net/sched/sch_taprio.c

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