net/sched/sch_taprio.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* net/sched/sch_taprio.c        Time Aware Priority Scheduler
   4  *
   5  * Authors:     Vinicius Costa Gomes <vinicius.gomes@intel.com>
   6  *
   7  */
   8
   9 #include <linux/ethtool.h>
  10 #include <linux/ethtool_netlink.h>
  11 #include <linux/types.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/string.h>
  15 #include <linux/list.h>
  16 #include <linux/errno.h>
  17 #include <linux/skbuff.h>
  18 #include <linux/math64.h>
  19 #include <linux/module.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/time.h>
  23 #include <net/netlink.h>
  24 #include <net/pkt_sched.h>
  25 #include <net/pkt_cls.h>
  26 #include <net/sch_generic.h>
  27 #include <net/sock.h>
  28 #include <net/tcp.h>
  29
  30 #include "sch_mqprio_lib.h"
  31
  32 static LIST_HEAD(taprio_list);
  33 static struct static_key_false taprio_have_broken_mqprio;
  34 static struct static_key_false taprio_have_working_mqprio;
  35
  36 #define TAPRIO_ALL_GATES_OPEN -1
  37
  38 #define TXTIME_ASSIST_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST)
  39 #define FULL_OFFLOAD_IS_ENABLED(flags) ((flags) & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD)
  40 #define TAPRIO_FLAGS_INVALID U32_MAX
  41
  42 struct sched_entry {
  43         /* Durations between this GCL entry and the GCL entry where the
  44          * respective traffic class gate closes
  45          */
  46         u64 gate_duration[TC_MAX_QUEUE];
  47         atomic_t budget[TC_MAX_QUEUE];
  48         /* The qdisc makes some effort so that no packet leaves
  49          * after this time
  50          */
  51         ktime_t gate_close_time[TC_MAX_QUEUE];
  52         struct list_head list;
  53         /* Used to calculate when to advance the schedule */
  54         ktime_t end_time;
  55         ktime_t next_txtime;
  56         int index;
  57         u32 gate_mask;
  58         u32 interval;
  59         u8 command;
  60 };
  61
  62 struct sched_gate_list {
  63         /* Longest non-zero contiguous gate durations per traffic class,
  64          * or 0 if a traffic class gate never opens during the schedule.
  65          */
  66         u64 max_open_gate_duration[TC_MAX_QUEUE];
  67         u32 max_frm_len[TC_MAX_QUEUE]; /* for the fast path */
  68         u32 max_sdu[TC_MAX_QUEUE]; /* for dump */
  69         struct rcu_head rcu;
  70         struct list_head entries;
  71         size_t num_entries;
  72         ktime_t cycle_end_time;
  73         s64 cycle_time;
  74         s64 cycle_time_extension;
  75         s64 base_time;
  76 };
  77
  78 struct taprio_sched {
  79         struct Qdisc **qdiscs;
  80         struct Qdisc *root;
  81         u32 flags;
  82         enum tk_offsets tk_offset;
  83         int clockid;
  84         bool offloaded;
  85         bool detected_mqprio;
  86         bool broken_mqprio;
  87         atomic64_t picos_per_byte; /* Using picoseconds because for 10Gbps+
  88                                     * speeds it's sub-nanoseconds per byte
  89                                     */
  90
  91         /* Protects the update side of the RCU protected current_entry */
  92         spinlock_t current_entry_lock;
  93         struct sched_entry __rcu *current_entry;
  94         struct sched_gate_list __rcu *oper_sched;
  95         struct sched_gate_list __rcu *admin_sched;
  96         struct hrtimer advance_timer;
  97         struct list_head taprio_list;
  98         int cur_txq[TC_MAX_QUEUE];
  99         u32 max_sdu[TC_MAX_QUEUE]; /* save info from the user */
 100         u32 fp[TC_QOPT_MAX_QUEUE]; /* only for dump and offloading */
 101         u32 txtime_delay;
 102 };
 103
 104 struct __tc_taprio_qopt_offload {
 105         refcount_t users;
 106         struct tc_taprio_qopt_offload offload;
 107 };
 108
 109 static void taprio_calculate_gate_durations(struct taprio_sched *q,
 110                                             struct sched_gate_list *sched)
 111 {
 112         struct net_device *dev = qdisc_dev(q->root);
 113         int num_tc = netdev_get_num_tc(dev);
 114         struct sched_entry *entry, *cur;
 115         int tc;
 116
 117         list_for_each_entry(entry, &sched->entries, list) {
 118                 u32 gates_still_open = entry->gate_mask;
 119
 120                 /* For each traffic class, calculate each open gate duration,
 121                  * starting at this schedule entry and ending at the schedule
 122                  * entry containing a gate close event for that TC.
 123                  */
 124                 cur = entry;
 125
 126                 do {
 127                         if (!gates_still_open)
 128                                 break;
 129
 130                         for (tc = 0; tc < num_tc; tc++) {
 131                                 if (!(gates_still_open & BIT(tc)))
 132                                         continue;
 133
 134                                 if (cur->gate_mask & BIT(tc))
 135                                         entry->gate_duration[tc] += cur->interval;
 136                                 else
 137                                         gates_still_open &= ~BIT(tc);
 138                         }
 139
 140                         cur = list_next_entry_circular(cur, &sched->entries, list);
 141                 } while (cur != entry);
 142
 143                 /* Keep track of the maximum gate duration for each traffic
 144                  * class, taking care to not confuse a traffic class which is
 145                  * temporarily closed with one that is always closed.
 146                  */
 147                 for (tc = 0; tc < num_tc; tc++)
 148                         if (entry->gate_duration[tc] &&
 149                             sched->max_open_gate_duration[tc] < entry->gate_duration[tc])
 150                                 sched->max_open_gate_duration[tc] = entry->gate_duration[tc];
 151         }
 152 }
 153
 154 static bool taprio_entry_allows_tx(ktime_t skb_end_time,
 155                                    struct sched_entry *entry, int tc)
 156 {
 157         return ktime_before(skb_end_time, entry->gate_close_time[tc]);
 158 }
 159
 160 static ktime_t sched_base_time(const struct sched_gate_list *sched)
 161 {
 162         if (!sched)
 163                 return KTIME_MAX;
 164
 165         return ns_to_ktime(sched->base_time);
 166 }
 167
 168 static ktime_t taprio_mono_to_any(const struct taprio_sched *q, ktime_t mono)
 169 {
 170         /* This pairs with WRITE_ONCE() in taprio_parse_clockid() */
 171         enum tk_offsets tk_offset = READ_ONCE(q->tk_offset);
 172
 173         switch (tk_offset) {
 174         case TK_OFFS_MAX:
 175                 return mono;
 176         default:
 177                 return ktime_mono_to_any(mono, tk_offset);
 178         }
 179 }
 180
 181 static ktime_t taprio_get_time(const struct taprio_sched *q)
 182 {
 183         return taprio_mono_to_any(q, ktime_get());
 184 }
 185
 186 static void taprio_free_sched_cb(struct rcu_head *head)
 187 {
 188         struct sched_gate_list *sched = container_of(head, struct sched_gate_list, rcu);
 189         struct sched_entry *entry, *n;
 190
 191         list_for_each_entry_safe(entry, n, &sched->entries, list) {
 192                 list_del(&entry->list);
 193                 kfree(entry);
 194         }
 195
 196         kfree(sched);
 197 }
 198
 199 static void switch_schedules(struct taprio_sched *q,
 200                              struct sched_gate_list **admin,
 201                              struct sched_gate_list **oper)
 202 {
 203         rcu_assign_pointer(q->oper_sched, *admin);
 204         rcu_assign_pointer(q->admin_sched, NULL);
 205
 206         if (*oper)
 207                 call_rcu(&(*oper)->rcu, taprio_free_sched_cb);
 208
 209         *oper = *admin;
 210         *admin = NULL;
 211 }
 212
 213 /* Get how much time has been already elapsed in the current cycle. */
 214 static s32 get_cycle_time_elapsed(struct sched_gate_list *sched, ktime_t time)
 215 {
 216         ktime_t time_since_sched_start;
 217         s32 time_elapsed;
 218
 219         time_since_sched_start = ktime_sub(time, sched->base_time);
 220         div_s64_rem(time_since_sched_start, sched->cycle_time, &time_elapsed);
 221
 222         return time_elapsed;
 223 }
 224
 225 static ktime_t get_interval_end_time(struct sched_gate_list *sched,
 226                                      struct sched_gate_list *admin,
 227                                      struct sched_entry *entry,
 228                                      ktime_t intv_start)
 229 {
 230         s32 cycle_elapsed = get_cycle_time_elapsed(sched, intv_start);
 231         ktime_t intv_end, cycle_ext_end, cycle_end;
 232
 233         cycle_end = ktime_add_ns(intv_start, sched->cycle_time - cycle_elapsed);
 234         intv_end = ktime_add_ns(intv_start, entry->interval);
 235         cycle_ext_end = ktime_add(cycle_end, sched->cycle_time_extension);
 236
 237         if (ktime_before(intv_end, cycle_end))
 238                 return intv_end;
 239         else if (admin && admin != sched &&
 240                  ktime_after(admin->base_time, cycle_end) &&
 241                  ktime_before(admin->base_time, cycle_ext_end))
 242                 return admin->base_time;
 243         else
 244                 return cycle_end;
 245 }
 246
 247 static int length_to_duration(struct taprio_sched *q, int len)
 248 {
 249         return div_u64(len * atomic64_read(&q->picos_per_byte), PSEC_PER_NSEC);
 250 }
 251
 252 static int duration_to_length(struct taprio_sched *q, u64 duration)
 253 {
 254         return div_u64(duration * PSEC_PER_NSEC, atomic64_read(&q->picos_per_byte));
 255 }
 256
 257 /* Sets sched->max_sdu[] and sched->max_frm_len[] to the minimum between the
 258  * q->max_sdu[] requested by the user and the max_sdu dynamically determined by
 259  * the maximum open gate durations at the given link speed.
 260  */
 261 static void taprio_update_queue_max_sdu(struct taprio_sched *q,
 262                                         struct sched_gate_list *sched,
 263                                         struct qdisc_size_table *stab)
 264 {
 265         struct net_device *dev = qdisc_dev(q->root);
 266         int num_tc = netdev_get_num_tc(dev);
 267         u32 max_sdu_from_user;
 268         u32 max_sdu_dynamic;
 269         u32 max_sdu;
 270         int tc;
 271
 272         for (tc = 0; tc < num_tc; tc++) {
 273                 max_sdu_from_user = q->max_sdu[tc] ?: U32_MAX;
 274
 275                 /* TC gate never closes => keep the queueMaxSDU
 276                  * selected by the user
 277                  */
 278                 if (sched->max_open_gate_duration[tc] == sched->cycle_time) {
 279                         max_sdu_dynamic = U32_MAX;
 280                 } else {
 281                         u32 max_frm_len;
 282
 283                         max_frm_len = duration_to_length(q, sched->max_open_gate_duration[tc]);
 284                         /* Compensate for L1 overhead from size table,
 285                          * but don't let the frame size go negative
 286                          */
 287                         if (stab) {
 288                                 max_frm_len -= stab->szopts.overhead;
 289                                 max_frm_len = max_t(int, max_frm_len,
 290                                                     dev->hard_header_len + 1);
 291                         }
 292                         max_sdu_dynamic = max_frm_len - dev->hard_header_len;
 293                         if (max_sdu_dynamic > dev->max_mtu)
 294                                 max_sdu_dynamic = U32_MAX;
 295                 }
 296
 297                 max_sdu = min(max_sdu_dynamic, max_sdu_from_user);
 298
 299                 if (max_sdu != U32_MAX) {
 300                         sched->max_frm_len[tc] = max_sdu + dev->hard_header_len;
 301                         sched->max_sdu[tc] = max_sdu;
 302                 } else {
 303                         sched->max_frm_len[tc] = U32_MAX; /* never oversized */
 304                         sched->max_sdu[tc] = 0;
 305                 }
 306         }
 307 }
 308
 309 /* Returns the entry corresponding to next available interval. If
 310  * validate_interval is set, it only validates whether the timestamp occurs
 311  * when the gate corresponding to the skb's traffic class is open.
 312  */
 313 static struct sched_entry *find_entry_to_transmit(struct sk_buff *skb,
 314                                                   struct Qdisc *sch,
 315                                                   struct sched_gate_list *sched,
 316                                                   struct sched_gate_list *admin,
 317                                                   ktime_t time,
 318                                                   ktime_t *interval_start,
 319                                                   ktime_t *interval_end,
 320                                                   bool validate_interval)
 321 {
 322         ktime_t curr_intv_start, curr_intv_end, cycle_end, packet_transmit_time;
 323         ktime_t earliest_txtime = KTIME_MAX, txtime, cycle, transmit_end_time;
 324         struct sched_entry *entry = NULL, *entry_found = NULL;
 325         struct taprio_sched *q = qdisc_priv(sch);
 326         struct net_device *dev = qdisc_dev(sch);
 327         bool entry_available = false;
 328         s32 cycle_elapsed;
 329         int tc, n;
 330
 331         tc = netdev_get_prio_tc_map(dev, skb->priority);
 332         packet_transmit_time = length_to_duration(q, qdisc_pkt_len(skb));
 333
 334         *interval_start = 0;
 335         *interval_end = 0;
 336
 337         if (!sched)
 338                 return NULL;
 339
 340         cycle = sched->cycle_time;
 341         cycle_elapsed = get_cycle_time_elapsed(sched, time);
 342         curr_intv_end = ktime_sub_ns(time, cycle_elapsed);
 343         cycle_end = ktime_add_ns(curr_intv_end, cycle);
 344
 345         list_for_each_entry(entry, &sched->entries, list) {
 346                 curr_intv_start = curr_intv_end;
 347                 curr_intv_end = get_interval_end_time(sched, admin, entry,
 348                                                       curr_intv_start);
 349
 350                 if (ktime_after(curr_intv_start, cycle_end))
 351                         break;
 352
 353                 if (!(entry->gate_mask & BIT(tc)) ||
 354                     packet_transmit_time > entry->interval)
 355                         continue;
 356
 357                 txtime = entry->next_txtime;
 358
 359                 if (ktime_before(txtime, time) || validate_interval) {
 360                         transmit_end_time = ktime_add_ns(time, packet_transmit_time);
 361                         if ((ktime_before(curr_intv_start, time) &&
 362                              ktime_before(transmit_end_time, curr_intv_end)) ||
 363                             (ktime_after(curr_intv_start, time) && !validate_interval)) {
 364                                 entry_found = entry;
 365                                 *interval_start = curr_intv_start;
 366                                 *interval_end = curr_intv_end;
 367                                 break;
 368                         } else if (!entry_available && !validate_interval) {
 369                                 /* Here, we are just trying to find out the
 370                                  * first available interval in the next cycle.
 371                                  */
 372                                 entry_available = true;
 373                                 entry_found = entry;
 374                                 *interval_start = ktime_add_ns(curr_intv_start, cycle);
 375                                 *interval_end = ktime_add_ns(curr_intv_end, cycle);
 376                         }
 377                 } else if (ktime_before(txtime, earliest_txtime) &&
 378                            !entry_available) {
 379                         earliest_txtime = txtime;
 380                         entry_found = entry;
 381                         n = div_s64(ktime_sub(txtime, curr_intv_start), cycle);
 382                         *interval_start = ktime_add(curr_intv_start, n * cycle);
 383                         *interval_end = ktime_add(curr_intv_end, n * cycle);
 384                 }
 385         }
 386
 387         return entry_found;
 388 }
 389
 390 static bool is_valid_interval(struct sk_buff *skb, struct Qdisc *sch)
 391 {
 392         struct taprio_sched *q = qdisc_priv(sch);
 393         struct sched_gate_list *sched, *admin;
 394         ktime_t interval_start, interval_end;
 395         struct sched_entry *entry;
 396
 397         rcu_read_lock();
 398         sched = rcu_dereference(q->oper_sched);
 399         admin = rcu_dereference(q->admin_sched);
 400
 401         entry = find_entry_to_transmit(skb, sch, sched, admin, skb->tstamp,
 402                                        &interval_start, &interval_end, true);
 403         rcu_read_unlock();
 404
 405         return entry;
 406 }
 407
 408 static bool taprio_flags_valid(u32 flags)
 409 {
 410         /* Make sure no other flag bits are set. */
 411         if (flags & ~(TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST |
 412                       TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 413                 return false;
 414         /* txtime-assist and full offload are mutually exclusive */
 415         if ((flags & TCA_TAPRIO_ATTR_FLAG_TXTIME_ASSIST) &&
 416             (flags & TCA_TAPRIO_ATTR_FLAG_FULL_OFFLOAD))
 417                 return false;
 418         return true;
 419 }
 420
 421 /* This returns the tstamp value set by TCP in terms of the set clock. */
 422 static ktime_t get_tcp_tstamp(struct taprio_sched *q, struct sk_buff *skb)
 423 {
 424         unsigned int offset = skb_network_offset(skb);
 425         const struct ipv6hdr *ipv6h;
 426         const struct iphdr *iph;
 427         struct ipv6hdr _ipv6h;
 428
 429         ipv6h = skb_header_pointer(skb, offset, sizeof(_ipv6h), &_ipv6h);
 430         if (!ipv6h)
 431                 return 0;
 432
 433         if (ipv6h->version == 4) {
 434                 iph = (struct iphdr *)ipv6h;
 435                 offset += iph->ihl * 4;
 436
 437                 /* special-case 6in4 tunnelling, as that is a common way to get
 438                  * v6 connectivity in the home
 439                  */
 440                 if (iph->protocol == IPPROTO_IPV6) {
 441                         ipv6h = skb_header_pointer(skb, offset,
 442                                                    sizeof(_ipv6h), &_ipv6h);
 443
 444                         if (!ipv6h || ipv6h->nexthdr != IPPROTO_TCP)
 445                                 return 0;
 446                 } else if (iph->protocol != IPPROTO_TCP) {
 447                         return 0;
 448                 }
 449         } else if (ipv6h->version == 6 && ipv6h->nexthdr != IPPROTO_TCP) {
 450                 return 0;
 451         }
 452
 453         return taprio_mono_to_any(q, skb->skb_mstamp_ns);
 454 }
 455
 456 /* There are a few scenarios where we will have to modify the txtime from
 457  * what is read from next_txtime in sched_entry. They are:
 458  * 1. If txtime is in the past,
 459  *    a. The gate for the traffic class is currently open and packet can be
 460  *       transmitted before it closes, schedule the packet right away.
 461  *    b. If the gate corresponding to the traffic class is going to open later
 462  *       in the cycle, set the txtime of packet to the interval start.
 463  * 2. If txtime is in the future, there are packets corresponding to the
 464  *    current traffic class waiting to be transmitted. So, the following
 465  *    possibilities exist:
 466  *    a. We can transmit the packet before the window containing the txtime
 467  *       closes.
 468  *    b. The window might close before the transmission can be completed
 469  *       successfully. So, schedule the packet in the next open window.
 470  */
 471 static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)
 472 {
 473         ktime_t transmit_end_time, interval_end, interval_start, tcp_tstamp;
 474         struct taprio_sched *q = qdisc_priv(sch);
 475         struct sched_gate_list *sched, *admin;
 476         ktime_t minimum_time, now, txtime;
 477         int len, packet_transmit_time;
 478         struct sched_entry *entry;
 479         bool sched_changed;
 480
 481         now = taprio_get_time(q);
 482         minimum_time = ktime_add_ns(now, q->txtime_delay);
 483
 484         tcp_tstamp = get_tcp_tstamp(q, skb);
 485         minimum_time = max_t(ktime_t, minimum_time, tcp_tstamp);
 486
 487         rcu_read_lock();
 488         admin = rcu_dereference(q->admin_sched);
 489         sched = rcu_dereference(q->oper_sched);
 490         if (admin && ktime_after(minimum_time, admin->base_time))
 491                 switch_schedules(q, &admin, &sched);
 492
 493         /* Until the schedule starts, all the queues are open */
 494         if (!sched || ktime_before(minimum_time, sched->base_time)) {
 495                 txtime = minimum_time;
 496                 goto done;
 497         }
 498
 499         len = qdisc_pkt_len(skb);
 500         packet_transmit_time = length_to_duration(q, len);
 501
 502         do {
 503                 sched_changed = false;
 504
 505                 entry = find_entry_to_transmit(skb, sch, sched, admin,
 506                                                minimum_time,
 507                                                &interval_start, &interval_end,
 508                                                false);
 509                 if (!entry) {
 510                         txtime = 0;
 511                         goto done;
 512                 }
 513
 514                 txtime = entry->next_txtime;
 515                 txtime = max_t(ktime_t, txtime, minimum_time);
 516                 txtime = max_t(ktime_t, txtime, interval_start);
 517
 518                 if (admin && admin != sched &&
 519                     ktime_after(txtime, admin->base_time)) {
 520                         sched = admin;
 521                         sched_changed = true;
 522                         continue;
 523                 }
 524
 525                 transmit_end_time = ktime_add(txtime, packet_transmit_time);
 526                 minimum_time = transmit_end_time;
 527
 528                 /* Update the txtime of current entry to the next time it's
 529                  * interval starts.
 530                  */
 531                 if (ktime_after(transmit_end_time, interval_end))
 532                         entry->next_txtime = ktime_add(interval_start, sched->cycle_time);
 533         } while (sched_changed || ktime_after(transmit_end_time, interval_end));
 534
 535         entry->next_txtime = transmit_end_time;
 536
 537 done:
 538         rcu_read_unlock();
 539         return txtime;
 540 }
 541
 542 /* Devices with full offload are expected to honor this in hardware */
 543 static bool taprio_skb_exceeds_queue_max_sdu(struct Qdisc *sch,
 544                                              struct sk_buff *skb)
 545 {
 546         struct taprio_sched *q = qdisc_priv(sch);
 547         struct net_device *dev = qdisc_dev(sch);
 548         struct sched_gate_list *sched;
 549         int prio = skb->priority;
 550         bool exceeds = false;
 551         u8 tc;
 552
 553         tc = netdev_get_prio_tc_map(dev, prio);
 554
 555         rcu_read_lock();
 556         sched = rcu_dereference(q->oper_sched);
 557         if (sched && skb->len > sched->max_frm_len[tc])
 558                 exceeds = true;
 559         rcu_read_unlock();
 560
 561         return exceeds;
 562 }
 563
 564 static int taprio_enqueue_one(struct sk_buff *skb, struct Qdisc *sch,
 565                               struct Qdisc *child, struct sk_buff **to_free)
 566 {
 567         struct taprio_sched *q = qdisc_priv(sch);
 568
 569         /* sk_flags are only safe to use on full sockets. */
 570         if (skb->sk && sk_fullsock(skb->sk) && sock_flag(skb->sk, SOCK_TXTIME)) {
 571                 if (!is_valid_interval(skb, sch))
 572                         return qdisc_drop(skb, sch, to_free);
 573         } else if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
 574                 skb->tstamp = get_packet_txtime(skb, sch);
 575                 if (!skb->tstamp)
 576                         return qdisc_drop(skb, sch, to_free);
 577         }
 578
 579         qdisc_qstats_backlog_inc(sch, skb);
 580         sch->q.qlen++;
 581
 582         return qdisc_enqueue(skb, child, to_free);
 583 }
 584
 585 static int taprio_enqueue_segmented(struct sk_buff *skb, struct Qdisc *sch,
 586                                     struct Qdisc *child,
 587                                     struct sk_buff **to_free)
 588 {
 589         unsigned int slen = 0, numsegs = 0, len = qdisc_pkt_len(skb);
 590         netdev_features_t features = netif_skb_features(skb);
 591         struct sk_buff *segs, *nskb;
 592         int ret;
 593
 594         segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 595         if (IS_ERR_OR_NULL(segs))
 596                 return qdisc_drop(skb, sch, to_free);
 597
 598         skb_list_walk_safe(segs, segs, nskb) {
 599                 skb_mark_not_on_list(segs);
 600                 qdisc_skb_cb(segs)->pkt_len = segs->len;
 601                 slen += segs->len;
 602
 603                 /* FIXME: we should be segmenting to a smaller size
 604                  * rather than dropping these
 605                  */
 606                 if (taprio_skb_exceeds_queue_max_sdu(sch, segs))
 607                         ret = qdisc_drop(segs, sch, to_free);
 608                 else
 609                         ret = taprio_enqueue_one(segs, sch, child, to_free);
 610
 611                 if (ret != NET_XMIT_SUCCESS) {
 612                         if (net_xmit_drop_count(ret))
 613                                 qdisc_qstats_drop(sch);
 614                 } else {
 615                         numsegs++;
 616                 }
 617         }
 618
 619         if (numsegs > 1)
 620                 qdisc_tree_reduce_backlog(sch, 1 - numsegs, len - slen);
 621         consume_skb(skb);
 622
 623         return numsegs > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
 624 }
 625
 626 /* Will not be called in the full offload case, since the TX queues are
 627  * attached to the Qdisc created using qdisc_create_dflt()
 628  */
 629 static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 630                           struct sk_buff **to_free)
 631 {
 632         struct taprio_sched *q = qdisc_priv(sch);
 633         struct Qdisc *child;
 634         int queue;
 635
 636         queue = skb_get_queue_mapping(skb);
 637
 638         child = q->qdiscs[queue];
 639         if (unlikely(!child))
 640                 return qdisc_drop(skb, sch, to_free);
 641
 642         if (taprio_skb_exceeds_queue_max_sdu(sch, skb)) {
 643                 /* Large packets might not be transmitted when the transmission
 644                  * duration exceeds any configured interval. Therefore, segment
 645                  * the skb into smaller chunks. Drivers with full offload are
 646                  * expected to handle this in hardware.
 647                  */
 648                 if (skb_is_gso(skb))
 649                         return taprio_enqueue_segmented(skb, sch, child,
 650                                                         to_free);
 651
 652                 return qdisc_drop(skb, sch, to_free);
 653         }
 654
 655         return taprio_enqueue_one(skb, sch, child, to_free);
 656 }
 657
 658 static struct sk_buff *taprio_peek(struct Qdisc *sch)
 659 {
 660         WARN_ONCE(1, "taprio only supports operating as root qdisc, peek() not implemented");
 661         return NULL;
 662 }
 663
 664 static void taprio_set_budgets(struct taprio_sched *q,
 665                                struct sched_gate_list *sched,
 666                                struct sched_entry *entry)
 667 {
 668         struct net_device *dev = qdisc_dev(q->root);
 669         int num_tc = netdev_get_num_tc(dev);
 670         int tc, budget;
 671
 672         for (tc = 0; tc < num_tc; tc++) {
 673                 /* Traffic classes which never close have infinite budget */
 674                 if (entry->gate_duration[tc] == sched->cycle_time)
 675                         budget = INT_MAX;
 676                 else
 677                         budget = div64_u64((u64)entry->gate_duration[tc] * PSEC_PER_NSEC,
 678                                            atomic64_read(&q->picos_per_byte));
 679
 680                 atomic_set(&entry->budget[tc], budget);
 681         }
 682 }
 683
 684 /* When an skb is sent, it consumes from the budget of all traffic classes */
 685 static int taprio_update_budgets(struct sched_entry *entry, size_t len,
 686                                  int tc_consumed, int num_tc)
 687 {
 688         int tc, budget, new_budget = 0;
 689
 690         for (tc = 0; tc < num_tc; tc++) {
 691                 budget = atomic_read(&entry->budget[tc]);
 692                 /* Don't consume from infinite budget */
 693                 if (budget == INT_MAX) {
 694                         if (tc == tc_consumed)
 695                                 new_budget = budget;
 696                         continue;
 697                 }
 698
 699                 if (tc == tc_consumed)
 700                         new_budget = atomic_sub_return(len, &entry->budget[tc]);
 701                 else
 702                         atomic_sub(len, &entry->budget[tc]);
 703         }
 704
 705         return new_budget;
 706 }
 707
 708 static struct sk_buff *taprio_dequeue_from_txq(struct Qdisc *sch, int txq,
 709                                                struct sched_entry *entry,
 710                                                u32 gate_mask)
 711 {
 712         struct taprio_sched *q = qdisc_priv(sch);
 713         struct net_device *dev = qdisc_dev(sch);
 714         struct Qdisc *child = q->qdiscs[txq];
 715         int num_tc = netdev_get_num_tc(dev);
 716         struct sk_buff *skb;
 717         ktime_t guard;
 718         int prio;
 719         int len;
 720         u8 tc;
 721
 722         if (unlikely(!child))
 723                 return NULL;
 724
 725         if (TXTIME_ASSIST_IS_ENABLED(q->flags))
 726                 goto skip_peek_checks;
 727
 728         skb = child->ops->peek(child);
 729         if (!skb)
 730                 return NULL;
 731
 732         prio = skb->priority;
 733         tc = netdev_get_prio_tc_map(dev, prio);
 734
 735         if (!(gate_mask & BIT(tc)))
 736                 return NULL;
 737
 738         len = qdisc_pkt_len(skb);
 739         guard = ktime_add_ns(taprio_get_time(q), length_to_duration(q, len));
 740
 741         /* In the case that there's no gate entry, there's no
 742          * guard band ...
 743          */
 744         if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 745             !taprio_entry_allows_tx(guard, entry, tc))
 746                 return NULL;
 747
 748         /* ... and no budget. */
 749         if (gate_mask != TAPRIO_ALL_GATES_OPEN &&
 750             taprio_update_budgets(entry, len, tc, num_tc) < 0)
 751                 return NULL;
 752
 753 skip_peek_checks:
 754         skb = child->ops->dequeue(child);
 755         if (unlikely(!skb))
 756                 return NULL;
 757
 758         qdisc_bstats_update(sch, skb);
 759         qdisc_qstats_backlog_dec(sch, skb);
 760         sch->q.qlen--;
 761
 762         return skb;
 763 }
 764
 765 static void taprio_next_tc_txq(struct net_device *dev, int tc, int *txq)
 766 {
 767         int offset = dev->tc_to_txq[tc].offset;
 768         int count = dev->tc_to_txq[tc].count;
 769
 770         (*txq)++;
 771         if (*txq == offset + count)
 772                 *txq = offset;
 773 }
 774
 775 /* Prioritize higher traffic classes, and select among TXQs belonging to the
 776  * same TC using round robin
 777  */
 778 static struct sk_buff *taprio_dequeue_tc_priority(struct Qdisc *sch,
 779                                                   struct sched_entry *entry,
 780                                                   u32 gate_mask)
 781 {
 782         struct taprio_sched *q = qdisc_priv(sch);
 783         struct net_device *dev = qdisc_dev(sch);
 784         int num_tc = netdev_get_num_tc(dev);
 785         struct sk_buff *skb;
 786         int tc;
 787
 788         for (tc = num_tc - 1; tc >= 0; tc--) {
 789                 int first_txq = q->cur_txq[tc];
 790
 791                 if (!(gate_mask & BIT(tc)))
 792                         continue;
 793
 794                 do {
 795                         skb = taprio_dequeue_from_txq(sch, q->cur_txq[tc],
 796                                                       entry, gate_mask);
 797
 798                         taprio_next_tc_txq(dev, tc, &q->cur_txq[tc]);
 799
 800                         if (q->cur_txq[tc] >= dev->num_tx_queues)
 801                                 q->cur_txq[tc] = first_txq;
 802
 803                         if (skb)
 804                                 return skb;
 805                 } while (q->cur_txq[tc] != first_txq);
 806         }
 807
 808         return NULL;
 809 }
 810
 811 /* Broken way of prioritizing smaller TXQ indices and ignoring the traffic
 812  * class other than to determine whether the gate is open or not
 813  */
 814 static struct sk_buff *taprio_dequeue_txq_priority(struct Qdisc *sch,
 815                                                    struct sched_entry *entry,
 816                                                    u32 gate_mask)
 817 {
 818         struct net_device *dev = qdisc_dev(sch);
 819         struct sk_buff *skb;
 820         int i;
 821
 822         for (i = 0; i < dev->num_tx_queues; i++) {
 823                 skb = taprio_dequeue_from_txq(sch, i, entry, gate_mask);
 824                 if (skb)
 825                         return skb;
 826         }
 827
 828         return NULL;
 829 }
 830
 831 /* Will not be called in the full offload case, since the TX queues are
 832  * attached to the Qdisc created using qdisc_create_dflt()
 833  */
 834 static struct sk_buff *taprio_dequeue(struct Qdisc *sch)
 835 {
 836         struct taprio_sched *q = qdisc_priv(sch);
 837         struct sk_buff *skb = NULL;
 838         struct sched_entry *entry;
 839         u32 gate_mask;
 840
 841         rcu_read_lock();
 842         entry = rcu_dereference(q->current_entry);
 843         /* if there's no entry, it means that the schedule didn't
 844          * start yet, so force all gates to be open, this is in
 845          * accordance to IEEE 802.1Qbv-2015 Section 8.6.9.4.5
 846          * "AdminGateStates"
 847          */
 848         gate_mask = entry ? entry->gate_mask : TAPRIO_ALL_GATES_OPEN;
 849         if (!gate_mask)
 850                 goto done;
 851
 852         if (static_branch_unlikely(&taprio_have_broken_mqprio) &&
 853             !static_branch_likely(&taprio_have_working_mqprio)) {
 854                 /* Single NIC kind which is broken */
 855                 skb = taprio_dequeue_txq_priority(sch, entry, gate_mask);
 856         } else if (static_branch_likely(&taprio_have_working_mqprio) &&
 857                    !static_branch_unlikely(&taprio_have_broken_mqprio)) {
 858                 /* Single NIC kind which prioritizes properly */
 859                 skb = taprio_dequeue_tc_priority(sch, entry, gate_mask);
 860         } else {
 861                 /* Mixed NIC kinds present in system, need dynamic testing */
 862                 if (q->broken_mqprio)
 863                         skb = taprio_dequeue_txq_priority(sch, entry, gate_mask);
 864                 else
 865                         skb = taprio_dequeue_tc_priority(sch, entry, gate_mask);
 866         }
 867
 868 done:
 869         rcu_read_unlock();
 870
 871         return skb;
 872 }
 873
 874 static bool should_restart_cycle(const struct sched_gate_list *oper,
 875                                  const struct sched_entry *entry)
 876 {
 877         if (list_is_last(&entry->list, &oper->entries))
 878                 return true;
 879
 880         if (ktime_compare(entry->end_time, oper->cycle_end_time) == 0)
 881                 return true;
 882
 883         return false;
 884 }
 885
 886 static bool should_change_schedules(const struct sched_gate_list *admin,
 887                                     const struct sched_gate_list *oper,
 888                                     ktime_t end_time)
 889 {
 890         ktime_t next_base_time, extension_time;
 891
 892         if (!admin)
 893                 return false;
 894
 895         next_base_time = sched_base_time(admin);
 896
 897         /* This is the simple case, the end_time would fall after
 898          * the next schedule base_time.
 899          */
 900         if (ktime_compare(next_base_time, end_time) <= 0)
 901                 return true;
 902
 903         /* This is the cycle_time_extension case, if the end_time
 904          * plus the amount that can be extended would fall after the
 905          * next schedule base_time, we can extend the current schedule
 906          * for that amount.
 907          */
 908         extension_time = ktime_add_ns(end_time, oper->cycle_time_extension);
 909
 910         /* FIXME: the IEEE 802.1Q-2018 Specification isn't clear about
 911          * how precisely the extension should be made. So after
 912          * conformance testing, this logic may change.
 913          */
 914         if (ktime_compare(next_base_time, extension_time) <= 0)
 915                 return true;
 916
 917         return false;
 918 }
 919
 920 static enum hrtimer_restart advance_sched(struct hrtimer *timer)
 921 {
 922         struct taprio_sched *q = container_of(timer, struct taprio_sched,
 923                                               advance_timer);
 924         struct net_device *dev = qdisc_dev(q->root);
 925         struct sched_gate_list *oper, *admin;
 926         int num_tc = netdev_get_num_tc(dev);
 927         struct sched_entry *entry, *next;
 928         struct Qdisc *sch = q->root;
 929         ktime_t end_time;
 930         int tc;
 931
 932         spin_lock(&q->current_entry_lock);
 933         entry = rcu_dereference_protected(q->current_entry,
 934                                           lockdep_is_held(&q->current_entry_lock));
 935         oper = rcu_dereference_protected(q->oper_sched,
 936                                          lockdep_is_held(&q->current_entry_lock));
 937         admin = rcu_dereference_protected(q->admin_sched,
 938                                           lockdep_is_held(&q->current_entry_lock));
 939
 940         if (!oper)
 941                 switch_schedules(q, &admin, &oper);
 942
 943         /* This can happen in two cases: 1. this is the very first run
 944          * of this function (i.e. we weren't running any schedule
 945          * previously); 2. The previous schedule just ended. The first
 946          * entry of all schedules are pre-calculated during the
 947          * schedule initialization.
 948          */
 949         if (unlikely(!entry || entry->end_time == oper->base_time)) {
 950                 next = list_first_entry(&oper->entries, struct sched_entry,
 951                                         list);
 952                 end_time = next->end_time;
 953                 goto first_run;
 954         }
 955
 956         if (should_restart_cycle(oper, entry)) {
 957                 next = list_first_entry(&oper->entries, struct sched_entry,
 958                                         list);
 959                 oper->cycle_end_time = ktime_add_ns(oper->cycle_end_time,
 960                                                     oper->cycle_time);
 961         } else {
 962                 next = list_next_entry(entry, list);
 963         }
 964
 965         end_time = ktime_add_ns(entry->end_time, next->interval);
 966         end_time = min_t(ktime_t, end_time, oper->cycle_end_time);
 967
 968         for (tc = 0; tc < num_tc; tc++) {
 969                 if (next->gate_duration[tc] == oper->cycle_time)
 970                         next->gate_close_time[tc] = KTIME_MAX;
 971                 else
 972                         next->gate_close_time[tc] = ktime_add_ns(entry->end_time,
 973                                                                  next->gate_duration[tc]);
 974         }
 975
 976         if (should_change_schedules(admin, oper, end_time)) {
 977                 /* Set things so the next time this runs, the new
 978                  * schedule runs.
 979                  */
 980                 end_time = sched_base_time(admin);
 981                 switch_schedules(q, &admin, &oper);
 982         }
 983
 984         next->end_time = end_time;
 985         taprio_set_budgets(q, oper, next);
 986
 987 first_run:
 988         rcu_assign_pointer(q->current_entry, next);
 989         spin_unlock(&q->current_entry_lock);
 990
 991         hrtimer_set_expires(&q->advance_timer, end_time);
 992
 993         rcu_read_lock();
 994         __netif_schedule(sch);
 995         rcu_read_unlock();
 996
 997         return HRTIMER_RESTART;
 998 }
 999
1000 static const struct nla_policy entry_policy[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = {
1001         [TCA_TAPRIO_SCHED_ENTRY_INDEX]     = { .type = NLA_U32 },
1002         [TCA_TAPRIO_SCHED_ENTRY_CMD]       = { .type = NLA_U8 },
1003         [TCA_TAPRIO_SCHED_ENTRY_GATE_MASK] = { .type = NLA_U32 },
1004         [TCA_TAPRIO_SCHED_ENTRY_INTERVAL]  = { .type = NLA_U32 },
1005 };
1006
1007 static const struct nla_policy taprio_tc_policy[TCA_TAPRIO_TC_ENTRY_MAX + 1] = {
1008         [TCA_TAPRIO_TC_ENTRY_INDEX]        = { .type = NLA_U32 },
1009         [TCA_TAPRIO_TC_ENTRY_MAX_SDU]      = { .type = NLA_U32 },
1010         [TCA_TAPRIO_TC_ENTRY_FP]           = NLA_POLICY_RANGE(NLA_U32,
1011                                                               TC_FP_EXPRESS,
1012                                                               TC_FP_PREEMPTIBLE),
1013 };
1014
1015 static const struct nla_policy taprio_policy[TCA_TAPRIO_ATTR_MAX + 1] = {
1016         [TCA_TAPRIO_ATTR_PRIOMAP]              = {
1017                 .len = sizeof(struct tc_mqprio_qopt)
1018         },
1019         [TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST]           = { .type = NLA_NESTED },
1020         [TCA_TAPRIO_ATTR_SCHED_BASE_TIME]            = { .type = NLA_S64 },
1021         [TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]         = { .type = NLA_NESTED },
1022         [TCA_TAPRIO_ATTR_SCHED_CLOCKID]              = { .type = NLA_S32 },
1023         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]           = { .type = NLA_S64 },
1024         [TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION] = { .type = NLA_S64 },
1025         [TCA_TAPRIO_ATTR_FLAGS]                      = { .type = NLA_U32 },
1026         [TCA_TAPRIO_ATTR_TXTIME_DELAY]               = { .type = NLA_U32 },
1027         [TCA_TAPRIO_ATTR_TC_ENTRY]                   = { .type = NLA_NESTED },
1028 };
1029
1030 static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
1031                             struct sched_entry *entry,
1032                             struct netlink_ext_ack *extack)
1033 {
1034         int min_duration = length_to_duration(q, ETH_ZLEN);
1035         u32 interval = 0;
1036
1037         if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
1038                 entry->command = nla_get_u8(
1039                         tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
1040
1041         if (tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK])
1042                 entry->gate_mask = nla_get_u32(
1043                         tb[TCA_TAPRIO_SCHED_ENTRY_GATE_MASK]);
1044
1045         if (tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL])
1046                 interval = nla_get_u32(
1047                         tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
1048
1049         /* The interval should allow at least the minimum ethernet
1050          * frame to go out.
1051          */
1052         if (interval < min_duration) {
1053                 NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
1054                 return -EINVAL;
1055         }
1056
1057         entry->interval = interval;
1058
1059         return 0;
1060 }
1061
1062 static int parse_sched_entry(struct taprio_sched *q, struct nlattr *n,
1063                              struct sched_entry *entry, int index,
1064                              struct netlink_ext_ack *extack)
1065 {
1066         struct nlattr *tb[TCA_TAPRIO_SCHED_ENTRY_MAX + 1] = { };
1067         int err;
1068
1069         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_SCHED_ENTRY_MAX, n,
1070                                           entry_policy, NULL);
1071         if (err < 0) {
1072                 NL_SET_ERR_MSG(extack, "Could not parse nested entry");
1073                 return -EINVAL;
1074         }
1075
1076         entry->index = index;
1077
1078         return fill_sched_entry(q, tb, entry, extack);
1079 }
1080
1081 static int parse_sched_list(struct taprio_sched *q, struct nlattr *list,
1082                             struct sched_gate_list *sched,
1083                             struct netlink_ext_ack *extack)
1084 {
1085         struct nlattr *n;
1086         int err, rem;
1087         int i = 0;
1088
1089         if (!list)
1090                 return -EINVAL;
1091
1092         nla_for_each_nested(n, list, rem) {
1093                 struct sched_entry *entry;
1094
1095                 if (nla_type(n) != TCA_TAPRIO_SCHED_ENTRY) {
1096                         NL_SET_ERR_MSG(extack, "Attribute is not of type 'entry'");
1097                         continue;
1098                 }
1099
1100                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1101                 if (!entry) {
1102                         NL_SET_ERR_MSG(extack, "Not enough memory for entry");
1103                         return -ENOMEM;
1104                 }
1105
1106                 err = parse_sched_entry(q, n, entry, i, extack);
1107                 if (err < 0) {
1108                         kfree(entry);
1109                         return err;
1110                 }
1111
1112                 list_add_tail(&entry->list, &sched->entries);
1113                 i++;
1114         }
1115
1116         sched->num_entries = i;
1117
1118         return i;
1119 }
1120
1121 static int parse_taprio_schedule(struct taprio_sched *q, struct nlattr **tb,
1122                                  struct sched_gate_list *new,
1123                                  struct netlink_ext_ack *extack)
1124 {
1125         int err = 0;
1126
1127         if (tb[TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY]) {
1128                 NL_SET_ERR_MSG(extack, "Adding a single entry is not supported");
1129                 return -ENOTSUPP;
1130         }
1131
1132         if (tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME])
1133                 new->base_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_BASE_TIME]);
1134
1135         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION])
1136                 new->cycle_time_extension = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION]);
1137
1138         if (tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME])
1139                 new->cycle_time = nla_get_s64(tb[TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME]);
1140
1141         if (tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST])
1142                 err = parse_sched_list(q, tb[TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST],
1143                                        new, extack);
1144         if (err < 0)
1145                 return err;
1146
1147         if (!new->cycle_time) {
1148                 struct sched_entry *entry;
1149                 ktime_t cycle = 0;
1150
1151                 list_for_each_entry(entry, &new->entries, list)
1152                         cycle = ktime_add_ns(cycle, entry->interval);
1153
1154                 if (!cycle) {
1155                         NL_SET_ERR_MSG(extack, "'cycle_time' can never be 0");
1156                         return -EINVAL;
1157                 }
1158
1159                 new->cycle_time = cycle;
1160         }
1161
1162         taprio_calculate_gate_durations(q, new);
1163
1164         return 0;
1165 }
1166
1167 static int taprio_parse_mqprio_opt(struct net_device *dev,
1168                                    struct tc_mqprio_qopt *qopt,
1169                                    struct netlink_ext_ack *extack,
1170                                    u32 taprio_flags)
1171 {
1172         bool allow_overlapping_txqs = TXTIME_ASSIST_IS_ENABLED(taprio_flags);
1173
1174         if (!qopt && !dev->num_tc) {
1175                 NL_SET_ERR_MSG(extack, "'mqprio' configuration is necessary");
1176                 return -EINVAL;
1177         }
1178
1179         /* If num_tc is already set, it means that the user already
1180          * configured the mqprio part
1181          */
1182         if (dev->num_tc)
1183                 return 0;
1184
1185         /* taprio imposes that traffic classes map 1:n to tx queues */
1186         if (qopt->num_tc > dev->num_tx_queues) {
1187                 NL_SET_ERR_MSG(extack, "Number of traffic classes is greater than number of HW queues");
1188                 return -EINVAL;
1189         }
1190
1191         /* For some reason, in txtime-assist mode, we allow TXQ ranges for
1192          * different TCs to overlap, and just validate the TXQ ranges.
1193          */
1194         return mqprio_validate_qopt(dev, qopt, true, allow_overlapping_txqs,
1195                                     extack);
1196 }
1197
1198 static int taprio_get_start_time(struct Qdisc *sch,
1199                                  struct sched_gate_list *sched,
1200                                  ktime_t *start)
1201 {
1202         struct taprio_sched *q = qdisc_priv(sch);
1203         ktime_t now, base, cycle;
1204         s64 n;
1205
1206         base = sched_base_time(sched);
1207         now = taprio_get_time(q);
1208
1209         if (ktime_after(base, now)) {
1210                 *start = base;
1211                 return 0;
1212         }
1213
1214         cycle = sched->cycle_time;
1215
1216         /* The qdisc is expected to have at least one sched_entry.  Moreover,
1217          * any entry must have 'interval' > 0. Thus if the cycle time is zero,
1218          * something went really wrong. In that case, we should warn about this
1219          * inconsistent state and return error.
1220          */
1221         if (WARN_ON(!cycle))
1222                 return -EFAULT;
1223
1224         /* Schedule the start time for the beginning of the next
1225          * cycle.
1226          */
1227         n = div64_s64(ktime_sub_ns(now, base), cycle);
1228         *start = ktime_add_ns(base, (n + 1) * cycle);
1229         return 0;
1230 }
1231
1232 static void setup_first_end_time(struct taprio_sched *q,
1233                                  struct sched_gate_list *sched, ktime_t base)
1234 {
1235         struct net_device *dev = qdisc_dev(q->root);
1236         int num_tc = netdev_get_num_tc(dev);
1237         struct sched_entry *first;
1238         ktime_t cycle;
1239         int tc;
1240
1241         first = list_first_entry(&sched->entries,
1242                                  struct sched_entry, list);
1243
1244         cycle = sched->cycle_time;
1245
1246         /* FIXME: find a better place to do this */
1247         sched->cycle_end_time = ktime_add_ns(base, cycle);
1248
1249         first->end_time = ktime_add_ns(base, first->interval);
1250         taprio_set_budgets(q, sched, first);
1251
1252         for (tc = 0; tc < num_tc; tc++) {
1253                 if (first->gate_duration[tc] == sched->cycle_time)
1254                         first->gate_close_time[tc] = KTIME_MAX;
1255                 else
1256                         first->gate_close_time[tc] = ktime_add_ns(base, first->gate_duration[tc]);
1257         }
1258
1259         rcu_assign_pointer(q->current_entry, NULL);
1260 }
1261
1262 static void taprio_start_sched(struct Qdisc *sch,
1263                                ktime_t start, struct sched_gate_list *new)
1264 {
1265         struct taprio_sched *q = qdisc_priv(sch);
1266         ktime_t expires;
1267
1268         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1269                 return;
1270
1271         expires = hrtimer_get_expires(&q->advance_timer);
1272         if (expires == 0)
1273                 expires = KTIME_MAX;
1274
1275         /* If the new schedule starts before the next expiration, we
1276          * reprogram it to the earliest one, so we change the admin
1277          * schedule to the operational one at the right time.
1278          */
1279         start = min_t(ktime_t, start, expires);
1280
1281         hrtimer_start(&q->advance_timer, start, HRTIMER_MODE_ABS);
1282 }
1283
1284 static void taprio_set_picos_per_byte(struct net_device *dev,
1285                                       struct taprio_sched *q)
1286 {
1287         struct ethtool_link_ksettings ecmd;
1288         int speed = SPEED_10;
1289         int picos_per_byte;
1290         int err;
1291
1292         err = __ethtool_get_link_ksettings(dev, &ecmd);
1293         if (err < 0)
1294                 goto skip;
1295
1296         if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
1297                 speed = ecmd.base.speed;
1298
1299 skip:
1300         picos_per_byte = (USEC_PER_SEC * 8) / speed;
1301
1302         atomic64_set(&q->picos_per_byte, picos_per_byte);
1303         netdev_dbg(dev, "taprio: set %s's picos_per_byte to: %lld, linkspeed: %d\n",
1304                    dev->name, (long long)atomic64_read(&q->picos_per_byte),
1305                    ecmd.base.speed);
1306 }
1307
1308 static int taprio_dev_notifier(struct notifier_block *nb, unsigned long event,
1309                                void *ptr)
1310 {
1311         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1312         struct sched_gate_list *oper, *admin;
1313         struct qdisc_size_table *stab;
1314         struct taprio_sched *q;
1315
1316         ASSERT_RTNL();
1317
1318         if (event != NETDEV_UP && event != NETDEV_CHANGE)
1319                 return NOTIFY_DONE;
1320
1321         list_for_each_entry(q, &taprio_list, taprio_list) {
1322                 if (dev != qdisc_dev(q->root))
1323                         continue;
1324
1325                 taprio_set_picos_per_byte(dev, q);
1326
1327                 stab = rtnl_dereference(q->root->stab);
1328
1329                 oper = rtnl_dereference(q->oper_sched);
1330                 if (oper)
1331                         taprio_update_queue_max_sdu(q, oper, stab);
1332
1333                 admin = rtnl_dereference(q->admin_sched);
1334                 if (admin)
1335                         taprio_update_queue_max_sdu(q, admin, stab);
1336
1337                 break;
1338         }
1339
1340         return NOTIFY_DONE;
1341 }
1342
1343 static void setup_txtime(struct taprio_sched *q,
1344                          struct sched_gate_list *sched, ktime_t base)
1345 {
1346         struct sched_entry *entry;
1347         u32 interval = 0;
1348
1349         list_for_each_entry(entry, &sched->entries, list) {
1350                 entry->next_txtime = ktime_add_ns(base, interval);
1351                 interval += entry->interval;
1352         }
1353 }
1354
1355 static struct tc_taprio_qopt_offload *taprio_offload_alloc(int num_entries)
1356 {
1357         struct __tc_taprio_qopt_offload *__offload;
1358
1359         __offload = kzalloc(struct_size(__offload, offload.entries, num_entries),
1360                             GFP_KERNEL);
1361         if (!__offload)
1362                 return NULL;
1363
1364         refcount_set(&__offload->users, 1);
1365
1366         return &__offload->offload;
1367 }
1368
1369 struct tc_taprio_qopt_offload *taprio_offload_get(struct tc_taprio_qopt_offload
1370                                                   *offload)
1371 {
1372         struct __tc_taprio_qopt_offload *__offload;
1373
1374         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1375                                  offload);
1376
1377         refcount_inc(&__offload->users);
1378
1379         return offload;
1380 }
1381 EXPORT_SYMBOL_GPL(taprio_offload_get);
1382
1383 void taprio_offload_free(struct tc_taprio_qopt_offload *offload)
1384 {
1385         struct __tc_taprio_qopt_offload *__offload;
1386
1387         __offload = container_of(offload, struct __tc_taprio_qopt_offload,
1388                                  offload);
1389
1390         if (!refcount_dec_and_test(&__offload->users))
1391                 return;
1392
1393         kfree(__offload);
1394 }
1395 EXPORT_SYMBOL_GPL(taprio_offload_free);
1396
1397 /* The function will only serve to keep the pointers to the "oper" and "admin"
1398  * schedules valid in relation to their base times, so when calling dump() the
1399  * users looks at the right schedules.
1400  * When using full offload, the admin configuration is promoted to oper at the
1401  * base_time in the PHC time domain.  But because the system time is not
1402  * necessarily in sync with that, we can't just trigger a hrtimer to call
1403  * switch_schedules at the right hardware time.
1404  * At the moment we call this by hand right away from taprio, but in the future
1405  * it will be useful to create a mechanism for drivers to notify taprio of the
1406  * offload state (PENDING, ACTIVE, INACTIVE) so it can be visible in dump().
1407  * This is left as TODO.
1408  */
1409 static void taprio_offload_config_changed(struct taprio_sched *q)
1410 {
1411         struct sched_gate_list *oper, *admin;
1412
1413         oper = rtnl_dereference(q->oper_sched);
1414         admin = rtnl_dereference(q->admin_sched);
1415
1416         switch_schedules(q, &admin, &oper);
1417 }
1418
1419 static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
1420 {
1421         u32 i, queue_mask = 0;
1422
1423         for (i = 0; i < dev->num_tc; i++) {
1424                 u32 offset, count;
1425
1426                 if (!(tc_mask & BIT(i)))
1427                         continue;
1428
1429                 offset = dev->tc_to_txq[i].offset;
1430                 count = dev->tc_to_txq[i].count;
1431
1432                 queue_mask |= GENMASK(offset + count - 1, offset);
1433         }
1434
1435         return queue_mask;
1436 }
1437
1438 static void taprio_sched_to_offload(struct net_device *dev,
1439                                     struct sched_gate_list *sched,
1440                                     struct tc_taprio_qopt_offload *offload,
1441                                     const struct tc_taprio_caps *caps)
1442 {
1443         struct sched_entry *entry;
1444         int i = 0;
1445
1446         offload->base_time = sched->base_time;
1447         offload->cycle_time = sched->cycle_time;
1448         offload->cycle_time_extension = sched->cycle_time_extension;
1449
1450         list_for_each_entry(entry, &sched->entries, list) {
1451                 struct tc_taprio_sched_entry *e = &offload->entries[i];
1452
1453                 e->command = entry->command;
1454                 e->interval = entry->interval;
1455                 if (caps->gate_mask_per_txq)
1456                         e->gate_mask = tc_map_to_queue_mask(dev,
1457                                                             entry->gate_mask);
1458                 else
1459                         e->gate_mask = entry->gate_mask;
1460
1461                 i++;
1462         }
1463
1464         offload->num_entries = i;
1465 }
1466
1467 static void taprio_detect_broken_mqprio(struct taprio_sched *q)
1468 {
1469         struct net_device *dev = qdisc_dev(q->root);
1470         struct tc_taprio_caps caps;
1471
1472         qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1473                                  &caps, sizeof(caps));
1474
1475         q->broken_mqprio = caps.broken_mqprio;
1476         if (q->broken_mqprio)
1477                 static_branch_inc(&taprio_have_broken_mqprio);
1478         else
1479                 static_branch_inc(&taprio_have_working_mqprio);
1480
1481         q->detected_mqprio = true;
1482 }
1483
1484 static void taprio_cleanup_broken_mqprio(struct taprio_sched *q)
1485 {
1486         if (!q->detected_mqprio)
1487                 return;
1488
1489         if (q->broken_mqprio)
1490                 static_branch_dec(&taprio_have_broken_mqprio);
1491         else
1492                 static_branch_dec(&taprio_have_working_mqprio);
1493 }
1494
1495 static int taprio_enable_offload(struct net_device *dev,
1496                                  struct taprio_sched *q,
1497                                  struct sched_gate_list *sched,
1498                                  struct netlink_ext_ack *extack)
1499 {
1500         const struct net_device_ops *ops = dev->netdev_ops;
1501         struct tc_taprio_qopt_offload *offload;
1502         struct tc_taprio_caps caps;
1503         int tc, err = 0;
1504
1505         if (!ops->ndo_setup_tc) {
1506                 NL_SET_ERR_MSG(extack,
1507                                "Device does not support taprio offload");
1508                 return -EOPNOTSUPP;
1509         }
1510
1511         qdisc_offload_query_caps(dev, TC_SETUP_QDISC_TAPRIO,
1512                                  &caps, sizeof(caps));
1513
1514         if (!caps.supports_queue_max_sdu) {
1515                 for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
1516                         if (q->max_sdu[tc]) {
1517                                 NL_SET_ERR_MSG_MOD(extack,
1518                                                    "Device does not handle queueMaxSDU");
1519                                 return -EOPNOTSUPP;
1520                         }
1521                 }
1522         }
1523
1524         offload = taprio_offload_alloc(sched->num_entries);
1525         if (!offload) {
1526                 NL_SET_ERR_MSG(extack,
1527                                "Not enough memory for enabling offload mode");
1528                 return -ENOMEM;
1529         }
1530         offload->enable = 1;
1531         offload->extack = extack;
1532         mqprio_qopt_reconstruct(dev, &offload->mqprio.qopt);
1533         offload->mqprio.extack = extack;
1534         taprio_sched_to_offload(dev, sched, offload, &caps);
1535         mqprio_fp_to_offload(q->fp, &offload->mqprio);
1536
1537         for (tc = 0; tc < TC_MAX_QUEUE; tc++)
1538                 offload->max_sdu[tc] = q->max_sdu[tc];
1539
1540         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1541         if (err < 0) {
1542                 NL_SET_ERR_MSG_WEAK(extack,
1543                                     "Device failed to setup taprio offload");
1544                 goto done;
1545         }
1546
1547         q->offloaded = true;
1548
1549 done:
1550         /* The offload structure may linger around via a reference taken by the
1551          * device driver, so clear up the netlink extack pointer so that the
1552          * driver isn't tempted to dereference data which stopped being valid
1553          */
1554         offload->extack = NULL;
1555         offload->mqprio.extack = NULL;
1556         taprio_offload_free(offload);
1557
1558         return err;
1559 }
1560
1561 static int taprio_disable_offload(struct net_device *dev,
1562                                   struct taprio_sched *q,
1563                                   struct netlink_ext_ack *extack)
1564 {
1565         const struct net_device_ops *ops = dev->netdev_ops;
1566         struct tc_taprio_qopt_offload *offload;
1567         int err;
1568
1569         if (!q->offloaded)
1570                 return 0;
1571
1572         offload = taprio_offload_alloc(0);
1573         if (!offload) {
1574                 NL_SET_ERR_MSG(extack,
1575                                "Not enough memory to disable offload mode");
1576                 return -ENOMEM;
1577         }
1578         offload->enable = 0;
1579
1580         err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TAPRIO, offload);
1581         if (err < 0) {
1582                 NL_SET_ERR_MSG(extack,
1583                                "Device failed to disable offload");
1584                 goto out;
1585         }
1586
1587         q->offloaded = false;
1588
1589 out:
1590         taprio_offload_free(offload);
1591
1592         return err;
1593 }
1594
1595 /* If full offload is enabled, the only possible clockid is the net device's
1596  * PHC. For that reason, specifying a clockid through netlink is incorrect.
1597  * For txtime-assist, it is implicitly assumed that the device's PHC is kept
1598  * in sync with the specified clockid via a user space daemon such as phc2sys.
1599  * For both software taprio and txtime-assist, the clockid is used for the
1600  * hrtimer that advances the schedule and hence mandatory.
1601  */
1602 static int taprio_parse_clockid(struct Qdisc *sch, struct nlattr **tb,
1603                                 struct netlink_ext_ack *extack)
1604 {
1605         struct taprio_sched *q = qdisc_priv(sch);
1606         struct net_device *dev = qdisc_dev(sch);
1607         int err = -EINVAL;
1608
1609         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1610                 const struct ethtool_ops *ops = dev->ethtool_ops;
1611                 struct ethtool_ts_info info = {
1612                         .cmd = ETHTOOL_GET_TS_INFO,
1613                         .phc_index = -1,
1614                 };
1615
1616                 if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1617                         NL_SET_ERR_MSG(extack,
1618                                        "The 'clockid' cannot be specified for full offload");
1619                         goto out;
1620                 }
1621
1622                 if (ops && ops->get_ts_info)
1623                         err = ops->get_ts_info(dev, &info);
1624
1625                 if (err || info.phc_index < 0) {
1626                         NL_SET_ERR_MSG(extack,
1627                                        "Device does not have a PTP clock");
1628                         err = -ENOTSUPP;
1629                         goto out;
1630                 }
1631         } else if (tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]) {
1632                 int clockid = nla_get_s32(tb[TCA_TAPRIO_ATTR_SCHED_CLOCKID]);
1633                 enum tk_offsets tk_offset;
1634
1635                 /* We only support static clockids and we don't allow
1636                  * for it to be modified after the first init.
1637                  */
1638                 if (clockid < 0 ||
1639                     (q->clockid != -1 && q->clockid != clockid)) {
1640                         NL_SET_ERR_MSG(extack,
1641                                        "Changing the 'clockid' of a running schedule is not supported");
1642                         err = -ENOTSUPP;
1643                         goto out;
1644                 }
1645
1646                 switch (clockid) {
1647                 case CLOCK_REALTIME:
1648                         tk_offset = TK_OFFS_REAL;
1649                         break;
1650                 case CLOCK_MONOTONIC:
1651                         tk_offset = TK_OFFS_MAX;
1652                         break;
1653                 case CLOCK_BOOTTIME:
1654                         tk_offset = TK_OFFS_BOOT;
1655                         break;
1656                 case CLOCK_TAI:
1657                         tk_offset = TK_OFFS_TAI;
1658                         break;
1659                 default:
1660                         NL_SET_ERR_MSG(extack, "Invalid 'clockid'");
1661                         err = -EINVAL;
1662                         goto out;
1663                 }
1664                 /* This pairs with READ_ONCE() in taprio_mono_to_any */
1665                 WRITE_ONCE(q->tk_offset, tk_offset);
1666
1667                 q->clockid = clockid;
1668         } else {
1669                 NL_SET_ERR_MSG(extack, "Specifying a 'clockid' is mandatory");
1670                 goto out;
1671         }
1672
1673         /* Everything went ok, return success. */
1674         err = 0;
1675
1676 out:
1677         return err;
1678 }
1679
1680 static int taprio_parse_tc_entry(struct Qdisc *sch,
1681                                  struct nlattr *opt,
1682                                  u32 max_sdu[TC_QOPT_MAX_QUEUE],
1683                                  u32 fp[TC_QOPT_MAX_QUEUE],
1684                                  unsigned long *seen_tcs,
1685                                  struct netlink_ext_ack *extack)
1686 {
1687         struct nlattr *tb[TCA_TAPRIO_TC_ENTRY_MAX + 1] = { };
1688         struct net_device *dev = qdisc_dev(sch);
1689         int err, tc;
1690         u32 val;
1691
1692         err = nla_parse_nested(tb, TCA_TAPRIO_TC_ENTRY_MAX, opt,
1693                                taprio_tc_policy, extack);
1694         if (err < 0)
1695                 return err;
1696
1697         if (!tb[TCA_TAPRIO_TC_ENTRY_INDEX]) {
1698                 NL_SET_ERR_MSG_MOD(extack, "TC entry index missing");
1699                 return -EINVAL;
1700         }
1701
1702         tc = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_INDEX]);
1703         if (tc >= TC_QOPT_MAX_QUEUE) {
1704                 NL_SET_ERR_MSG_MOD(extack, "TC entry index out of range");
1705                 return -ERANGE;
1706         }
1707
1708         if (*seen_tcs & BIT(tc)) {
1709                 NL_SET_ERR_MSG_MOD(extack, "Duplicate TC entry");
1710                 return -EINVAL;
1711         }
1712
1713         *seen_tcs |= BIT(tc);
1714
1715         if (tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]) {
1716                 val = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_MAX_SDU]);
1717                 if (val > dev->max_mtu) {
1718                         NL_SET_ERR_MSG_MOD(extack, "TC max SDU exceeds device max MTU");
1719                         return -ERANGE;
1720                 }
1721
1722                 max_sdu[tc] = val;
1723         }
1724
1725         if (tb[TCA_TAPRIO_TC_ENTRY_FP])
1726                 fp[tc] = nla_get_u32(tb[TCA_TAPRIO_TC_ENTRY_FP]);
1727
1728         return 0;
1729 }
1730
1731 static int taprio_parse_tc_entries(struct Qdisc *sch,
1732                                    struct nlattr *opt,
1733                                    struct netlink_ext_ack *extack)
1734 {
1735         struct taprio_sched *q = qdisc_priv(sch);
1736         struct net_device *dev = qdisc_dev(sch);
1737         u32 max_sdu[TC_QOPT_MAX_QUEUE];
1738         bool have_preemption = false;
1739         unsigned long seen_tcs = 0;
1740         u32 fp[TC_QOPT_MAX_QUEUE];
1741         struct nlattr *n;
1742         int tc, rem;
1743         int err = 0;
1744
1745         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1746                 max_sdu[tc] = q->max_sdu[tc];
1747                 fp[tc] = q->fp[tc];
1748         }
1749
1750         nla_for_each_nested(n, opt, rem) {
1751                 if (nla_type(n) != TCA_TAPRIO_ATTR_TC_ENTRY)
1752                         continue;
1753
1754                 err = taprio_parse_tc_entry(sch, n, max_sdu, fp, &seen_tcs,
1755                                             extack);
1756                 if (err)
1757                         return err;
1758         }
1759
1760         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
1761                 q->max_sdu[tc] = max_sdu[tc];
1762                 q->fp[tc] = fp[tc];
1763                 if (fp[tc] != TC_FP_EXPRESS)
1764                         have_preemption = true;
1765         }
1766
1767         if (have_preemption) {
1768                 if (!FULL_OFFLOAD_IS_ENABLED(q->flags)) {
1769                         NL_SET_ERR_MSG(extack,
1770                                        "Preemption only supported with full offload");
1771                         return -EOPNOTSUPP;
1772                 }
1773
1774                 if (!ethtool_dev_mm_supported(dev)) {
1775                         NL_SET_ERR_MSG(extack,
1776                                        "Device does not support preemption");
1777                         return -EOPNOTSUPP;
1778                 }
1779         }
1780
1781         return err;
1782 }
1783
1784 static int taprio_mqprio_cmp(const struct net_device *dev,
1785                              const struct tc_mqprio_qopt *mqprio)
1786 {
1787         int i;
1788
1789         if (!mqprio || mqprio->num_tc != dev->num_tc)
1790                 return -1;
1791
1792         for (i = 0; i < mqprio->num_tc; i++)
1793                 if (dev->tc_to_txq[i].count != mqprio->count[i] ||
1794                     dev->tc_to_txq[i].offset != mqprio->offset[i])
1795                         return -1;
1796
1797         for (i = 0; i <= TC_BITMASK; i++)
1798                 if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])
1799                         return -1;
1800
1801         return 0;
1802 }
1803
1804 /* The semantics of the 'flags' argument in relation to 'change()'
1805  * requests, are interpreted following two rules (which are applied in
1806  * this order): (1) an omitted 'flags' argument is interpreted as
1807  * zero; (2) the 'flags' of a "running" taprio instance cannot be
1808  * changed.
1809  */
1810 static int taprio_new_flags(const struct nlattr *attr, u32 old,
1811                             struct netlink_ext_ack *extack)
1812 {
1813         u32 new = 0;
1814
1815         if (attr)
1816                 new = nla_get_u32(attr);
1817
1818         if (old != TAPRIO_FLAGS_INVALID && old != new) {
1819                 NL_SET_ERR_MSG_MOD(extack, "Changing 'flags' of a running schedule is not supported");
1820                 return -EOPNOTSUPP;
1821         }
1822
1823         if (!taprio_flags_valid(new)) {
1824                 NL_SET_ERR_MSG_MOD(extack, "Specified 'flags' are not valid");
1825                 return -EINVAL;
1826         }
1827
1828         return new;
1829 }
1830
1831 static int taprio_change(struct Qdisc *sch, struct nlattr *opt,
1832                          struct netlink_ext_ack *extack)
1833 {
1834         struct qdisc_size_table *stab = rtnl_dereference(sch->stab);
1835         struct nlattr *tb[TCA_TAPRIO_ATTR_MAX + 1] = { };
1836         struct sched_gate_list *oper, *admin, *new_admin;
1837         struct taprio_sched *q = qdisc_priv(sch);
1838         struct net_device *dev = qdisc_dev(sch);
1839         struct tc_mqprio_qopt *mqprio = NULL;
1840         unsigned long flags;
1841         ktime_t start;
1842         int i, err;
1843
1844         err = nla_parse_nested_deprecated(tb, TCA_TAPRIO_ATTR_MAX, opt,
1845                                           taprio_policy, extack);
1846         if (err < 0)
1847                 return err;
1848
1849         if (tb[TCA_TAPRIO_ATTR_PRIOMAP])
1850                 mqprio = nla_data(tb[TCA_TAPRIO_ATTR_PRIOMAP]);
1851
1852         err = taprio_new_flags(tb[TCA_TAPRIO_ATTR_FLAGS],
1853                                q->flags, extack);
1854         if (err < 0)
1855                 return err;
1856
1857         q->flags = err;
1858
1859         err = taprio_parse_mqprio_opt(dev, mqprio, extack, q->flags);
1860         if (err < 0)
1861                 return err;
1862
1863         err = taprio_parse_tc_entries(sch, opt, extack);
1864         if (err)
1865                 return err;
1866
1867         new_admin = kzalloc(sizeof(*new_admin), GFP_KERNEL);
1868         if (!new_admin) {
1869                 NL_SET_ERR_MSG(extack, "Not enough memory for a new schedule");
1870                 return -ENOMEM;
1871         }
1872         INIT_LIST_HEAD(&new_admin->entries);
1873
1874         oper = rtnl_dereference(q->oper_sched);
1875         admin = rtnl_dereference(q->admin_sched);
1876
1877         /* no changes - no new mqprio settings */
1878         if (!taprio_mqprio_cmp(dev, mqprio))
1879                 mqprio = NULL;
1880
1881         if (mqprio && (oper || admin)) {
1882                 NL_SET_ERR_MSG(extack, "Changing the traffic mapping of a running schedule is not supported");
1883                 err = -ENOTSUPP;
1884                 goto free_sched;
1885         }
1886
1887         if (mqprio) {
1888                 err = netdev_set_num_tc(dev, mqprio->num_tc);
1889                 if (err)
1890                         goto free_sched;
1891                 for (i = 0; i < mqprio->num_tc; i++) {
1892                         netdev_set_tc_queue(dev, i,
1893                                             mqprio->count[i],
1894                                             mqprio->offset[i]);
1895                         q->cur_txq[i] = mqprio->offset[i];
1896                 }
1897
1898                 /* Always use supplied priority mappings */
1899                 for (i = 0; i <= TC_BITMASK; i++)
1900                         netdev_set_prio_tc_map(dev, i,
1901                                                mqprio->prio_tc_map[i]);
1902         }
1903
1904         err = parse_taprio_schedule(q, tb, new_admin, extack);
1905         if (err < 0)
1906                 goto free_sched;
1907
1908         if (new_admin->num_entries == 0) {
1909                 NL_SET_ERR_MSG(extack, "There should be at least one entry in the schedule");
1910                 err = -EINVAL;
1911                 goto free_sched;
1912         }
1913
1914         err = taprio_parse_clockid(sch, tb, extack);
1915         if (err < 0)
1916                 goto free_sched;
1917
1918         taprio_set_picos_per_byte(dev, q);
1919         taprio_update_queue_max_sdu(q, new_admin, stab);
1920
1921         if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1922                 err = taprio_enable_offload(dev, q, new_admin, extack);
1923         else
1924                 err = taprio_disable_offload(dev, q, extack);
1925         if (err)
1926                 goto free_sched;
1927
1928         /* Protects against enqueue()/dequeue() */
1929         spin_lock_bh(qdisc_lock(sch));
1930
1931         if (tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]) {
1932                 if (!TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1933                         NL_SET_ERR_MSG_MOD(extack, "txtime-delay can only be set when txtime-assist mode is enabled");
1934                         err = -EINVAL;
1935                         goto unlock;
1936                 }
1937
1938                 q->txtime_delay = nla_get_u32(tb[TCA_TAPRIO_ATTR_TXTIME_DELAY]);
1939         }
1940
1941         if (!TXTIME_ASSIST_IS_ENABLED(q->flags) &&
1942             !FULL_OFFLOAD_IS_ENABLED(q->flags) &&
1943             !hrtimer_active(&q->advance_timer)) {
1944                 hrtimer_init(&q->advance_timer, q->clockid, HRTIMER_MODE_ABS);
1945                 q->advance_timer.function = advance_sched;
1946         }
1947
1948         err = taprio_get_start_time(sch, new_admin, &start);
1949         if (err < 0) {
1950                 NL_SET_ERR_MSG(extack, "Internal error: failed get start time");
1951                 goto unlock;
1952         }
1953
1954         setup_txtime(q, new_admin, start);
1955
1956         if (TXTIME_ASSIST_IS_ENABLED(q->flags)) {
1957                 if (!oper) {
1958                         rcu_assign_pointer(q->oper_sched, new_admin);
1959                         err = 0;
1960                         new_admin = NULL;
1961                         goto unlock;
1962                 }
1963
1964                 rcu_assign_pointer(q->admin_sched, new_admin);
1965                 if (admin)
1966                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1967         } else {
1968                 setup_first_end_time(q, new_admin, start);
1969
1970                 /* Protects against advance_sched() */
1971                 spin_lock_irqsave(&q->current_entry_lock, flags);
1972
1973                 taprio_start_sched(sch, start, new_admin);
1974
1975                 rcu_assign_pointer(q->admin_sched, new_admin);
1976                 if (admin)
1977                         call_rcu(&admin->rcu, taprio_free_sched_cb);
1978
1979                 spin_unlock_irqrestore(&q->current_entry_lock, flags);
1980
1981                 if (FULL_OFFLOAD_IS_ENABLED(q->flags))
1982                         taprio_offload_config_changed(q);
1983         }
1984
1985         new_admin = NULL;
1986         err = 0;
1987
1988         if (!stab)
1989                 NL_SET_ERR_MSG_MOD(extack,
1990                                    "Size table not specified, frame length estimations may be inaccurate");
1991
1992 unlock:
1993         spin_unlock_bh(qdisc_lock(sch));
1994
1995 free_sched:
1996         if (new_admin)
1997                 call_rcu(&new_admin->rcu, taprio_free_sched_cb);
1998
1999         return err;
2000 }
2001
2002 static void taprio_reset(struct Qdisc *sch)
2003 {
2004         struct taprio_sched *q = qdisc_priv(sch);
2005         struct net_device *dev = qdisc_dev(sch);
2006         int i;
2007
2008         hrtimer_cancel(&q->advance_timer);
2009
2010         if (q->qdiscs) {
2011                 for (i = 0; i < dev->num_tx_queues; i++)
2012                         if (q->qdiscs[i])
2013                                 qdisc_reset(q->qdiscs[i]);
2014         }
2015 }
2016
2017 static void taprio_destroy(struct Qdisc *sch)
2018 {
2019         struct taprio_sched *q = qdisc_priv(sch);
2020         struct net_device *dev = qdisc_dev(sch);
2021         struct sched_gate_list *oper, *admin;
2022         unsigned int i;
2023
2024         list_del(&q->taprio_list);
2025
2026         /* Note that taprio_reset() might not be called if an error
2027          * happens in qdisc_create(), after taprio_init() has been called.
2028          */
2029         hrtimer_cancel(&q->advance_timer);
2030         qdisc_synchronize(sch);
2031
2032         taprio_disable_offload(dev, q, NULL);
2033
2034         if (q->qdiscs) {
2035                 for (i = 0; i < dev->num_tx_queues; i++)
2036                         qdisc_put(q->qdiscs[i]);
2037
2038                 kfree(q->qdiscs);
2039         }
2040         q->qdiscs = NULL;
2041
2042         netdev_reset_tc(dev);
2043
2044         oper = rtnl_dereference(q->oper_sched);
2045         admin = rtnl_dereference(q->admin_sched);
2046
2047         if (oper)
2048                 call_rcu(&oper->rcu, taprio_free_sched_cb);
2049
2050         if (admin)
2051                 call_rcu(&admin->rcu, taprio_free_sched_cb);
2052
2053         taprio_cleanup_broken_mqprio(q);
2054 }
2055
2056 static int taprio_init(struct Qdisc *sch, struct nlattr *opt,
2057                        struct netlink_ext_ack *extack)
2058 {
2059         struct taprio_sched *q = qdisc_priv(sch);
2060         struct net_device *dev = qdisc_dev(sch);
2061         int i, tc;
2062
2063         spin_lock_init(&q->current_entry_lock);
2064
2065         hrtimer_init(&q->advance_timer, CLOCK_TAI, HRTIMER_MODE_ABS);
2066         q->advance_timer.function = advance_sched;
2067
2068         q->root = sch;
2069
2070         /* We only support static clockids. Use an invalid value as default
2071          * and get the valid one on taprio_change().
2072          */
2073         q->clockid = -1;
2074         q->flags = TAPRIO_FLAGS_INVALID;
2075
2076         list_add(&q->taprio_list, &taprio_list);
2077
2078         if (sch->parent != TC_H_ROOT) {
2079                 NL_SET_ERR_MSG_MOD(extack, "Can only be attached as root qdisc");
2080                 return -EOPNOTSUPP;
2081         }
2082
2083         if (!netif_is_multiqueue(dev)) {
2084                 NL_SET_ERR_MSG_MOD(extack, "Multi-queue device is required");
2085                 return -EOPNOTSUPP;
2086         }
2087
2088         /* pre-allocate qdisc, attachment can't fail */
2089         q->qdiscs = kcalloc(dev->num_tx_queues,
2090                             sizeof(q->qdiscs[0]),
2091                             GFP_KERNEL);
2092
2093         if (!q->qdiscs)
2094                 return -ENOMEM;
2095
2096         if (!opt)
2097                 return -EINVAL;
2098
2099         for (i = 0; i < dev->num_tx_queues; i++) {
2100                 struct netdev_queue *dev_queue;
2101                 struct Qdisc *qdisc;
2102
2103                 dev_queue = netdev_get_tx_queue(dev, i);
2104                 qdisc = qdisc_create_dflt(dev_queue,
2105                                           &pfifo_qdisc_ops,
2106                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
2107                                                     TC_H_MIN(i + 1)),
2108                                           extack);
2109                 if (!qdisc)
2110                         return -ENOMEM;
2111
2112                 if (i < dev->real_num_tx_queues)
2113                         qdisc_hash_add(qdisc, false);
2114
2115                 q->qdiscs[i] = qdisc;
2116         }
2117
2118         for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++)
2119                 q->fp[tc] = TC_FP_EXPRESS;
2120
2121         taprio_detect_broken_mqprio(q);
2122
2123         return taprio_change(sch, opt, extack);
2124 }
2125
2126 static void taprio_attach(struct Qdisc *sch)
2127 {
2128         struct taprio_sched *q = qdisc_priv(sch);
2129         struct net_device *dev = qdisc_dev(sch);
2130         unsigned int ntx;
2131
2132         /* Attach underlying qdisc */
2133         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
2134                 struct Qdisc *qdisc = q->qdiscs[ntx];
2135                 struct Qdisc *old;
2136
2137                 if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2138                         qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2139                         old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
2140                 } else {
2141                         old = dev_graft_qdisc(qdisc->dev_queue, sch);
2142                         qdisc_refcount_inc(sch);
2143                 }
2144                 if (old)
2145                         qdisc_put(old);
2146         }
2147
2148         /* access to the child qdiscs is not needed in offload mode */
2149         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2150                 kfree(q->qdiscs);
2151                 q->qdiscs = NULL;
2152         }
2153 }
2154
2155 static struct netdev_queue *taprio_queue_get(struct Qdisc *sch,
2156                                              unsigned long cl)
2157 {
2158         struct net_device *dev = qdisc_dev(sch);
2159         unsigned long ntx = cl - 1;
2160
2161         if (ntx >= dev->num_tx_queues)
2162                 return NULL;
2163
2164         return netdev_get_tx_queue(dev, ntx);
2165 }
2166
2167 static int taprio_graft(struct Qdisc *sch, unsigned long cl,
2168                         struct Qdisc *new, struct Qdisc **old,
2169                         struct netlink_ext_ack *extack)
2170 {
2171         struct taprio_sched *q = qdisc_priv(sch);
2172         struct net_device *dev = qdisc_dev(sch);
2173         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2174
2175         if (!dev_queue)
2176                 return -EINVAL;
2177
2178         if (dev->flags & IFF_UP)
2179                 dev_deactivate(dev);
2180
2181         if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
2182                 *old = dev_graft_qdisc(dev_queue, new);
2183         } else {
2184                 *old = q->qdiscs[cl - 1];
2185                 q->qdiscs[cl - 1] = new;
2186         }
2187
2188         if (new)
2189                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
2190
2191         if (dev->flags & IFF_UP)
2192                 dev_activate(dev);
2193
2194         return 0;
2195 }
2196
2197 static int dump_entry(struct sk_buff *msg,
2198                       const struct sched_entry *entry)
2199 {
2200         struct nlattr *item;
2201
2202         item = nla_nest_start_noflag(msg, TCA_TAPRIO_SCHED_ENTRY);
2203         if (!item)
2204                 return -ENOSPC;
2205
2206         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INDEX, entry->index))
2207                 goto nla_put_failure;
2208
2209         if (nla_put_u8(msg, TCA_TAPRIO_SCHED_ENTRY_CMD, entry->command))
2210                 goto nla_put_failure;
2211
2212         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_GATE_MASK,
2213                         entry->gate_mask))
2214                 goto nla_put_failure;
2215
2216         if (nla_put_u32(msg, TCA_TAPRIO_SCHED_ENTRY_INTERVAL,
2217                         entry->interval))
2218                 goto nla_put_failure;
2219
2220         return nla_nest_end(msg, item);
2221
2222 nla_put_failure:
2223         nla_nest_cancel(msg, item);
2224         return -1;
2225 }
2226
2227 static int dump_schedule(struct sk_buff *msg,
2228                          const struct sched_gate_list *root)
2229 {
2230         struct nlattr *entry_list;
2231         struct sched_entry *entry;
2232
2233         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_BASE_TIME,
2234                         root->base_time, TCA_TAPRIO_PAD))
2235                 return -1;
2236
2237         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME,
2238                         root->cycle_time, TCA_TAPRIO_PAD))
2239                 return -1;
2240
2241         if (nla_put_s64(msg, TCA_TAPRIO_ATTR_SCHED_CYCLE_TIME_EXTENSION,
2242                         root->cycle_time_extension, TCA_TAPRIO_PAD))
2243                 return -1;
2244
2245         entry_list = nla_nest_start_noflag(msg,
2246                                            TCA_TAPRIO_ATTR_SCHED_ENTRY_LIST);
2247         if (!entry_list)
2248                 goto error_nest;
2249
2250         list_for_each_entry(entry, &root->entries, list) {
2251                 if (dump_entry(msg, entry) < 0)
2252                         goto error_nest;
2253         }
2254
2255         nla_nest_end(msg, entry_list);
2256         return 0;
2257
2258 error_nest:
2259         nla_nest_cancel(msg, entry_list);
2260         return -1;
2261 }
2262
2263 static int taprio_dump_tc_entries(struct sk_buff *skb,
2264                                   struct taprio_sched *q,
2265                                   struct sched_gate_list *sched)
2266 {
2267         struct nlattr *n;
2268         int tc;
2269
2270         for (tc = 0; tc < TC_MAX_QUEUE; tc++) {
2271                 n = nla_nest_start(skb, TCA_TAPRIO_ATTR_TC_ENTRY);
2272                 if (!n)
2273                         return -EMSGSIZE;
2274
2275                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_INDEX, tc))
2276                         goto nla_put_failure;
2277
2278                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_MAX_SDU,
2279                                 sched->max_sdu[tc]))
2280                         goto nla_put_failure;
2281
2282                 if (nla_put_u32(skb, TCA_TAPRIO_TC_ENTRY_FP, q->fp[tc]))
2283                         goto nla_put_failure;
2284
2285                 nla_nest_end(skb, n);
2286         }
2287
2288         return 0;
2289
2290 nla_put_failure:
2291         nla_nest_cancel(skb, n);
2292         return -EMSGSIZE;
2293 }
2294
2295 static int taprio_dump(struct Qdisc *sch, struct sk_buff *skb)
2296 {
2297         struct taprio_sched *q = qdisc_priv(sch);
2298         struct net_device *dev = qdisc_dev(sch);
2299         struct sched_gate_list *oper, *admin;
2300         struct tc_mqprio_qopt opt = { 0 };
2301         struct nlattr *nest, *sched_nest;
2302
2303         oper = rtnl_dereference(q->oper_sched);
2304         admin = rtnl_dereference(q->admin_sched);
2305
2306         mqprio_qopt_reconstruct(dev, &opt);
2307
2308         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2309         if (!nest)
2310                 goto start_error;
2311
2312         if (nla_put(skb, TCA_TAPRIO_ATTR_PRIOMAP, sizeof(opt), &opt))
2313                 goto options_error;
2314
2315         if (!FULL_OFFLOAD_IS_ENABLED(q->flags) &&
2316             nla_put_s32(skb, TCA_TAPRIO_ATTR_SCHED_CLOCKID, q->clockid))
2317                 goto options_error;
2318
2319         if (q->flags && nla_put_u32(skb, TCA_TAPRIO_ATTR_FLAGS, q->flags))
2320                 goto options_error;
2321
2322         if (q->txtime_delay &&
2323             nla_put_u32(skb, TCA_TAPRIO_ATTR_TXTIME_DELAY, q->txtime_delay))
2324                 goto options_error;
2325
2326         if (oper && taprio_dump_tc_entries(skb, q, oper))
2327                 goto options_error;
2328
2329         if (oper && dump_schedule(skb, oper))
2330                 goto options_error;
2331
2332         if (!admin)
2333                 goto done;
2334
2335         sched_nest = nla_nest_start_noflag(skb, TCA_TAPRIO_ATTR_ADMIN_SCHED);
2336         if (!sched_nest)
2337                 goto options_error;
2338
2339         if (dump_schedule(skb, admin))
2340                 goto admin_error;
2341
2342         nla_nest_end(skb, sched_nest);
2343
2344 done:
2345         return nla_nest_end(skb, nest);
2346
2347 admin_error:
2348         nla_nest_cancel(skb, sched_nest);
2349
2350 options_error:
2351         nla_nest_cancel(skb, nest);
2352
2353 start_error:
2354         return -ENOSPC;
2355 }
2356
2357 static struct Qdisc *taprio_leaf(struct Qdisc *sch, unsigned long cl)
2358 {
2359         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2360
2361         if (!dev_queue)
2362                 return NULL;
2363
2364         return rtnl_dereference(dev_queue->qdisc_sleeping);
2365 }
2366
2367 static unsigned long taprio_find(struct Qdisc *sch, u32 classid)
2368 {
2369         unsigned int ntx = TC_H_MIN(classid);
2370
2371         if (!taprio_queue_get(sch, ntx))
2372                 return 0;
2373         return ntx;
2374 }
2375
2376 static int taprio_dump_class(struct Qdisc *sch, unsigned long cl,
2377                              struct sk_buff *skb, struct tcmsg *tcm)
2378 {
2379         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2380
2381         tcm->tcm_parent = TC_H_ROOT;
2382         tcm->tcm_handle |= TC_H_MIN(cl);
2383         tcm->tcm_info = rtnl_dereference(dev_queue->qdisc_sleeping)->handle;
2384
2385         return 0;
2386 }
2387
2388 static int taprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
2389                                    struct gnet_dump *d)
2390         __releases(d->lock)
2391         __acquires(d->lock)
2392 {
2393         struct netdev_queue *dev_queue = taprio_queue_get(sch, cl);
2394
2395         sch = rtnl_dereference(dev_queue->qdisc_sleeping);
2396         if (gnet_stats_copy_basic(d, NULL, &sch->bstats, true) < 0 ||
2397             qdisc_qstats_copy(d, sch) < 0)
2398                 return -1;
2399         return 0;
2400 }
2401
2402 static void taprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2403 {
2404         struct net_device *dev = qdisc_dev(sch);
2405         unsigned long ntx;
2406
2407         if (arg->stop)
2408                 return;
2409
2410         arg->count = arg->skip;
2411         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
2412                 if (!tc_qdisc_stats_dump(sch, ntx + 1, arg))
2413                         break;
2414         }
2415 }
2416
2417 static struct netdev_queue *taprio_select_queue(struct Qdisc *sch,
2418                                                 struct tcmsg *tcm)
2419 {
2420         return taprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
2421 }
2422
2423 static const struct Qdisc_class_ops taprio_class_ops = {
2424         .graft          = taprio_graft,
2425         .leaf           = taprio_leaf,
2426         .find           = taprio_find,
2427         .walk           = taprio_walk,
2428         .dump           = taprio_dump_class,
2429         .dump_stats     = taprio_dump_class_stats,
2430         .select_queue   = taprio_select_queue,
2431 };
2432
2433 static struct Qdisc_ops taprio_qdisc_ops __read_mostly = {
2434         .cl_ops         = &taprio_class_ops,
2435         .id             = "taprio",
2436         .priv_size      = sizeof(struct taprio_sched),
2437         .init           = taprio_init,
2438         .change         = taprio_change,
2439         .destroy        = taprio_destroy,
2440         .reset          = taprio_reset,
2441         .attach         = taprio_attach,
2442         .peek           = taprio_peek,
2443         .dequeue        = taprio_dequeue,
2444         .enqueue        = taprio_enqueue,
2445         .dump           = taprio_dump,
2446         .owner          = THIS_MODULE,
2447 };
2448
2449 static struct notifier_block taprio_device_notifier = {
2450         .notifier_call = taprio_dev_notifier,
2451 };
2452
2453 static int __init taprio_module_init(void)
2454 {
2455         int err = register_netdevice_notifier(&taprio_device_notifier);
2456
2457         if (err)
2458                 return err;
2459
2460         return register_qdisc(&taprio_qdisc_ops);
2461 }
2462
2463 static void __exit taprio_module_exit(void)
2464 {
2465         unregister_qdisc(&taprio_qdisc_ops);
2466         unregister_netdevice_notifier(&taprio_device_notifier);
2467 }
2468
2469 module_init(taprio_module_init);
2470 module_exit(taprio_module_exit);
2471 MODULE_LICENSE("GPL");