target/linux/generic-2.6/patches/200-sched_esfq.patch

   1 diff -urN linux-2.6.19.old/include/linux/pkt_sched.h linux-2.6.19.dev/include/linux/pkt_sched.h
   2 --- linux-2.6.19.old/include/linux/pkt_sched.h  2006-11-29 22:57:37.000000000 +0100
   3 +++ linux-2.6.19.dev/include/linux/pkt_sched.h  2006-12-14 03:13:51.000000000 +0100
   4 @@ -146,8 +146,35 @@
   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 -urN linux-2.6.19.old/net/sched/Kconfig linux-2.6.19.dev/net/sched/Kconfig
  41 --- linux-2.6.19.old/net/sched/Kconfig  2006-11-29 22:57:37.000000000 +0100
  42 +++ linux-2.6.19.dev/net/sched/Kconfig  2006-12-14 03:13:51.000000000 +0100
  43 @@ -185,6 +185,28 @@
  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 "True Link Equalizer (TEQL)"
  71         ---help---
  72 diff -urN linux-2.6.19.old/net/sched/Makefile linux-2.6.19.dev/net/sched/Makefile
  73 --- linux-2.6.19.old/net/sched/Makefile 2006-11-29 22:57:37.000000000 +0100
  74 +++ linux-2.6.19.dev/net/sched/Makefile 2006-12-14 03:13:51.000000000 +0100
  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 -urN linux-2.6.19.old/net/sched/sch_esfq.c linux-2.6.19.dev/net/sched/sch_esfq.c
  84 --- linux-2.6.19.old/net/sched/sch_esfq.c       1970-01-01 01:00:00.000000000 +0100
  85 +++ linux-2.6.19.dev/net/sched/sch_esfq.c       2006-12-14 03:13:51.000000000 +0100
  86 @@ -0,0 +1,644 @@
  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/autoconf.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_SCTP ||
 255 +                    iph->protocol == IPPROTO_DCCP ||
 256 +                    iph->protocol == IPPROTO_ESP))
 257 +                       h2 ^= *(((u32*)iph) + iph->ihl);
 258 +               break;
 259 +       }
 260 +       case __constant_htons(ETH_P_IPV6):
 261 +       {
 262 +               struct ipv6hdr *iph = skb->nh.ipv6h;
 263 +               h = iph->daddr.s6_addr32[3];
 264 +               hs = iph->saddr.s6_addr32[3];
 265 +               nfm = skb->nfmark;
 266 +               h2 = hs^iph->nexthdr;
 267 +               if (iph->nexthdr == IPPROTO_TCP ||
 268 +                   iph->nexthdr == IPPROTO_UDP ||
 269 +                   iph->nexthdr == IPPROTO_SCTP ||
 270 +                   iph->nexthdr == IPPROTO_DCCP ||
 271 +                   iph->nexthdr == IPPROTO_ESP)
 272 +                       h2 ^= *(u32*)&iph[1];
 273 +               break;
 274 +       }
 275 +       default:
 276 +               h = (u32)(unsigned long)skb->dst;
 277 +               hs = (u32)(unsigned long)skb->sk;
 278 +               nfm = skb->nfmark;
 279 +               h2 = hs^skb->protocol;
 280 +       }
 281 +       switch(q->hash_kind)
 282 +       {
 283 +       case TCA_SFQ_HASH_CLASSIC:
 284 +               return esfq_fold_hash_classic(q, h, h2);
 285 +       case TCA_SFQ_HASH_DST:
 286 +               return esfq_hash_u32(q,h);
 287 +       case TCA_SFQ_HASH_DSTDIR:
 288 +               return esfq_hash_direct(q, ntohl(h));
 289 +       case TCA_SFQ_HASH_SRC:
 290 +               return esfq_hash_u32(q,hs);
 291 +       case TCA_SFQ_HASH_SRCDIR:
 292 +               return esfq_hash_direct(q, ntohl(hs));
 293 +#ifdef CONFIG_NETFILTER
 294 +       case TCA_SFQ_HASH_FWMARK:
 295 +               return esfq_hash_u32(q,nfm);
 296 +       case TCA_SFQ_HASH_FWMARKDIR:
 297 +               return esfq_hash_direct(q,nfm);
 298 +#endif
 299 +       default:
 300 +               if (net_ratelimit())
 301 +                       printk(KERN_WARNING "ESFQ: Unknown hash method. Falling back to classic.\n");
 302 +       }
 303 +       return esfq_fold_hash_classic(q, h, h2);
 304 +}
 305 +
 306 +static inline void esfq_link(struct esfq_sched_data *q, esfq_index x)
 307 +{
 308 +       esfq_index p, n;
 309 +       int d = q->qs[x].qlen + q->depth;
 310 +
 311 +       p = d;
 312 +       n = q->dep[d].next;
 313 +       q->dep[x].next = n;
 314 +       q->dep[x].prev = p;
 315 +       q->dep[p].next = q->dep[n].prev = x;
 316 +}
 317 +
 318 +static inline void esfq_dec(struct esfq_sched_data *q, esfq_index x)
 319 +{
 320 +       esfq_index p, n;
 321 +
 322 +       n = q->dep[x].next;
 323 +       p = q->dep[x].prev;
 324 +       q->dep[p].next = n;
 325 +       q->dep[n].prev = p;
 326 +
 327 +       if (n == p && q->max_depth == q->qs[x].qlen + 1)
 328 +               q->max_depth--;
 329 +
 330 +       esfq_link(q, x);
 331 +}
 332 +
 333 +static inline void esfq_inc(struct esfq_sched_data *q, esfq_index x)
 334 +{
 335 +       esfq_index p, n;
 336 +       int d;
 337 +
 338 +       n = q->dep[x].next;
 339 +       p = q->dep[x].prev;
 340 +       q->dep[p].next = n;
 341 +       q->dep[n].prev = p;
 342 +       d = q->qs[x].qlen;
 343 +       if (q->max_depth < d)
 344 +               q->max_depth = d;
 345 +
 346 +       esfq_link(q, x);
 347 +}
 348 +
 349 +static unsigned int esfq_drop(struct Qdisc *sch)
 350 +{
 351 +       struct esfq_sched_data *q = qdisc_priv(sch);
 352 +       esfq_index d = q->max_depth;
 353 +       struct sk_buff *skb;
 354 +       unsigned int len;
 355 +
 356 +       /* Queue is full! Find the longest slot and
 357 +          drop a packet from it */
 358 +
 359 +       if (d > 1) {
 360 +               esfq_index x = q->dep[d+q->depth].next;
 361 +               skb = q->qs[x].prev;
 362 +               len = skb->len;
 363 +               __skb_unlink(skb, &q->qs[x]);
 364 +               kfree_skb(skb);
 365 +               esfq_dec(q, x);
 366 +               sch->q.qlen--;
 367 +               sch->qstats.drops++;
 368 +               return len;
 369 +       }
 370 +
 371 +       if (d == 1) {
 372 +               /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
 373 +               d = q->next[q->tail];
 374 +               q->next[q->tail] = q->next[d];
 375 +               q->allot[q->next[d]] += q->quantum;
 376 +               skb = q->qs[d].prev;
 377 +               len = skb->len;
 378 +               __skb_unlink(skb, &q->qs[d]);
 379 +               kfree_skb(skb);
 380 +               esfq_dec(q, d);
 381 +               sch->q.qlen--;
 382 +               q->ht[q->hash[d]] = q->depth;
 383 +               sch->qstats.drops++;
 384 +               return len;
 385 +       }
 386 +
 387 +       return 0;
 388 +}
 389 +
 390 +static int
 391 +esfq_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 392 +{
 393 +       struct esfq_sched_data *q = qdisc_priv(sch);
 394 +       unsigned hash = esfq_hash(q, skb);
 395 +       unsigned depth = q->depth;
 396 +       esfq_index x;
 397 +
 398 +       x = q->ht[hash];
 399 +       if (x == depth) {
 400 +               q->ht[hash] = x = q->dep[depth].next;
 401 +               q->hash[x] = hash;
 402 +       }
 403 +       __skb_queue_tail(&q->qs[x], skb);
 404 +       esfq_inc(q, x);
 405 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 406 +               if (q->tail == depth) { /* It is the first flow */
 407 +                       q->tail = x;
 408 +                       q->next[x] = x;
 409 +                       q->allot[x] = q->quantum;
 410 +               } else {
 411 +                       q->next[x] = q->next[q->tail];
 412 +                       q->next[q->tail] = x;
 413 +                       q->tail = x;
 414 +               }
 415 +       }
 416 +       if (++sch->q.qlen < q->limit-1) {
 417 +               sch->bstats.bytes += skb->len;
 418 +               sch->bstats.packets++;
 419 +               return 0;
 420 +       }
 421 +
 422 +       esfq_drop(sch);
 423 +       return NET_XMIT_CN;
 424 +}
 425 +
 426 +static int
 427 +esfq_requeue(struct sk_buff *skb, struct Qdisc* sch)
 428 +{
 429 +       struct esfq_sched_data *q = qdisc_priv(sch);
 430 +       unsigned hash = esfq_hash(q, skb);
 431 +       unsigned depth = q->depth;
 432 +       esfq_index x;
 433 +
 434 +       x = q->ht[hash];
 435 +       if (x == depth) {
 436 +               q->ht[hash] = x = q->dep[depth].next;
 437 +               q->hash[x] = hash;
 438 +       }
 439 +       __skb_queue_head(&q->qs[x], skb);
 440 +       esfq_inc(q, x);
 441 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 442 +               if (q->tail == depth) { /* It is the first flow */
 443 +                       q->tail = x;
 444 +                       q->next[x] = x;
 445 +                       q->allot[x] = q->quantum;
 446 +               } else {
 447 +                       q->next[x] = q->next[q->tail];
 448 +                       q->next[q->tail] = x;
 449 +                       q->tail = x;
 450 +               }
 451 +       }
 452 +       if (++sch->q.qlen < q->limit - 1) {
 453 +               sch->qstats.requeues++;
 454 +               return 0;
 455 +       }
 456 +
 457 +       sch->qstats.drops++;
 458 +       esfq_drop(sch);
 459 +       return NET_XMIT_CN;
 460 +}
 461 +
 462 +
 463 +
 464 +
 465 +static struct sk_buff *
 466 +esfq_dequeue(struct Qdisc* sch)
 467 +{
 468 +       struct esfq_sched_data *q = qdisc_priv(sch);
 469 +       struct sk_buff *skb;
 470 +       unsigned depth = q->depth;
 471 +       esfq_index a, old_a;
 472 +
 473 +       /* No active slots */
 474 +       if (q->tail == depth)
 475 +               return NULL;
 476 +
 477 +       a = old_a = q->next[q->tail];
 478 +
 479 +       /* Grab packet */
 480 +       skb = __skb_dequeue(&q->qs[a]);
 481 +       esfq_dec(q, a);
 482 +       sch->q.qlen--;
 483 +
 484 +       /* Is the slot empty? */
 485 +       if (q->qs[a].qlen == 0) {
 486 +               q->ht[q->hash[a]] = depth;
 487 +               a = q->next[a];
 488 +               if (a == old_a) {
 489 +                       q->tail = depth;
 490 +                       return skb;
 491 +               }
 492 +               q->next[q->tail] = a;
 493 +               q->allot[a] += q->quantum;
 494 +       } else if ((q->allot[a] -= skb->len) <= 0) {
 495 +               q->tail = a;
 496 +               a = q->next[a];
 497 +               q->allot[a] += q->quantum;
 498 +       }
 499 +
 500 +       return skb;
 501 +}
 502 +
 503 +static void
 504 +esfq_reset(struct Qdisc* sch)
 505 +{
 506 +       struct sk_buff *skb;
 507 +
 508 +       while ((skb = esfq_dequeue(sch)) != NULL)
 509 +               kfree_skb(skb);
 510 +}
 511 +
 512 +static void esfq_perturbation(unsigned long arg)
 513 +{
 514 +       struct Qdisc *sch = (struct Qdisc*)arg;
 515 +       struct esfq_sched_data *q = qdisc_priv(sch);
 516 +
 517 +       q->perturbation = net_random()&0x1F;
 518 +
 519 +       if (q->perturb_period) {
 520 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 521 +               add_timer(&q->perturb_timer);
 522 +       }
 523 +}
 524 +
 525 +static int esfq_change(struct Qdisc *sch, struct rtattr *opt)
 526 +{
 527 +       struct esfq_sched_data *q = qdisc_priv(sch);
 528 +       struct tc_esfq_qopt *ctl = RTA_DATA(opt);
 529 +       int old_perturb = q->perturb_period;
 530 +
 531 +       if (opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 532 +               return -EINVAL;
 533 +
 534 +       sch_tree_lock(sch);
 535 +       q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 536 +       q->perturb_period = ctl->perturb_period*HZ;
 537 +//     q->hash_divisor = ctl->divisor;
 538 +//     q->tail = q->limit = q->depth = ctl->flows;
 539 +
 540 +       if (ctl->limit)
 541 +               q->limit = min_t(u32, ctl->limit, q->depth);
 542 +
 543 +       if (ctl->hash_kind) {
 544 +               q->hash_kind = ctl->hash_kind;
 545 +               if (q->hash_kind !=  TCA_SFQ_HASH_CLASSIC)
 546 +                       q->perturb_period = 0;
 547 +       }
 548 +
 549 +       // is sch_tree_lock enough to do this ?
 550 +       while (sch->q.qlen >= q->limit-1)
 551 +               esfq_drop(sch);
 552 +
 553 +       if (old_perturb)
 554 +               del_timer(&q->perturb_timer);
 555 +       if (q->perturb_period) {
 556 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 557 +               add_timer(&q->perturb_timer);
 558 +       } else {
 559 +               q->perturbation = 0;
 560 +       }
 561 +       sch_tree_unlock(sch);
 562 +       return 0;
 563 +}
 564 +
 565 +static int esfq_init(struct Qdisc *sch, struct rtattr *opt)
 566 +{
 567 +       struct esfq_sched_data *q = qdisc_priv(sch);
 568 +       struct tc_esfq_qopt *ctl;
 569 +       esfq_index p = ~0UL/2;
 570 +       int i;
 571 +
 572 +       if (opt && opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 573 +               return -EINVAL;
 574 +
 575 +       init_timer(&q->perturb_timer);
 576 +       q->perturb_timer.data = (unsigned long)sch;
 577 +       q->perturb_timer.function = esfq_perturbation;
 578 +       q->perturbation = 0;
 579 +       q->hash_kind = TCA_SFQ_HASH_CLASSIC;
 580 +       q->max_depth = 0;
 581 +       q->dyn_min = ~0U; /* maximum value for this type */
 582 +       q->dyn_max = 0;  /* dyn_min/dyn_max will be set properly upon first packet */
 583 +       if (opt == NULL) {
 584 +               q->quantum = psched_mtu(sch->dev);
 585 +               q->perturb_period = 0;
 586 +               q->hash_divisor = 1024;
 587 +               q->tail = q->limit = q->depth = 128;
 588 +
 589 +       } else {
 590 +               ctl = RTA_DATA(opt);
 591 +               q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 592 +               q->perturb_period = ctl->perturb_period*HZ;
 593 +               q->hash_divisor = ctl->divisor ? : 1024;
 594 +               q->tail = q->limit = q->depth = ctl->flows ? : 128;
 595 +
 596 +               if ( q->depth > p - 1 )
 597 +                       return -EINVAL;
 598 +
 599 +               if (ctl->limit)
 600 +                       q->limit = min_t(u32, ctl->limit, q->depth);
 601 +
 602 +               if (ctl->hash_kind) {
 603 +                       q->hash_kind = ctl->hash_kind;
 604 +               }
 605 +
 606 +               if (q->perturb_period) {
 607 +                       q->perturb_timer.expires = jiffies + q->perturb_period;
 608 +                       add_timer(&q->perturb_timer);
 609 +               }
 610 +       }
 611 +
 612 +       q->ht = kmalloc(q->hash_divisor*sizeof(esfq_index), GFP_KERNEL);
 613 +       if (!q->ht)
 614 +               goto err_case;
 615 +
 616 +       q->dep = kmalloc((1+q->depth*2)*sizeof(struct esfq_head), GFP_KERNEL);
 617 +       if (!q->dep)
 618 +               goto err_case;
 619 +       q->next = kmalloc(q->depth*sizeof(esfq_index), GFP_KERNEL);
 620 +       if (!q->next)
 621 +               goto err_case;
 622 +
 623 +       q->allot = kmalloc(q->depth*sizeof(short), GFP_KERNEL);
 624 +       if (!q->allot)
 625 +               goto err_case;
 626 +       q->hash = kmalloc(q->depth*sizeof(unsigned short), GFP_KERNEL);
 627 +       if (!q->hash)
 628 +               goto err_case;
 629 +       q->qs = kmalloc(q->depth*sizeof(struct sk_buff_head), GFP_KERNEL);
 630 +       if (!q->qs)
 631 +               goto err_case;
 632 +
 633 +       for (i=0; i< q->hash_divisor; i++)
 634 +               q->ht[i] = q->depth;
 635 +       for (i=0; i<q->depth; i++) {
 636 +               skb_queue_head_init(&q->qs[i]);
 637 +               q->dep[i+q->depth].next = i+q->depth;
 638 +               q->dep[i+q->depth].prev = i+q->depth;
 639 +       }
 640 +
 641 +       for (i=0; i<q->depth; i++)
 642 +               esfq_link(q, i);
 643 +       return 0;
 644 +err_case:
 645 +       del_timer(&q->perturb_timer);
 646 +       if (q->ht)
 647 +               kfree(q->ht);
 648 +       if (q->dep)
 649 +               kfree(q->dep);
 650 +       if (q->next)
 651 +               kfree(q->next);
 652 +       if (q->allot)
 653 +               kfree(q->allot);
 654 +       if (q->hash)
 655 +               kfree(q->hash);
 656 +       if (q->qs)
 657 +               kfree(q->qs);
 658 +       return -ENOBUFS;
 659 +}
 660 +
 661 +static void esfq_destroy(struct Qdisc *sch)
 662 +{
 663 +       struct esfq_sched_data *q = qdisc_priv(sch);
 664 +       del_timer(&q->perturb_timer);
 665 +       if(q->ht)
 666 +               kfree(q->ht);
 667 +       if(q->dep)
 668 +               kfree(q->dep);
 669 +       if(q->next)
 670 +               kfree(q->next);
 671 +       if(q->allot)
 672 +               kfree(q->allot);
 673 +       if(q->hash)
 674 +               kfree(q->hash);
 675 +       if(q->qs)
 676 +               kfree(q->qs);
 677 +}
 678 +
 679 +static int esfq_dump(struct Qdisc *sch, struct sk_buff *skb)
 680 +{
 681 +       struct esfq_sched_data *q = qdisc_priv(sch);
 682 +       unsigned char    *b = skb->tail;
 683 +       struct tc_esfq_qopt opt;
 684 +
 685 +       opt.quantum = q->quantum;
 686 +       opt.perturb_period = q->perturb_period/HZ;
 687 +
 688 +       opt.limit = q->limit;
 689 +       opt.divisor = q->hash_divisor;
 690 +       opt.flows = q->depth;
 691 +       opt.hash_kind = q->hash_kind;
 692 +
 693 +       RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 694 +
 695 +       return skb->len;
 696 +
 697 +rtattr_failure:
 698 +       skb_trim(skb, b - skb->data);
 699 +       return -1;
 700 +}
 701 +
 702 +static struct Qdisc_ops esfq_qdisc_ops =
 703 +{
 704 +       .next           =       NULL,
 705 +       .cl_ops         =       NULL,
 706 +       .id             =       "esfq",
 707 +       .priv_size      =       sizeof(struct esfq_sched_data),
 708 +       .enqueue        =       esfq_enqueue,
 709 +       .dequeue        =       esfq_dequeue,
 710 +       .requeue        =       esfq_requeue,
 711 +       .drop           =       esfq_drop,
 712 +       .init           =       esfq_init,
 713 +       .reset          =       esfq_reset,
 714 +       .destroy        =       esfq_destroy,
 715 +       .change         =       NULL, /* esfq_change - needs more work */
 716 +       .dump           =       esfq_dump,
 717 +       .owner          =       THIS_MODULE,
 718 +};
 719 +
 720 +static int __init esfq_module_init(void)
 721 +{
 722 +       return register_qdisc(&esfq_qdisc_ops);
 723 +}
 724 +static void __exit esfq_module_exit(void)
 725 +{
 726 +       unregister_qdisc(&esfq_qdisc_ops);
 727 +}
 728 +module_init(esfq_module_init)
 729 +module_exit(esfq_module_exit)
 730 +MODULE_LICENSE("GPL");