esfq-kernel.patch

   1 diff -Naur linux-2.6.13.3.old/include/linux/pkt_sched.h linux-2.6.13.3/include/linux/pkt_sched.h
   2 --- linux-2.6.13.3.old/include/linux/pkt_sched.h        2005-10-23 17:39:16.000000000 -0700
   3 +++ linux-2.6.13.3/include/linux/pkt_sched.h    2005-10-21 18:13:37.000000000 -0700
   4 @@ -145,8 +145,35 @@
   5   *
   6   *     The only reason for this is efficiency, it is possible
   7   *     to change these parameters in compile time.
   8 + *
   9 + *     If you need to play with these values use esfq instead.
  10   */
  11
  12 +/* ESFQ section */
  13 +
  14 +enum
  15 +{
  16 +        /* traditional */
  17 +       TCA_SFQ_HASH_CLASSIC,
  18 +       TCA_SFQ_HASH_DST,
  19 +       TCA_SFQ_HASH_SRC,
  20 +       TCA_SFQ_HASH_FWMARK,
  21 +        /* direct */
  22 +       TCA_SFQ_HASH_DSTDIR,
  23 +       TCA_SFQ_HASH_SRCDIR,
  24 +       TCA_SFQ_HASH_FWMARKDIR,
  25 +};
  26 +
  27 +struct tc_esfq_qopt
  28 +{
  29 +       unsigned        quantum;        /* Bytes per round allocated to flow */
  30 +       int             perturb_period; /* Period of hash perturbation */
  31 +       __u32           limit;          /* Maximal packets in queue */
  32 +       unsigned        divisor;        /* Hash divisor  */
  33 +       unsigned        flows;          /* Maximal number of flows  */
  34 +       unsigned        hash_kind;      /* Hash function to use for flow identification */
  35 +};
  36 +
  37  /* RED section */
  38
  39  enum
  40 diff -Naur linux-2.6.13.3.old/net/sched/Kconfig linux-2.6.13.3/net/sched/Kconfig
  41 --- linux-2.6.13.3.old/net/sched/Kconfig        2005-10-23 17:39:16.000000000 -0700
  42 +++ linux-2.6.13.3/net/sched/Kconfig    2005-10-09 16:56:27.000000000 -0700
  43 @@ -191,6 +191,28 @@
  44           To compile this code as a module, choose M here: the
  45           module will be called sch_sfq.
  46
  47 +config NET_SCH_ESFQ
  48 +       tristate "ESFQ queue"
  49 +       depends on NET_SCHED
  50 +       ---help---
  51 +         Say Y here if you want to use the Enhanced Stochastic Fairness
  52 +         Queueing (ESFQ) packet scheduling algorithm for some of your network
  53 +         devices or as a leaf discipline for a classful qdisc such as HTB or
  54 +         CBQ (see the top of <file:net/sched/sch_esfq.c> for details and
  55 +         references to the SFQ algorithm).
  56 +
  57 +         This is an enchanced SFQ version which allows you to control some
  58 +         hardcoded values in the SFQ scheduler: queue depth, hash table size,
  59 +         and queues limit.
  60 +
  61 +         ESFQ also adds control to the hash function used to identify packet
  62 +         flows. The original SFQ hashes by individual flow (TCP session or UDP
  63 +         stream); ESFQ can hash by src or dst IP as well, which can be more
  64 +         fair to users in some networking situations.
  65 +
  66 +         To compile this code as a module, choose M here: the
  67 +         module will be called sch_esfq.
  68 +
  69  config NET_SCH_TEQL
  70         tristate "TEQL queue"
  71         depends on NET_SCHED
  72 diff -Naur linux-2.6.13.3.old/net/sched/Makefile linux-2.6.13.3/net/sched/Makefile
  73 --- linux-2.6.13.3.old/net/sched/Makefile       2005-10-23 17:39:16.000000000 -0700
  74 +++ linux-2.6.13.3/net/sched/Makefile   2005-10-09 16:56:27.000000000 -0700
  75 @@ -23,6 +23,7 @@
  76  obj-$(CONFIG_NET_SCH_INGRESS)  += sch_ingress.o
  77  obj-$(CONFIG_NET_SCH_DSMARK)   += sch_dsmark.o
  78  obj-$(CONFIG_NET_SCH_SFQ)      += sch_sfq.o
  79 +obj-$(CONFIG_NET_SCH_ESFQ)     += sch_esfq.o
  80  obj-$(CONFIG_NET_SCH_TBF)      += sch_tbf.o
  81  obj-$(CONFIG_NET_SCH_TEQL)     += sch_teql.o
  82  obj-$(CONFIG_NET_SCH_PRIO)     += sch_prio.o
  83 diff -Naur linux-2.6.13.3.old/net/sched/sch_esfq.c linux-2.6.13.3/net/sched/sch_esfq.c
  84 --- linux-2.6.13.3.old/net/sched/sch_esfq.c     1969-12-31 16:00:00.000000000 -0800
  85 +++ linux-2.6.13.3/net/sched/sch_esfq.c 2005-10-23 17:32:19.000000000 -0700
  86 @@ -0,0 +1,639 @@
  87 +/*
  88 + * net/sched/sch_esfq.c        Extended Stochastic Fairness Queueing discipline.
  89 + *
  90 + *             This program is free software; you can redistribute it and/or
  91 + *             modify it under the terms of the GNU General Public License
  92 + *             as published by the Free Software Foundation; either version
  93 + *             2 of the License, or (at your option) any later version.
  94 + *
  95 + * Authors:    Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  96 + *
  97 + * Changes:    Alexander Atanasov, <alex@ssi.bg>
  98 + *             Added dynamic depth,limit,divisor,hash_kind options.
  99 + *             Added dst and src hashes.
 100 + *
 101 + *             Alexander Clouter, <alex@digriz.org.uk>
 102 + *             Ported ESFQ to Linux 2.6.
 103 + *
 104 + *             Corey Hickey, <bugfood-c@fatooh.org>
 105 + *             Maintenance of the Linux 2.6 port.
 106 + *             Added fwmark hash (thanks to Robert Kurjata)
 107 + *             Added direct hashing for src, dst, and fwmark.
 108 + *
 109 + */
 110 +
 111 +#include <linux/config.h>
 112 +#include <linux/module.h>
 113 +#include <asm/uaccess.h>
 114 +#include <asm/system.h>
 115 +#include <linux/bitops.h>
 116 +#include <linux/types.h>
 117 +#include <linux/kernel.h>
 118 +#include <linux/jiffies.h>
 119 +#include <linux/string.h>
 120 +#include <linux/mm.h>
 121 +#include <linux/socket.h>
 122 +#include <linux/sockios.h>
 123 +#include <linux/in.h>
 124 +#include <linux/errno.h>
 125 +#include <linux/interrupt.h>
 126 +#include <linux/if_ether.h>
 127 +#include <linux/inet.h>
 128 +#include <linux/netdevice.h>
 129 +#include <linux/etherdevice.h>
 130 +#include <linux/notifier.h>
 131 +#include <linux/init.h>
 132 +#include <net/ip.h>
 133 +#include <linux/ipv6.h>
 134 +#include <net/route.h>
 135 +#include <linux/skbuff.h>
 136 +#include <net/sock.h>
 137 +#include <net/pkt_sched.h>
 138 +
 139 +
 140 +/*     Stochastic Fairness Queuing algorithm.
 141 +       For more comments look at sch_sfq.c.
 142 +       The difference is that you can change limit, depth,
 143 +       hash table size and choose 7 hash types.
 144 +
 145 +       classic:        same as in sch_sfq.c
 146 +       dst:            destination IP address
 147 +       src:            source IP address
 148 +       fwmark:         netfilter mark value
 149 +       dst_direct:
 150 +       src_direct:
 151 +       fwmark_direct:  direct hashing of the above sources
 152 +
 153 +       TODO:
 154 +               make sfq_change work.
 155 +*/
 156 +
 157 +
 158 +/* This type should contain at least SFQ_DEPTH*2 values */
 159 +typedef unsigned int esfq_index;
 160 +
 161 +struct esfq_head
 162 +{
 163 +       esfq_index      next;
 164 +       esfq_index      prev;
 165 +};
 166 +
 167 +struct esfq_sched_data
 168 +{
 169 +/* Parameters */
 170 +       int             perturb_period;
 171 +       unsigned        quantum;        /* Allotment per round: MUST BE >= MTU */
 172 +       int             limit;
 173 +       unsigned        depth;
 174 +       unsigned        hash_divisor;
 175 +       unsigned        hash_kind;
 176 +/* Variables */
 177 +       struct timer_list perturb_timer;
 178 +       int             perturbation;
 179 +       esfq_index      tail;           /* Index of current slot in round */
 180 +       esfq_index      max_depth;      /* Maximal depth */
 181 +
 182 +       esfq_index      *ht;                    /* Hash table */
 183 +       esfq_index      *next;                  /* Active slots link */
 184 +       short           *allot;                 /* Current allotment per slot */
 185 +       unsigned short  *hash;                  /* Hash value indexed by slots */
 186 +       struct sk_buff_head     *qs;            /* Slot queue */
 187 +       struct esfq_head        *dep;           /* Linked list of slots, indexed by depth */
 188 +       unsigned        dyn_min;        /* For dynamic divisor adjustment; minimum value seen */
 189 +       unsigned        dyn_max;        /*                                 maximum value seen */
 190 +       unsigned        dyn_range;      /*                                 saved range */
 191 +};
 192 +
 193 +static __inline__ unsigned esfq_hash_u32(struct esfq_sched_data *q,u32 h)
 194 +{
 195 +       int pert = q->perturbation;
 196 +
 197 +       if (pert)
 198 +               h = (h<<pert) ^ (h>>(0x1F - pert));
 199 +
 200 +       h = ntohl(h) * 2654435761UL;
 201 +       return h & (q->hash_divisor-1);
 202 +}
 203 +
 204 +/* Hash input values directly into the "nearest" slot, taking into account the
 205 + * range of input values seen. This is most useful when the hash table is at
 206 + * least as large as the range of possible values. */
 207 +static __inline__ unsigned esfq_hash_direct(struct esfq_sched_data *q, u32 h)
 208 +{
 209 +       /* adjust minimum and maximum */
 210 +       if (h < q->dyn_min || h > q->dyn_max) {
 211 +               q->dyn_min = h < q->dyn_min ? h : q->dyn_min;
 212 +               q->dyn_max = h > q->dyn_max ? h : q->dyn_max;
 213 +
 214 +               /* find new range */
 215 +               if ((q->dyn_range = q->dyn_max - q->dyn_min) >= q->hash_divisor)
 216 +                       printk(KERN_WARNING "ESFQ: (direct hash) Input range %u is larger than hash "
 217 +                                       "table. See ESFQ README for details.\n", q->dyn_range);
 218 +       }
 219 +
 220 +       /* hash input values into slot numbers */
 221 +       if (q->dyn_min == q->dyn_max)
 222 +               return 0; /* only one value seen; avoid division by 0 */
 223 +       else
 224 +               return (h - q->dyn_min) * (q->hash_divisor - 1) / q->dyn_range;
 225 +}
 226 +
 227 +static __inline__ unsigned esfq_fold_hash_classic(struct esfq_sched_data *q, u32 h, u32 h1)
 228 +{
 229 +       int pert = q->perturbation;
 230 +
 231 +       /* Have we any rotation primitives? If not, WHY? */
 232 +       h ^= (h1<<pert) ^ (h1>>(0x1F - pert));
 233 +       h ^= h>>10;
 234 +       return h & (q->hash_divisor-1);
 235 +}
 236 +
 237 +static unsigned esfq_hash(struct esfq_sched_data *q, struct sk_buff *skb)
 238 +{
 239 +       u32 h, h2;
 240 +       u32 hs;
 241 +       u32 nfm;
 242 +
 243 +       switch (skb->protocol) {
 244 +       case __constant_htons(ETH_P_IP):
 245 +       {
 246 +               struct iphdr *iph = skb->nh.iph;
 247 +               h = iph->daddr;
 248 +               hs = iph->saddr;
 249 +               nfm = skb->nfmark;
 250 +               h2 = hs^iph->protocol;
 251 +               if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) &&
 252 +                   (iph->protocol == IPPROTO_TCP ||
 253 +                    iph->protocol == IPPROTO_UDP ||
 254 +                    iph->protocol == IPPROTO_ESP))
 255 +                       h2 ^= *(((u32*)iph) + iph->ihl);
 256 +               break;
 257 +       }
 258 +       case __constant_htons(ETH_P_IPV6):
 259 +       {
 260 +               struct ipv6hdr *iph = skb->nh.ipv6h;
 261 +               h = iph->daddr.s6_addr32[3];
 262 +               hs = iph->saddr.s6_addr32[3];
 263 +               nfm = skb->nfmark;
 264 +               h2 = hs^iph->nexthdr;
 265 +               if (iph->nexthdr == IPPROTO_TCP ||
 266 +                   iph->nexthdr == IPPROTO_UDP ||
 267 +                   iph->nexthdr == IPPROTO_ESP)
 268 +                       h2 ^= *(u32*)&iph[1];
 269 +               break;
 270 +       }
 271 +       default:
 272 +               h = (u32)(unsigned long)skb->dst;
 273 +               hs = (u32)(unsigned long)skb->sk;
 274 +               nfm = skb->nfmark;
 275 +               h2 = hs^skb->protocol;
 276 +       }
 277 +       switch(q->hash_kind)
 278 +       {
 279 +       case TCA_SFQ_HASH_CLASSIC:
 280 +               return esfq_fold_hash_classic(q, h, h2);
 281 +       case TCA_SFQ_HASH_DST:
 282 +               return esfq_hash_u32(q,h);
 283 +       case TCA_SFQ_HASH_DSTDIR:
 284 +               return esfq_hash_direct(q, ntohl(h));
 285 +       case TCA_SFQ_HASH_SRC:
 286 +               return esfq_hash_u32(q,hs);
 287 +       case TCA_SFQ_HASH_SRCDIR:
 288 +               return esfq_hash_direct(q, ntohl(hs));
 289 +#ifdef CONFIG_NETFILTER
 290 +       case TCA_SFQ_HASH_FWMARK:
 291 +               return esfq_hash_u32(q,nfm);
 292 +       case TCA_SFQ_HASH_FWMARKDIR:
 293 +               return esfq_hash_direct(q,nfm);
 294 +#endif
 295 +       default:
 296 +               if (net_ratelimit())
 297 +                       printk(KERN_WARNING "ESFQ: Unknown hash method. Falling back to classic.\n");
 298 +       }
 299 +       return esfq_fold_hash_classic(q, h, h2);
 300 +}
 301 +
 302 +static inline void esfq_link(struct esfq_sched_data *q, esfq_index x)
 303 +{
 304 +       esfq_index p, n;
 305 +       int d = q->qs[x].qlen + q->depth;
 306 +
 307 +       p = d;
 308 +       n = q->dep[d].next;
 309 +       q->dep[x].next = n;
 310 +       q->dep[x].prev = p;
 311 +       q->dep[p].next = q->dep[n].prev = x;
 312 +}
 313 +
 314 +static inline void esfq_dec(struct esfq_sched_data *q, esfq_index x)
 315 +{
 316 +       esfq_index p, n;
 317 +
 318 +       n = q->dep[x].next;
 319 +       p = q->dep[x].prev;
 320 +       q->dep[p].next = n;
 321 +       q->dep[n].prev = p;
 322 +
 323 +       if (n == p && q->max_depth == q->qs[x].qlen + 1)
 324 +               q->max_depth--;
 325 +
 326 +       esfq_link(q, x);
 327 +}
 328 +
 329 +static inline void esfq_inc(struct esfq_sched_data *q, esfq_index x)
 330 +{
 331 +       esfq_index p, n;
 332 +       int d;
 333 +
 334 +       n = q->dep[x].next;
 335 +       p = q->dep[x].prev;
 336 +       q->dep[p].next = n;
 337 +       q->dep[n].prev = p;
 338 +       d = q->qs[x].qlen;
 339 +       if (q->max_depth < d)
 340 +               q->max_depth = d;
 341 +
 342 +       esfq_link(q, x);
 343 +}
 344 +
 345 +static unsigned int esfq_drop(struct Qdisc *sch)
 346 +{
 347 +       struct esfq_sched_data *q = qdisc_priv(sch);
 348 +       esfq_index d = q->max_depth;
 349 +       struct sk_buff *skb;
 350 +       unsigned int len;
 351 +
 352 +       /* Queue is full! Find the longest slot and
 353 +          drop a packet from it */
 354 +
 355 +       if (d > 1) {
 356 +               esfq_index x = q->dep[d+q->depth].next;
 357 +               skb = q->qs[x].prev;
 358 +               len = skb->len;
 359 +               __skb_unlink(skb, &q->qs[x]);
 360 +               kfree_skb(skb);
 361 +               esfq_dec(q, x);
 362 +               sch->q.qlen--;
 363 +               sch->qstats.drops++;
 364 +               return len;
 365 +       }
 366 +
 367 +       if (d == 1) {
 368 +               /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
 369 +               d = q->next[q->tail];
 370 +               q->next[q->tail] = q->next[d];
 371 +               q->allot[q->next[d]] += q->quantum;
 372 +               skb = q->qs[d].prev;
 373 +               len = skb->len;
 374 +               __skb_unlink(skb, &q->qs[d]);
 375 +               kfree_skb(skb);
 376 +               esfq_dec(q, d);
 377 +               sch->q.qlen--;
 378 +               q->ht[q->hash[d]] = q->depth;
 379 +               sch->qstats.drops++;
 380 +               return len;
 381 +       }
 382 +
 383 +       return 0;
 384 +}
 385 +
 386 +static int
 387 +esfq_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 388 +{
 389 +       struct esfq_sched_data *q = qdisc_priv(sch);
 390 +       unsigned hash = esfq_hash(q, skb);
 391 +       unsigned depth = q->depth;
 392 +       esfq_index x;
 393 +
 394 +       x = q->ht[hash];
 395 +       if (x == depth) {
 396 +               q->ht[hash] = x = q->dep[depth].next;
 397 +               q->hash[x] = hash;
 398 +       }
 399 +       __skb_queue_tail(&q->qs[x], skb);
 400 +       esfq_inc(q, x);
 401 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 402 +               if (q->tail == depth) { /* It is the first flow */
 403 +                       q->tail = x;
 404 +                       q->next[x] = x;
 405 +                       q->allot[x] = q->quantum;
 406 +               } else {
 407 +                       q->next[x] = q->next[q->tail];
 408 +                       q->next[q->tail] = x;
 409 +                       q->tail = x;
 410 +               }
 411 +       }
 412 +       if (++sch->q.qlen < q->limit-1) {
 413 +               sch->bstats.bytes += skb->len;
 414 +               sch->bstats.packets++;
 415 +               return 0;
 416 +       }
 417 +
 418 +       esfq_drop(sch);
 419 +       return NET_XMIT_CN;
 420 +}
 421 +
 422 +static int
 423 +esfq_requeue(struct sk_buff *skb, struct Qdisc* sch)
 424 +{
 425 +       struct esfq_sched_data *q = qdisc_priv(sch);
 426 +       unsigned hash = esfq_hash(q, skb);
 427 +       unsigned depth = q->depth;
 428 +       esfq_index x;
 429 +
 430 +       x = q->ht[hash];
 431 +       if (x == depth) {
 432 +               q->ht[hash] = x = q->dep[depth].next;
 433 +               q->hash[x] = hash;
 434 +       }
 435 +       __skb_queue_head(&q->qs[x], skb);
 436 +       esfq_inc(q, x);
 437 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 438 +               if (q->tail == depth) { /* It is the first flow */
 439 +                       q->tail = x;
 440 +                       q->next[x] = x;
 441 +                       q->allot[x] = q->quantum;
 442 +               } else {
 443 +                       q->next[x] = q->next[q->tail];
 444 +                       q->next[q->tail] = x;
 445 +                       q->tail = x;
 446 +               }
 447 +       }
 448 +       if (++sch->q.qlen < q->limit - 1) {
 449 +               sch->qstats.requeues++;
 450 +               return 0;
 451 +       }
 452 +
 453 +       sch->qstats.drops++;
 454 +       esfq_drop(sch);
 455 +       return NET_XMIT_CN;
 456 +}
 457 +
 458 +
 459 +
 460 +
 461 +static struct sk_buff *
 462 +esfq_dequeue(struct Qdisc* sch)
 463 +{
 464 +       struct esfq_sched_data *q = qdisc_priv(sch);
 465 +       struct sk_buff *skb;
 466 +       unsigned depth = q->depth;
 467 +       esfq_index a, old_a;
 468 +
 469 +       /* No active slots */
 470 +       if (q->tail == depth)
 471 +               return NULL;
 472 +
 473 +       a = old_a = q->next[q->tail];
 474 +
 475 +       /* Grab packet */
 476 +       skb = __skb_dequeue(&q->qs[a]);
 477 +       esfq_dec(q, a);
 478 +       sch->q.qlen--;
 479 +
 480 +       /* Is the slot empty? */
 481 +       if (q->qs[a].qlen == 0) {
 482 +               q->ht[q->hash[a]] = depth;
 483 +               a = q->next[a];
 484 +               if (a == old_a) {
 485 +                       q->tail = depth;
 486 +                       return skb;
 487 +               }
 488 +               q->next[q->tail] = a;
 489 +               q->allot[a] += q->quantum;
 490 +       } else if ((q->allot[a] -= skb->len) <= 0) {
 491 +               q->tail = a;
 492 +               a = q->next[a];
 493 +               q->allot[a] += q->quantum;
 494 +       }
 495 +
 496 +       return skb;
 497 +}
 498 +
 499 +static void
 500 +esfq_reset(struct Qdisc* sch)
 501 +{
 502 +       struct sk_buff *skb;
 503 +
 504 +       while ((skb = esfq_dequeue(sch)) != NULL)
 505 +               kfree_skb(skb);
 506 +}
 507 +
 508 +static void esfq_perturbation(unsigned long arg)
 509 +{
 510 +       struct Qdisc *sch = (struct Qdisc*)arg;
 511 +       struct esfq_sched_data *q = qdisc_priv(sch);
 512 +
 513 +       q->perturbation = net_random()&0x1F;
 514 +
 515 +       if (q->perturb_period) {
 516 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 517 +               add_timer(&q->perturb_timer);
 518 +       }
 519 +}
 520 +
 521 +static int esfq_change(struct Qdisc *sch, struct rtattr *opt)
 522 +{
 523 +       struct esfq_sched_data *q = qdisc_priv(sch);
 524 +       struct tc_esfq_qopt *ctl = RTA_DATA(opt);
 525 +       int old_perturb = q->perturb_period;
 526 +
 527 +       if (opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 528 +               return -EINVAL;
 529 +
 530 +       sch_tree_lock(sch);
 531 +       q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 532 +       q->perturb_period = ctl->perturb_period*HZ;
 533 +//     q->hash_divisor = ctl->divisor;
 534 +//     q->tail = q->limit = q->depth = ctl->flows;
 535 +
 536 +       if (ctl->limit)
 537 +               q->limit = min_t(u32, ctl->limit, q->depth);
 538 +
 539 +       if (ctl->hash_kind) {
 540 +               q->hash_kind = ctl->hash_kind;
 541 +               if (q->hash_kind !=  TCA_SFQ_HASH_CLASSIC)
 542 +                       q->perturb_period = 0;
 543 +       }
 544 +
 545 +       // is sch_tree_lock enough to do this ?
 546 +       while (sch->q.qlen >= q->limit-1)
 547 +               esfq_drop(sch);
 548 +
 549 +       if (old_perturb)
 550 +               del_timer(&q->perturb_timer);
 551 +       if (q->perturb_period) {
 552 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 553 +               add_timer(&q->perturb_timer);
 554 +       } else {
 555 +               q->perturbation = 0;
 556 +       }
 557 +       sch_tree_unlock(sch);
 558 +       return 0;
 559 +}
 560 +
 561 +static int esfq_init(struct Qdisc *sch, struct rtattr *opt)
 562 +{
 563 +       struct esfq_sched_data *q = qdisc_priv(sch);
 564 +       struct tc_esfq_qopt *ctl;
 565 +       esfq_index p = ~0UL/2;
 566 +       int i;
 567 +
 568 +       if (opt && opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 569 +               return -EINVAL;
 570 +
 571 +       init_timer(&q->perturb_timer);
 572 +       q->perturb_timer.data = (unsigned long)sch;
 573 +       q->perturb_timer.function = esfq_perturbation;
 574 +       q->perturbation = 0;
 575 +       q->hash_kind = TCA_SFQ_HASH_CLASSIC;
 576 +       q->max_depth = 0;
 577 +       q->dyn_min = ~0U; /* maximum value for this type */
 578 +       q->dyn_max = 0;  /* dyn_min/dyn_max will be set properly upon first packet */
 579 +       if (opt == NULL) {
 580 +               q->quantum = psched_mtu(sch->dev);
 581 +               q->perturb_period = 0;
 582 +               q->hash_divisor = 1024;
 583 +               q->tail = q->limit = q->depth = 128;
 584 +
 585 +       } else {
 586 +               ctl = RTA_DATA(opt);
 587 +               q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 588 +               q->perturb_period = ctl->perturb_period*HZ;
 589 +               q->hash_divisor = ctl->divisor ? : 1024;
 590 +               q->tail = q->limit = q->depth = ctl->flows ? : 128;
 591 +
 592 +               if ( q->depth > p - 1 )
 593 +                       return -EINVAL;
 594 +
 595 +               if (ctl->limit)
 596 +                       q->limit = min_t(u32, ctl->limit, q->depth);
 597 +
 598 +               if (ctl->hash_kind) {
 599 +                       q->hash_kind = ctl->hash_kind;
 600 +               }
 601 +
 602 +               if (q->perturb_period) {
 603 +                       q->perturb_timer.expires = jiffies + q->perturb_period;
 604 +                       add_timer(&q->perturb_timer);
 605 +               }
 606 +       }
 607 +
 608 +       q->ht = kmalloc(q->hash_divisor*sizeof(esfq_index), GFP_KERNEL);
 609 +       if (!q->ht)
 610 +               goto err_case;
 611 +
 612 +       q->dep = kmalloc((1+q->depth*2)*sizeof(struct esfq_head), GFP_KERNEL);
 613 +       if (!q->dep)
 614 +               goto err_case;
 615 +       q->next = kmalloc(q->depth*sizeof(esfq_index), GFP_KERNEL);
 616 +       if (!q->next)
 617 +               goto err_case;
 618 +
 619 +       q->allot = kmalloc(q->depth*sizeof(short), GFP_KERNEL);
 620 +       if (!q->allot)
 621 +               goto err_case;
 622 +       q->hash = kmalloc(q->depth*sizeof(unsigned short), GFP_KERNEL);
 623 +       if (!q->hash)
 624 +               goto err_case;
 625 +       q->qs = kmalloc(q->depth*sizeof(struct sk_buff_head), GFP_KERNEL);
 626 +       if (!q->qs)
 627 +               goto err_case;
 628 +
 629 +       for (i=0; i< q->hash_divisor; i++)
 630 +               q->ht[i] = q->depth;
 631 +       for (i=0; i<q->depth; i++) {
 632 +               skb_queue_head_init(&q->qs[i]);
 633 +               q->dep[i+q->depth].next = i+q->depth;
 634 +               q->dep[i+q->depth].prev = i+q->depth;
 635 +       }
 636 +
 637 +       for (i=0; i<q->depth; i++)
 638 +               esfq_link(q, i);
 639 +       return 0;
 640 +err_case:
 641 +       if (q->ht)
 642 +               kfree(q->ht);
 643 +       if (q->dep)
 644 +               kfree(q->dep);
 645 +       if (q->next)
 646 +               kfree(q->next);
 647 +       if (q->allot)
 648 +               kfree(q->allot);
 649 +       if (q->hash)
 650 +               kfree(q->hash);
 651 +       if (q->qs)
 652 +               kfree(q->qs);
 653 +       return -ENOBUFS;
 654 +}
 655 +
 656 +static void esfq_destroy(struct Qdisc *sch)
 657 +{
 658 +       struct esfq_sched_data *q = qdisc_priv(sch);
 659 +       del_timer(&q->perturb_timer);
 660 +       if(q->ht)
 661 +               kfree(q->ht);
 662 +       if(q->dep)
 663 +               kfree(q->dep);
 664 +       if(q->next)
 665 +               kfree(q->next);
 666 +       if(q->allot)
 667 +               kfree(q->allot);
 668 +       if(q->hash)
 669 +               kfree(q->hash);
 670 +       if(q->qs)
 671 +               kfree(q->qs);
 672 +}
 673 +
 674 +static int esfq_dump(struct Qdisc *sch, struct sk_buff *skb)
 675 +{
 676 +       struct esfq_sched_data *q = qdisc_priv(sch);
 677 +       unsigned char    *b = skb->tail;
 678 +       struct tc_esfq_qopt opt;
 679 +
 680 +       opt.quantum = q->quantum;
 681 +       opt.perturb_period = q->perturb_period/HZ;
 682 +
 683 +       opt.limit = q->limit;
 684 +       opt.divisor = q->hash_divisor;
 685 +       opt.flows = q->depth;
 686 +       opt.hash_kind = q->hash_kind;
 687 +
 688 +       RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 689 +
 690 +       return skb->len;
 691 +
 692 +rtattr_failure:
 693 +       skb_trim(skb, b - skb->data);
 694 +       return -1;
 695 +}
 696 +
 697 +static struct Qdisc_ops esfq_qdisc_ops =
 698 +{
 699 +       .next           =       NULL,
 700 +       .cl_ops         =       NULL,
 701 +       .id             =       "esfq",
 702 +       .priv_size      =       sizeof(struct esfq_sched_data),
 703 +       .enqueue        =       esfq_enqueue,
 704 +       .dequeue        =       esfq_dequeue,
 705 +       .requeue        =       esfq_requeue,
 706 +       .drop           =       esfq_drop,
 707 +       .init           =       esfq_init,
 708 +       .reset          =       esfq_reset,
 709 +       .destroy        =       esfq_destroy,
 710 +       .change         =       NULL, /* esfq_change - needs more work */
 711 +       .dump           =       esfq_dump,
 712 +       .owner          =       THIS_MODULE,
 713 +};
 714 +
 715 +static int __init esfq_module_init(void)
 716 +{
 717 +       return register_qdisc(&esfq_qdisc_ops);
 718 +}
 719 +static void __exit esfq_module_exit(void)
 720 +{
 721 +       unregister_qdisc(&esfq_qdisc_ops);
 722 +}
 723 +module_init(esfq_module_init)
 724 +module_exit(esfq_module_exit)
 725 +MODULE_LICENSE("GPL");