interval_start, tcp_tstamp;
 348         struct taprio_sched *q = qdisc_priv(sch);
 349         struct sched_gate_list *sched, *admin;
 350         ktime_t minimum_time, now, txtime;
 351         int len, packet_transmit_time;
 352         struct sched_entry *entry;
 353         bool sched_changed;
 354
 355         now = taprio_get_time(q);
 356         minimum_time = ktime_add_ns(now, q->txtime_delay);
 357
 358         tcp_tstamp = get_tcp_tstamp(q, skb);
 359         minimum_time = max_t(ktime_t, minimum_time, tcp_tstamp);
 360
 361         rcu_read_lock();
 362         admin = rcu_dereference(q->admin_sched);
 363         sched = rcu_dereference(q->oper_sched);
 364         if (admin && ktime_after(minimum_time, admin->base_time))
 365                 switch_schedules(q, &admin, &sched);
 366
 367         /* Until the schedule starts, all the queues are open */
 368         if (!sched || ktime_before(minimum_time, sched->base_time)) {
 369                 txtime = minimum_time;
 370                 goto done;
 371         }
 372
 373         len = qdisc_pkt_len(skb);
 374         packet_transmit_time = length_to_duration(q, len);
 375
 376         do {
 377                 sched_changed = false;
 378
 379                 entry = find_entry_to_transmit(skb, sch, sched, admin,
 380                                                minimum_time,
 381                                                &interval_start, &interval_end,
 382                                                false);
 383                 if (!entry) {
 384                         txtime = 0;
 385                         goto done;
 386                 }
 387
 388                 txtime = entry->next_txtime;
 389                 txtime = max_t(ktime_t, txtime, minimum_time);
 390                 txtime = max_t(ktime_t, txtime, interval_start);
 391
 392                 if (admin && admin != sched &&
 393                     ktime_after(txtime, admin->base_time)) {
 394                         sched = admin;
 395                         sched_changed = true;
 396                         continue;
 397                 }
 398
 399                 transmit_end_time = ktime_add(txtime, packet_transmit_time);
 400                 minimum_time = transmit_end_time;
 401
 402                 /* Update the txtime of current entry to the next time it's
 403                  * interval starts.
 404                  */
 405                 if (ktime_after(transmit_end_time, interval_end))
 406                         entry->next_txtime = ktime_add(interval_start, sched->cycle_time);
 407         } while (sched_changed || ktime_after(transmit_end_time, interval_end));
 408
 409         entry->next_txtime = transmit_end_time;
 410
 411 done:
 412         rcu_read_unlock();
 413         return txtime;
 414 }
 415
 416 static int taprio_enqueue_one(struct sk_buff *skb, struct Qdisc *sch,
 417                               struct Qdisc *child, struct sk_buff **to_free)
 418 {
 419         struct taprio_sched *q = qdisc_priv(sch);
 420         struct net_device *dev = qdisc_dev(sch);
 421         int prio = skb->priority;
 422         u8 tc;
 423
 424         /* sk_flags are only safe to use on full sockets. */
 425         if (skb->sk && sk_fullsock(skb->sk) && sock_flag(skb->sk, SOCK_TXTIME)) {
 426                 if (!is_valid_interval(skb, sch))
 427                         return qdisc_drop(skb, sch, to_free);
 428         } else if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
 429                 skb->tstamp = get_packet_txtime(skb, sch);
 430                 if (!skb->tstamp)
 431                         return qdisc_drop(skb, sch, to_free);
 432         }
 433
 434         /* Devices with full offload are expected to honor this in hardware */
 435         tc = netdev_get_prio_tc_map(dev, prio);
 436         if (skb->len > q->max_frm_len[tc])
 437                 return qdisc_drop(skb, sch, to_free);
 438
 439         qdisc_qstats_backlog_inc(sch, skb);
 440         sch->q.qlen++;
 441
 442         return qdisc_enqueue(skb, child, to_free);
 443 }
 444
 445 /* Will not be called in the full offload case, since the TX queues are
 446  * attached to the Qdisc created using qdisc_create_dflt()
 447  */
 448 static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 449                           struct sk_buff **to_free)
 450 {
 451         struct taprio_sched *q = qdisc_priv(sch);
 452         struct Qdisc *child;
 453         int queue;
 454
 455         queue = skb_get_queue_mapping(skb);
 456
 457         child = q->qdiscs[queue];
 458         if (unlikely(!child))
 459                 return qdisc_drop(skb, sch, to_free);
 460
 461         /* Large packets might not be transmitted when the transmission duration
 462          * exceeds any configured interval. Therefore, segment the skb into
 463          * smaller chunks. Drivers with full offload are expected to handle
 464          * this in hardware.
 465          */
 466         if (skb_is_gso(skb)) {
 467                 unsigned int slen = 0, numsegs = 0, len = qdisc_pkt_len(skb);
 468                 netdev_features_t features = netif_skb_features(skb);
 469                 struct sk_buff *segs, *nskb;
 470                 int ret;
 471
 472                 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 473                 if (IS_ERR_OR_NULL(segs))
 474                         return qdisc_drop(skb, sch, to_free);
 475
 476                 skb_list_walk_safe(segs, segs, nskb) {
 477                         skb_mark_not_on_list(segs);
 478                         qdisc_skb_cb(segs)->pkt_len = segs->len;
 479                         slen += segs->len;
 480
 481                         ret = taprio_enqueue_one(segs, sch, child, to_free);
 482                         if (ret != NET_XMIT_SUCCESS) {
 483                                 if (net_xmit_drop_count(ret))
 484                                         qdisc_qstats_drop(sch);
 485                         } else {
 486                                 numsegs++;
 487                         }
 488                 }
 489
 490                 if (numsegs > 1)
 491                         qdisc_tree_reduce_backlog(sch, 1 - numsegs, len - slen);
 492                 consume_skb(skb);
 493
 494                 return numsegs > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
 495         }
 496
 497         return taprio_enqueue_one(skb, sch, child, to_free);
 498 }
 499
 500 /* Will not be called in the full offload case, since the TX queues are
 501  * attached to the Qdisc created using qdisc_create_dflt()
 502  */
 503 static struct sk_buff *taprio_peek(struct Qdisc *sch)
 504 {
 505         struct taprio_sched *q = qdisc_priv(sch);
 506         struct net_device *dev = qdisc_dev(sch);
 507         struct sched_entry *entry;
 508         struct sk_buff *skb;
 509         u32 gate_mask;
 510         int i;
 511
 512         rcu_read_lock();
 513         entry = rcu_dereference(q->current_entry);
 514         gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
 515         rcu_read_unlock();
 516
 517         if (!gate_mask)
 518                 return NULL;
 519
 520         for (i = 0; i < dev->num_tx_queues; i++) {
 521                 struct Qdisc *child = q->qdiscs[i];
 522                 int prio;
 523                 u8 tc;
 524
 525                 if (unlikely(!child))
 526                         continue;
 527
 528                 skb = child->ops->peek(child);
 529                 if (!skb)
 530                         continue;
 531
 532                 if (TXTIME_ASSIST_IS_ENABLED(q->flags))
 533                         return skb;
 534
 535                 prio = skb->priority;
 536                 tc = netdev_get_prio_tc_map(dev, prio);
 537
 538                 if (!(gate_mask & BIT(tc)))
 539                         continue;
 540
 541                 return skb;
 542         }
 543
 544         return NULL;
 545 }
 546
 547 static void taprio_set_budget(struct taprio_sched *q, struct sched_entry *entry)
 548 {
 549         atomic_set(&entry->budget,
 550                    div64_u64((u64)entry->interval * PSEC_PER_NSEC,
 551                              atomic64_read(&q->picos_per_byte)));
 552 }
 553
 554 /* Will not be called in the full offload case, since the TX queues are
 555  * attached to the Qdisc created using qdisc_create_dflt()
 556  */
 557 static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
 558 {
 559         struct taprio_sched *q = qdisc_priv(sch);
 560         struct net_device *dev = qdisc_dev(sch);
 561         struct sk_buff *skb = NULL;
 562         struct sched_entry *entry;
 563         u32 gate_mask;
 564         int i;
 565
 566         rcu_read_lock();
 567         entry = rcu_dereference(q->current_entry);
 568         /* if there's no entry, it means that the schedule didn't
 569          * start yet, so force all gates to be open, this is in
 570          * accordance to IEEE 802.1Qbv-2015 Section 8.6.9.4.5
 571          * "AdminGateStates"
 572          */
 573         gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
 574
 575         if (!gate_mask)
 576                 goto done;
 577
 578         for (i = 0; i < dev->num_tx_queues; i++) {
 579                 struct Qdisc *child = q->qdiscs[i];
 580                 ktime_t guard;
 581                 int prio;
 582                 int len;
 583                 u8 tc;
 584
 585                 if (unlikely(!child))
 586                         continue;
 587
 588                 if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
 589                         skb = child->ops->dequeue(child);
 590                         if (!skb)
 591                                 continue;
 592                         goto skb_found;
 593                 }
 594
 595                 skb = child->ops->peek(child);
 596                 if (!skb)
 597                         continue;
 598
 599                 prio = skb->priority;
 600                 tc = netdev_get_prio_tc_map(dev, prio);
 601
 602                 if (!(gate_mask & BIT(tc))) {
 603                         skb = NULL;
 604                         continue;
 605                 }
 606
 607                 len = qdisc_pkt_len(skb);
 608                 guard = ktime_add_ns(taprio_get_time(q),
 609                                      length_to_duration(q, len));
 610
 611                 /* In the case that there's no gate entry, there's no
 612                  * guard band ...
 613                  */
 614                 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 615                     ktime_after(guard, entry->close_time)) {
 616                         skb = NULL;
 617                         continue;
 618                 }
 619
 620                 /* ... and no budget. */
 621                 if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 622                     atomic_sub_return(len, &entry->budget) < 0) {
 623                         skb = NULL;
 624                         continue;
 625                 }
 626
 627                 skb = child->ops->dequeue(child);
 628                 if (unlikely(!skb))
 629                         goto done;
 630
 631 skb_found:
 632                 qdisc_bstats_update(sch, skb);
 633                 qdisc_qstats_backlog_dec(sch, skb);
 634                 sch->q.qlen--;
 635
 636                 goto done;
 637         }
 638
 639 done:
 640         rcu_read_unlock();
 641
 642         return skb;
 643 }
 644
 645 static bool should_restart_cycle(const struct sched_gate_list *oper,
 646                                  const struct sched_entry *entry)
 647 {
 648         if (list_is_last(&entry->list, &oper->entries))
 649                 return true;
 650
 651         if (ktime_compare(entry->close_time, oper->cycle_close_time) == 0)
 652                 return true;
 653
 654         return false;
 655 }
 656
 657 static bool should_change_schedules(const struct sched_gate_list *admin,
 658                                     const struct sched_gate_list *oper,
 659                                     ktime_t close_time)
 660 {
 661         ktime_t next_base_time, extension_time;
 662
 663         if (!admin)
 664                 return false;
 665
 666         next_base_time = sched_base_time(admin);
 667
 668         /* This is the simple case, the close_time would fall after
 669          * the next schedule base_time.
 670          */
 671         if (ktime_compare(next_base_time, close_time) <= 0)
 672                 return true;
 673
 674         /* This is the cycle_time_extension case, if the close_time
 675          * plus the amount that can be extended would fall after the
 676          * next schedule base_time, we can extend the current schedule
 677          * for that amount.
 678          */
 679         extension_time = ktime_add_ns(close_time, oper->cycle_time_extension);
 680
 681         /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
 682          * how precisely the extension should be made. So after
 683          * conformance testing, this logic may change.
 684          */
 685         if (ktime_compare(next_base_time, extension_time) <= 0)
 686                 return true;
 687
 688         return false;
 689 }
 690
 691 static enum hrtimer_restart advance_sched(struct hrtimer *timer)
 692 {
 693         struct taprio_sched *q = container_of(timer, struct taprio_sched,
 694                                               advance_timer);
 695         struct sched_gate_list *oper, *admin;
 696         struct sched_entry *entry, *next;
 697         struct Qdisc *sch = q->root;
 698         ktime_t close_time;
 699
 700         spin_lock(&q->current_entry_lock);
 701         entry = rcu_dereference_protected(q->current_entry,
 702                                           lockdep_is_held(&q->current_entry_lock));
 703         oper = rcu_dereference_protected(q->oper_sched,
 704                                          lockdep_is_held(&q->current_entry_lock));
 705         admin = rcu_dereference_protected(q->admin_sched,
 706                                           lockdep_is_held(&q->current_entry_lock));
 707
 708         if (!oper)
 709                 switch_schedules(q, &admin, &oper);
 710
 711         /* This can happen in two cases: 1. this is the very first run
 712          * of this function (i.e. we weren't running any schedule
 713          * previously); 2. The previous schedule just ended. The first
 714          * entry of all schedules are pre-calculated during the
 715          * schedule initialization.
 716          */
 717         if (unlikely(!entry || entry->close_time == oper->base_time)) {
 718                 next = list_first_entry(&oper->entries, struct sched_entry,
 719                                         list);
 720                 close_time = next->close_time;
 721                 goto first_run;
 722         }
 723
 724         if (should_restart_cycle(oper, entry)) {
 725                 next = list_first_entry(&oper->entries, struct sched_entry,
 726                                         list);
 727                 oper->cycle_close_time = ktime_add_ns(oper->cycle_close_time,
 728                                                       oper->cycle_time);
 729         } else {
 730                 next = list_next_entry(entry, list);
 731         }
 732
 733         close_time = ktime_add_ns(entry->close_time, next->interval);
 734         close_time = min_t(ktime_t, close_time, oper->cycle_close_time);
 735
 736         if (should_change_schedules(admin, oper, close_time)) {
 737                 /* Set things so the next time this runs, the new
 738                  * schedule runs.
 739                  */
 740                 close_time = sched_base_time(admin);
 741                 switch_schedules(q, &admin, &oper);
 742         }
 743
 744         next->close_time = close_time;
 745         taprio_set_budget(q, next);
 746
 747 first_run:
 748         rcu_assign_pointer(q->current_entry, next);
 749         spin_unlock(&q->current_entry_lock);
 750
 751         hrtimer_set_expires(&q->advance_timer, close_time);
 752
 753         rcu_read_lock();
 754         __netif_schedule(sch);
 755         rcu_read_unlock();
 756
 757         return HRTIMER_RESTART;
 758 }
 759
 760 static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
 761         [TCA_TAPRIO_SCHED_ENTRY_INDEX]     = { .type = NLA_U32 },
 762         [TCA_TAPRIO_SCHED_ENTRY_CMD]       = { .type = NLA_U8 },
 763         [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
 764         [TCA_TAPRIO_SCHED_ENTRY_INTERVAL]  = { .type = NLA_U32 },
 765 };
 766
 767 static const struct nla_policy taprio_tc_policy[TCA_TAPRIO_TC_ENTRY_MAX + 1] = {
 768         [TCA_TAPRIO_TC_ENTRY_INDEX]        = { .type = NLA_U32 },
 769         [TCA_TAPRIO_TC_ENTRY_MAX_SDU]      = { .type = NLA_U32 },
 770 };
 771
 772 static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
 773         [TCA_TAPRIO_ATTR_PRIOMAP]              = {
 774                 .len = sizeof(struct tc_mqprio_qopt)
 775         },
 776         [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST]           = { .type = NLA_NESTED },
 777         [TCA_TAPRIO_ATTR_SCHED_BASE_TIME]            = { .type = NLA_S64 },
 778         [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]         = { .type = NLA_NESTED },
 779         [TCA_TAPRIO_ATTR_SCHED_CLOCKID]              = { .type = NLA_S32 },
 780         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]           = { .type = NLA_S64 },
 781         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
 782         [TCA_TAPRIO_ATTR_FLAGS]                      = { .type = NLA_U32 },
 783         [TCA_TAPRIO_ATTR_TXTIME_DELAY]               = { .type = NLA_U32 },
 784         [TCA_TAPRIO_ATTR_TC_ENTRY]                   = { .type = NLA_NESTED },
 785 };
 786
 787 static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
 788                             struct sched_entry *entry,
 789                             struct netlink_ext_ack *extack)
 790 {
 791         int min_duration = length_to_duration(q, ETH_ZLEN);
 792         u32 interval = 0;
 793
 794         if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
 795                 entry->command = nla_get_u8(
 796                         tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
 797
 798         if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
 799                 entry->gate_mask = nla_get_u32(
 800                         tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
 801
 802         if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
 803                 interval = nla_get_u32(
 804                         tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
 805
 806         /* The interval should allow at least the minimum ethernet
 807          * frame to go out.
 808          */
 809         if (interval < min_duration) {
 810                 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
 811                 return -EINVAL;
 812         }
 813
 814         entry->interval = interval;
 815
 816         return 0;
 817 }
 818
 819 static int parse_sched_entry(struct taprio_sched *q, struct nlattr *n,
 820                              struct sched_entry *entry, int index,
 821                              struct netlink_ext_ack *extack)
 822 {
 823         struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
 824         int err;
 825
 826         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
 827                                           entry_policy, NULL);
 828         if (err < 0) {
 829                 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
 830                 return -EINVAL;
 831         }
 832
 833         entry->index = index;
 834
 835         return fill_sched_entry(q, tb, entry, extack);
 836 }
 837
 838 static int parse_sched_list(struct taprio_sched *q, struct nlattr *list,
 839                             struct sched_gate_list *sched,
 840                             struct netlink_ext_ack *extack)
 841 {
 842         struct nlattr *n;
 843         int err, rem;
 844         int i = 0;
 845
 846         if (!list)
 847                 return -EINVAL;
 848
 849         nla_for_each_nested(n, list, rem) {
 850                 struct sched_entry *entry;
 851
 852                 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
 853                         NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
 854                         continue;
 855                 }
 856
 857                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 858                 if (!entry) {
 859                         NL_SET_ERR_MSG(extack, "Not enough memory for entry");
 860                         return -ENOMEM;
 861                 }
 862
 863                 err = parse_sched_entry(q, n, entry, i, extack);
 864                 if (err < 0) {
 865                         kfree(entry);
 866                         return err;
 867                 }
 868
 869                 list_add_tail(&entry->list, &sched->entries);
 870                 i++;
 871         }
 872
 873         sched->num_entries = i;
 874
 875         return i;
 876 }
 877
 878 static int parse_taprio_schedule(struct taprio_sched *q, struct nlattr **tb,
 879                                  struct sched_gate_list *new,
 880                                  struct netlink_ext_ack *extack)
 881 {
 882         int err = 0;
 883
 884         if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
 885                 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
 886                 return -ENOTSUPP;
 887         }
 888
 889         if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
 890                 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
 891
 892         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
 893                 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
 894
 895         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
 896                 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
 897
 898         if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
 899                 err = parse_sched_list(q, tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST],
 900                                        new, extack);
 901         if (err < 0)
 902                 return err;
 903
 904         if (!new->cycle_time) {
 905                 struct sched_entry *entry;
 906                 ktime_t cycle = 0;
 907
 908                 list_for_each_entry(entry, &new->entries, list)
 909                         cycle = ktime_add_ns(cycle, entry->interval);
 910
 911                 if (!cycle) {
 912                         NL_SET_ERR_MSG(extack, "'cycle_time' can never be 0");
 913                         return -EINVAL;
 914                 }
 915
 916                 new->cycle_time = cycle;
 917         }
 918
 919         return 0;
 920 }
 921
 922 static int taprio_parse_mqprio_opt(struct net_device *dev,
 923                                    struct tc_mqprio_qopt *qopt,
 924                                    struct netlink_ext_ack *extack,
 925                                    u32 taprio_flags)
 926 {
 927         int i, j;
 928
 929         if (!qopt && !dev->num_tc) {
 930                 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
 931                 return -EINVAL;
 932         }
 933
 934         /* If num_tc is already set, it means that the user already
 935          * configured the mqprio part
 936          */
 937         if (dev->num_tc)
 938                 return 0;
 939
 940         /* Verify num_tc is not out of max range */
 941         if (qopt->num_tc > TC_MAX_QUEUE) {
 942                 NL_SET_ERR_MSG(extack, "Number of traffic classes is outside valid range");
 943                 return -EINVAL;
 944         }
 945
 946         /* taprio imposes that traffic classes map 1:n to tx queues */
 947         if (qopt->num_tc > dev->num_tx_queues) {
 948                 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
 949                 return -EINVAL;
 950         }
 951
 952         /* Verify priority mapping uses valid tcs */
 953         for (i = 0; i <= TC_BITMASK; i++) {
 954                 if (qopt->prio_tc_map[i] >= qopt->num_tc) {
 955                         NL_SET_ERR_MSG(extack, "Invalid traffic class in priority to traffic class mapping");
 956                         return -EINVAL;
 957                 }
 958         }
 959
 960         for (i = 0; i < qopt->num_tc; i++) {
 961                 unsigned int last = qopt->offset[i] + qopt->count[i];
 962
 963                 /* Verify the queue count is in tx range being equal to the
 964                  * real_num_tx_queues indicates the last queue is in use.
 965                  */
 966                 if (qopt->offset[i] >= dev->num_tx_queues ||
 967                     !qopt->count[i] ||
 968                     last > dev->real_num_tx_queues) {
 969                         NL_SET_ERR_MSG(extack, "Invalid queue in traffic class to queue mapping");
 970                         return -EINVAL;
 971                 }
 972
 973                 if (TXTIME_ASSIST_IS_ENABLED(taprio_flags))
 974                         continue;
 975
 976                 /* Verify that the offset and counts do not overlap */
 977                 for (j = i + 1; j < qopt->num_tc; j++) {
 978                         if (last > qopt->offset[j]) {
 979                                 NL_SET_ERR_MSG(extack, "Detected overlap in the traffic class to queue mapping");
 980                                 return -EINVAL;
 981                         }
 982                 }
 983         }
 984
 985         return 0;
 986 }
 987
 988 static int taprio_get_start_time(struct Qdisc *sch,
 989                                  struct sched_gate_list *sched,
 990                                  ktime_t *start)
 991 {
 992         struct taprio_sched *q = qdisc_priv(sch);
 993         ktime_t now, base, cycle;
 994         s64 n;
 995
 996         base = sched_base_time(sched);
 997         now = taprio_get_time(q);
 998
 999         if (ktime_after(base, now)) {
1000                 *start = base;
1001                 return 0;
1002         }
1003
1004         cycle = sched->cycle_time;
1005
1006         /* The qdisc is expected to have at least one sched_entry.  Moreover,
1007          * any entry must have 'interval' > 0. Thus if the cycle time is zero,
1008          * something went really wrong. In that case, we should warn about this
1009          * inconsistent state and return error.
1010          */
1011         if (WARN_ON(!cycle))
1012                 return -EFAULT;
1013
1014         /* Schedule the start time for the beginning of the next
1015          * cycle.
1016          */
1017         n = div64_s64(ktime_sub_ns(now, base), cycle);
1018         *start = ktime_add_ns(base, (n + 1) * cycle);
1019         return 0;
1020 }
1021
1022 static void setup_first_close_time(struct taprio_sched *q,
1023                                    struct sched_gate_list *sched, ktime_t base)
1024 {
1025         struct sched_entry *first;
1026         ktime_t cycle;
1027
1028         first = list_first_entry(&sched->entries,
1029                                  struct sched_entry, list);
1030
1031         cycle = sched->cycle_time;
1032
1033         /* FIXME: find a better place to do this */
1034         sched->cycle_close_time = ktime_add_ns(base, cycle);
1035
1036         first->close_time = ktime_add_ns(base, first->interval);
1037         taprio_set_budget(q, first);
1038         rcu_assign_pointer(q->current_entry, NULL);
1039 }
1040
1041 static void taprio_start_sched(struct Qdisc *sch,
1042                                ktime_t start, struct sched_gate_list *new)
1043 {
1044         struct taprio_sched *q = qdisc_priv(sch);
1045         ktime_t expires;
1046
1047         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1048                 return;
1049
1050         expires = hrtimer_get_expires(&q->advance_timer);
1051         if (expires == 0)
1052                 expires = KTIME_MAX;
1053
1054         /* If the new schedule starts before the next expiration, we
1055          * reprogram it to the earliest one, so we change the admin
1056          * schedule to the operational one at the right time.
1057          */
1058         start = min_t(ktime_t, start, expires);
1059
1060         hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1061 }
1062
1063 static void taprio_set_picos_per_byte(struct net_device *dev,
1064                                       struct taprio_sched *q)
1065 {
1066         struct ethtool_link_ksettings ecmd;
1067         int speed = SPEED_10;
1068         int picos_per_byte;
1069         int err;
1070
1071         err = __ethtool_get_link_ksettings(dev, &ecmd);
1072         if (err < 0)
1073                 goto skip;
1074
1075         if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1076                 speed = ecmd.base.speed;
1077
1078 skip:
1079         picos_per_byte = (USEC_PER_SEC * 8) / speed;
1080
1081         atomic64_set(&q->picos_per_byte, picos_per_byte);
1082         netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1083                    dev->name, (long long)atomic64_read(&q->picos_per_byte),
1084                    ecmd.base.speed);
1085 }
1086
1087 static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1088                                void *ptr)
1089 {
1090         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1091         struct taprio_sched *q;
1092
1093         ASSERT_RTNL();
1094
1095         if (event != NETDEV_UP && event != NETDEV_CHANGE)
1096                 return NOTIFY_DONE;
1097
1098         list_for_each_entry(q, &taprio_list, taprio_list) {
1099                 if (dev != qdisc_dev(q->root))
1100                         continue;
1101
1102                 taprio_set_picos_per_byte(dev, q);
1103                 break;
1104         }
1105
1106         return NOTIFY_DONE;
1107 }
1108
1109 static void setup_txtime(struct taprio_sched *q,
1110                          struct sched_gate_list *sched, ktime_t base)
1111 {
1112         struct sched_entry *entry;
1113         u32 interval = 0;
1114
1115         list_for_each_entry(entry, &sched->entries, list) {
1116                 entry->next_txtime = ktime_add_ns(base, interval);
1117                 interval += entry->interval;
1118         }
1119 }
1120
1121 static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1122 {
1123         struct __tc_taprio_qopt_offload *__offload;
1124
1125         __offload = kzalloc(struct_size(__offload, offload.entries, num_entries),
1126                             GFP_KERNEL);
1127         if (!__offload)
1128                 return NULL;
1129
1130         refcount_set(&__offload->users, 1);
1131
1132         return &__offload->offload;
1133 }
1134
1135 struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1136                                                   *offload)
1137 {
1138         struct __tc_taprio_qopt_offload *__offload;
1139
1140         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1141                                  offload);
1142
1143         refcount_inc(&__offload->users);
1144
1145         return offload;
1146 }
1147 EXPORT_SYMBOL_GPL(taprio_offload_get);
1148
1149 void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1150 {
1151         struct __tc_taprio_qopt_offload *__offload;
1152
1153         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1154                                  offload);
1155
1156         if (!refcount_dec_and_test(&__offload->users))
1157                 return;
1158
1159         kfree(__offload);
1160 }
1161 EXPORT_SYMBOL_GPL(taprio_offload_free);
1162
1163 /* The function will only serve to keep the pointers to the "oper" and "admin"
1164  * schedules valid in relation to their base times, so when calling dump() the
1165  * users looks at the right schedules.
1166  * When using full offload, the admin configuration is promoted to oper at the
1167  * base_time in the PHC time domain.  But because the system time is not
1168  * necessarily in sync with that, we can't just trigger a hrtimer to call
1169  * switch_schedules at the right hardware time.
1170  * At the moment we call this by hand right away from taprio, but in the future
1171  * it will be useful to create a mechanism for drivers to notify taprio of the
1172  * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1173  * This is left as TODO.
1174  */
1175 static void taprio_offload_config_changed(struct taprio_sched *q)
1176 {
1177         struct sched_gate_list *oper, *admin;
1178
1179         oper = rtnl_dereference(q->oper_sched);
1180         admin = rtnl_dereference(q->admin_sched);
1181
1182         switch_schedules(q, &admin, &oper);
1183 }
1184
1185 static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
1186 {
1187         u32 i, queue_mask = 0;
1188
1189         for (i = 0; i < dev->num_tc; i++) {
1190                 u32 offset, count;
1191
1192                 if (!(tc_mask & BIT(i)))
1193                         continue;
1194
1195                 offset = dev->tc_to_txq[i].offset;
1196                 count = dev->tc_to_txq[i].count;
1197
1198                 queue_mask |= GENMASK(offset + count - 1, offset);
1199         }
1200
1201         return queue_mask;
1202 }
1203
1204 static void taprio_sched_to_offload(struct net_device *dev,
1205                                     struct sched_gate_list *sched,
1206                                     struct tc_taprio_qopt_offload *offload)
1207 {
1208         struct sched_entry *entry;
1209         int i = 0;
1210
1211         offload->base_time = sched->base_time;
1212         offload->cycle_time = sched->cycle_time;
1213         offload->cycle_time_extension = sched->cycle_time_extension;
1214
1215         list_for_each_entry(entry, &sched->entries, list) {
1216                 struct tc_taprio_sched_entry *e = &offload->entries[i];
1217
1218                 e->command = entry->command;
1219                 e->interval = entry->interval;
1220                 e->gate_mask = tc_map_to_queue_mask(dev, entry->gate_mask);
1221
1222                 i++;
1223         }
1224
1225         offload->num_entries = i;
1226 }
1227
1228 static int taprio_enable_offload(struct net_device *dev,
1229                                  struct taprio_sched *q,
1230                                  struct sched_gate_list *sched,
1231                                  struct netlink_ext_ack *extack)
1232 {
1233         const struct net_device_ops *ops = dev->netdev_ops;
1234         struct tc_taprio_qopt_offload *offload;
1235         struct tc_taprio_caps caps;
1236         int tc, err = 0;
1237
1238         if (!ops->ndo_setup_tc) {
1239                 NL_SET_ERR_MSG(extack,
1240                                "Device does not support taprio offload");
1241                 return -EOPNOTSUPP;
1242         }
1243
1244         qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1245                                  &caps, sizeof(caps));
1246
1247         if (!caps.supports_queue_max_sdu) {
1248                 for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
1249                         if (q->max_sdu[tc]) {
1250                                 NL_SET_ERR_MSG_MOD(extack,
1251                                                    "Device does not handle queueMaxSDU");
1252                                 return -EOPNOTSUPP;
1253                         }
1254                 }
1255         }
1256
1257         offload = taprio_offload_alloc(sched->num_entries);
1258         if (!offload) {
1259                 NL_SET_ERR_MSG(extack,
1260                                "Not enough memory for enabling offload mode");
1261                 return -ENOMEM;
1262         }
1263         offload->enable = 1;
1264         taprio_sched_to_offload(dev, sched, offload);
1265
1266         for (tc = 0; tc < TC_MAX_QUEUE; tc++)
1267                 offload->max_sdu[tc] = q->max_sdu[tc];
1268
1269         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1270         if (err < 0) {
1271                 NL_SET_ERR_MSG(extack,
1272                                "Device failed to setup taprio offload");
1273                 goto done;
1274         }
1275
1276         q->offloaded = true;
1277
1278 done:
1279         taprio_offload_free(offload);
1280
1281         return err;
1282 }
1283
1284 static int taprio_disable_offload(struct net_device *dev,
1285                                   struct taprio_sched *q,
1286                                   struct netlink_ext_ack *extack)
1287 {
1288         const struct net_device_ops *ops = dev->netdev_ops;
1289         struct tc_taprio_qopt_offload *offload;
1290         int err;
1291
1292         if (!q->offloaded)
1293                 return 0;
1294
1295         offload = taprio_offload_alloc(0);
1296         if (!offload) {
1297                 NL_SET_ERR_MSG(extack,
1298                                "Not enough memory to disable offload mode");
1299                 return -ENOMEM;
1300         }
1301         offload->enable = 0;
1302
1303         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1304         if (err < 0) {
1305                 NL_SET_ERR_MSG(extack,
1306                                "Device failed to disable offload");
1307                 goto out;
1308         }
1309
1310         q->offloaded = false;
1311
1312 out:
1313         taprio_offload_free(offload);
1314
1315         return err;
1316 }
1317
1318 /* If full offload is enabled, the only possible clockid is the net device's
1319  * PHC. For that reason, specifying a clockid through netlink is incorrect.
1320  * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1321  * in sync with the specified clockid via a user space daemon such as phc2sys.
1322  * For both software taprio and txtime-assist, the clockid is used for the
1323  * hrtimer that advances the schedule and hence mandatory.
1324  */
1325 static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1326                                 struct netlink_ext_ack *extack)
1327 {
1328         struct taprio_sched *q = qdisc_priv(sch);
1329         struct net_device *dev = qdisc_dev(sch);
1330         int err = -EINVAL;
1331
1332         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1333                 const struct ethtool_ops *ops = dev->ethtool_ops;
1334                 struct ethtool_ts_info info = {
1335                         .cmd = ETHTOOL_GET_TS_INFO,
1336                         .phc_index = -1,
1337                 };
1338
1339                 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1340                         NL_SET_ERR_MSG(extack,
1341                                        "The 'clockid' cannot be specified for full offload");
1342                         goto out;
1343                 }
1344
1345                 if (ops && ops->get_ts_info)
1346                         err = ops->get_ts_info(dev, &info);
1347
1348                 if (err || info.phc_index < 0) {
1349                         NL_SET_ERR_MSG(extack,
1350                                        "Device does not have a PTP clock");
1351                         err = -ENOTSUPP;
1352                         goto out;
1353                 }
1354         } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1355                 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1356                 enum tk_offsets tk_offset;
1357
1358                 /* We only support static clockids and we don't allow
1359                  * for it to be modified after the first init.
1360                  */
1361                 if (clockid < 0 ||
1362                     (q->clockid != -1 && q->clockid != clockid)) {
1363                         NL_SET_ERR_MSG(extack,
1364                                        "Changing the 'clockid' of a running schedule is not supported");
1365                         err = -ENOTSUPP;
1366                         goto out;
1367                 }
1368
1369                 switch (clockid) {
1370                 case CLOCK_REALTIME:
1371                         tk_offset = TK_OFFS_REAL;
1372                         break;
1373                 case CLOCK_MONOTONIC:
1374                         tk_offset = TK_OFFS_MAX;
1375                         break;
1376                 case CLOCK_BOOTTIME:
1377                         tk_offset = TK_OFFS_BOOT;
1378                         break;
1379                 case CLOCK_TAI:
1380                         tk_offset = TK_OFFS_TAI;
1381                         break;
1382                 default:
1383                         NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1384                         err = -EINVAL;
1385                         goto out;
1386                 }
1387                 /* This pairs with READ_ONCE() in taprio_mono_to_any */
1388                 WRITE_ONCE(q->tk_offset, tk_offset);
1389
1390                 q->clockid = clockid;
1391         } else {
1392                 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1393                 goto out;
1394         }
1395
1396         /* Everything went ok, return success. */
1397         err = 0;
1398
1399 out:
1400         return err;
1401 }
1402
1403 static int taprio_parse_tc_entry(struct Qdisc *sch,
1404                                  struct nlattr *opt,
1405                                  u32 max_sdu[TC_QOPT_MAX_QUEUE],
1406                                  unsigned long *seen_tcs,
1407                                  struct netlink_ext_ack *extack)
1408 {
1409         struct nlattr *tb[TCA_TAPRIO_TC_ENTRY_MAX + 1] = { };
1410         struct net_device *dev = qdisc_dev(sch);
1411         u32 val = 0;
1412         int err, tc;
1413
1414         err = nla_parse_nested(tb, TCA_TAPRIO_TC_ENTRY_MAX, opt,
1415                                taprio_tc_policy, extack);
1416         if (err < 0)
1417                 return err;
1418
1419         if (!tb[TCA_TAPRIO_TC_ENTRY_INDEX]) {
1420                 NL_SET_ERR_MSG_MOD(extack, "TC entry index missing");
1421                 return -EINVAL;
1422         }
1423
1424         tc = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_INDEX]);
1425         if (tc >= TC_QOPT_MAX_QUEUE) {
1426                 NL_SET_ERR_MSG_MOD(extack, "TC entry index out of range");
1427                 return -ERANGE;
1428         }
1429
1430         if (*seen_tcs & BIT(tc)) {
1431                 NL_SET_ERR_MSG_MOD(extack, "Duplicate TC entry");
1432                 return -EINVAL;
1433         }
1434
1435         *seen_tcs |= BIT(tc);
1436
1437         if (tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU])
1438                 val = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]);
1439
1440         if (val > dev->max_mtu) {
1441                 NL_SET_ERR_MSG_MOD(extack, "TC max SDU exceeds device max MTU");
1442                 return -ERANGE;
1443         }
1444
1445         max_sdu[tc] = val;
1446
1447         return 0;
1448 }
1449
1450 static int taprio_parse_tc_entries(struct Qdisc *sch,
1451                                    struct nlattr *opt,
1452                                    struct netlink_ext_ack *extack)
1453 {
1454         struct taprio_sched *q = qdisc_priv(sch);
1455         struct net_device *dev = qdisc_dev(sch);
1456         u32 max_sdu[TC_QOPT_MAX_QUEUE];
1457         unsigned long seen_tcs = 0;
1458         struct nlattr *n;
1459         int tc, rem;
1460         int err = 0;
1461
1462         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++)
1463                 max_sdu[tc] = q->max_sdu[tc];
1464
1465         nla_for_each_nested(n, opt, rem) {
1466                 if (nla_type(n) != TCA_TAPRIO_ATTR_TC_ENTRY)
1467                         continue;
1468
1469                 err = taprio_parse_tc_entry(sch, n, max_sdu, &seen_tcs, extack);
1470                 if (err)
1471                         goto out;
1472         }
1473
1474         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1475                 q->max_sdu[tc] = max_sdu[tc];
1476                 if (max_sdu[tc])
1477                         q->max_frm_len[tc] = max_sdu[tc] + dev->hard_header_len;
1478                 else
1479                         q->max_frm_len[tc] = U32_MAX; /* never oversized */
1480         }
1481
1482 out:
1483         return err;
1484 }
1485
1486 static int taprio_mqprio_cmp(const struct net_device *dev,
1487                              const struct tc_mqprio_qopt *mqprio)
1488 {
1489         int i;
1490
1491         if (!mqprio || mqprio->num_tc != dev->num_tc)
1492                 return -1;
1493
1494         for (i = 0; i < mqprio->num_tc; i++)
1495                 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1496                     dev->tc_to_txq[i].offset != mqprio->offset[i])
1497                         return -1;
1498
1499         for (i = 0; i <= TC_BITMASK; i++)
1500                 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1501                         return -1;
1502
1503         return 0;
1504 }
1505
1506 /* The semantics of the 'flags' argument in relation to 'change()'
1507  * requests, are interpreted following two rules (which are applied in
1508  * this order): (1) an omitted 'flags' argument is interpreted as
1509  * zero; (2) the 'flags' of a "running" taprio instance cannot be
1510  * changed.
1511  */
1512 static int taprio_new_flags(const struct nlattr *attr, u32 old,
1513                             struct netlink_ext_ack *extack)
1514 {
1515         u32 new = 0;
1516
1517         if (attr)
1518                 new = nla_get_u32(attr);
1519
1520         if (old != TAPRIO_FLAGS_INVALID && old != new) {
1521                 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1522                 return -EOPNOTSUPP;
1523         }
1524
1525         if (!taprio_flags_valid(new)) {
1526                 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1527                 return -EINVAL;
1528         }
1529
1530         return new;
1531 }
1532
1533 static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1534                          struct netlink_ext_ack *extack)
1535 {
1536         struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1537         struct sched_gate_list *oper, *admin, *new_admin;
1538         struct taprio_sched *q = qdisc_priv(sch);
1539         struct net_device *dev = qdisc_dev(sch);
1540         struct tc_mqprio_qopt *mqprio = NULL;
1541         unsigned long flags;
1542         ktime_t start;
1543         int i, err;
1544
1545         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1546                                           taprio_policy, extack);
1547         if (err < 0)
1548                 return err;
1549
1550         if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1551                 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1552
1553         err = taprio_new_flags(tb[TCA_TAPRIO_ATTR_FLAGS],
1554                                q->flags, extack);
1555         if (err < 0)
1556                 return err;
1557
1558         q->flags = err;
1559
1560         err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1561         if (err < 0)
1562                 return err;
1563
1564         err = taprio_parse_tc_entries(sch, opt, extack);
1565         if (err)
1566                 return err;
1567
1568         new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1569         if (!new_admin) {
1570                 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1571                 return -ENOMEM;
1572         }
1573         INIT_LIST_HEAD(&new_admin->entries);
1574
1575         oper = rtnl_dereference(q->oper_sched);
1576         admin = rtnl_dereference(q->admin_sched);
1577
1578         /* no changes - no new mqprio settings */
1579         if (!taprio_mqprio_cmp(dev, mqprio))
1580                 mqprio = NULL;
1581
1582         if (mqprio && (oper || admin)) {
1583                 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1584                 err = -ENOTSUPP;
1585                 goto free_sched;
1586         }
1587
1588         err = parse_taprio_schedule(q, tb, new_admin, extack);
1589         if (err < 0)
1590                 goto free_sched;
1591
1592         if (new_admin->num_entries == 0) {
1593                 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1594                 err = -EINVAL;
1595                 goto free_sched;
1596         }
1597
1598         err = taprio_parse_clockid(sch, tb, extack);
1599         if (err < 0)
1600                 goto free_sched;
1601
1602         taprio_set_picos_per_byte(dev, q);
1603
1604         if (mqprio) {
1605                 err = netdev_set_num_tc(dev, mqprio->num_tc);
1606                 if (err)
1607                         goto free_sched;
1608                 for (i = 0; i < mqprio->num_tc; i++)
1609                         netdev_set_tc_queue(dev, i,
1610                                             mqprio->count[i],
1611                                             mqprio->offset[i]);
1612
1613                 /* Always use supplied priority mappings */
1614                 for (i = 0; i <= TC_BITMASK; i++)
1615                         netdev_set_prio_tc_map(dev, i,
1616                                                mqprio->prio_tc_map[i]);
1617         }
1618
1619         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1620                 err = taprio_enable_offload(dev, q, new_admin, extack);
1621         else
1622                 err = taprio_disable_offload(dev, q, extack);
1623         if (err)
1624                 goto free_sched;
1625
1626         /* Protects against enqueue()/dequeue() */
1627         spin_lock_bh(qdisc_lock(sch));
1628
1629         if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1630                 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1631                         NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1632                         err = -EINVAL;
1633                         goto unlock;
1634                 }
1635
1636                 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1637         }
1638
1639         if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1640             !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1641             !hrtimer_active(&q->advance_timer)) {
1642                 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1643                 q->advance_timer.function = advance_sched;
1644         }
1645
1646         err = taprio_get_start_time(sch, new_admin, &start);
1647         if (err < 0) {
1648                 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1649                 goto unlock;
1650         }
1651
1652         setup_txtime(q, new_admin, start);
1653
1654         if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1655                 if (!oper) {
1656                         rcu_assign_pointer(q->oper_sched, new_admin);
1657                         err = 0;
1658                         new_admin = NULL;
1659                         goto unlock;
1660                 }
1661
1662                 rcu_assign_pointer(q->admin_sched, new_admin);
1663                 if (admin)
1664                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1665         } else {
1666                 setup_first_close_time(q, new_admin, start);
1667
1668                 /* Protects against advance_sched() */
1669                 spin_lock_irqsave(&q->current_entry_lock, flags);
1670
1671                 taprio_start_sched(sch, start, new_admin);
1672
1673                 rcu_assign_pointer(q->admin_sched, new_admin);
1674                 if (admin)
1675                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1676
1677                 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1678
1679                 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1680                         taprio_offload_config_changed(q);
1681         }
1682
1683         new_admin = NULL;
1684         err = 0;
1685
1686 unlock:
1687         spin_unlock_bh(qdisc_lock(sch));
1688
1689 free_sched:
1690         if (new_admin)
1691                 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
1692
1693         return err;
1694 }
1695
1696 static void taprio_reset(struct Qdisc *sch)
1697 {
1698         struct taprio_sched *q = qdisc_priv(sch);
1699         struct net_device *dev = qdisc_dev(sch);
1700         int i;
1701
1702         hrtimer_cancel(&q->advance_timer);
1703         if (q->qdiscs) {
1704                 for (i = 0; i < dev->num_tx_queues; i++)
1705                         if (q->qdiscs[i])
1706                                 qdisc_reset(q->qdiscs[i]);
1707         }
1708 }
1709
1710 static void taprio_destroy(struct Qdisc *sch)
1711 {
1712         struct taprio_sched *q = qdisc_priv(sch);
1713         struct net_device *dev = qdisc_dev(sch);
1714         struct sched_gate_list *oper, *admin;
1715         unsigned int i;
1716
1717         list_del(&q->taprio_list);
1718
1719         /* Note that taprio_reset() might not be called if an error
1720          * happens in qdisc_create(), after taprio_init() has been called.
1721          */
1722         hrtimer_cancel(&q->advance_timer);
1723
1724         taprio_disable_offload(dev, q, NULL);
1725
1726         if (q->qdiscs) {
1727                 for (i = 0; i < dev->num_tx_queues; i++)
1728                         qdisc_put(q->qdiscs[i]);
1729
1730                 kfree(q->qdiscs);
1731         }
1732         q->qdiscs = NULL;
1733
1734         netdev_reset_tc(dev);
1735
1736         oper = rtnl_dereference(q->oper_sched);
1737         admin = rtnl_dereference(q->admin_sched);
1738
1739         if (oper)
1740                 call_rcu(&oper->rcu, taprio_free_sched_cb);
1741
1742         if (admin)
1743                 call_rcu(&admin->rcu, taprio_free_sched_cb);
1744 }
1745
1746 static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
1747                        struct netlink_ext_ack *extack)
1748 {
1749         struct taprio_sched *q = qdisc_priv(sch);
1750         struct net_device *dev = qdisc_dev(sch);
1751         int i;
1752
1753         spin_lock_init(&q->current_entry_lock);
1754
1755         hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
1756         q->advance_timer.function = advance_sched;
1757
1758         q->root = sch;
1759
1760         /* We only support static clockids. Use an invalid value as default
1761          * and get the valid one on taprio_change().
1762          */
1763         q->clockid = -1;
1764         q->flags = TAPRIO_FLAGS_INVALID;
1765
1766         list_add(&q->taprio_list, &taprio_list);
1767
1768         if (sch->parent != TC_H_ROOT) {
1769                 NL_SET_ERR_MSG_MOD(extack, "Can only be attached as root qdisc");
1770                 return -EOPNOTSUPP;
1771         }
1772
1773         if (!netif_is_multiqueue(dev)) {
1774                 NL_SET_ERR_MSG_MOD(extack, "Multi-queue device is required");
1775                 return -EOPNOTSUPP;
1776         }
1777
1778         /* pre-allocate qdisc, attachment can't fail */
1779         q->qdiscs = kcalloc(dev->num_tx_queues,
1780                             sizeof(q->qdiscs[0]),
1781                             GFP_KERNEL);
1782
1783         if (!q->qdiscs)
1784                 return -ENOMEM;
1785
1786         if (!opt)
1787                 return -EINVAL;
1788
1789         for (i = 0; i < dev->num_tx_queues; i++) {
1790                 struct netdev_queue *dev_queue;
1791                 struct Qdisc *qdisc;
1792
1793                 dev_queue = netdev_get_tx_queue(dev, i);
1794                 qdisc = qdisc_create_dflt(dev_queue,
1795                                           &pfifo_qdisc_ops,
1796                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
1797                                                     TC_H_MIN(i + 1)),
1798                                           extack);
1799                 if (!qdisc)
1800                         return -ENOMEM;
1801
1802                 if (i < dev->real_num_tx_queues)
1803                         qdisc_hash_add(qdisc, false);
1804
1805                 q->qdiscs[i] = qdisc;
1806         }
1807
1808         return taprio_change(sch, opt, extack);
1809 }
1810
1811 static void taprio_attach(struct Qdisc *sch)
1812 {
1813         struct taprio_sched *q = qdisc_priv(sch);
1814         struct net_device *dev = qdisc_dev(sch);
1815         unsigned int ntx;
1816
1817         /* Attach underlying qdisc */
1818         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
1819                 struct Qdisc *qdisc = q->qdiscs[ntx];
1820                 struct Qdisc *old;
1821
1822                 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1823                         qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1824                         old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
1825                 } else {
1826                         old = dev_graft_qdisc(qdisc->dev_queue, sch);
1827                         qdisc_refcount_inc(sch);
1828                 }
1829                 if (old)
1830                         qdisc_put(old);
1831         }
1832
1833         /* access to the child qdiscs is not needed in offload mode */
1834         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1835                 kfree(q->qdiscs);
1836                 q->qdiscs = NULL;
1837         }
1838 }
1839
1840 static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
1841                                              unsigned long cl)
1842 {
1843         struct net_device *dev = qdisc_dev(sch);
1844         unsigned long ntx = cl - 1;
1845
1846         if (ntx >= dev->num_tx_queues)
1847                 return NULL;
1848
1849         return netdev_get_tx_queue(dev, ntx);
1850 }
1851
1852 static int taprio_graft(struct Qdisc *sch, unsigned long cl,
1853                         struct Qdisc *new, struct Qdisc **old,
1854                         struct netlink_ext_ack *extack)
1855 {
1856         struct taprio_sched *q = qdisc_priv(sch);
1857         struct net_device *dev = qdisc_dev(sch);
1858         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
1859
1860         if (!dev_queue)
1861                 return -EINVAL;
1862
1863         if (dev->flags & IFF_UP)
1864                 dev_deactivate(dev);
1865
1866         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1867                 *old = dev_graft_qdisc(dev_queue, new);
1868         } else {
1869                 *old = q->qdiscs[cl - 1];
1870                 q->qdiscs[cl - 1] = new;
1871         }
1872
1873         if (new)
1874                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1875
1876         if (dev->flags & IFF_UP)
1877                 dev_activate(dev);
1878
1879         return 0;
1880 }
1881
1882 static int dump_entry(struct sk_buff *msg,
1883                       const struct sched_entry *entry)
1884 {
1885         struct nlattr *item;
1886
1887         item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
1888         if (!item)
1889                 return -ENOSPC;
1890
1891         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
1892                 goto nla_put_failure;
1893
1894         if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
1895                 goto nla_put_failure;
1896
1897         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
1898                         entry->gate_mask))
1899                 goto nla_put_failure;
1900
1901         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
1902                         entry->interval))
1903                 goto nla_put_failure;
1904
1905         return nla_nest_end(msg, item);
1906
1907 nla_put_failure:
1908         nla_nest_cancel(msg, item);
1909         return -1;
1910 }
1911
1912 static int dump_schedule(struct sk_buff *msg,
1913                          const struct sched_gate_list *root)
1914 {
1915         struct nlattr *entry_list;
1916         struct sched_entry *entry;
1917
1918         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
1919                         root->base_time, TCA_TAPRIO_PAD))
1920                 return -1;
1921
1922         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
1923                         root->cycle_time, TCA_TAPRIO_PAD))
1924                 return -1;
1925
1926         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
1927                         root->cycle_time_extension, TCA_TAPRIO_PAD))
1928                 return -1;
1929
1930         entry_list = nla_nest_start_noflag(msg,
1931                                            TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
1932         if (!entry_list)
1933                 goto error_nest;
1934
1935         list_for_each_entry(entry, &root->entries, list) {
1936                 if (dump_entry(msg, entry) < 0)
1937                         goto error_nest;
1938         }
1939
1940         nla_nest_end(msg, entry_list);
1941         return 0;
1942
1943 error_nest:
1944         nla_nest_cancel(msg, entry_list);
1945         return -1;
1946 }
1947
1948 static int taprio_dump_tc_entries(struct taprio_sched *q, struct sk_buff *skb)
1949 {
1950         struct nlattr *n;
1951         int tc;
1952
1953         for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
1954                 n = nla_nest_start(skb, TCA_TAPRIO_ATTR_TC_ENTRY);
1955                 if (!n)
1956                         return -EMSGSIZE;
1957
1958                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_INDEX, tc))
1959                         goto nla_put_failure;
1960
1961                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_MAX_SDU,
1962                                 q->max_sdu[tc]))
1963                         goto nla_put_failure;
1964
1965                 nla_nest_end(skb, n);
1966         }
1967
1968         return 0;
1969
1970 nla_put_failure:
1971         nla_nest_cancel(skb, n);
1972         return -EMSGSIZE;
1973 }
1974
1975 static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
1976 {
1977         struct taprio_sched *q = qdisc_priv(sch);
1978         struct net_device *dev = qdisc_dev(sch);
1979         struct sched_gate_list *oper, *admin;
1980         struct tc_mqprio_qopt opt = { 0 };
1981         struct nlattr *nest, *sched_nest;
1982         unsigned int i;
1983
1984         oper = rtnl_dereference(q->oper_sched);
1985         admin = rtnl_dereference(q->admin_sched);
1986
1987         opt.num_tc = netdev_get_num_tc(dev);
1988         memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
1989
1990         for (i = 0; i < netdev_get_num_tc(dev); i++) {
1991                 opt.count[i] = dev->tc_to_txq[i].count;
1992                 opt.offset[i] = dev->tc_to_txq[i].offset;
1993         }
1994
1995         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1996         if (!nest)
1997                 goto start_error;
1998
1999         if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
2000                 goto options_error;
2001
2002         if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
2003             nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
2004                 goto options_error;
2005
2006         if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
2007                 goto options_error;
2008
2009         if (q->txtime_delay &&
2010             nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
2011                 goto options_error;
2012
2013         if (taprio_dump_tc_entries(q, skb))
2014                 goto options_error;
2015
2016         if (oper && dump_schedule(skb, oper))
2017                 goto options_error;
2018
2019         if (!admin)
2020                 goto done;
2021
2022         sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
2023         if (!sched_nest)
2024                 goto options_error;
2025
2026         if (dump_schedule(skb, admin))
2027                 goto admin_error;
2028
2029         nla_nest_end(skb, sched_nest);
2030
2031 done:
2032         return nla_nest_end(skb, nest);
2033
2034 admin_error:
2035         nla_nest_cancel(skb, sched_nest);
2036
2037 options_error:
2038         nla_nest_cancel(skb, nest);
2039
2040 start_error:
2041         return -ENOSPC;
2042 }
2043
2044 static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
2045 {
2046         struct taprio_sched *q = qdisc_priv(sch);
2047         struct net_device *dev = qdisc_dev(sch);
2048         unsigned int ntx = cl - 1;
2049
2050         if (ntx >= dev->num_tx_queues)
2051                 return NULL;
2052
2053         return q->qdiscs[ntx];
2054 }
2055
2056 static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
2057 {
2058         unsigned int ntx = TC_H_MIN(classid);
2059
2060         if (!taprio_queue_get(sch, ntx))
2061                 return 0;
2062         return ntx;
2063 }
2064
2065 static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
2066                              struct sk_buff *skb, struct tcmsg *tcm)
2067 {
2068         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2069
2070         tcm->tcm_parent = TC_H_ROOT;
2071         tcm->tcm_handle |= TC_H_MIN(cl);
2072         tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
2073
2074         return 0;
2075 }
2076
2077 static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
2078                                    struct gnet_dump *d)
2079         __releases(d->lock)
2080         __acquires(d->lock)
2081 {
2082         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2083
2084         sch = dev_queue->qdisc_sleeping;
2085         if (gnet_stats_copy_basic(d, NULL, &sch->bstats, true) < 0 ||
2086             qdisc_qstats_copy(d, sch) < 0)
2087                 return -1;
2088         return 0;
2089 }
2090
2091 static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2092 {
2093         struct net_device *dev = qdisc_dev(sch);
2094         unsigned long ntx;
2095
2096         if (arg->stop)
2097                 return;
2098
2099         arg->count = arg->skip;
2100         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
2101                 if (!tc_qdisc_stats_dump(sch, ntx + 1, arg))
2102                         break;
2103         }
2104 }
2105
2106 static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
2107                                                 struct tcmsg *tcm)
2108 {
2109         return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
2110 }
2111
2112 static const struct Qdisc_class_ops taprio_class_ops = {
2113         .graft          = taprio_graft,
2114         .leaf           = taprio_leaf,
2115         .find           = taprio_find,
2116         .walk           = taprio_walk,
2117         .dump           = taprio_dump_class,
2118         .dump_stats     = taprio_dump_class_stats,
2119         .select_queue   = taprio_select_queue,
2120 };
2121
2122 static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
2123         .cl_ops         = &taprio_class_ops,
2124         .id             = "taprio",
2125         .priv_size      = sizeof(struct taprio_sched),
2126         .init           = taprio_init,
2127         .change         = taprio_change,
2128         .destroy        = taprio_destroy,
2129         .reset          = taprio_reset,
2130         .attach         = taprio_attach,
2131         .peek           = taprio_peek,
2132         .dequeue        = taprio_dequeue,
2133         .enqueue        = taprio_enqueue,
2134         .dump           = taprio_dump,
2135         .owner          = THIS_MODULE,
2136 };
2137
2138 static struct notifier_block taprio_device_notifier = {
2139         .notifier_call = taprio_dev_notifier,
2140 };
2141
2142 static int __init taprio_module_init(void)
2143 {
2144         int err = register_netdevice_notifier(&taprio_device_notifier);
2145
2146         if (err)
2147                 return err;
2148
2149         return register_qdisc(&taprio_qdisc_ops);
2150 }
2151
2152 static void __exit taprio_module_exit(void)
2153 {
2154         unregister_qdisc(&taprio_qdisc_ops);
2155         unregister_netdevice_notifier(&taprio_device_notifier);
2156 }
2157
2158 module_init(taprio_module_init);
2159 module_exit(taprio_module_exit);
2160 MODULE_LICENSE("GPL");