target/linux/generic-2.6/patches-2.6.23/200-sched_esfq.patch

   1 Index: linux-2.6.23-rc6/include/linux/pkt_sched.h
   2 ===================================================================
   3 --- linux-2.6.23-rc6.orig/include/linux/pkt_sched.h     2007-09-21 16:23:53.000000000 +0800
   4 +++ linux-2.6.23-rc6/include/linux/pkt_sched.h  2007-09-21 16:24:04.000000000 +0800
   5 @@ -155,8 +155,40 @@
   6   *
   7   *     The only reason for this is efficiency, it is possible
   8   *     to change these parameters in compile time.
   9 + *
  10 + *     If you need to play with these values use esfq instead.
  11   */
  12
  13 +/* ESFQ section */
  14 +
  15 +enum
  16 +{
  17 +        /* traditional */
  18 +       TCA_SFQ_HASH_CLASSIC,
  19 +       TCA_SFQ_HASH_DST,
  20 +       TCA_SFQ_HASH_SRC,
  21 +       TCA_SFQ_HASH_FWMARK,
  22 +        /* direct */
  23 +       TCA_SFQ_HASH_DSTDIR,
  24 +       TCA_SFQ_HASH_SRCDIR,
  25 +       TCA_SFQ_HASH_FWMARKDIR,
  26 +       /* conntrack */
  27 +       TCA_SFQ_HASH_CTORIGDST,
  28 +       TCA_SFQ_HASH_CTORIGSRC,
  29 +       TCA_SFQ_HASH_CTREPLDST,
  30 +       TCA_SFQ_HASH_CTREPLSRC,
  31 +};
  32 +
  33 +struct tc_esfq_qopt
  34 +{
  35 +       unsigned        quantum;        /* Bytes per round allocated to flow */
  36 +       int             perturb_period; /* Period of hash perturbation */
  37 +       __u32           limit;          /* Maximal packets in queue */
  38 +       unsigned        divisor;        /* Hash divisor  */
  39 +       unsigned        flows;          /* Maximal number of flows  */
  40 +       unsigned        hash_kind;      /* Hash function to use for flow identification */
  41 +};
  42 +
  43  /* RED section */
  44
  45  enum
  46 Index: linux-2.6.23-rc6/net/sched/Kconfig
  47 ===================================================================
  48 --- linux-2.6.23-rc6.orig/net/sched/Kconfig     2007-09-21 16:23:53.000000000 +0800
  49 +++ linux-2.6.23-rc6/net/sched/Kconfig  2007-09-21 16:24:04.000000000 +0800
  50 @@ -144,6 +144,26 @@
  51           To compile this code as a module, choose M here: the
  52           module will be called sch_sfq.
  53
  54 +config NET_SCH_ESFQ
  55 +       tristate "Enhanced Stochastic Fairness Queueing (ESFQ)"
  56 +       ---help---
  57 +         Say Y here if you want to use the Enhanced Stochastic Fairness
  58 +         Queueing (ESFQ) packet scheduling algorithm for some of your network
  59 +         devices or as a leaf discipline for a classful qdisc such as HTB or
  60 +         CBQ (see the top of <file:net/sched/sch_esfq.c> for details and
  61 +         references to the SFQ algorithm).
  62 +
  63 +         This is an enchanced SFQ version which allows you to control some
  64 +         hardcoded values in the SFQ scheduler.
  65 +
  66 +         ESFQ also adds control of the hash function used to identify packet
  67 +         flows. The original SFQ discipline hashes by connection; ESFQ add
  68 +         several other hashing methods, such as by src IP or by dst IP, which
  69 +         can be more fair to users in some networking situations.
  70 +
  71 +         To compile this code as a module, choose M here: the
  72 +         module will be called sch_esfq.
  73 +
  74  config NET_SCH_TEQL
  75         tristate "True Link Equalizer (TEQL)"
  76         ---help---
  77 Index: linux-2.6.23-rc6/net/sched/Makefile
  78 ===================================================================
  79 --- linux-2.6.23-rc6.orig/net/sched/Makefile    2007-09-21 16:23:53.000000000 +0800
  80 +++ linux-2.6.23-rc6/net/sched/Makefile 2007-09-21 16:24:04.000000000 +0800
  81 @@ -22,6 +22,7 @@
  82  obj-$(CONFIG_NET_SCH_INGRESS)  += sch_ingress.o
  83  obj-$(CONFIG_NET_SCH_DSMARK)   += sch_dsmark.o
  84  obj-$(CONFIG_NET_SCH_SFQ)      += sch_sfq.o
  85 +obj-$(CONFIG_NET_SCH_ESFQ)     += sch_esfq.o
  86  obj-$(CONFIG_NET_SCH_TBF)      += sch_tbf.o
  87  obj-$(CONFIG_NET_SCH_TEQL)     += sch_teql.o
  88  obj-$(CONFIG_NET_SCH_PRIO)     += sch_prio.o
  89 Index: linux-2.6.23-rc6/net/sched/sch_esfq.c
  90 ===================================================================
  91 --- /dev/null   1970-01-01 00:00:00.000000000 +0000
  92 +++ linux-2.6.23-rc6/net/sched/sch_esfq.c       2007-09-21 16:24:04.000000000 +0800
  93 @@ -0,0 +1,704 @@
  94 +/*
  95 + * net/sched/sch_esfq.c        Extended Stochastic Fairness Queueing discipline.
  96 + *
  97 + *             This program is free software; you can redistribute it and/or
  98 + *             modify it under the terms of the GNU General Public License
  99 + *             as published by the Free Software Foundation; either version
 100 + *             2 of the License, or (at your option) any later version.
 101 + *
 102 + * Authors:    Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
 103 + *
 104 + * Changes:    Alexander Atanasov, <alex@ssi.bg>
 105 + *             Added dynamic depth,limit,divisor,hash_kind options.
 106 + *             Added dst and src hashes.
 107 + *
 108 + *             Alexander Clouter, <alex@digriz.org.uk>
 109 + *             Ported ESFQ to Linux 2.6.
 110 + *
 111 + *             Corey Hickey, <bugfood-c@fatooh.org>
 112 + *             Maintenance of the Linux 2.6 port.
 113 + *             Added fwmark hash (thanks to Robert Kurjata).
 114 + *             Added direct hashing for src, dst, and fwmark.
 115 + *             Added usage of jhash.
 116 + *
 117 + */
 118 +
 119 +#include <linux/module.h>
 120 +#include <asm/uaccess.h>
 121 +#include <asm/system.h>
 122 +#include <linux/bitops.h>
 123 +#include <linux/types.h>
 124 +#include <linux/kernel.h>
 125 +#include <linux/jiffies.h>
 126 +#include <linux/string.h>
 127 +#include <linux/mm.h>
 128 +#include <linux/socket.h>
 129 +#include <linux/sockios.h>
 130 +#include <linux/in.h>
 131 +#include <linux/errno.h>
 132 +#include <linux/interrupt.h>
 133 +#include <linux/if_ether.h>
 134 +#include <linux/inet.h>
 135 +#include <linux/netdevice.h>
 136 +#include <linux/etherdevice.h>
 137 +#include <linux/notifier.h>
 138 +#include <linux/init.h>
 139 +#include <net/ip.h>
 140 +#include <linux/ipv6.h>
 141 +#include <net/route.h>
 142 +#include <linux/skbuff.h>
 143 +#include <net/sock.h>
 144 +#include <net/pkt_sched.h>
 145 +#include <linux/jhash.h>
 146 +
 147 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 148 +#include <net/netfilter/nf_conntrack.h>
 149 +#endif
 150 +
 151 +/*     Stochastic Fairness Queuing algorithm.
 152 +       For more comments look at sch_sfq.c.
 153 +       The difference is that you can change limit, depth,
 154 +       hash table size and choose alternate hash types.
 155 +
 156 +       classic:        same as in sch_sfq.c
 157 +       dst:            destination IP address
 158 +       src:            source IP address
 159 +       fwmark:         netfilter mark value
 160 +       dst_direct:
 161 +       src_direct:
 162 +       fwmark_direct:  direct hashing of the above sources
 163 +       ctorigdst:      original destination IP address
 164 +       ctorigsrc:      original source IP address
 165 +       ctrepldst:      reply destination IP address
 166 +       ctreplsrc:      reply source IP
 167 +
 168 +*/
 169 +
 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 +       unsigned        dyn_min;        /* For dynamic divisor adjustment; minimum value seen */
 202 +       unsigned        dyn_max;        /*                                 maximum value seen */
 203 +       unsigned        dyn_range;      /*                                 saved range */
 204 +};
 205 +
 206 +/* This contains the info we will hash. */
 207 +struct esfq_packet_info
 208 +{
 209 +       u32     proto;          /* protocol or port */
 210 +       u32     src;            /* source from packet header */
 211 +       u32     dst;            /* destination from packet header */
 212 +       u32     ctorigsrc;      /* original source from conntrack */
 213 +       u32     ctorigdst;      /* original destination from conntrack */
 214 +       u32     ctreplsrc;      /* reply source from conntrack */
 215 +       u32     ctrepldst;      /* reply destination from conntrack */
 216 +       u32     mark;           /* netfilter mark (fwmark) */
 217 +};
 218 +
 219 +/* Hash input values directly into the "nearest" slot, taking into account the
 220 + * range of input values seen. This is most useful when the hash table is at
 221 + * least as large as the range of possible values.
 222 + * Note: this functionality was added before the change to using jhash, and may
 223 + * no longer be useful. */
 224 +static __inline__ unsigned esfq_hash_direct(struct esfq_sched_data *q, u32 h)
 225 +{
 226 +       /* adjust minimum and maximum */
 227 +       if (h < q->dyn_min || h > q->dyn_max) {
 228 +               q->dyn_min = h < q->dyn_min ? h : q->dyn_min;
 229 +               q->dyn_max = h > q->dyn_max ? h : q->dyn_max;
 230 +
 231 +               /* find new range */
 232 +               if ((q->dyn_range = q->dyn_max - q->dyn_min) >= q->hash_divisor)
 233 +                       printk(KERN_WARNING "ESFQ: (direct hash) Input range %u is larger than hash "
 234 +                                       "table. See ESFQ README for details.\n", q->dyn_range);
 235 +       }
 236 +
 237 +       /* hash input values into slot numbers */
 238 +       if (q->dyn_min == q->dyn_max)
 239 +               return 0; /* only one value seen; avoid division by 0 */
 240 +       else
 241 +               return (h - q->dyn_min) * (q->hash_divisor - 1) / q->dyn_range;
 242 +}
 243 +
 244 +static __inline__ unsigned esfq_jhash_1word(struct esfq_sched_data *q,u32 a)
 245 +{
 246 +       return jhash_1word(a, q->perturbation) & (q->hash_divisor-1);
 247 +}
 248 +
 249 +static __inline__ unsigned esfq_jhash_2words(struct esfq_sched_data *q, u32 a, u32 b)
 250 +{
 251 +       return jhash_2words(a, b, q->perturbation) & (q->hash_divisor-1);
 252 +}
 253 +
 254 +static __inline__ unsigned esfq_jhash_3words(struct esfq_sched_data *q, u32 a, u32 b, u32 c)
 255 +{
 256 +       return jhash_3words(a, b, c, q->perturbation) & (q->hash_divisor-1);
 257 +}
 258 +
 259 +
 260 +static unsigned esfq_hash(struct esfq_sched_data *q, struct sk_buff *skb)
 261 +{
 262 +       struct esfq_packet_info info;
 263 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 264 +       enum ip_conntrack_info ctinfo;
 265 +       struct nf_conn *ct = nf_ct_get(skb, &ctinfo);
 266 +#endif
 267 +
 268 +       switch (skb->protocol) {
 269 +       case __constant_htons(ETH_P_IP):
 270 +       {
 271 +               struct iphdr *iph = ip_hdr(skb);
 272 +               info.dst = iph->daddr;
 273 +               info.src = iph->saddr;
 274 +               if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) &&
 275 +                   (iph->protocol == IPPROTO_TCP ||
 276 +                    iph->protocol == IPPROTO_UDP ||
 277 +                    iph->protocol == IPPROTO_SCTP ||
 278 +                    iph->protocol == IPPROTO_DCCP ||
 279 +                    iph->protocol == IPPROTO_ESP))
 280 +                       info.proto = *(((u32*)iph) + iph->ihl);
 281 +               else
 282 +                       info.proto = iph->protocol;
 283 +               break;
 284 +       }
 285 +       case __constant_htons(ETH_P_IPV6):
 286 +       {
 287 +               struct ipv6hdr *iph = ipv6_hdr(skb);
 288 +               /* Hash ipv6 addresses into a u32. This isn't ideal,
 289 +                * but the code is simple. */
 290 +               info.dst = jhash2(iph->daddr.s6_addr32, 4, q->perturbation);
 291 +               info.src = jhash2(iph->saddr.s6_addr32, 4, q->perturbation);
 292 +               if (iph->nexthdr == IPPROTO_TCP ||
 293 +                   iph->nexthdr == IPPROTO_UDP ||
 294 +                   iph->nexthdr == IPPROTO_SCTP ||
 295 +                   iph->nexthdr == IPPROTO_DCCP ||
 296 +                   iph->nexthdr == IPPROTO_ESP)
 297 +                       info.proto = *(u32*)&iph[1];
 298 +               else
 299 +                       info.proto = iph->nexthdr;
 300 +               break;
 301 +       }
 302 +       default:
 303 +               info.dst   = (u32)(unsigned long)skb->dst;
 304 +               info.src   = (u32)(unsigned long)skb->sk;
 305 +               info.proto = skb->protocol;
 306 +       }
 307 +
 308 +       info.mark = skb->mark;
 309 +
 310 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 311 +       /* defaults if there is no conntrack info */
 312 +       info.ctorigsrc = info.src;
 313 +       info.ctorigdst = info.dst;
 314 +       info.ctreplsrc = info.dst;
 315 +       info.ctrepldst = info.src;
 316 +       /* collect conntrack info */
 317 +       if (ct && ct != &nf_conntrack_untracked) {
 318 +               if (skb->protocol == __constant_htons(ETH_P_IP)) {
 319 +                       info.ctorigsrc = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip;
 320 +                       info.ctorigdst = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip;
 321 +                       info.ctreplsrc = ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip;
 322 +                       info.ctrepldst = ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip;
 323 +               }
 324 +               else if (skb->protocol == __constant_htons(ETH_P_IPV6)) {
 325 +                       /* Again, hash ipv6 addresses into a single u32. */
 326 +                       info.ctorigsrc = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip6, 4, q->perturbation);
 327 +                       info.ctorigdst = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip6, 4, q->perturbation);
 328 +                       info.ctreplsrc = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip6, 4, q->perturbation);
 329 +                       info.ctrepldst = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip6, 4, q->perturbation);
 330 +               }
 331 +
 332 +       }
 333 +#endif
 334 +
 335 +       switch(q->hash_kind)
 336 +       {
 337 +       case TCA_SFQ_HASH_CLASSIC:
 338 +               return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 339 +       case TCA_SFQ_HASH_DST:
 340 +               return esfq_jhash_1word(q, info.dst);
 341 +       case TCA_SFQ_HASH_DSTDIR:
 342 +               return esfq_hash_direct(q, ntohl(info.dst));
 343 +       case TCA_SFQ_HASH_SRC:
 344 +               return esfq_jhash_1word(q, info.src);
 345 +       case TCA_SFQ_HASH_SRCDIR:
 346 +               return esfq_hash_direct(q, ntohl(info.src));
 347 +       case TCA_SFQ_HASH_FWMARK:
 348 +               return esfq_jhash_1word(q, info.mark);
 349 +       case TCA_SFQ_HASH_FWMARKDIR:
 350 +               return esfq_hash_direct(q, info.mark);
 351 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 352 +       case TCA_SFQ_HASH_CTORIGDST:
 353 +               return esfq_jhash_1word(q, info.ctorigdst);
 354 +       case TCA_SFQ_HASH_CTORIGSRC:
 355 +               return esfq_jhash_1word(q, info.ctorigsrc);
 356 +       case TCA_SFQ_HASH_CTREPLDST:
 357 +               return esfq_jhash_1word(q, info.ctrepldst);
 358 +       case TCA_SFQ_HASH_CTREPLSRC:
 359 +               return esfq_jhash_1word(q, info.ctreplsrc);
 360 +#endif
 361 +       default:
 362 +               if (net_ratelimit())
 363 +                       printk(KERN_WARNING "ESFQ: Unknown hash method. Falling back to classic.\n");
 364 +       }
 365 +       return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 366 +}
 367 +
 368 +static inline void esfq_link(struct esfq_sched_data *q, esfq_index x)
 369 +{
 370 +       esfq_index p, n;
 371 +       int d = q->qs[x].qlen + q->depth;
 372 +
 373 +       p = d;
 374 +       n = q->dep[d].next;
 375 +       q->dep[x].next = n;
 376 +       q->dep[x].prev = p;
 377 +       q->dep[p].next = q->dep[n].prev = x;
 378 +}
 379 +
 380 +static inline void esfq_dec(struct esfq_sched_data *q, esfq_index x)
 381 +{
 382 +       esfq_index p, n;
 383 +
 384 +       n = q->dep[x].next;
 385 +       p = q->dep[x].prev;
 386 +       q->dep[p].next = n;
 387 +       q->dep[n].prev = p;
 388 +
 389 +       if (n == p && q->max_depth == q->qs[x].qlen + 1)
 390 +               q->max_depth--;
 391 +
 392 +       esfq_link(q, x);
 393 +}
 394 +
 395 +static inline void esfq_inc(struct esfq_sched_data *q, esfq_index x)
 396 +{
 397 +       esfq_index p, n;
 398 +       int d;
 399 +
 400 +       n = q->dep[x].next;
 401 +       p = q->dep[x].prev;
 402 +       q->dep[p].next = n;
 403 +       q->dep[n].prev = p;
 404 +       d = q->qs[x].qlen;
 405 +       if (q->max_depth < d)
 406 +               q->max_depth = d;
 407 +
 408 +       esfq_link(q, x);
 409 +}
 410 +
 411 +static unsigned int esfq_drop(struct Qdisc *sch)
 412 +{
 413 +       struct esfq_sched_data *q = qdisc_priv(sch);
 414 +       esfq_index d = q->max_depth;
 415 +       struct sk_buff *skb;
 416 +       unsigned int len;
 417 +
 418 +       /* Queue is full! Find the longest slot and
 419 +          drop a packet from it */
 420 +
 421 +       if (d > 1) {
 422 +               esfq_index x = q->dep[d+q->depth].next;
 423 +               skb = q->qs[x].prev;
 424 +               len = skb->len;
 425 +               __skb_unlink(skb, &q->qs[x]);
 426 +               kfree_skb(skb);
 427 +               esfq_dec(q, x);
 428 +               sch->q.qlen--;
 429 +               sch->qstats.drops++;
 430 +               sch->qstats.backlog -= len;
 431 +               return len;
 432 +       }
 433 +
 434 +       if (d == 1) {
 435 +               /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
 436 +               d = q->next[q->tail];
 437 +               q->next[q->tail] = q->next[d];
 438 +               q->allot[q->next[d]] += q->quantum;
 439 +               skb = q->qs[d].prev;
 440 +               len = skb->len;
 441 +               __skb_unlink(skb, &q->qs[d]);
 442 +               kfree_skb(skb);
 443 +               esfq_dec(q, d);
 444 +               sch->q.qlen--;
 445 +               q->ht[q->hash[d]] = q->depth;
 446 +               sch->qstats.drops++;
 447 +               sch->qstats.backlog -= len;
 448 +               return len;
 449 +       }
 450 +
 451 +       return 0;
 452 +}
 453 +
 454 +static int
 455 +esfq_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 456 +{
 457 +       struct esfq_sched_data *q = qdisc_priv(sch);
 458 +       unsigned hash = esfq_hash(q, skb);
 459 +       unsigned depth = q->depth;
 460 +       esfq_index x;
 461 +
 462 +       x = q->ht[hash];
 463 +       if (x == depth) {
 464 +               q->ht[hash] = x = q->dep[depth].next;
 465 +               q->hash[x] = hash;
 466 +       }
 467 +       sch->qstats.backlog += skb->len;
 468 +       __skb_queue_tail(&q->qs[x], skb);
 469 +       esfq_inc(q, x);
 470 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 471 +               if (q->tail == depth) { /* It is the first flow */
 472 +                       q->tail = x;
 473 +                       q->next[x] = x;
 474 +                       q->allot[x] = q->quantum;
 475 +               } else {
 476 +                       q->next[x] = q->next[q->tail];
 477 +                       q->next[q->tail] = x;
 478 +                       q->tail = x;
 479 +               }
 480 +       }
 481 +       if (++sch->q.qlen < q->limit-1) {
 482 +               sch->bstats.bytes += skb->len;
 483 +               sch->bstats.packets++;
 484 +               return 0;
 485 +       }
 486 +
 487 +       esfq_drop(sch);
 488 +       return NET_XMIT_CN;
 489 +}
 490 +
 491 +static int
 492 +esfq_requeue(struct sk_buff *skb, struct Qdisc* sch)
 493 +{
 494 +       struct esfq_sched_data *q = qdisc_priv(sch);
 495 +       unsigned hash = esfq_hash(q, skb);
 496 +       unsigned depth = q->depth;
 497 +       esfq_index x;
 498 +
 499 +       x = q->ht[hash];
 500 +       if (x == depth) {
 501 +               q->ht[hash] = x = q->dep[depth].next;
 502 +               q->hash[x] = hash;
 503 +       }
 504 +       sch->qstats.backlog += skb->len;
 505 +       __skb_queue_head(&q->qs[x], skb);
 506 +       esfq_inc(q, x);
 507 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 508 +               if (q->tail == depth) { /* It is the first flow */
 509 +                       q->tail = x;
 510 +                       q->next[x] = x;
 511 +                       q->allot[x] = q->quantum;
 512 +               } else {
 513 +                       q->next[x] = q->next[q->tail];
 514 +                       q->next[q->tail] = x;
 515 +                       q->tail = x;
 516 +               }
 517 +       }
 518 +       if (++sch->q.qlen < q->limit - 1) {
 519 +               sch->qstats.requeues++;
 520 +               return 0;
 521 +       }
 522 +
 523 +       sch->qstats.drops++;
 524 +       esfq_drop(sch);
 525 +       return NET_XMIT_CN;
 526 +}
 527 +
 528 +
 529 +
 530 +
 531 +static struct sk_buff *
 532 +esfq_dequeue(struct Qdisc* sch)
 533 +{
 534 +       struct esfq_sched_data *q = qdisc_priv(sch);
 535 +       struct sk_buff *skb;
 536 +       unsigned depth = q->depth;
 537 +       esfq_index a, old_a;
 538 +
 539 +       /* No active slots */
 540 +       if (q->tail == depth)
 541 +               return NULL;
 542 +
 543 +       a = old_a = q->next[q->tail];
 544 +
 545 +       /* Grab packet */
 546 +       skb = __skb_dequeue(&q->qs[a]);
 547 +       esfq_dec(q, a);
 548 +       sch->q.qlen--;
 549 +       sch->qstats.backlog -= skb->len;
 550 +
 551 +       /* Is the slot empty? */
 552 +       if (q->qs[a].qlen == 0) {
 553 +               q->ht[q->hash[a]] = depth;
 554 +               a = q->next[a];
 555 +               if (a == old_a) {
 556 +                       q->tail = depth;
 557 +                       return skb;
 558 +               }
 559 +               q->next[q->tail] = a;
 560 +               q->allot[a] += q->quantum;
 561 +       } else if ((q->allot[a] -= skb->len) <= 0) {
 562 +               q->tail = a;
 563 +               a = q->next[a];
 564 +               q->allot[a] += q->quantum;
 565 +       }
 566 +
 567 +       return skb;
 568 +}
 569 +
 570 +static void
 571 +esfq_reset(struct Qdisc* sch)
 572 +{
 573 +       struct sk_buff *skb;
 574 +
 575 +       while ((skb = esfq_dequeue(sch)) != NULL)
 576 +               kfree_skb(skb);
 577 +}
 578 +
 579 +static void esfq_perturbation(unsigned long arg)
 580 +{
 581 +       struct Qdisc *sch = (struct Qdisc*)arg;
 582 +       struct esfq_sched_data *q = qdisc_priv(sch);
 583 +
 584 +       q->perturbation = net_random()&0x1F;
 585 +
 586 +       if (q->perturb_period) {
 587 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 588 +               add_timer(&q->perturb_timer);
 589 +       }
 590 +}
 591 +
 592 +static int esfq_change(struct Qdisc *sch, struct rtattr *opt)
 593 +{
 594 +       struct esfq_sched_data *q = qdisc_priv(sch);
 595 +       struct tc_esfq_qopt *ctl = RTA_DATA(opt);
 596 +       int old_perturb = q->perturb_period;
 597 +
 598 +       if (opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 599 +               return -EINVAL;
 600 +
 601 +       sch_tree_lock(sch);
 602 +       q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 603 +       q->perturb_period = ctl->perturb_period*HZ;
 604 +//     q->hash_divisor = ctl->divisor;
 605 +//     q->tail = q->limit = q->depth = ctl->flows;
 606 +
 607 +       if (ctl->limit)
 608 +               q->limit = min_t(u32, ctl->limit, q->depth);
 609 +
 610 +       if (ctl->hash_kind) {
 611 +               q->hash_kind = ctl->hash_kind;
 612 +               if (q->hash_kind != TCA_SFQ_HASH_CLASSIC)
 613 +                       q->perturb_period = 0;
 614 +       }
 615 +
 616 +       // is sch_tree_lock enough to do this ?
 617 +       while (sch->q.qlen >= q->limit-1)
 618 +               esfq_drop(sch);
 619 +
 620 +       if (old_perturb)
 621 +               del_timer(&q->perturb_timer);
 622 +       if (q->perturb_period) {
 623 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 624 +               add_timer(&q->perturb_timer);
 625 +       } else {
 626 +               q->perturbation = 0;
 627 +       }
 628 +       sch_tree_unlock(sch);
 629 +       return 0;
 630 +}
 631 +
 632 +static int esfq_init(struct Qdisc *sch, struct rtattr *opt)
 633 +{
 634 +       struct esfq_sched_data *q = qdisc_priv(sch);
 635 +       struct tc_esfq_qopt *ctl;
 636 +       esfq_index p = ~0U/2;
 637 +       int i;
 638 +
 639 +       if (opt && opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 640 +               return -EINVAL;
 641 +
 642 +       init_timer(&q->perturb_timer);
 643 +       q->perturb_timer.data = (unsigned long)sch;
 644 +       q->perturb_timer.function = esfq_perturbation;
 645 +       q->perturbation = 0;
 646 +       q->hash_kind = TCA_SFQ_HASH_CLASSIC;
 647 +       q->max_depth = 0;
 648 +       q->dyn_min = ~0U; /* maximum value for this type */
 649 +       q->dyn_max = 0;  /* dyn_min/dyn_max will be set properly upon first packet */
 650 +       if (opt == NULL) {
 651 +               q->quantum = psched_mtu(sch->dev);
 652 +               q->perturb_period = 0;
 653 +               q->hash_divisor = 1024;
 654 +               q->tail = q->limit = q->depth = 128;
 655 +
 656 +       } else {
 657 +               ctl = RTA_DATA(opt);
 658 +               q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 659 +               q->perturb_period = ctl->perturb_period*HZ;
 660 +               q->hash_divisor = ctl->divisor ? : 1024;
 661 +               q->tail = q->limit = q->depth = ctl->flows ? : 128;
 662 +
 663 +               if ( q->depth > p - 1 )
 664 +                       return -EINVAL;
 665 +
 666 +               if (ctl->limit)
 667 +                       q->limit = min_t(u32, ctl->limit, q->depth);
 668 +
 669 +               if (ctl->hash_kind) {
 670 +                       q->hash_kind = ctl->hash_kind;
 671 +               }
 672 +
 673 +               if (q->perturb_period) {
 674 +                       q->perturb_timer.expires = jiffies + q->perturb_period;
 675 +                       add_timer(&q->perturb_timer);
 676 +               }
 677 +       }
 678 +
 679 +       q->ht = kmalloc(q->hash_divisor*sizeof(esfq_index), GFP_KERNEL);
 680 +       if (!q->ht)
 681 +               goto err_case;
 682 +
 683 +       q->dep = kmalloc((1+q->depth*2)*sizeof(struct esfq_head), GFP_KERNEL);
 684 +       if (!q->dep)
 685 +               goto err_case;
 686 +       q->next = kmalloc(q->depth*sizeof(esfq_index), GFP_KERNEL);
 687 +       if (!q->next)
 688 +               goto err_case;
 689 +
 690 +       q->allot = kmalloc(q->depth*sizeof(short), GFP_KERNEL);
 691 +       if (!q->allot)
 692 +               goto err_case;
 693 +       q->hash = kmalloc(q->depth*sizeof(unsigned short), GFP_KERNEL);
 694 +       if (!q->hash)
 695 +               goto err_case;
 696 +       q->qs = kmalloc(q->depth*sizeof(struct sk_buff_head), GFP_KERNEL);
 697 +       if (!q->qs)
 698 +               goto err_case;
 699 +
 700 +       for (i=0; i< q->hash_divisor; i++)
 701 +               q->ht[i] = q->depth;
 702 +       for (i=0; i<q->depth; i++) {
 703 +               skb_queue_head_init(&q->qs[i]);
 704 +               q->dep[i+q->depth].next = i+q->depth;
 705 +               q->dep[i+q->depth].prev = i+q->depth;
 706 +       }
 707 +
 708 +       for (i=0; i<q->depth; i++)
 709 +               esfq_link(q, i);
 710 +       return 0;
 711 +err_case:
 712 +       del_timer(&q->perturb_timer);
 713 +       if (q->ht)
 714 +               kfree(q->ht);
 715 +       if (q->dep)
 716 +               kfree(q->dep);
 717 +       if (q->next)
 718 +               kfree(q->next);
 719 +       if (q->allot)
 720 +               kfree(q->allot);
 721 +       if (q->hash)
 722 +               kfree(q->hash);
 723 +       if (q->qs)
 724 +               kfree(q->qs);
 725 +       return -ENOBUFS;
 726 +}
 727 +
 728 +static void esfq_destroy(struct Qdisc *sch)
 729 +{
 730 +       struct esfq_sched_data *q = qdisc_priv(sch);
 731 +       del_timer(&q->perturb_timer);
 732 +       if(q->ht)
 733 +               kfree(q->ht);
 734 +       if(q->dep)
 735 +               kfree(q->dep);
 736 +       if(q->next)
 737 +               kfree(q->next);
 738 +       if(q->allot)
 739 +               kfree(q->allot);
 740 +       if(q->hash)
 741 +               kfree(q->hash);
 742 +       if(q->qs)
 743 +               kfree(q->qs);
 744 +}
 745 +
 746 +static int esfq_dump(struct Qdisc *sch, struct sk_buff *skb)
 747 +{
 748 +       struct esfq_sched_data *q = qdisc_priv(sch);
 749 +       unsigned char    *b = skb->tail;
 750 +       struct tc_esfq_qopt opt;
 751 +
 752 +       opt.quantum = q->quantum;
 753 +       opt.perturb_period = q->perturb_period/HZ;
 754 +
 755 +       opt.limit = q->limit;
 756 +       opt.divisor = q->hash_divisor;
 757 +       opt.flows = q->depth;
 758 +       opt.hash_kind = q->hash_kind;
 759 +
 760 +       RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 761 +
 762 +       return skb->len;
 763 +
 764 +rtattr_failure:
 765 +       skb_trim(skb, b - skb->data);
 766 +       return -1;
 767 +}
 768 +
 769 +static struct Qdisc_ops esfq_qdisc_ops =
 770 +{
 771 +       .next           =       NULL,
 772 +       .cl_ops         =       NULL,
 773 +       .id             =       "esfq",
 774 +       .priv_size      =       sizeof(struct esfq_sched_data),
 775 +       .enqueue        =       esfq_enqueue,
 776 +       .dequeue        =       esfq_dequeue,
 777 +       .requeue        =       esfq_requeue,
 778 +       .drop           =       esfq_drop,
 779 +       .init           =       esfq_init,
 780 +       .reset          =       esfq_reset,
 781 +       .destroy        =       esfq_destroy,
 782 +       .change         =       NULL, /* esfq_change - needs more work */
 783 +       .dump           =       esfq_dump,
 784 +       .owner          =       THIS_MODULE,
 785 +};
 786 +
 787 +static int __init esfq_module_init(void)
 788 +{
 789 +       return register_qdisc(&esfq_qdisc_ops);
 790 +}
 791 +static void __exit esfq_module_exit(void)
 792 +{
 793 +       unregister_qdisc(&esfq_qdisc_ops);
 794 +}
 795 +module_init(esfq_module_init)
 796 +module_exit(esfq_module_exit)
 797 +MODULE_LICENSE("GPL");