kernel-esfq.patch

   1 diff -Naur linux-2.6.21.5.orig/include/linux/pkt_sched.h linux-2.6.21.5/include/linux/pkt_sched.h
   2 --- linux-2.6.21.5.orig/include/linux/pkt_sched.h       2007-06-11 11:37:06.000000000 -0700
   3 +++ linux-2.6.21.5/include/linux/pkt_sched.h    2007-06-22 22:53:46.000000000 -0700
   4 @@ -146,8 +146,37 @@
   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 +       /* conntrack */
  22 +       TCA_SFQ_HASH_CTORIGDST,
  23 +       TCA_SFQ_HASH_CTORIGSRC,
  24 +       TCA_SFQ_HASH_CTREPLDST,
  25 +       TCA_SFQ_HASH_CTREPLSRC,
  26 +       TCA_SFQ_HASH_CTNATCHG,
  27 +};
  28 +
  29 +struct tc_esfq_qopt
  30 +{
  31 +       unsigned        quantum;        /* Bytes per round allocated to flow */
  32 +       int             perturb_period; /* Period of hash perturbation */
  33 +       __u32           limit;          /* Maximal packets in queue */
  34 +       unsigned        divisor;        /* Hash divisor  */
  35 +       unsigned        flows;          /* Maximal number of flows  */
  36 +       unsigned        hash_kind;      /* Hash function to use for flow identification */
  37 +};
  38 +
  39  /* RED section */
  40
  41  enum
  42 diff -Naur linux-2.6.21.5.orig/net/sched/Kconfig linux-2.6.21.5/net/sched/Kconfig
  43 --- linux-2.6.21.5.orig/net/sched/Kconfig       2007-06-11 11:37:06.000000000 -0700
  44 +++ linux-2.6.21.5/net/sched/Kconfig    2007-06-23 14:11:02.000000000 -0700
  45 @@ -189,6 +189,37 @@
  46           To compile this code as a module, choose M here: the
  47           module will be called sch_sfq.
  48
  49 +config NET_SCH_ESFQ
  50 +       tristate "Enhanced Stochastic Fairness Queueing (ESFQ)"
  51 +       ---help---
  52 +         Say Y here if you want to use the Enhanced Stochastic Fairness
  53 +         Queueing (ESFQ) packet scheduling algorithm for some of your network
  54 +         devices or as a leaf discipline for a classful qdisc such as HTB or
  55 +         CBQ (see the top of <file:net/sched/sch_esfq.c> for details and
  56 +         references to the SFQ algorithm).
  57 +
  58 +         This is an enchanced SFQ version which allows you to control some
  59 +         hardcoded values in the SFQ scheduler.
  60 +
  61 +         ESFQ also adds control of the hash function used to identify packet
  62 +         flows. The original SFQ discipline hashes by connection; ESFQ add
  63 +         several other hashing methods, such as by src IP or by dst IP, which
  64 +         can be more 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_ESFQ_NFCT
  70 +       bool "Connection Tracking Hash Types"
  71 +       depends on NET_SCH_ESFQ && NF_CONNTRACK
  72 +       ---help---
  73 +         Say Y here to enable support for hashing based on netfilter connection
  74 +         tracking information. This is useful for a router that is also using
  75 +         NAT to connect privately-addressed hosts to the Internet. If you want
  76 +         to provide fair distribution of upstream bandwidth, ESFQ must use
  77 +         connection tracking information, since all outgoing packets will share
  78 +         the same source address.
  79 +
  80  config NET_SCH_TEQL
  81         tristate "True Link Equalizer (TEQL)"
  82         ---help---
  83 diff -Naur linux-2.6.21.5.orig/net/sched/Makefile linux-2.6.21.5/net/sched/Makefile
  84 --- linux-2.6.21.5.orig/net/sched/Makefile      2007-06-11 11:37:06.000000000 -0700
  85 +++ linux-2.6.21.5/net/sched/Makefile   2007-06-22 22:53:46.000000000 -0700
  86 @@ -23,6 +23,7 @@
  87  obj-$(CONFIG_NET_SCH_INGRESS)  += sch_ingress.o
  88  obj-$(CONFIG_NET_SCH_DSMARK)   += sch_dsmark.o
  89  obj-$(CONFIG_NET_SCH_SFQ)      += sch_sfq.o
  90 +obj-$(CONFIG_NET_SCH_ESFQ)     += sch_esfq.o
  91  obj-$(CONFIG_NET_SCH_TBF)      += sch_tbf.o
  92  obj-$(CONFIG_NET_SCH_TEQL)     += sch_teql.o
  93  obj-$(CONFIG_NET_SCH_PRIO)     += sch_prio.o
  94 diff -Naur linux-2.6.21.5.orig/net/sched/sch_esfq.c linux-2.6.21.5/net/sched/sch_esfq.c
  95 --- linux-2.6.21.5.orig/net/sched/sch_esfq.c    1969-12-31 16:00:00.000000000 -0800
  96 +++ linux-2.6.21.5/net/sched/sch_esfq.c 2007-06-23 19:18:00.000000000 -0700
  97 @@ -0,0 +1,702 @@
  98 +/*
  99 + * net/sched/sch_esfq.c        Extended Stochastic Fairness Queueing discipline.
 100 + *
 101 + *             This program is free software; you can redistribute it and/or
 102 + *             modify it under the terms of the GNU General Public License
 103 + *             as published by the Free Software Foundation; either version
 104 + *             2 of the License, or (at your option) any later version.
 105 + *
 106 + * Authors:    Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
 107 + *
 108 + * Changes:    Alexander Atanasov, <alex@ssi.bg>
 109 + *             Added dynamic depth,limit,divisor,hash_kind options.
 110 + *             Added dst and src hashes.
 111 + *
 112 + *             Alexander Clouter, <alex@digriz.org.uk>
 113 + *             Ported ESFQ to Linux 2.6.
 114 + *
 115 + *             Corey Hickey, <bugfood-c@fatooh.org>
 116 + *             Maintenance of the Linux 2.6 port.
 117 + *             Added fwmark hash (thanks to Robert Kurjata).
 118 + *             Added usage of jhash.
 119 + *             Added conntrack support.
 120 + *             Added ctnatchg hash (thanks to Ben Pfountz).
 121 + */
 122 +
 123 +#include <linux/module.h>
 124 +#include <asm/uaccess.h>
 125 +#include <asm/system.h>
 126 +#include <linux/bitops.h>
 127 +#include <linux/types.h>
 128 +#include <linux/kernel.h>
 129 +#include <linux/jiffies.h>
 130 +#include <linux/string.h>
 131 +#include <linux/mm.h>
 132 +#include <linux/socket.h>
 133 +#include <linux/sockios.h>
 134 +#include <linux/in.h>
 135 +#include <linux/errno.h>
 136 +#include <linux/interrupt.h>
 137 +#include <linux/if_ether.h>
 138 +#include <linux/inet.h>
 139 +#include <linux/netdevice.h>
 140 +#include <linux/etherdevice.h>
 141 +#include <linux/notifier.h>
 142 +#include <linux/init.h>
 143 +#include <net/ip.h>
 144 +#include <linux/ipv6.h>
 145 +#include <net/route.h>
 146 +#include <linux/skbuff.h>
 147 +#include <net/sock.h>
 148 +#include <net/pkt_sched.h>
 149 +#include <linux/jhash.h>
 150 +#include <net/netfilter/nf_conntrack.h>
 151 +
 152 +/*     Stochastic Fairness Queuing algorithm.
 153 +       For more comments look at sch_sfq.c.
 154 +       The difference is that you can change limit, depth,
 155 +       hash table size and choose alternate hash types.
 156 +
 157 +       classic:        same as in sch_sfq.c
 158 +       dst:            destination IP address
 159 +       src:            source IP address
 160 +       fwmark:         netfilter mark value
 161 +       ctorigdst:      original destination IP address
 162 +       ctorigsrc:      original source IP address
 163 +       ctrepldst:      reply destination IP address
 164 +       ctreplsrc:      reply source IP
 165 +
 166 +*/
 167 +
 168 +#define ESFQ_HEAD 0
 169 +#define ESFQ_TAIL 1
 170 +
 171 +/* This type should contain at least SFQ_DEPTH*2 values */
 172 +typedef unsigned int esfq_index;
 173 +
 174 +struct esfq_head
 175 +{
 176 +       esfq_index      next;
 177 +       esfq_index      prev;
 178 +};
 179 +
 180 +struct esfq_sched_data
 181 +{
 182 +/* Parameters */
 183 +       int             perturb_period;
 184 +       unsigned        quantum;        /* Allotment per round: MUST BE >= MTU */
 185 +       int             limit;
 186 +       unsigned        depth;
 187 +       unsigned        hash_divisor;
 188 +       unsigned        hash_kind;
 189 +/* Variables */
 190 +       struct timer_list perturb_timer;
 191 +       int             perturbation;
 192 +       esfq_index      tail;           /* Index of current slot in round */
 193 +       esfq_index      max_depth;      /* Maximal depth */
 194 +
 195 +       esfq_index      *ht;                    /* Hash table */
 196 +       esfq_index      *next;                  /* Active slots link */
 197 +       short           *allot;                 /* Current allotment per slot */
 198 +       unsigned short  *hash;                  /* Hash value indexed by slots */
 199 +       struct sk_buff_head     *qs;            /* Slot queue */
 200 +       struct esfq_head        *dep;           /* Linked list of slots, indexed by depth */
 201 +};
 202 +
 203 +/* This contains the info we will hash. */
 204 +struct esfq_packet_info
 205 +{
 206 +       u32     proto;          /* protocol or port */
 207 +       u32     src;            /* source from packet header */
 208 +       u32     dst;            /* destination from packet header */
 209 +       u32     ctorigsrc;      /* original source from conntrack */
 210 +       u32     ctorigdst;      /* original destination from conntrack */
 211 +       u32     ctreplsrc;      /* reply source from conntrack */
 212 +       u32     ctrepldst;      /* reply destination from conntrack */
 213 +       u32     mark;           /* netfilter mark (fwmark) */
 214 +};
 215 +
 216 +static __inline__ unsigned esfq_jhash_1word(struct esfq_sched_data *q,u32 a)
 217 +{
 218 +       return jhash_1word(a, q->perturbation) & (q->hash_divisor-1);
 219 +}
 220 +
 221 +static __inline__ unsigned esfq_jhash_2words(struct esfq_sched_data *q, u32 a, u32 b)
 222 +{
 223 +       return jhash_2words(a, b, q->perturbation) & (q->hash_divisor-1);
 224 +}
 225 +
 226 +static __inline__ unsigned esfq_jhash_3words(struct esfq_sched_data *q, u32 a, u32 b, u32 c)
 227 +{
 228 +       return jhash_3words(a, b, c, q->perturbation) & (q->hash_divisor-1);
 229 +}
 230 +
 231 +static unsigned esfq_hash(struct esfq_sched_data *q, struct sk_buff *skb)
 232 +{
 233 +       struct esfq_packet_info info;
 234 +#ifdef CONFIG_NET_SCH_ESFQ_NFCT
 235 +       enum ip_conntrack_info ctinfo;
 236 +       struct nf_conn *ct = nf_ct_get(skb, &ctinfo);
 237 +#endif
 238 +
 239 +       switch (skb->protocol) {
 240 +       case __constant_htons(ETH_P_IP):
 241 +       {
 242 +               struct iphdr *iph = skb->nh.iph;
 243 +               info.dst = iph->daddr;
 244 +               info.src = iph->saddr;
 245 +               if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) &&
 246 +                   (iph->protocol == IPPROTO_TCP ||
 247 +                    iph->protocol == IPPROTO_UDP ||
 248 +                    iph->protocol == IPPROTO_SCTP ||
 249 +                    iph->protocol == IPPROTO_DCCP ||
 250 +                    iph->protocol == IPPROTO_ESP))
 251 +                       info.proto = *(((u32*)iph) + iph->ihl);
 252 +               else
 253 +                       info.proto = iph->protocol;
 254 +               break;
 255 +       }
 256 +       case __constant_htons(ETH_P_IPV6):
 257 +       {
 258 +               struct ipv6hdr *iph = skb->nh.ipv6h;
 259 +               /* Hash ipv6 addresses into a u32. This isn't ideal,
 260 +                * but the code is simple. */
 261 +               info.dst = jhash2(iph->daddr.s6_addr32, 4, q->perturbation);
 262 +               info.src = jhash2(iph->saddr.s6_addr32, 4, q->perturbation);
 263 +               if (iph->nexthdr == IPPROTO_TCP ||
 264 +                   iph->nexthdr == IPPROTO_UDP ||
 265 +                   iph->nexthdr == IPPROTO_SCTP ||
 266 +                   iph->nexthdr == IPPROTO_DCCP ||
 267 +                   iph->nexthdr == IPPROTO_ESP)
 268 +                       info.proto = *(u32*)&iph[1];
 269 +               else
 270 +                       info.proto = iph->nexthdr;
 271 +               break;
 272 +       }
 273 +       default:
 274 +               info.dst   = (u32)(unsigned long)skb->dst;
 275 +               info.src   = (u32)(unsigned long)skb->sk;
 276 +               info.proto = skb->protocol;
 277 +       }
 278 +
 279 +       info.mark = skb->mark;
 280 +
 281 +#ifdef CONFIG_NET_SCH_ESFQ_NFCT
 282 +       /* defaults if there is no conntrack info */
 283 +       info.ctorigsrc = info.src;
 284 +       info.ctorigdst = info.dst;
 285 +       info.ctreplsrc = info.dst;
 286 +       info.ctrepldst = info.src;
 287 +       /* collect conntrack info */
 288 +       if (ct && ct != &nf_conntrack_untracked) {
 289 +               if (skb->protocol == __constant_htons(ETH_P_IP)) {
 290 +                       info.ctorigsrc = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip;
 291 +                       info.ctorigdst = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip;
 292 +                       info.ctreplsrc = ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip;
 293 +                       info.ctrepldst = ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip;
 294 +               }
 295 +               else if (skb->protocol == __constant_htons(ETH_P_IPV6)) {
 296 +                       /* Again, hash ipv6 addresses into a single u32. */
 297 +                       info.ctorigsrc = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip6, 4, q->perturbation);
 298 +                       info.ctorigdst = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip6, 4, q->perturbation);
 299 +                       info.ctreplsrc = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip6, 4, q->perturbation);
 300 +                       info.ctrepldst = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip6, 4, q->perturbation);
 301 +               }
 302 +
 303 +       }
 304 +#endif
 305 +
 306 +       switch(q->hash_kind) {
 307 +       case TCA_SFQ_HASH_CLASSIC:
 308 +               return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 309 +       case TCA_SFQ_HASH_DST:
 310 +               return esfq_jhash_1word(q, info.dst);
 311 +       case TCA_SFQ_HASH_SRC:
 312 +               return esfq_jhash_1word(q, info.src);
 313 +       case TCA_SFQ_HASH_FWMARK:
 314 +               return esfq_jhash_1word(q, info.mark);
 315 +#ifdef CONFIG_NET_SCH_ESFQ_NFCT
 316 +       case TCA_SFQ_HASH_CTORIGDST:
 317 +               return esfq_jhash_1word(q, info.ctorigdst);
 318 +       case TCA_SFQ_HASH_CTORIGSRC:
 319 +               return esfq_jhash_1word(q, info.ctorigsrc);
 320 +       case TCA_SFQ_HASH_CTREPLDST:
 321 +               return esfq_jhash_1word(q, info.ctrepldst);
 322 +       case TCA_SFQ_HASH_CTREPLSRC:
 323 +               return esfq_jhash_1word(q, info.ctreplsrc);
 324 +       case TCA_SFQ_HASH_CTNATCHG:
 325 +       {
 326 +               if (info.ctorigdst == info.ctreplsrc)
 327 +                       return esfq_jhash_1word(q, info.ctorigsrc);
 328 +               return esfq_jhash_1word(q, info.ctreplsrc);
 329 +       }
 330 +#endif
 331 +       default:
 332 +               if (net_ratelimit())
 333 +                       printk(KERN_WARNING "ESFQ: Unknown hash method. Falling back to classic.\n");
 334 +       }
 335 +       return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 336 +}
 337 +
 338 +static inline void esfq_link(struct esfq_sched_data *q, esfq_index x)
 339 +{
 340 +       esfq_index p, n;
 341 +       int d = q->qs[x].qlen + q->depth;
 342 +
 343 +       p = d;
 344 +       n = q->dep[d].next;
 345 +       q->dep[x].next = n;
 346 +       q->dep[x].prev = p;
 347 +       q->dep[p].next = q->dep[n].prev = x;
 348 +}
 349 +
 350 +static inline void esfq_dec(struct esfq_sched_data *q, esfq_index x)
 351 +{
 352 +       esfq_index p, n;
 353 +
 354 +       n = q->dep[x].next;
 355 +       p = q->dep[x].prev;
 356 +       q->dep[p].next = n;
 357 +       q->dep[n].prev = p;
 358 +
 359 +       if (n == p && q->max_depth == q->qs[x].qlen + 1)
 360 +               q->max_depth--;
 361 +
 362 +       esfq_link(q, x);
 363 +}
 364 +
 365 +static inline void esfq_inc(struct esfq_sched_data *q, esfq_index x)
 366 +{
 367 +       esfq_index p, n;
 368 +       int d;
 369 +
 370 +       n = q->dep[x].next;
 371 +       p = q->dep[x].prev;
 372 +       q->dep[p].next = n;
 373 +       q->dep[n].prev = p;
 374 +       d = q->qs[x].qlen;
 375 +       if (q->max_depth < d)
 376 +               q->max_depth = d;
 377 +
 378 +       esfq_link(q, x);
 379 +}
 380 +
 381 +static unsigned int esfq_drop(struct Qdisc *sch)
 382 +{
 383 +       struct esfq_sched_data *q = qdisc_priv(sch);
 384 +       esfq_index d = q->max_depth;
 385 +       struct sk_buff *skb;
 386 +       unsigned int len;
 387 +
 388 +       /* Queue is full! Find the longest slot and
 389 +          drop a packet from it */
 390 +
 391 +       if (d > 1) {
 392 +               esfq_index x = q->dep[d+q->depth].next;
 393 +               skb = q->qs[x].prev;
 394 +               len = skb->len;
 395 +               __skb_unlink(skb, &q->qs[x]);
 396 +               kfree_skb(skb);
 397 +               esfq_dec(q, x);
 398 +               sch->q.qlen--;
 399 +               sch->qstats.drops++;
 400 +               sch->qstats.backlog -= len;
 401 +               return len;
 402 +       }
 403 +
 404 +       if (d == 1) {
 405 +               /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
 406 +               d = q->next[q->tail];
 407 +               q->next[q->tail] = q->next[d];
 408 +               q->allot[q->next[d]] += q->quantum;
 409 +               skb = q->qs[d].prev;
 410 +               len = skb->len;
 411 +               __skb_unlink(skb, &q->qs[d]);
 412 +               kfree_skb(skb);
 413 +               esfq_dec(q, d);
 414 +               sch->q.qlen--;
 415 +               q->ht[q->hash[d]] = q->depth;
 416 +               sch->qstats.drops++;
 417 +               sch->qstats.backlog -= len;
 418 +               return len;
 419 +       }
 420 +
 421 +       return 0;
 422 +}
 423 +
 424 +static void esfq_q_enqueue(struct sk_buff *skb, struct esfq_sched_data *q, unsigned int end)
 425 +{
 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 +
 436 +       if (end == ESFQ_TAIL)
 437 +               __skb_queue_tail(&q->qs[x], skb);
 438 +       else
 439 +               __skb_queue_head(&q->qs[x], skb);
 440 +
 441 +       esfq_inc(q, x);
 442 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 443 +               if (q->tail == depth) { /* It is the first flow */
 444 +                       q->tail = x;
 445 +                       q->next[x] = x;
 446 +                       q->allot[x] = q->quantum;
 447 +               } else {
 448 +                       q->next[x] = q->next[q->tail];
 449 +                       q->next[q->tail] = x;
 450 +                       q->tail = x;
 451 +               }
 452 +       }
 453 +}
 454 +
 455 +static int esfq_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 456 +{
 457 +       struct esfq_sched_data *q = qdisc_priv(sch);
 458 +       esfq_q_enqueue(skb, q, ESFQ_TAIL);
 459 +       sch->qstats.backlog += skb->len;
 460 +       if (++sch->q.qlen < q->limit-1) {
 461 +               sch->bstats.bytes += skb->len;
 462 +               sch->bstats.packets++;
 463 +               return 0;
 464 +       }
 465 +
 466 +       sch->qstats.drops++;
 467 +       esfq_drop(sch);
 468 +       return NET_XMIT_CN;
 469 +}
 470 +
 471 +
 472 +static int esfq_requeue(struct sk_buff *skb, struct Qdisc* sch)
 473 +{
 474 +       struct esfq_sched_data *q = qdisc_priv(sch);
 475 +       esfq_q_enqueue(skb, q, ESFQ_HEAD);
 476 +       sch->qstats.backlog += skb->len;
 477 +       if (++sch->q.qlen < q->limit - 1) {
 478 +               sch->qstats.requeues++;
 479 +               return 0;
 480 +       }
 481 +
 482 +       sch->qstats.drops++;
 483 +       esfq_drop(sch);
 484 +       return NET_XMIT_CN;
 485 +}
 486 +
 487 +static struct sk_buff *esfq_q_dequeue(struct esfq_sched_data *q)
 488 +{
 489 +       struct sk_buff *skb;
 490 +       unsigned depth = q->depth;
 491 +       esfq_index a, old_a;
 492 +
 493 +       /* No active slots */
 494 +       if (q->tail == depth)
 495 +               return NULL;
 496 +
 497 +       a = old_a = q->next[q->tail];
 498 +
 499 +       /* Grab packet */
 500 +       skb = __skb_dequeue(&q->qs[a]);
 501 +       esfq_dec(q, a);
 502 +
 503 +       /* Is the slot empty? */
 504 +       if (q->qs[a].qlen == 0) {
 505 +               q->ht[q->hash[a]] = depth;
 506 +               a = q->next[a];
 507 +               if (a == old_a) {
 508 +                       q->tail = depth;
 509 +                       return skb;
 510 +               }
 511 +               q->next[q->tail] = a;
 512 +               q->allot[a] += q->quantum;
 513 +       } else if ((q->allot[a] -= skb->len) <= 0) {
 514 +               q->tail = a;
 515 +               a = q->next[a];
 516 +               q->allot[a] += q->quantum;
 517 +       }
 518 +
 519 +       return skb;
 520 +}
 521 +
 522 +static struct sk_buff *esfq_dequeue(struct Qdisc* sch)
 523 +{
 524 +       struct esfq_sched_data *q = qdisc_priv(sch);
 525 +       struct sk_buff *skb;
 526 +
 527 +       skb = esfq_q_dequeue(q);
 528 +       if (skb == NULL)
 529 +               return NULL;
 530 +       sch->q.qlen--;
 531 +       sch->qstats.backlog -= skb->len;
 532 +       return skb;
 533 +}
 534 +
 535 +static void esfq_q_destroy(struct esfq_sched_data *q)
 536 +{
 537 +       del_timer(&q->perturb_timer);
 538 +       if(q->ht)
 539 +               kfree(q->ht);
 540 +       if(q->dep)
 541 +               kfree(q->dep);
 542 +       if(q->next)
 543 +               kfree(q->next);
 544 +       if(q->allot)
 545 +               kfree(q->allot);
 546 +       if(q->hash)
 547 +               kfree(q->hash);
 548 +       if(q->qs)
 549 +               kfree(q->qs);
 550 +}
 551 +
 552 +static void esfq_destroy(struct Qdisc *sch)
 553 +{
 554 +       struct esfq_sched_data *q = qdisc_priv(sch);
 555 +       esfq_q_destroy(q);
 556 +}
 557 +
 558 +
 559 +static void esfq_reset(struct Qdisc* sch)
 560 +{
 561 +       struct sk_buff *skb;
 562 +
 563 +       while ((skb = esfq_dequeue(sch)) != NULL)
 564 +               kfree_skb(skb);
 565 +}
 566 +
 567 +static void esfq_perturbation(unsigned long arg)
 568 +{
 569 +       struct Qdisc *sch = (struct Qdisc*)arg;
 570 +       struct esfq_sched_data *q = qdisc_priv(sch);
 571 +
 572 +       q->perturbation = net_random()&0x1F;
 573 +
 574 +       if (q->perturb_period) {
 575 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 576 +               add_timer(&q->perturb_timer);
 577 +       }
 578 +}
 579 +
 580 +static unsigned int esfq_check_hash(unsigned int kind)
 581 +{
 582 +       switch (kind) {
 583 +       case TCA_SFQ_HASH_CTORIGDST:
 584 +       case TCA_SFQ_HASH_CTORIGSRC:
 585 +       case TCA_SFQ_HASH_CTREPLDST:
 586 +       case TCA_SFQ_HASH_CTREPLSRC:
 587 +       case TCA_SFQ_HASH_CTNATCHG:
 588 +#ifndef CONFIG_NET_SCH_ESFQ_NFCT
 589 +       {
 590 +               if (net_ratelimit())
 591 +                       printk(KERN_WARNING "ESFQ: Conntrack hash types disabled in kernel config. Falling back to classic.\n");
 592 +               return TCA_SFQ_HASH_CLASSIC;
 593 +       }
 594 +#endif
 595 +       case TCA_SFQ_HASH_CLASSIC:
 596 +       case TCA_SFQ_HASH_DST:
 597 +       case TCA_SFQ_HASH_SRC:
 598 +       case TCA_SFQ_HASH_FWMARK:
 599 +               return kind;
 600 +       default:
 601 +       {
 602 +               if (net_ratelimit())
 603 +                       printk(KERN_WARNING "ESFQ: Unknown hash type. Falling back to classic.\n");
 604 +               return TCA_SFQ_HASH_CLASSIC;
 605 +       }
 606 +       }
 607 +}
 608 +
 609 +static int esfq_q_init(struct esfq_sched_data *q, struct rtattr *opt)
 610 +{
 611 +       struct tc_esfq_qopt *ctl = RTA_DATA(opt);
 612 +       esfq_index p = ~0U/2;
 613 +       int i;
 614 +
 615 +       if (opt && opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 616 +               return -EINVAL;
 617 +
 618 +       q->perturbation = 0;
 619 +       q->hash_kind = TCA_SFQ_HASH_CLASSIC;
 620 +       q->max_depth = 0;
 621 +       if (opt == NULL) {
 622 +               q->perturb_period = 0;
 623 +               q->hash_divisor = 1024;
 624 +               q->tail = q->limit = q->depth = 128;
 625 +
 626 +       } else {
 627 +               struct tc_esfq_qopt *ctl = RTA_DATA(opt);
 628 +               if (ctl->quantum)
 629 +                       q->quantum = ctl->quantum;
 630 +               q->perturb_period = ctl->perturb_period*HZ;
 631 +               q->hash_divisor = ctl->divisor ? : 1024;
 632 +               q->tail = q->limit = q->depth = ctl->flows ? : 128;
 633 +
 634 +               if ( q->depth > p - 1 )
 635 +                       return -EINVAL;
 636 +
 637 +               if (ctl->limit)
 638 +                       q->limit = min_t(u32, ctl->limit, q->depth);
 639 +
 640 +               if (ctl->hash_kind) {
 641 +                       q->hash_kind = esfq_check_hash(ctl->hash_kind);
 642 +               }
 643 +       }
 644 +
 645 +       q->ht = kmalloc(q->hash_divisor*sizeof(esfq_index), GFP_KERNEL);
 646 +       if (!q->ht)
 647 +               goto err_case;
 648 +       q->dep = kmalloc((1+q->depth*2)*sizeof(struct esfq_head), GFP_KERNEL);
 649 +       if (!q->dep)
 650 +               goto err_case;
 651 +       q->next = kmalloc(q->depth*sizeof(esfq_index), GFP_KERNEL);
 652 +       if (!q->next)
 653 +               goto err_case;
 654 +       q->allot = kmalloc(q->depth*sizeof(short), GFP_KERNEL);
 655 +       if (!q->allot)
 656 +               goto err_case;
 657 +       q->hash = kmalloc(q->depth*sizeof(unsigned short), GFP_KERNEL);
 658 +       if (!q->hash)
 659 +               goto err_case;
 660 +       q->qs = kmalloc(q->depth*sizeof(struct sk_buff_head), GFP_KERNEL);
 661 +       if (!q->qs)
 662 +               goto err_case;
 663 +
 664 +       for (i=0; i< q->hash_divisor; i++)
 665 +               q->ht[i] = q->depth;
 666 +       for (i=0; i<q->depth; i++) {
 667 +               skb_queue_head_init(&q->qs[i]);
 668 +               q->dep[i+q->depth].next = i+q->depth;
 669 +               q->dep[i+q->depth].prev = i+q->depth;
 670 +       }
 671 +
 672 +       for (i=0; i<q->depth; i++)
 673 +               esfq_link(q, i);
 674 +       return 0;
 675 +err_case:
 676 +       esfq_q_destroy(q);
 677 +       return -ENOBUFS;
 678 +}
 679 +
 680 +static int esfq_init(struct Qdisc *sch, struct rtattr *opt)
 681 +{
 682 +       struct esfq_sched_data *q = qdisc_priv(sch);
 683 +       int err;
 684 +
 685 +       q->quantum = psched_mtu(sch->dev); /* default */
 686 +       if ((err = esfq_q_init(q, opt)))
 687 +               return err;
 688 +
 689 +       init_timer(&q->perturb_timer);
 690 +       q->perturb_timer.data = (unsigned long)sch;
 691 +       q->perturb_timer.function = esfq_perturbation;
 692 +       if (q->perturb_period) {
 693 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 694 +               add_timer(&q->perturb_timer);
 695 +       }
 696 +
 697 +       return 0;
 698 +}
 699 +
 700 +static int esfq_change(struct Qdisc *sch, struct rtattr *opt)
 701 +{
 702 +       struct esfq_sched_data *q = qdisc_priv(sch);
 703 +       struct esfq_sched_data new;
 704 +       struct sk_buff *skb;
 705 +       int err;
 706 +
 707 +       /* set up new queue */
 708 +       memset(&new, 0, sizeof(struct esfq_sched_data));
 709 +       new.quantum = psched_mtu(sch->dev); /* default */
 710 +       if ((err = esfq_q_init(&new, opt)))
 711 +               return err;
 712 +
 713 +       /* copy all packets from the old queue to the new queue */
 714 +       sch_tree_lock(sch);
 715 +       while ((skb = esfq_q_dequeue(q)) != NULL)
 716 +               esfq_q_enqueue(skb, &new, ESFQ_TAIL);
 717 +
 718 +       /* clean up the old queue */
 719 +       esfq_q_destroy(q);
 720 +
 721 +       /* copy elements of the new queue into the old queue */
 722 +       q->perturb_period = new.perturb_period;
 723 +       q->quantum        = new.quantum;
 724 +       q->limit          = new.limit;
 725 +       q->depth          = new.depth;
 726 +       q->hash_divisor   = new.hash_divisor;
 727 +       q->hash_kind      = new.hash_kind;
 728 +       q->tail           = new.tail;
 729 +       q->max_depth      = new.max_depth;
 730 +       q->ht    = new.ht;
 731 +       q->dep   = new.dep;
 732 +       q->next  = new.next;
 733 +       q->allot = new.allot;
 734 +       q->hash  = new.hash;
 735 +       q->qs    = new.qs;
 736 +
 737 +       /* finish up */
 738 +       if (q->perturb_period) {
 739 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 740 +               add_timer(&q->perturb_timer);
 741 +       } else {
 742 +               q->perturbation = 0;
 743 +       }
 744 +       sch_tree_unlock(sch);
 745 +       return 0;
 746 +}
 747 +
 748 +static int esfq_dump(struct Qdisc *sch, struct sk_buff *skb)
 749 +{
 750 +       struct esfq_sched_data *q = qdisc_priv(sch);
 751 +       unsigned char *b = skb->tail;
 752 +       struct tc_esfq_qopt opt;
 753 +
 754 +       opt.quantum = q->quantum;
 755 +       opt.perturb_period = q->perturb_period/HZ;
 756 +
 757 +       opt.limit = q->limit;
 758 +       opt.divisor = q->hash_divisor;
 759 +       opt.flows = q->depth;
 760 +       opt.hash_kind = q->hash_kind;
 761 +
 762 +       RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 763 +
 764 +       return skb->len;
 765 +
 766 +rtattr_failure:
 767 +       skb_trim(skb, b - skb->data);
 768 +       return -1;
 769 +}
 770 +
 771 +static struct Qdisc_ops esfq_qdisc_ops =
 772 +{
 773 +       .next           =       NULL,
 774 +       .cl_ops         =       NULL,
 775 +       .id             =       "esfq",
 776 +       .priv_size      =       sizeof(struct esfq_sched_data),
 777 +       .enqueue        =       esfq_enqueue,
 778 +       .dequeue        =       esfq_dequeue,
 779 +       .requeue        =       esfq_requeue,
 780 +       .drop           =       esfq_drop,
 781 +       .init           =       esfq_init,
 782 +       .reset          =       esfq_reset,
 783 +       .destroy        =       esfq_destroy,
 784 +       .change         =       esfq_change,
 785 +       .dump           =       esfq_dump,
 786 +       .owner          =       THIS_MODULE,
 787 +};
 788 +
 789 +static int __init esfq_module_init(void)
 790 +{
 791 +       return register_qdisc(&esfq_qdisc_ops);
 792 +}
 793 +static void __exit esfq_module_exit(void)
 794 +{
 795 +       unregister_qdisc(&esfq_qdisc_ops);
 796 +}
 797 +module_init(esfq_module_init)
 798 +module_exit(esfq_module_exit)
 799 +MODULE_LICENSE("GPL");