net/sched/sch_hhf.c

   1 /* net/sched/sch_hhf.c          Heavy-Hitter Filter (HHF)
   2  *
   3  * Copyright (C) 2013 Terry Lam <vtlam@google.com>
   4  * Copyright (C) 2013 Nandita Dukkipati <nanditad@google.com>
   5  */
   6
   7 #include <linux/jhash.h>
   8 #include <linux/jiffies.h>
   9 #include <linux/module.h>
  10 #include <linux/skbuff.h>
  11 #include <linux/vmalloc.h>
  12 #include <net/pkt_sched.h>
  13 #include <net/sock.h>
  14
  15 /*      Heavy-Hitter Filter (HHF)
  16  *
  17  * Principles :
  18  * Flows are classified into two buckets: non-heavy-hitter and heavy-hitter
  19  * buckets. Initially, a new flow starts as non-heavy-hitter. Once classified
  20  * as heavy-hitter, it is immediately switched to the heavy-hitter bucket.
  21  * The buckets are dequeued by a Weighted Deficit Round Robin (WDRR) scheduler,
  22  * in which the heavy-hitter bucket is served with less weight.
  23  * In other words, non-heavy-hitters (e.g., short bursts of critical traffic)
  24  * are isolated from heavy-hitters (e.g., persistent bulk traffic) and also have
  25  * higher share of bandwidth.
  26  *
  27  * To capture heavy-hitters, we use the "multi-stage filter" algorithm in the
  28  * following paper:
  29  * [EV02] C. Estan and G. Varghese, "New Directions in Traffic Measurement and
  30  * Accounting", in ACM SIGCOMM, 2002.
  31  *
  32  * Conceptually, a multi-stage filter comprises k independent hash functions
  33  * and k counter arrays. Packets are indexed into k counter arrays by k hash
  34  * functions, respectively. The counters are then increased by the packet sizes.
  35  * Therefore,
  36  *    - For a heavy-hitter flow: *all* of its k array counters must be large.
  37  *    - For a non-heavy-hitter flow: some of its k array counters can be large
  38  *      due to hash collision with other small flows; however, with high
  39  *      probability, not *all* k counters are large.
  40  *
  41  * By the design of the multi-stage filter algorithm, the false negative rate
  42  * (heavy-hitters getting away uncaptured) is zero. However, the algorithm is
  43  * susceptible to false positives (non-heavy-hitters mistakenly classified as
  44  * heavy-hitters).
  45  * Therefore, we also implement the following optimizations to reduce false
  46  * positives by avoiding unnecessary increment of the counter values:
  47  *    - Optimization O1: once a heavy-hitter is identified, its bytes are not
  48  *        accounted in the array counters. This technique is called "shielding"
  49  *        in Section 3.3.1 of [EV02].
  50  *    - Optimization O2: conservative update of counters
  51  *                       (Section 3.3.2 of [EV02]),
  52  *        New counter value = max {old counter value,
  53  *                                 smallest counter value + packet bytes}
  54  *
  55  * Finally, we refresh the counters periodically since otherwise the counter
  56  * values will keep accumulating.
  57  *
  58  * Once a flow is classified as heavy-hitter, we also save its per-flow state
  59  * in an exact-matching flow table so that its subsequent packets can be
  60  * dispatched to the heavy-hitter bucket accordingly.
  61  *
  62  *
  63  * At a high level, this qdisc works as follows:
  64  * Given a packet p:
  65  *   - If the flow-id of p (e.g., TCP 5-tuple) is already in the exact-matching
  66  *     heavy-hitter flow table, denoted table T, then send p to the heavy-hitter
  67  *     bucket.
  68  *   - Otherwise, forward p to the multi-stage filter, denoted filter F
  69  *        + If F decides that p belongs to a non-heavy-hitter flow, then send p
  70  *          to the non-heavy-hitter bucket.
  71  *        + Otherwise, if F decides that p belongs to a new heavy-hitter flow,
  72  *          then set up a new flow entry for the flow-id of p in the table T and
  73  *          send p to the heavy-hitter bucket.
  74  *
  75  * In this implementation:
  76  *   - T is a fixed-size hash-table with 1024 entries. Hash collision is
  77  *     resolved by linked-list chaining.
  78  *   - F has four counter arrays, each array containing 1024 32-bit counters.
  79  *     That means 4 * 1024 * 32 bits = 16KB of memory.
  80  *   - Since each array in F contains 1024 counters, 10 bits are sufficient to
  81  *     index into each array.
  82  *     Hence, instead of having four hash functions, we chop the 32-bit
  83  *     skb-hash into three 10-bit chunks, and the remaining 10-bit chunk is
  84  *     computed as XOR sum of those three chunks.
  85  *   - We need to clear the counter arrays periodically; however, directly
  86  *     memsetting 16KB of memory can lead to cache eviction and unwanted delay.
  87  *     So by representing each counter by a valid bit, we only need to reset
  88  *     4K of 1 bit (i.e. 512 bytes) instead of 16KB of memory.
  89  *   - The Deficit Round Robin engine is taken from fq_codel implementation
  90  *     (net/sched/sch_fq_codel.c). Note that wdrr_bucket corresponds to
  91  *     fq_codel_flow in fq_codel implementation.
  92  *
  93  */
  94
  95 /* Non-configurable parameters */
  96 #define HH_FLOWS_CNT     1024  /* number of entries in exact-matching table T */
  97 #define HHF_ARRAYS_CNT   4     /* number of arrays in multi-stage filter F */
  98 #define HHF_ARRAYS_LEN   1024  /* number of counters in each array of F */
  99 #define HHF_BIT_MASK_LEN 10    /* masking 10 bits */
 100 #define HHF_BIT_MASK     0x3FF /* bitmask of 10 bits */
 101
 102 #define WDRR_BUCKET_CNT  2     /* two buckets for Weighted DRR */
 103 enum wdrr_bucket_idx {
 104         WDRR_BUCKET_FOR_HH      = 0, /* bucket id for heavy-hitters */
 105         WDRR_BUCKET_FOR_NON_HH  = 1  /* bucket id for non-heavy-hitters */
 106 };
 107
 108 #define hhf_time_before(a, b)   \
 109         (typecheck(u32, a) && typecheck(u32, b) && ((s32)((a) - (b)) < 0))
 110
 111 /* Heavy-hitter per-flow state */
 112 struct hh_flow_state {
 113         u32              hash_id;       /* hash of flow-id (e.g. TCP 5-tuple) */
 114         u32              hit_timestamp; /* last time heavy-hitter was seen */
 115         struct list_head flowchain;     /* chaining under hash collision */
 116 };
 117
 118 /* Weighted Deficit Round Robin (WDRR) scheduler */
 119 struct wdrr_bucket {
 120         struct sk_buff    *head;
 121         struct sk_buff    *tail;
 122         struct list_head  bucketchain;
 123         int               deficit;
 124 };
 125
 126 struct hhf_sched_data {
 127         struct wdrr_bucket buckets[WDRR_BUCKET_CNT];
 128         u32                perturbation;   /* hash perturbation */
 129         u32                quantum;        /* psched_mtu(qdisc_dev(sch)); */
 130         u32                drop_overlimit; /* number of times max qdisc packet
 131                                             * limit was hit
 132                                             */
 133         struct list_head   *hh_flows;       /* table T (currently active HHs) */
 134         u32                hh_flows_limit;            /* max active HH allocs */
 135         u32                hh_flows_overlimit; /* num of disallowed HH allocs */
 136         u32                hh_flows_total_cnt;          /* total admitted HHs */
 137         u32                hh_flows_current_cnt;        /* total current HHs  */
 138         u32                *hhf_arrays[HHF_ARRAYS_CNT]; /* HH filter F */
 139         u32                hhf_arrays_reset_timestamp;  /* last time hhf_arrays
 140                                                          * was reset
 141                                                          */
 142         unsigned long      *hhf_valid_bits[HHF_ARRAYS_CNT]; /* shadow valid bits
 143                                                              * of hhf_arrays
 144                                                              */
 145         /* Similar to the "new_flows" vs. "old_flows" concept in fq_codel DRR */
 146         struct list_head   new_buckets; /* list of new buckets */
 147         struct list_head   old_buckets; /* list of old buckets */
 148
 149         /* Configurable HHF parameters */
 150         u32                hhf_reset_timeout; /* interval to reset counter
 151                                                * arrays in filter F
 152                                                * (default 40ms)
 153                                                */
 154         u32                hhf_admit_bytes;   /* counter thresh to classify as
 155                                                * HH (default 128KB).
 156                                                * With these default values,
 157                                                * 128KB / 40ms = 25 Mbps
 158                                                * i.e., we expect to capture HHs
 159                                                * sending > 25 Mbps.
 160                                                */
 161         u32                hhf_evict_timeout; /* aging threshold to evict idle
 162                                                * HHs out of table T. This should
 163                                                * be large enough to avoid
 164                                                * reordering during HH eviction.
 165                                                * (default 1s)
 166                                                */
 167         u32                hhf_non_hh_weight; /* WDRR weight for non-HHs
 168                                                * (default 2,
 169                                                *  i.e., non-HH : HH = 2 : 1)
 170                                                */
 171 };
 172
 173 static u32 hhf_time_stamp(void)
 174 {
 175         return jiffies;
 176 }
 177
 178 /* Looks up a heavy-hitter flow in a chaining list of table T. */
 179 static struct hh_flow_state *seek_list(const u32 hash,
 180                                        struct list_head *head,
 181                                        struct hhf_sched_data *q)
 182 {
 183         struct hh_flow_state *flow, *next;
 184         u32 now = hhf_time_stamp();
 185
 186         if (list_empty(head))
 187                 return NULL;
 188
 189         list_for_each_entry_safe(flow, next, head, flowchain) {
 190                 u32 prev = flow->hit_timestamp + q->hhf_evict_timeout;
 191
 192                 if (hhf_time_before(prev, now)) {
 193                         /* Delete expired heavy-hitters, but preserve one entry
 194                          * to avoid kzalloc() when next time this slot is hit.
 195                          */
 196                         if (list_is_last(&flow->flowchain, head))
 197                                 return NULL;
 198                         list_del(&flow->flowchain);
 199                         kfree(flow);
 200                         q->hh_flows_current_cnt--;
 201                 } else if (flow->hash_id == hash) {
 202                         return flow;
 203                 }
 204         }
 205         return NULL;
 206 }
 207
 208 /* Returns a flow state entry for a new heavy-hitter.  Either reuses an expired
 209  * entry or dynamically alloc a new entry.
 210  */
 211 static struct hh_flow_state *alloc_new_hh(struct list_head *head,
 212                                           struct hhf_sched_data *q)
 213 {
 214         struct hh_flow_state *flow;
 215         u32 now = hhf_time_stamp();
 216
 217         if (!list_empty(head)) {
 218                 /* Find an expired heavy-hitter flow entry. */
 219                 list_for_each_entry(flow, head, flowchain) {
 220                         u32 prev = flow->hit_timestamp + q->hhf_evict_timeout;
 221
 222                         if (hhf_time_before(prev, now))
 223                                 return flow;
 224                 }
 225         }
 226
 227         if (q->hh_flows_current_cnt >= q->hh_flows_limit) {
 228                 q->hh_flows_overlimit++;
 229                 return NULL;
 230         }
 231         /* Create new entry. */
 232         flow = kzalloc(sizeof(struct hh_flow_state), GFP_ATOMIC);
 233         if (!flow)
 234                 return NULL;
 235
 236         q->hh_flows_current_cnt++;
 237         INIT_LIST_HEAD(&flow->flowchain);
 238         list_add_tail(&flow->flowchain, head);
 239
 240         return flow;
 241 }
 242
 243 /* Assigns packets to WDRR buckets.  Implements a multi-stage filter to
 244  * classify heavy-hitters.
 245  */
 246 static enum wdrr_bucket_idx hhf_classify(struct sk_buff *skb, struct Qdisc *sch)
 247 {
 248         struct hhf_sched_data *q = qdisc_priv(sch);
 249         u32 tmp_hash, hash;
 250         u32 xorsum, filter_pos[HHF_ARRAYS_CNT], flow_pos;
 251         struct hh_flow_state *flow;
 252         u32 pkt_len, min_hhf_val;
 253         int i;
 254         u32 prev;
 255         u32 now = hhf_time_stamp();
 256
 257         /* Reset the HHF counter arrays if this is the right time. */
 258         prev = q->hhf_arrays_reset_timestamp + q->hhf_reset_timeout;
 259         if (hhf_time_before(prev, now)) {
 260                 for (i = 0; i < HHF_ARRAYS_CNT; i++)
 261                         bitmap_zero(q->hhf_valid_bits[i], HHF_ARRAYS_LEN);
 262                 q->hhf_arrays_reset_timestamp = now;
 263         }
 264
 265         /* Get hashed flow-id of the skb. */
 266         hash = skb_get_hash_perturb(skb, q->perturbation);
 267
 268         /* Check if this packet belongs to an already established HH flow. */
 269         flow_pos = hash & HHF_BIT_MASK;
 270         flow = seek_list(hash, &q->hh_flows[flow_pos], q);
 271         if (flow) { /* found its HH flow */
 272                 flow->hit_timestamp = now;
 273                 return WDRR_BUCKET_FOR_HH;
 274         }
 275
 276         /* Now pass the packet through the multi-stage filter. */
 277         tmp_hash = hash;
 278         xorsum = 0;
 279         for (i = 0; i < HHF_ARRAYS_CNT - 1; i++) {
 280                 /* Split the skb_hash into three 10-bit chunks. */
 281                 filter_pos[i] = tmp_hash & HHF_BIT_MASK;
 282                 xorsum ^= filter_pos[i];
 283                 tmp_hash >>= HHF_BIT_MASK_LEN;
 284         }
 285         /* The last chunk is computed as XOR sum of other chunks. */
 286         filter_pos[HHF_ARRAYS_CNT - 1] = xorsum ^ tmp_hash;
 287
 288         pkt_len = qdisc_pkt_len(skb);
 289         min_hhf_val = ~0U;
 290         for (i = 0; i < HHF_ARRAYS_CNT; i++) {
 291                 u32 val;
 292
 293                 if (!test_bit(filter_pos[i], q->hhf_valid_bits[i])) {
 294                         q->hhf_arrays[i][filter_pos[i]] = 0;
 295                         __set_bit(filter_pos[i], q->hhf_valid_bits[i]);
 296                 }
 297
 298                 val = q->hhf_arrays[i][filter_pos[i]] + pkt_len;
 299                 if (min_hhf_val > val)
 300                         min_hhf_val = val;
 301         }
 302
 303         /* Found a new HH iff all counter values > HH admit threshold. */
 304         if (min_hhf_val > q->hhf_admit_bytes) {
 305                 /* Just captured a new heavy-hitter. */
 306                 flow = alloc_new_hh(&q->hh_flows[flow_pos], q);
 307                 if (!flow) /* memory alloc problem */
 308                         return WDRR_BUCKET_FOR_NON_HH;
 309                 flow->hash_id = hash;
 310                 flow->hit_timestamp = now;
 311                 q->hh_flows_total_cnt++;
 312
 313                 /* By returning without updating counters in q->hhf_arrays,
 314                  * we implicitly implement "shielding" (see Optimization O1).
 315                  */
 316                 return WDRR_BUCKET_FOR_HH;
 317         }
 318
 319         /* Conservative update of HHF arrays (see Optimization O2). */
 320         for (i = 0; i < HHF_ARRAYS_CNT; i++) {
 321                 if (q->hhf_arrays[i][filter_pos[i]] < min_hhf_val)
 322                         q->hhf_arrays[i][filter_pos[i]] = min_hhf_val;
 323         }
 324         return WDRR_BUCKET_FOR_NON_HH;
 325 }
 326
 327 /* Removes one skb from head of bucket. */
 328 static struct sk_buff *dequeue_head(struct wdrr_bucket *bucket)
 329 {
 330         struct sk_buff *skb = bucket->head;
 331
 332         bucket->head = skb->next;
 333         skb->next = NULL;
 334         return skb;
 335 }
 336
 337 /* Tail-adds skb to bucket. */
 338 static void bucket_add(struct wdrr_bucket *bucket, struct sk_buff *skb)
 339 {
 340         if (bucket->head == NULL)
 341                 bucket->head = skb;
 342         else
 343                 bucket->tail->next = skb;
 344         bucket->tail = skb;
 345         skb->next = NULL;
 346 }
 347
 348 static unsigned int hhf_drop(struct Qdisc *sch, struct sk_buff **to_free)
 349 {
 350         struct hhf_sched_data *q = qdisc_priv(sch);
 351         struct wdrr_bucket *bucket;
 352
 353         /* Always try to drop from heavy-hitters first. */
 354         bucket = &q->buckets[WDRR_BUCKET_FOR_HH];
 355         if (!bucket->head)
 356                 bucket = &q->buckets[WDRR_BUCKET_FOR_NON_HH];
 357
 358         if (bucket->head) {
 359                 struct sk_buff *skb = dequeue_head(bucket);
 360
 361                 sch->q.qlen--;
 362                 qdisc_qstats_backlog_dec(sch, skb);
 363                 qdisc_drop(skb, sch, to_free);
 364         }
 365
 366         /* Return id of the bucket from which the packet was dropped. */
 367         return bucket - q->buckets;
 368 }
 369
 370 static int hhf_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 371                        struct sk_buff **to_free)
 372 {
 373         struct hhf_sched_data *q = qdisc_priv(sch);
 374         enum wdrr_bucket_idx idx;
 375         struct wdrr_bucket *bucket;
 376         unsigned int prev_backlog;
 377
 378         idx = hhf_classify(skb, sch);
 379
 380         bucket = &q->buckets[idx];
 381         bucket_add(bucket, skb);
 382         qdisc_qstats_backlog_inc(sch, skb);
 383
 384         if (list_empty(&bucket->bucketchain)) {
 385                 unsigned int weight;
 386
 387                 /* The logic of new_buckets vs. old_buckets is the same as
 388                  * new_flows vs. old_flows in the implementation of fq_codel,
 389                  * i.e., short bursts of non-HHs should have strict priority.
 390                  */
 391                 if (idx == WDRR_BUCKET_FOR_HH) {
 392                         /* Always move heavy-hitters to old bucket. */
 393                         weight = 1;
 394                         list_add_tail(&bucket->bucketchain, &q->old_buckets);
 395                 } else {
 396                         weight = q->hhf_non_hh_weight;
 397                         list_add_tail(&bucket->bucketchain, &q->new_buckets);
 398                 }
 399                 bucket->deficit = weight * q->quantum;
 400         }
 401         if (++sch->q.qlen <= sch->limit)
 402                 return NET_XMIT_SUCCESS;
 403
 404         prev_backlog = sch->qstats.backlog;
 405         q->drop_overlimit++;
 406         /* Return Congestion Notification only if we dropped a packet from this
 407          * bucket.
 408          */
 409         if (hhf_drop(sch, to_free) == idx)
 410                 return NET_XMIT_CN;
 411
 412         /* As we dropped a packet, better let upper stack know this. */
 413         qdisc_tree_reduce_backlog(sch, 1, prev_backlog - sch->qstats.backlog);
 414         return NET_XMIT_SUCCESS;
 415 }
 416
 417 static struct sk_buff *hhf_dequeue(struct Qdisc *sch)
 418 {
 419         struct hhf_sched_data *q = qdisc_priv(sch);
 420         struct sk_buff *skb = NULL;
 421         struct wdrr_bucket *bucket;
 422         struct list_head *head;
 423
 424 begin:
 425         head = &q->new_buckets;
 426         if (list_empty(head)) {
 427                 head = &q->old_buckets;
 428                 if (list_empty(head))
 429                         return NULL;
 430         }
 431         bucket = list_first_entry(head, struct wdrr_bucket, bucketchain);
 432
 433         if (bucket->deficit <= 0) {
 434                 int weight = (bucket - q->buckets == WDRR_BUCKET_FOR_HH) ?
 435                               1 : q->hhf_non_hh_weight;
 436
 437                 bucket->deficit += weight * q->quantum;
 438                 list_move_tail(&bucket->bucketchain, &q->old_buckets);
 439                 goto begin;
 440         }
 441
 442         if (bucket->head) {
 443                 skb = dequeue_head(bucket);
 444                 sch->q.qlen--;
 445                 qdisc_qstats_backlog_dec(sch, skb);
 446         }
 447
 448         if (!skb) {
 449                 /* Force a pass through old_buckets to prevent starvation. */
 450                 if ((head == &q->new_buckets) && !list_empty(&q->old_buckets))
 451                         list_move_tail(&bucket->bucketchain, &q->old_buckets);
 452                 else
 453                         list_del_init(&bucket->bucketchain);
 454                 goto begin;
 455         }
 456         qdisc_bstats_update(sch, skb);
 457         bucket->deficit -= qdisc_pkt_len(skb);
 458
 459         return skb;
 460 }
 461
 462 static void hhf_reset(struct Qdisc *sch)
 463 {
 464         struct sk_buff *skb;
 465
 466         while ((skb = hhf_dequeue(sch)) != NULL)
 467                 rtnl_kfree_skbs(skb, skb);
 468 }
 469
 470 static void hhf_destroy(struct Qdisc *sch)
 471 {
 472         int i;
 473         struct hhf_sched_data *q = qdisc_priv(sch);
 474
 475         for (i = 0; i < HHF_ARRAYS_CNT; i++) {
 476                 kvfree(q->hhf_arrays[i]);
 477                 kvfree(q->hhf_valid_bits[i]);
 478         }
 479
 480         if (!q->hh_flows)
 481                 return;
 482
 483         for (i = 0; i < HH_FLOWS_CNT; i++) {
 484                 struct hh_flow_state *flow, *next;
 485                 struct list_head *head = &q->hh_flows[i];
 486
 487                 if (list_empty(head))
 488                         continue;
 489                 list_for_each_entry_safe(flow, next, head, flowchain) {
 490                         list_del(&flow->flowchain);
 491                         kfree(flow);
 492                 }
 493         }
 494         kvfree(q->hh_flows);
 495 }
 496
 497 static const struct nla_policy hhf_policy[TCA_HHF_MAX + 1] = {
 498         [TCA_HHF_BACKLOG_LIMIT]  = { .type = NLA_U32 },
 499         [TCA_HHF_QUANTUM]        = { .type = NLA_U32 },
 500         [TCA_HHF_HH_FLOWS_LIMIT] = { .type = NLA_U32 },
 501         [TCA_HHF_RESET_TIMEOUT]  = { .type = NLA_U32 },
 502         [TCA_HHF_ADMIT_BYTES]    = { .type = NLA_U32 },
 503         [TCA_HHF_EVICT_TIMEOUT]  = { .type = NLA_U32 },
 504         [TCA_HHF_NON_HH_WEIGHT]  = { .type = NLA_U32 },
 505 };
 506
 507 static int hhf_change(struct Qdisc *sch, struct nlattr *opt,
 508                       struct netlink_ext_ack *extack)
 509 {
 510         struct hhf_sched_data *q = qdisc_priv(sch);
 511         struct nlattr *tb[TCA_HHF_MAX + 1];
 512         unsigned int qlen, prev_backlog;
 513         int err;
 514         u64 non_hh_quantum;
 515         u32 new_quantum = q->quantum;
 516         u32 new_hhf_non_hh_weight = q->hhf_non_hh_weight;
 517
 518         if (!opt)
 519                 return -EINVAL;
 520
 521         err = nla_parse_nested(tb, TCA_HHF_MAX, opt, hhf_policy, NULL);
 522         if (err < 0)
 523                 return err;
 524
 525         if (tb[TCA_HHF_QUANTUM])
 526                 new_quantum = nla_get_u32(tb[TCA_HHF_QUANTUM]);
 527
 528         if (tb[TCA_HHF_NON_HH_WEIGHT])
 529                 new_hhf_non_hh_weight = nla_get_u32(tb[TCA_HHF_NON_HH_WEIGHT]);
 530
 531         non_hh_quantum = (u64)new_quantum * new_hhf_non_hh_weight;
 532         if (non_hh_quantum > INT_MAX)
 533                 return -EINVAL;
 534
 535         sch_tree_lock(sch);
 536
 537         if (tb[TCA_HHF_BACKLOG_LIMIT])
 538                 sch->limit = nla_get_u32(tb[TCA_HHF_BACKLOG_LIMIT]);
 539
 540         q->quantum = new_quantum;
 541         q->hhf_non_hh_weight = new_hhf_non_hh_weight;
 542
 543         if (tb[TCA_HHF_HH_FLOWS_LIMIT])
 544                 q->hh_flows_limit = nla_get_u32(tb[TCA_HHF_HH_FLOWS_LIMIT]);
 545
 546         if (tb[TCA_HHF_RESET_TIMEOUT]) {
 547                 u32 us = nla_get_u32(tb[TCA_HHF_RESET_TIMEOUT]);
 548
 549                 q->hhf_reset_timeout = usecs_to_jiffies(us);
 550         }
 551
 552         if (tb[TCA_HHF_ADMIT_BYTES])
 553                 q->hhf_admit_bytes = nla_get_u32(tb[TCA_HHF_ADMIT_BYTES]);
 554
 555         if (tb[TCA_HHF_EVICT_TIMEOUT]) {
 556                 u32 us = nla_get_u32(tb[TCA_HHF_EVICT_TIMEOUT]);
 557
 558                 q->hhf_evict_timeout = usecs_to_jiffies(us);
 559         }
 560
 561         qlen = sch->q.qlen;
 562         prev_backlog = sch->qstats.backlog;
 563         while (sch->q.qlen > sch->limit) {
 564                 struct sk_buff *skb = hhf_dequeue(sch);
 565
 566                 rtnl_kfree_skbs(skb, skb);
 567         }
 568         qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen,
 569                                   prev_backlog - sch->qstats.backlog);
 570
 571         sch_tree_unlock(sch);
 572         return 0;
 573 }
 574
 575 static int hhf_init(struct Qdisc *sch, struct nlattr *opt,
 576                     struct netlink_ext_ack *extack)
 577 {
 578         struct hhf_sched_data *q = qdisc_priv(sch);
 579         int i;
 580
 581         sch->limit = 1000;
 582         q->quantum = psched_mtu(qdisc_dev(sch));
 583         q->perturbation = prandom_u32();
 584         INIT_LIST_HEAD(&q->new_buckets);
 585         INIT_LIST_HEAD(&q->old_buckets);
 586
 587         /* Configurable HHF parameters */
 588         q->hhf_reset_timeout = HZ / 25; /* 40  ms */
 589         q->hhf_admit_bytes = 131072;    /* 128 KB */
 590         q->hhf_evict_timeout = HZ;      /* 1  sec */
 591         q->hhf_non_hh_weight = 2;
 592
 593         if (opt) {
 594                 int err = hhf_change(sch, opt, extack);
 595
 596                 if (err)
 597                         return err;
 598         }
 599
 600         if (!q->hh_flows) {
 601                 /* Initialize heavy-hitter flow table. */
 602                 q->hh_flows = kvcalloc(HH_FLOWS_CNT, sizeof(struct list_head),
 603                                        GFP_KERNEL);
 604                 if (!q->hh_flows)
 605                         return -ENOMEM;
 606                 for (i = 0; i < HH_FLOWS_CNT; i++)
 607                         INIT_LIST_HEAD(&q->hh_flows[i]);
 608
 609                 /* Cap max active HHs at twice len of hh_flows table. */
 610                 q->hh_flows_limit = 2 * HH_FLOWS_CNT;
 611                 q->hh_flows_overlimit = 0;
 612                 q->hh_flows_total_cnt = 0;
 613                 q->hh_flows_current_cnt = 0;
 614
 615                 /* Initialize heavy-hitter filter arrays. */
 616                 for (i = 0; i < HHF_ARRAYS_CNT; i++) {
 617                         q->hhf_arrays[i] = kvcalloc(HHF_ARRAYS_LEN,
 618                                                     sizeof(u32),
 619                                                     GFP_KERNEL);
 620                         if (!q->hhf_arrays[i]) {
 621                                 /* Note: hhf_destroy() will be called
 622                                  * by our caller.
 623                                  */
 624                                 return -ENOMEM;
 625                         }
 626                 }
 627                 q->hhf_arrays_reset_timestamp = hhf_time_stamp();
 628
 629                 /* Initialize valid bits of heavy-hitter filter arrays. */
 630                 for (i = 0; i < HHF_ARRAYS_CNT; i++) {
 631                         q->hhf_valid_bits[i] = kvzalloc(HHF_ARRAYS_LEN /
 632                                                           BITS_PER_BYTE, GFP_KERNEL);
 633                         if (!q->hhf_valid_bits[i]) {
 634                                 /* Note: hhf_destroy() will be called
 635                                  * by our caller.
 636                                  */
 637                                 return -ENOMEM;
 638                         }
 639                 }
 640
 641                 /* Initialize Weighted DRR buckets. */
 642                 for (i = 0; i < WDRR_BUCKET_CNT; i++) {
 643                         struct wdrr_bucket *bucket = q->buckets + i;
 644
 645                         INIT_LIST_HEAD(&bucket->bucketchain);
 646                 }
 647         }
 648
 649         return 0;
 650 }
 651
 652 static int hhf_dump(struct Qdisc *sch, struct sk_buff *skb)
 653 {
 654         struct hhf_sched_data *q = qdisc_priv(sch);
 655         struct nlattr *opts;
 656
 657         opts = nla_nest_start(skb, TCA_OPTIONS);
 658         if (opts == NULL)
 659                 goto nla_put_failure;
 660
 661         if (nla_put_u32(skb, TCA_HHF_BACKLOG_LIMIT, sch->limit) ||
 662             nla_put_u32(skb, TCA_HHF_QUANTUM, q->quantum) ||
 663             nla_put_u32(skb, TCA_HHF_HH_FLOWS_LIMIT, q->hh_flows_limit) ||
 664             nla_put_u32(skb, TCA_HHF_RESET_TIMEOUT,
 665                         jiffies_to_usecs(q->hhf_reset_timeout)) ||
 666             nla_put_u32(skb, TCA_HHF_ADMIT_BYTES, q->hhf_admit_bytes) ||
 667             nla_put_u32(skb, TCA_HHF_EVICT_TIMEOUT,
 668                         jiffies_to_usecs(q->hhf_evict_timeout)) ||
 669             nla_put_u32(skb, TCA_HHF_NON_HH_WEIGHT, q->hhf_non_hh_weight))
 670                 goto nla_put_failure;
 671
 672         return nla_nest_end(skb, opts);
 673
 674 nla_put_failure:
 675         return -1;
 676 }
 677
 678 static int hhf_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 679 {
 680         struct hhf_sched_data *q = qdisc_priv(sch);
 681         struct tc_hhf_xstats st = {
 682                 .drop_overlimit = q->drop_overlimit,
 683                 .hh_overlimit   = q->hh_flows_overlimit,
 684                 .hh_tot_count   = q->hh_flows_total_cnt,
 685                 .hh_cur_count   = q->hh_flows_current_cnt,
 686         };
 687
 688         return gnet_stats_copy_app(d, &st, sizeof(st));
 689 }
 690
 691 static struct Qdisc_ops hhf_qdisc_ops __read_mostly = {
 692         .id             =       "hhf",
 693         .priv_size      =       sizeof(struct hhf_sched_data),
 694
 695         .enqueue        =       hhf_enqueue,
 696         .dequeue        =       hhf_dequeue,
 697         .peek           =       qdisc_peek_dequeued,
 698         .init           =       hhf_init,
 699         .reset          =       hhf_reset,
 700         .destroy        =       hhf_destroy,
 701         .change         =       hhf_change,
 702         .dump           =       hhf_dump,
 703         .dump_stats     =       hhf_dump_stats,
 704         .owner          =       THIS_MODULE,
 705 };
 706
 707 static int __init hhf_module_init(void)
 708 {
 709         return register_qdisc(&hhf_qdisc_ops);
 710 }
 711
 712 static void __exit hhf_module_exit(void)
 713 {
 714         unregister_qdisc(&hhf_qdisc_ops);
 715 }
 716
 717 module_init(hhf_module_init)
 718 module_exit(hhf_module_exit)
 719 MODULE_AUTHOR("Terry Lam");
 720 MODULE_AUTHOR("Nandita Dukkipati");
 721 MODULE_LICENSE("GPL");