net/sched/sch_taprio.c

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