target/linux/generic-2.6/patches-2.6.22/200-sched_esfq.patch

   1 --- a/include/linux/pkt_sched.h
   2 +++ b/include/linux/pkt_sched.h
   3 @@ -146,8 +146,40 @@
   4   *
   5   *     The only reason for this is efficiency, it is possible
   6   *     to change these parameters in compile time.
   7 + *
   8 + *     If you need to play with these values use esfq instead.
   9   */
  10
  11 +/* ESFQ section */
  12 +
  13 +enum
  14 +{
  15 +        /* traditional */
  16 +       TCA_SFQ_HASH_CLASSIC,
  17 +       TCA_SFQ_HASH_DST,
  18 +       TCA_SFQ_HASH_SRC,
  19 +       TCA_SFQ_HASH_FWMARK,
  20 +        /* direct */
  21 +       TCA_SFQ_HASH_DSTDIR,
  22 +       TCA_SFQ_HASH_SRCDIR,
  23 +       TCA_SFQ_HASH_FWMARKDIR,
  24 +       /* conntrack */
  25 +       TCA_SFQ_HASH_CTORIGDST,
  26 +       TCA_SFQ_HASH_CTORIGSRC,
  27 +       TCA_SFQ_HASH_CTREPLDST,
  28 +       TCA_SFQ_HASH_CTREPLSRC,
  29 +};
  30 +
  31 +struct tc_esfq_qopt
  32 +{
  33 +       unsigned        quantum;        /* Bytes per round allocated to flow */
  34 +       int             perturb_period; /* Period of hash perturbation */
  35 +       __u32           limit;          /* Maximal packets in queue */
  36 +       unsigned        divisor;        /* Hash divisor  */
  37 +       unsigned        flows;          /* Maximal number of flows  */
  38 +       unsigned        hash_kind;      /* Hash function to use for flow identification */
  39 +};
  40 +
  41  /* RED section */
  42
  43  enum
  44 --- a/net/sched/Kconfig
  45 +++ b/net/sched/Kconfig
  46 @@ -133,6 +133,26 @@
  47           To compile this code as a module, choose M here: the
  48           module will be called sch_sfq.
  49
  50 +config NET_SCH_ESFQ
  51 +       tristate "Enhanced Stochastic Fairness Queueing (ESFQ)"
  52 +       ---help---
  53 +         Say Y here if you want to use the Enhanced Stochastic Fairness
  54 +         Queueing (ESFQ) packet scheduling algorithm for some of your network
  55 +         devices or as a leaf discipline for a classful qdisc such as HTB or
  56 +         CBQ (see the top of <file:net/sched/sch_esfq.c> for details and
  57 +         references to the SFQ algorithm).
  58 +
  59 +         This is an enchanced SFQ version which allows you to control some
  60 +         hardcoded values in the SFQ scheduler.
  61 +
  62 +         ESFQ also adds control of the hash function used to identify packet
  63 +         flows. The original SFQ discipline hashes by connection; ESFQ add
  64 +         several other hashing methods, such as by src IP or by dst IP, which
  65 +         can be more fair to users in some networking situations.
  66 +
  67 +         To compile this code as a module, choose M here: the
  68 +         module will be called sch_esfq.
  69 +
  70  config NET_SCH_TEQL
  71         tristate "True Link Equalizer (TEQL)"
  72         ---help---
  73 --- a/net/sched/Makefile
  74 +++ b/net/sched/Makefile
  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 --- /dev/null
  84 +++ b/net/sched/sch_esfq.c
  85 @@ -0,0 +1,704 @@
  86 +/*
  87 + * net/sched/sch_esfq.c        Extended Stochastic Fairness Queueing discipline.
  88 + *
  89 + *             This program is free software; you can redistribute it and/or
  90 + *             modify it under the terms of the GNU General Public License
  91 + *             as published by the Free Software Foundation; either version
  92 + *             2 of the License, or (at your option) any later version.
  93 + *
  94 + * Authors:    Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  95 + *
  96 + * Changes:    Alexander Atanasov, <alex@ssi.bg>
  97 + *             Added dynamic depth,limit,divisor,hash_kind options.
  98 + *             Added dst and src hashes.
  99 + *
 100 + *             Alexander Clouter, <alex@digriz.org.uk>
 101 + *             Ported ESFQ to Linux 2.6.
 102 + *
 103 + *             Corey Hickey, <bugfood-c@fatooh.org>
 104 + *             Maintenance of the Linux 2.6 port.
 105 + *             Added fwmark hash (thanks to Robert Kurjata).
 106 + *             Added direct hashing for src, dst, and fwmark.
 107 + *             Added usage of jhash.
 108 + *
 109 + */
 110 +
 111 +#include <linux/module.h>
 112 +#include <asm/uaccess.h>
 113 +#include <asm/system.h>
 114 +#include <linux/bitops.h>
 115 +#include <linux/types.h>
 116 +#include <linux/kernel.h>
 117 +#include <linux/jiffies.h>
 118 +#include <linux/string.h>
 119 +#include <linux/mm.h>
 120 +#include <linux/socket.h>
 121 +#include <linux/sockios.h>
 122 +#include <linux/in.h>
 123 +#include <linux/errno.h>
 124 +#include <linux/interrupt.h>
 125 +#include <linux/if_ether.h>
 126 +#include <linux/inet.h>
 127 +#include <linux/netdevice.h>
 128 +#include <linux/etherdevice.h>
 129 +#include <linux/notifier.h>
 130 +#include <linux/init.h>
 131 +#include <net/ip.h>
 132 +#include <linux/ipv6.h>
 133 +#include <net/route.h>
 134 +#include <linux/skbuff.h>
 135 +#include <net/sock.h>
 136 +#include <net/pkt_sched.h>
 137 +#include <linux/jhash.h>
 138 +
 139 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 140 +#include <net/netfilter/nf_conntrack.h>
 141 +#endif
 142 +
 143 +/*     Stochastic Fairness Queuing algorithm.
 144 +       For more comments look at sch_sfq.c.
 145 +       The difference is that you can change limit, depth,
 146 +       hash table size and choose alternate hash types.
 147 +
 148 +       classic:        same as in sch_sfq.c
 149 +       dst:            destination IP address
 150 +       src:            source IP address
 151 +       fwmark:         netfilter mark value
 152 +       dst_direct:
 153 +       src_direct:
 154 +       fwmark_direct:  direct hashing of the above sources
 155 +       ctorigdst:      original destination IP address
 156 +       ctorigsrc:      original source IP address
 157 +       ctrepldst:      reply destination IP address
 158 +       ctreplsrc:      reply source IP
 159 +
 160 +*/
 161 +
 162 +
 163 +/* This type should contain at least SFQ_DEPTH*2 values */
 164 +typedef unsigned int esfq_index;
 165 +
 166 +struct esfq_head
 167 +{
 168 +       esfq_index      next;
 169 +       esfq_index      prev;
 170 +};
 171 +
 172 +struct esfq_sched_data
 173 +{
 174 +/* Parameters */
 175 +       int             perturb_period;
 176 +       unsigned        quantum;        /* Allotment per round: MUST BE >= MTU */
 177 +       int             limit;
 178 +       unsigned        depth;
 179 +       unsigned        hash_divisor;
 180 +       unsigned        hash_kind;
 181 +/* Variables */
 182 +       struct timer_list perturb_timer;
 183 +       int             perturbation;
 184 +       esfq_index      tail;           /* Index of current slot in round */
 185 +       esfq_index      max_depth;      /* Maximal depth */
 186 +
 187 +       esfq_index      *ht;                    /* Hash table */
 188 +       esfq_index      *next;                  /* Active slots link */
 189 +       short           *allot;                 /* Current allotment per slot */
 190 +       unsigned short  *hash;                  /* Hash value indexed by slots */
 191 +       struct sk_buff_head     *qs;            /* Slot queue */
 192 +       struct esfq_head        *dep;           /* Linked list of slots, indexed by depth */
 193 +       unsigned        dyn_min;        /* For dynamic divisor adjustment; minimum value seen */
 194 +       unsigned        dyn_max;        /*                                 maximum value seen */
 195 +       unsigned        dyn_range;      /*                                 saved range */
 196 +};
 197 +
 198 +/* This contains the info we will hash. */
 199 +struct esfq_packet_info
 200 +{
 201 +       u32     proto;          /* protocol or port */
 202 +       u32     src;            /* source from packet header */
 203 +       u32     dst;            /* destination from packet header */
 204 +       u32     ctorigsrc;      /* original source from conntrack */
 205 +       u32     ctorigdst;      /* original destination from conntrack */
 206 +       u32     ctreplsrc;      /* reply source from conntrack */
 207 +       u32     ctrepldst;      /* reply destination from conntrack */
 208 +       u32     mark;           /* netfilter mark (fwmark) */
 209 +};
 210 +
 211 +/* Hash input values directly into the "nearest" slot, taking into account the
 212 + * range of input values seen. This is most useful when the hash table is at
 213 + * least as large as the range of possible values.
 214 + * Note: this functionality was added before the change to using jhash, and may
 215 + * no longer be useful. */
 216 +static __inline__ unsigned esfq_hash_direct(struct esfq_sched_data *q, u32 h)
 217 +{
 218 +       /* adjust minimum and maximum */
 219 +       if (h < q->dyn_min || h > q->dyn_max) {
 220 +               q->dyn_min = h < q->dyn_min ? h : q->dyn_min;
 221 +               q->dyn_max = h > q->dyn_max ? h : q->dyn_max;
 222 +
 223 +               /* find new range */
 224 +               if ((q->dyn_range = q->dyn_max - q->dyn_min) >= q->hash_divisor)
 225 +                       printk(KERN_WARNING "ESFQ: (direct hash) Input range %u is larger than hash "
 226 +                                       "table. See ESFQ README for details.\n", q->dyn_range);
 227 +       }
 228 +
 229 +       /* hash input values into slot numbers */
 230 +       if (q->dyn_min == q->dyn_max)
 231 +               return 0; /* only one value seen; avoid division by 0 */
 232 +       else
 233 +               return (h - q->dyn_min) * (q->hash_divisor - 1) / q->dyn_range;
 234 +}
 235 +
 236 +static __inline__ unsigned esfq_jhash_1word(struct esfq_sched_data *q,u32 a)
 237 +{
 238 +       return jhash_1word(a, q->perturbation) & (q->hash_divisor-1);
 239 +}
 240 +
 241 +static __inline__ unsigned esfq_jhash_2words(struct esfq_sched_data *q, u32 a, u32 b)
 242 +{
 243 +       return jhash_2words(a, b, q->perturbation) & (q->hash_divisor-1);
 244 +}
 245 +
 246 +static __inline__ unsigned esfq_jhash_3words(struct esfq_sched_data *q, u32 a, u32 b, u32 c)
 247 +{
 248 +       return jhash_3words(a, b, c, q->perturbation) & (q->hash_divisor-1);
 249 +}
 250 +
 251 +
 252 +static unsigned esfq_hash(struct esfq_sched_data *q, struct sk_buff *skb)
 253 +{
 254 +       struct esfq_packet_info info;
 255 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 256 +       enum ip_conntrack_info ctinfo;
 257 +       struct nf_conn *ct = nf_ct_get(skb, &ctinfo);
 258 +#endif
 259 +
 260 +       switch (skb->protocol) {
 261 +       case __constant_htons(ETH_P_IP):
 262 +       {
 263 +               struct iphdr *iph = ip_hdr(skb);
 264 +               info.dst = iph->daddr;
 265 +               info.src = iph->saddr;
 266 +               if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) &&
 267 +                   (iph->protocol == IPPROTO_TCP ||
 268 +                    iph->protocol == IPPROTO_UDP ||
 269 +                    iph->protocol == IPPROTO_SCTP ||
 270 +                    iph->protocol == IPPROTO_DCCP ||
 271 +                    iph->protocol == IPPROTO_ESP))
 272 +                       info.proto = *(((u32*)iph) + iph->ihl);
 273 +               else
 274 +                       info.proto = iph->protocol;
 275 +               break;
 276 +       }
 277 +       case __constant_htons(ETH_P_IPV6):
 278 +       {
 279 +               struct ipv6hdr *iph = ipv6_hdr(skb);
 280 +               /* Hash ipv6 addresses into a u32. This isn't ideal,
 281 +                * but the code is simple. */
 282 +               info.dst = jhash2(iph->daddr.s6_addr32, 4, q->perturbation);
 283 +               info.src = jhash2(iph->saddr.s6_addr32, 4, q->perturbation);
 284 +               if (iph->nexthdr == IPPROTO_TCP ||
 285 +                   iph->nexthdr == IPPROTO_UDP ||
 286 +                   iph->nexthdr == IPPROTO_SCTP ||
 287 +                   iph->nexthdr == IPPROTO_DCCP ||
 288 +                   iph->nexthdr == IPPROTO_ESP)
 289 +                       info.proto = *(u32*)&iph[1];
 290 +               else
 291 +                       info.proto = iph->nexthdr;
 292 +               break;
 293 +       }
 294 +       default:
 295 +               info.dst   = (u32)(unsigned long)skb->dst;
 296 +               info.src   = (u32)(unsigned long)skb->sk;
 297 +               info.proto = skb->protocol;
 298 +       }
 299 +
 300 +       info.mark = skb->mark;
 301 +
 302 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 303 +       /* defaults if there is no conntrack info */
 304 +       info.ctorigsrc = info.src;
 305 +       info.ctorigdst = info.dst;
 306 +       info.ctreplsrc = info.dst;
 307 +       info.ctrepldst = info.src;
 308 +       /* collect conntrack info */
 309 +       if (ct && ct != &nf_conntrack_untracked) {
 310 +               if (skb->protocol == __constant_htons(ETH_P_IP)) {
 311 +                       info.ctorigsrc = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip;
 312 +                       info.ctorigdst = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip;
 313 +                       info.ctreplsrc = ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip;
 314 +                       info.ctrepldst = ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip;
 315 +               }
 316 +               else if (skb->protocol == __constant_htons(ETH_P_IPV6)) {
 317 +                       /* Again, hash ipv6 addresses into a single u32. */
 318 +                       info.ctorigsrc = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip6, 4, q->perturbation);
 319 +                       info.ctorigdst = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip6, 4, q->perturbation);
 320 +                       info.ctreplsrc = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip6, 4, q->perturbation);
 321 +                       info.ctrepldst = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip6, 4, q->perturbation);
 322 +               }
 323 +
 324 +       }
 325 +#endif
 326 +
 327 +       switch(q->hash_kind)
 328 +       {
 329 +       case TCA_SFQ_HASH_CLASSIC:
 330 +               return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 331 +       case TCA_SFQ_HASH_DST:
 332 +               return esfq_jhash_1word(q, info.dst);
 333 +       case TCA_SFQ_HASH_DSTDIR:
 334 +               return esfq_hash_direct(q, ntohl(info.dst));
 335 +       case TCA_SFQ_HASH_SRC:
 336 +               return esfq_jhash_1word(q, info.src);
 337 +       case TCA_SFQ_HASH_SRCDIR:
 338 +               return esfq_hash_direct(q, ntohl(info.src));
 339 +       case TCA_SFQ_HASH_FWMARK:
 340 +               return esfq_jhash_1word(q, info.mark);
 341 +       case TCA_SFQ_HASH_FWMARKDIR:
 342 +               return esfq_hash_direct(q, info.mark);
 343 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 344 +       case TCA_SFQ_HASH_CTORIGDST:
 345 +               return esfq_jhash_1word(q, info.ctorigdst);
 346 +       case TCA_SFQ_HASH_CTORIGSRC:
 347 +               return esfq_jhash_1word(q, info.ctorigsrc);
 348 +       case TCA_SFQ_HASH_CTREPLDST:
 349 +               return esfq_jhash_1word(q, info.ctrepldst);
 350 +       case TCA_SFQ_HASH_CTREPLSRC:
 351 +               return esfq_jhash_1word(q, info.ctreplsrc);
 352 +#endif
 353 +       default:
 354 +               if (net_ratelimit())
 355 +                       printk(KERN_WARNING "ESFQ: Unknown hash method. Falling back to classic.\n");
 356 +       }
 357 +       return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 358 +}
 359 +
 360 +static inline void esfq_link(struct esfq_sched_data *q, esfq_index x)
 361 +{
 362 +       esfq_index p, n;
 363 +       int d = q->qs[x].qlen + q->depth;
 364 +
 365 +       p = d;
 366 +       n = q->dep[d].next;
 367 +       q->dep[x].next = n;
 368 +       q->dep[x].prev = p;
 369 +       q->dep[p].next = q->dep[n].prev = x;
 370 +}
 371 +
 372 +static inline void esfq_dec(struct esfq_sched_data *q, esfq_index x)
 373 +{
 374 +       esfq_index p, n;
 375 +
 376 +       n = q->dep[x].next;
 377 +       p = q->dep[x].prev;
 378 +       q->dep[p].next = n;
 379 +       q->dep[n].prev = p;
 380 +
 381 +       if (n == p && q->max_depth == q->qs[x].qlen + 1)
 382 +               q->max_depth--;
 383 +
 384 +       esfq_link(q, x);
 385 +}
 386 +
 387 +static inline void esfq_inc(struct esfq_sched_data *q, esfq_index x)
 388 +{
 389 +       esfq_index p, n;
 390 +       int d;
 391 +
 392 +       n = q->dep[x].next;
 393 +       p = q->dep[x].prev;
 394 +       q->dep[p].next = n;
 395 +       q->dep[n].prev = p;
 396 +       d = q->qs[x].qlen;
 397 +       if (q->max_depth < d)
 398 +               q->max_depth = d;
 399 +
 400 +       esfq_link(q, x);
 401 +}
 402 +
 403 +static unsigned int esfq_drop(struct Qdisc *sch)
 404 +{
 405 +       struct esfq_sched_data *q = qdisc_priv(sch);
 406 +       esfq_index d = q->max_depth;
 407 +       struct sk_buff *skb;
 408 +       unsigned int len;
 409 +
 410 +       /* Queue is full! Find the longest slot and
 411 +          drop a packet from it */
 412 +
 413 +       if (d > 1) {
 414 +               esfq_index x = q->dep[d+q->depth].next;
 415 +               skb = q->qs[x].prev;
 416 +               len = skb->len;
 417 +               __skb_unlink(skb, &q->qs[x]);
 418 +               kfree_skb(skb);
 419 +               esfq_dec(q, x);
 420 +               sch->q.qlen--;
 421 +               sch->qstats.drops++;
 422 +               sch->qstats.backlog -= len;
 423 +               return len;
 424 +       }
 425 +
 426 +       if (d == 1) {
 427 +               /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
 428 +               d = q->next[q->tail];
 429 +               q->next[q->tail] = q->next[d];
 430 +               q->allot[q->next[d]] += q->quantum;
 431 +               skb = q->qs[d].prev;
 432 +               len = skb->len;
 433 +               __skb_unlink(skb, &q->qs[d]);
 434 +               kfree_skb(skb);
 435 +               esfq_dec(q, d);
 436 +               sch->q.qlen--;
 437 +               q->ht[q->hash[d]] = q->depth;
 438 +               sch->qstats.drops++;
 439 +               sch->qstats.backlog -= len;
 440 +               return len;
 441 +       }
 442 +
 443 +       return 0;
 444 +}
 445 +
 446 +static int
 447 +esfq_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 448 +{
 449 +       struct esfq_sched_data *q = qdisc_priv(sch);
 450 +       unsigned hash = esfq_hash(q, skb);
 451 +       unsigned depth = q->depth;
 452 +       esfq_index x;
 453 +
 454 +       x = q->ht[hash];
 455 +       if (x == depth) {
 456 +               q->ht[hash] = x = q->dep[depth].next;
 457 +               q->hash[x] = hash;
 458 +       }
 459 +       sch->qstats.backlog += skb->len;
 460 +       __skb_queue_tail(&q->qs[x], skb);
 461 +       esfq_inc(q, x);
 462 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 463 +               if (q->tail == depth) { /* It is the first flow */
 464 +                       q->tail = x;
 465 +                       q->next[x] = x;
 466 +                       q->allot[x] = q->quantum;
 467 +               } else {
 468 +                       q->next[x] = q->next[q->tail];
 469 +                       q->next[q->tail] = x;
 470 +                       q->tail = x;
 471 +               }
 472 +       }
 473 +       if (++sch->q.qlen < q->limit-1) {
 474 +               sch->bstats.bytes += skb->len;
 475 +               sch->bstats.packets++;
 476 +               return 0;
 477 +       }
 478 +
 479 +       esfq_drop(sch);
 480 +       return NET_XMIT_CN;
 481 +}
 482 +
 483 +static int
 484 +esfq_requeue(struct sk_buff *skb, struct Qdisc* sch)
 485 +{
 486 +       struct esfq_sched_data *q = qdisc_priv(sch);
 487 +       unsigned hash = esfq_hash(q, skb);
 488 +       unsigned depth = q->depth;
 489 +       esfq_index x;
 490 +
 491 +       x = q->ht[hash];
 492 +       if (x == depth) {
 493 +               q->ht[hash] = x = q->dep[depth].next;
 494 +               q->hash[x] = hash;
 495 +       }
 496 +       sch->qstats.backlog += skb->len;
 497 +       __skb_queue_head(&q->qs[x], skb);
 498 +       esfq_inc(q, x);
 499 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 500 +               if (q->tail == depth) { /* It is the first flow */
 501 +                       q->tail = x;
 502 +                       q->next[x] = x;
 503 +                       q->allot[x] = q->quantum;
 504 +               } else {
 505 +                       q->next[x] = q->next[q->tail];
 506 +                       q->next[q->tail] = x;
 507 +                       q->tail = x;
 508 +               }
 509 +       }
 510 +       if (++sch->q.qlen < q->limit - 1) {
 511 +               sch->qstats.requeues++;
 512 +               return 0;
 513 +       }
 514 +
 515 +       sch->qstats.drops++;
 516 +       esfq_drop(sch);
 517 +       return NET_XMIT_CN;
 518 +}
 519 +
 520 +
 521 +
 522 +
 523 +static struct sk_buff *
 524 +esfq_dequeue(struct Qdisc* sch)
 525 +{
 526 +       struct esfq_sched_data *q = qdisc_priv(sch);
 527 +       struct sk_buff *skb;
 528 +       unsigned depth = q->depth;
 529 +       esfq_index a, old_a;
 530 +
 531 +       /* No active slots */
 532 +       if (q->tail == depth)
 533 +               return NULL;
 534 +
 535 +       a = old_a = q->next[q->tail];
 536 +
 537 +       /* Grab packet */
 538 +       skb = __skb_dequeue(&q->qs[a]);
 539 +       esfq_dec(q, a);
 540 +       sch->q.qlen--;
 541 +       sch->qstats.backlog -= skb->len;
 542 +
 543 +       /* Is the slot empty? */
 544 +       if (q->qs[a].qlen == 0) {
 545 +               q->ht[q->hash[a]] = depth;
 546 +               a = q->next[a];
 547 +               if (a == old_a) {
 548 +                       q->tail = depth;
 549 +                       return skb;
 550 +               }
 551 +               q->next[q->tail] = a;
 552 +               q->allot[a] += q->quantum;
 553 +       } else if ((q->allot[a] -= skb->len) <= 0) {
 554 +               q->tail = a;
 555 +               a = q->next[a];
 556 +               q->allot[a] += q->quantum;
 557 +       }
 558 +
 559 +       return skb;
 560 +}
 561 +
 562 +static void
 563 +esfq_reset(struct Qdisc* sch)
 564 +{
 565 +       struct sk_buff *skb;
 566 +
 567 +       while ((skb = esfq_dequeue(sch)) != NULL)
 568 +               kfree_skb(skb);
 569 +}
 570 +
 571 +static void esfq_perturbation(unsigned long arg)
 572 +{
 573 +       struct Qdisc *sch = (struct Qdisc*)arg;
 574 +       struct esfq_sched_data *q = qdisc_priv(sch);
 575 +
 576 +       q->perturbation = net_random()&0x1F;
 577 +
 578 +       if (q->perturb_period) {
 579 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 580 +               add_timer(&q->perturb_timer);
 581 +       }
 582 +}
 583 +
 584 +static int esfq_change(struct Qdisc *sch, struct rtattr *opt)
 585 +{
 586 +       struct esfq_sched_data *q = qdisc_priv(sch);
 587 +       struct tc_esfq_qopt *ctl = RTA_DATA(opt);
 588 +       int old_perturb = q->perturb_period;
 589 +
 590 +       if (opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 591 +               return -EINVAL;
 592 +
 593 +       sch_tree_lock(sch);
 594 +       q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 595 +       q->perturb_period = ctl->perturb_period*HZ;
 596 +//     q->hash_divisor = ctl->divisor;
 597 +//     q->tail = q->limit = q->depth = ctl->flows;
 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 +               if (q->hash_kind != TCA_SFQ_HASH_CLASSIC)
 605 +                       q->perturb_period = 0;
 606 +       }
 607 +
 608 +       // is sch_tree_lock enough to do this ?
 609 +       while (sch->q.qlen >= q->limit-1)
 610 +               esfq_drop(sch);
 611 +
 612 +       if (old_perturb)
 613 +               del_timer(&q->perturb_timer);
 614 +       if (q->perturb_period) {
 615 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 616 +               add_timer(&q->perturb_timer);
 617 +       } else {
 618 +               q->perturbation = 0;
 619 +       }
 620 +       sch_tree_unlock(sch);
 621 +       return 0;
 622 +}
 623 +
 624 +static int esfq_init(struct Qdisc *sch, struct rtattr *opt)
 625 +{
 626 +       struct esfq_sched_data *q = qdisc_priv(sch);
 627 +       struct tc_esfq_qopt *ctl;
 628 +       esfq_index p = ~0U/2;
 629 +       int i;
 630 +
 631 +       if (opt && opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 632 +               return -EINVAL;
 633 +
 634 +       init_timer(&q->perturb_timer);
 635 +       q->perturb_timer.data = (unsigned long)sch;
 636 +       q->perturb_timer.function = esfq_perturbation;
 637 +       q->perturbation = 0;
 638 +       q->hash_kind = TCA_SFQ_HASH_CLASSIC;
 639 +       q->max_depth = 0;
 640 +       q->dyn_min = ~0U; /* maximum value for this type */
 641 +       q->dyn_max = 0;  /* dyn_min/dyn_max will be set properly upon first packet */
 642 +       if (opt == NULL) {
 643 +               q->quantum = psched_mtu(sch->dev);
 644 +               q->perturb_period = 0;
 645 +               q->hash_divisor = 1024;
 646 +               q->tail = q->limit = q->depth = 128;
 647 +
 648 +       } else {
 649 +               ctl = RTA_DATA(opt);
 650 +               q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 651 +               q->perturb_period = ctl->perturb_period*HZ;
 652 +               q->hash_divisor = ctl->divisor ? : 1024;
 653 +               q->tail = q->limit = q->depth = ctl->flows ? : 128;
 654 +
 655 +               if ( q->depth > p - 1 )
 656 +                       return -EINVAL;
 657 +
 658 +               if (ctl->limit)
 659 +                       q->limit = min_t(u32, ctl->limit, q->depth);
 660 +
 661 +               if (ctl->hash_kind) {
 662 +                       q->hash_kind = ctl->hash_kind;
 663 +               }
 664 +
 665 +               if (q->perturb_period) {
 666 +                       q->perturb_timer.expires = jiffies + q->perturb_period;
 667 +                       add_timer(&q->perturb_timer);
 668 +               }
 669 +       }
 670 +
 671 +       q->ht = kmalloc(q->hash_divisor*sizeof(esfq_index), GFP_KERNEL);
 672 +       if (!q->ht)
 673 +               goto err_case;
 674 +
 675 +       q->dep = kmalloc((1+q->depth*2)*sizeof(struct esfq_head), GFP_KERNEL);
 676 +       if (!q->dep)
 677 +               goto err_case;
 678 +       q->next = kmalloc(q->depth*sizeof(esfq_index), GFP_KERNEL);
 679 +       if (!q->next)
 680 +               goto err_case;
 681 +
 682 +       q->allot = kmalloc(q->depth*sizeof(short), GFP_KERNEL);
 683 +       if (!q->allot)
 684 +               goto err_case;
 685 +       q->hash = kmalloc(q->depth*sizeof(unsigned short), GFP_KERNEL);
 686 +       if (!q->hash)
 687 +               goto err_case;
 688 +       q->qs = kmalloc(q->depth*sizeof(struct sk_buff_head), GFP_KERNEL);
 689 +       if (!q->qs)
 690 +               goto err_case;
 691 +
 692 +       for (i=0; i< q->hash_divisor; i++)
 693 +               q->ht[i] = q->depth;
 694 +       for (i=0; i<q->depth; i++) {
 695 +               skb_queue_head_init(&q->qs[i]);
 696 +               q->dep[i+q->depth].next = i+q->depth;
 697 +               q->dep[i+q->depth].prev = i+q->depth;
 698 +       }
 699 +
 700 +       for (i=0; i<q->depth; i++)
 701 +               esfq_link(q, i);
 702 +       return 0;
 703 +err_case:
 704 +       del_timer(&q->perturb_timer);
 705 +       if (q->ht)
 706 +               kfree(q->ht);
 707 +       if (q->dep)
 708 +               kfree(q->dep);
 709 +       if (q->next)
 710 +               kfree(q->next);
 711 +       if (q->allot)
 712 +               kfree(q->allot);
 713 +       if (q->hash)
 714 +               kfree(q->hash);
 715 +       if (q->qs)
 716 +               kfree(q->qs);
 717 +       return -ENOBUFS;
 718 +}
 719 +
 720 +static void esfq_destroy(struct Qdisc *sch)
 721 +{
 722 +       struct esfq_sched_data *q = qdisc_priv(sch);
 723 +       del_timer(&q->perturb_timer);
 724 +       if(q->ht)
 725 +               kfree(q->ht);
 726 +       if(q->dep)
 727 +               kfree(q->dep);
 728 +       if(q->next)
 729 +               kfree(q->next);
 730 +       if(q->allot)
 731 +               kfree(q->allot);
 732 +       if(q->hash)
 733 +               kfree(q->hash);
 734 +       if(q->qs)
 735 +               kfree(q->qs);
 736 +}
 737 +
 738 +static int esfq_dump(struct Qdisc *sch, struct sk_buff *skb)
 739 +{
 740 +       struct esfq_sched_data *q = qdisc_priv(sch);
 741 +       unsigned char    *b = skb->tail;
 742 +       struct tc_esfq_qopt opt;
 743 +
 744 +       opt.quantum = q->quantum;
 745 +       opt.perturb_period = q->perturb_period/HZ;
 746 +
 747 +       opt.limit = q->limit;
 748 +       opt.divisor = q->hash_divisor;
 749 +       opt.flows = q->depth;
 750 +       opt.hash_kind = q->hash_kind;
 751 +
 752 +       RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 753 +
 754 +       return skb->len;
 755 +
 756 +rtattr_failure:
 757 +       skb_trim(skb, b - skb->data);
 758 +       return -1;
 759 +}
 760 +
 761 +static struct Qdisc_ops esfq_qdisc_ops =
 762 +{
 763 +       .next           =       NULL,
 764 +       .cl_ops         =       NULL,
 765 +       .id             =       "esfq",
 766 +       .priv_size      =       sizeof(struct esfq_sched_data),
 767 +       .enqueue        =       esfq_enqueue,
 768 +       .dequeue        =       esfq_dequeue,
 769 +       .requeue        =       esfq_requeue,
 770 +       .drop           =       esfq_drop,
 771 +       .init           =       esfq_init,
 772 +       .reset          =       esfq_reset,
 773 +       .destroy        =       esfq_destroy,
 774 +       .change         =       NULL, /* esfq_change - needs more work */
 775 +       .dump           =       esfq_dump,
 776 +       .owner          =       THIS_MODULE,
 777 +};
 778 +
 779 +static int __init esfq_module_init(void)
 780 +{
 781 +       return register_qdisc(&esfq_qdisc_ops);
 782 +}
 783 +static void __exit esfq_module_exit(void)
 784 +{
 785 +       unregister_qdisc(&esfq_qdisc_ops);
 786 +}
 787 +module_init(esfq_module_init)
 788 +module_exit(esfq_module_exit)
 789 +MODULE_LICENSE("GPL");