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

   1 diff -urN linux-2.6.21.1.old/include/linux/pkt_sched.h linux-2.6.21.1.dev/include/linux/pkt_sched.h
   2 --- linux-2.6.21.1.old/include/linux/pkt_sched.h        2007-04-27 23:49:26.000000000 +0200
   3 +++ linux-2.6.21.1.dev/include/linux/pkt_sched.h        2007-05-26 20:43:12.530587320 +0200
   4 @@ -146,8 +146,40 @@
   5   *
   6   *     The only reason for this is efficiency, it is possible
   7   *     to change these parameters in compile time.
   8 + *
   9 + *     If you need to play with these values use esfq instead.
  10   */
  11
  12 +/* ESFQ section */
  13 +
  14 +enum
  15 +{
  16 +        /* traditional */
  17 +       TCA_SFQ_HASH_CLASSIC,
  18 +       TCA_SFQ_HASH_DST,
  19 +       TCA_SFQ_HASH_SRC,
  20 +       TCA_SFQ_HASH_FWMARK,
  21 +        /* direct */
  22 +       TCA_SFQ_HASH_DSTDIR,
  23 +       TCA_SFQ_HASH_SRCDIR,
  24 +       TCA_SFQ_HASH_FWMARKDIR,
  25 +       /* conntrack */
  26 +       TCA_SFQ_HASH_CTORIGDST,
  27 +       TCA_SFQ_HASH_CTORIGSRC,
  28 +       TCA_SFQ_HASH_CTREPLDST,
  29 +       TCA_SFQ_HASH_CTREPLSRC,
  30 +};
  31 +
  32 +struct tc_esfq_qopt
  33 +{
  34 +       unsigned        quantum;        /* Bytes per round allocated to flow */
  35 +       int             perturb_period; /* Period of hash perturbation */
  36 +       __u32           limit;          /* Maximal packets in queue */
  37 +       unsigned        divisor;        /* Hash divisor  */
  38 +       unsigned        flows;          /* Maximal number of flows  */
  39 +       unsigned        hash_kind;      /* Hash function to use for flow identification */
  40 +};
  41 +
  42  /* RED section */
  43
  44  enum
  45 diff -urN linux-2.6.21.1.old/net/sched/Kconfig linux-2.6.21.1.dev/net/sched/Kconfig
  46 --- linux-2.6.21.1.old/net/sched/Kconfig        2007-04-27 23:49:26.000000000 +0200
  47 +++ linux-2.6.21.1.dev/net/sched/Kconfig        2007-05-26 20:43:12.572580936 +0200
  48 @@ -133,6 +133,26 @@
  49           To compile this code as a module, choose M here: the
  50           module will be called sch_sfq.
  51
  52 +config NET_SCH_ESFQ
  53 +       tristate "Enhanced Stochastic Fairness Queueing (ESFQ)"
  54 +       ---help---
  55 +         Say Y here if you want to use the Enhanced Stochastic Fairness
  56 +         Queueing (ESFQ) packet scheduling algorithm for some of your network
  57 +         devices or as a leaf discipline for a classful qdisc such as HTB or
  58 +         CBQ (see the top of <file:net/sched/sch_esfq.c> for details and
  59 +         references to the SFQ algorithm).
  60 +
  61 +         This is an enchanced SFQ version which allows you to control some
  62 +         hardcoded values in the SFQ scheduler.
  63 +
  64 +         ESFQ also adds control of the hash function used to identify packet
  65 +         flows. The original SFQ discipline hashes by connection; ESFQ add
  66 +         several other hashing methods, such as by src IP or by dst IP, which
  67 +         can be more fair to users in some networking situations.
  68 +
  69 +         To compile this code as a module, choose M here: the
  70 +         module will be called sch_esfq.
  71 +
  72  config NET_SCH_TEQL
  73         tristate "True Link Equalizer (TEQL)"
  74         ---help---
  75 diff -urN linux-2.6.21.1.old/net/sched/Makefile linux-2.6.21.1.dev/net/sched/Makefile
  76 --- linux-2.6.21.1.old/net/sched/Makefile       2007-04-27 23:49:26.000000000 +0200
  77 +++ linux-2.6.21.1.dev/net/sched/Makefile       2007-05-26 20:43:12.577580176 +0200
  78 @@ -23,6 +23,7 @@
  79  obj-$(CONFIG_NET_SCH_INGRESS)  += sch_ingress.o
  80  obj-$(CONFIG_NET_SCH_DSMARK)   += sch_dsmark.o
  81  obj-$(CONFIG_NET_SCH_SFQ)      += sch_sfq.o
  82 +obj-$(CONFIG_NET_SCH_ESFQ)     += sch_esfq.o
  83  obj-$(CONFIG_NET_SCH_TBF)      += sch_tbf.o
  84  obj-$(CONFIG_NET_SCH_TEQL)     += sch_teql.o
  85  obj-$(CONFIG_NET_SCH_PRIO)     += sch_prio.o
  86 diff -urN linux-2.6.21.1.old/net/sched/sch_esfq.c linux-2.6.21.1.dev/net/sched/sch_esfq.c
  87 --- linux-2.6.21.1.old/net/sched/sch_esfq.c     1970-01-01 01:00:00.000000000 +0100
  88 +++ linux-2.6.21.1.dev/net/sched/sch_esfq.c     2007-05-26 20:43:12.578580024 +0200
  89 @@ -0,0 +1,704 @@
  90 +/*
  91 + * net/sched/sch_esfq.c        Extended Stochastic Fairness Queueing discipline.
  92 + *
  93 + *             This program is free software; you can redistribute it and/or
  94 + *             modify it under the terms of the GNU General Public License
  95 + *             as published by the Free Software Foundation; either version
  96 + *             2 of the License, or (at your option) any later version.
  97 + *
  98 + * Authors:    Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  99 + *
 100 + * Changes:    Alexander Atanasov, <alex@ssi.bg>
 101 + *             Added dynamic depth,limit,divisor,hash_kind options.
 102 + *             Added dst and src hashes.
 103 + *
 104 + *             Alexander Clouter, <alex@digriz.org.uk>
 105 + *             Ported ESFQ to Linux 2.6.
 106 + *
 107 + *             Corey Hickey, <bugfood-c@fatooh.org>
 108 + *             Maintenance of the Linux 2.6 port.
 109 + *             Added fwmark hash (thanks to Robert Kurjata).
 110 + *             Added direct hashing for src, dst, and fwmark.
 111 + *             Added usage of jhash.
 112 + *
 113 + */
 114 +
 115 +#include <linux/module.h>
 116 +#include <asm/uaccess.h>
 117 +#include <asm/system.h>
 118 +#include <linux/bitops.h>
 119 +#include <linux/types.h>
 120 +#include <linux/kernel.h>
 121 +#include <linux/jiffies.h>
 122 +#include <linux/string.h>
 123 +#include <linux/mm.h>
 124 +#include <linux/socket.h>
 125 +#include <linux/sockios.h>
 126 +#include <linux/in.h>
 127 +#include <linux/errno.h>
 128 +#include <linux/interrupt.h>
 129 +#include <linux/if_ether.h>
 130 +#include <linux/inet.h>
 131 +#include <linux/netdevice.h>
 132 +#include <linux/etherdevice.h>
 133 +#include <linux/notifier.h>
 134 +#include <linux/init.h>
 135 +#include <net/ip.h>
 136 +#include <linux/ipv6.h>
 137 +#include <net/route.h>
 138 +#include <linux/skbuff.h>
 139 +#include <net/sock.h>
 140 +#include <net/pkt_sched.h>
 141 +#include <linux/jhash.h>
 142 +
 143 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 144 +#include <net/netfilter/nf_conntrack.h>
 145 +#endif
 146 +
 147 +/*     Stochastic Fairness Queuing algorithm.
 148 +       For more comments look at sch_sfq.c.
 149 +       The difference is that you can change limit, depth,
 150 +       hash table size and choose alternate hash types.
 151 +
 152 +       classic:        same as in sch_sfq.c
 153 +       dst:            destination IP address
 154 +       src:            source IP address
 155 +       fwmark:         netfilter mark value
 156 +       dst_direct:
 157 +       src_direct:
 158 +       fwmark_direct:  direct hashing of the above sources
 159 +       ctorigdst:      original destination IP address
 160 +       ctorigsrc:      original source IP address
 161 +       ctrepldst:      reply destination IP address
 162 +       ctreplsrc:      reply source IP
 163 +
 164 +*/
 165 +
 166 +
 167 +/* This type should contain at least SFQ_DEPTH*2 values */
 168 +typedef unsigned int esfq_index;
 169 +
 170 +struct esfq_head
 171 +{
 172 +       esfq_index      next;
 173 +       esfq_index      prev;
 174 +};
 175 +
 176 +struct esfq_sched_data
 177 +{
 178 +/* Parameters */
 179 +       int             perturb_period;
 180 +       unsigned        quantum;        /* Allotment per round: MUST BE >= MTU */
 181 +       int             limit;
 182 +       unsigned        depth;
 183 +       unsigned        hash_divisor;
 184 +       unsigned        hash_kind;
 185 +/* Variables */
 186 +       struct timer_list perturb_timer;
 187 +       int             perturbation;
 188 +       esfq_index      tail;           /* Index of current slot in round */
 189 +       esfq_index      max_depth;      /* Maximal depth */
 190 +
 191 +       esfq_index      *ht;                    /* Hash table */
 192 +       esfq_index      *next;                  /* Active slots link */
 193 +       short           *allot;                 /* Current allotment per slot */
 194 +       unsigned short  *hash;                  /* Hash value indexed by slots */
 195 +       struct sk_buff_head     *qs;            /* Slot queue */
 196 +       struct esfq_head        *dep;           /* Linked list of slots, indexed by depth */
 197 +       unsigned        dyn_min;        /* For dynamic divisor adjustment; minimum value seen */
 198 +       unsigned        dyn_max;        /*                                 maximum value seen */
 199 +       unsigned        dyn_range;      /*                                 saved range */
 200 +};
 201 +
 202 +/* This contains the info we will hash. */
 203 +struct esfq_packet_info
 204 +{
 205 +       u32     proto;          /* protocol or port */
 206 +       u32     src;            /* source from packet header */
 207 +       u32     dst;            /* destination from packet header */
 208 +       u32     ctorigsrc;      /* original source from conntrack */
 209 +       u32     ctorigdst;      /* original destination from conntrack */
 210 +       u32     ctreplsrc;      /* reply source from conntrack */
 211 +       u32     ctrepldst;      /* reply destination from conntrack */
 212 +       u32     mark;           /* netfilter mark (fwmark) */
 213 +};
 214 +
 215 +/* Hash input values directly into the "nearest" slot, taking into account the
 216 + * range of input values seen. This is most useful when the hash table is at
 217 + * least as large as the range of possible values.
 218 + * Note: this functionality was added before the change to using jhash, and may
 219 + * no longer be useful. */
 220 +static __inline__ unsigned esfq_hash_direct(struct esfq_sched_data *q, u32 h)
 221 +{
 222 +       /* adjust minimum and maximum */
 223 +       if (h < q->dyn_min || h > q->dyn_max) {
 224 +               q->dyn_min = h < q->dyn_min ? h : q->dyn_min;
 225 +               q->dyn_max = h > q->dyn_max ? h : q->dyn_max;
 226 +
 227 +               /* find new range */
 228 +               if ((q->dyn_range = q->dyn_max - q->dyn_min) >= q->hash_divisor)
 229 +                       printk(KERN_WARNING "ESFQ: (direct hash) Input range %u is larger than hash "
 230 +                                       "table. See ESFQ README for details.\n", q->dyn_range);
 231 +       }
 232 +
 233 +       /* hash input values into slot numbers */
 234 +       if (q->dyn_min == q->dyn_max)
 235 +               return 0; /* only one value seen; avoid division by 0 */
 236 +       else
 237 +               return (h - q->dyn_min) * (q->hash_divisor - 1) / q->dyn_range;
 238 +}
 239 +
 240 +static __inline__ unsigned esfq_jhash_1word(struct esfq_sched_data *q,u32 a)
 241 +{
 242 +       return jhash_1word(a, q->perturbation) & (q->hash_divisor-1);
 243 +}
 244 +
 245 +static __inline__ unsigned esfq_jhash_2words(struct esfq_sched_data *q, u32 a, u32 b)
 246 +{
 247 +       return jhash_2words(a, b, q->perturbation) & (q->hash_divisor-1);
 248 +}
 249 +
 250 +static __inline__ unsigned esfq_jhash_3words(struct esfq_sched_data *q, u32 a, u32 b, u32 c)
 251 +{
 252 +       return jhash_3words(a, b, c, q->perturbation) & (q->hash_divisor-1);
 253 +}
 254 +
 255 +
 256 +static unsigned esfq_hash(struct esfq_sched_data *q, struct sk_buff *skb)
 257 +{
 258 +       struct esfq_packet_info info;
 259 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 260 +       enum ip_conntrack_info ctinfo;
 261 +       struct nf_conn *ct = nf_ct_get(skb, &ctinfo);
 262 +#endif
 263 +
 264 +       switch (skb->protocol) {
 265 +       case __constant_htons(ETH_P_IP):
 266 +       {
 267 +               struct iphdr *iph = ip_hdr(skb);
 268 +               info.dst = iph->daddr;
 269 +               info.src = iph->saddr;
 270 +               if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) &&
 271 +                   (iph->protocol == IPPROTO_TCP ||
 272 +                    iph->protocol == IPPROTO_UDP ||
 273 +                    iph->protocol == IPPROTO_SCTP ||
 274 +                    iph->protocol == IPPROTO_DCCP ||
 275 +                    iph->protocol == IPPROTO_ESP))
 276 +                       info.proto = *(((u32*)iph) + iph->ihl);
 277 +               else
 278 +                       info.proto = iph->protocol;
 279 +               break;
 280 +       }
 281 +       case __constant_htons(ETH_P_IPV6):
 282 +       {
 283 +               struct ipv6hdr *iph = ipv6_hdr(skb);
 284 +               /* Hash ipv6 addresses into a u32. This isn't ideal,
 285 +                * but the code is simple. */
 286 +               info.dst = jhash2(iph->daddr.s6_addr32, 4, q->perturbation);
 287 +               info.src = jhash2(iph->saddr.s6_addr32, 4, q->perturbation);
 288 +               if (iph->nexthdr == IPPROTO_TCP ||
 289 +                   iph->nexthdr == IPPROTO_UDP ||
 290 +                   iph->nexthdr == IPPROTO_SCTP ||
 291 +                   iph->nexthdr == IPPROTO_DCCP ||
 292 +                   iph->nexthdr == IPPROTO_ESP)
 293 +                       info.proto = *(u32*)&iph[1];
 294 +               else
 295 +                       info.proto = iph->nexthdr;
 296 +               break;
 297 +       }
 298 +       default:
 299 +               info.dst   = (u32)(unsigned long)skb->dst;
 300 +               info.src   = (u32)(unsigned long)skb->sk;
 301 +               info.proto = skb->protocol;
 302 +       }
 303 +
 304 +       info.mark = skb->mark;
 305 +
 306 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 307 +       /* defaults if there is no conntrack info */
 308 +       info.ctorigsrc = info.src;
 309 +       info.ctorigdst = info.dst;
 310 +       info.ctreplsrc = info.dst;
 311 +       info.ctrepldst = info.src;
 312 +       /* collect conntrack info */
 313 +       if (ct && ct != &nf_conntrack_untracked) {
 314 +               if (skb->protocol == __constant_htons(ETH_P_IP)) {
 315 +                       info.ctorigsrc = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip;
 316 +                       info.ctorigdst = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip;
 317 +                       info.ctreplsrc = ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip;
 318 +                       info.ctrepldst = ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip;
 319 +               }
 320 +               else if (skb->protocol == __constant_htons(ETH_P_IPV6)) {
 321 +                       /* Again, hash ipv6 addresses into a single u32. */
 322 +                       info.ctorigsrc = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip6, 4, q->perturbation);
 323 +                       info.ctorigdst = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip6, 4, q->perturbation);
 324 +                       info.ctreplsrc = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip6, 4, q->perturbation);
 325 +                       info.ctrepldst = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip6, 4, q->perturbation);
 326 +               }
 327 +
 328 +       }
 329 +#endif
 330 +
 331 +       switch(q->hash_kind)
 332 +       {
 333 +       case TCA_SFQ_HASH_CLASSIC:
 334 +               return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 335 +       case TCA_SFQ_HASH_DST:
 336 +               return esfq_jhash_1word(q, info.dst);
 337 +       case TCA_SFQ_HASH_DSTDIR:
 338 +               return esfq_hash_direct(q, ntohl(info.dst));
 339 +       case TCA_SFQ_HASH_SRC:
 340 +               return esfq_jhash_1word(q, info.src);
 341 +       case TCA_SFQ_HASH_SRCDIR:
 342 +               return esfq_hash_direct(q, ntohl(info.src));
 343 +       case TCA_SFQ_HASH_FWMARK:
 344 +               return esfq_jhash_1word(q, info.mark);
 345 +       case TCA_SFQ_HASH_FWMARKDIR:
 346 +               return esfq_hash_direct(q, info.mark);
 347 +#ifdef CONFIG_NF_CONNTRACK_ENABLED
 348 +       case TCA_SFQ_HASH_CTORIGDST:
 349 +               return esfq_jhash_1word(q, info.ctorigdst);
 350 +       case TCA_SFQ_HASH_CTORIGSRC:
 351 +               return esfq_jhash_1word(q, info.ctorigsrc);
 352 +       case TCA_SFQ_HASH_CTREPLDST:
 353 +               return esfq_jhash_1word(q, info.ctrepldst);
 354 +       case TCA_SFQ_HASH_CTREPLSRC:
 355 +               return esfq_jhash_1word(q, info.ctreplsrc);
 356 +#endif
 357 +       default:
 358 +               if (net_ratelimit())
 359 +                       printk(KERN_WARNING "ESFQ: Unknown hash method. Falling back to classic.\n");
 360 +       }
 361 +       return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 362 +}
 363 +
 364 +static inline void esfq_link(struct esfq_sched_data *q, esfq_index x)
 365 +{
 366 +       esfq_index p, n;
 367 +       int d = q->qs[x].qlen + q->depth;
 368 +
 369 +       p = d;
 370 +       n = q->dep[d].next;
 371 +       q->dep[x].next = n;
 372 +       q->dep[x].prev = p;
 373 +       q->dep[p].next = q->dep[n].prev = x;
 374 +}
 375 +
 376 +static inline void esfq_dec(struct esfq_sched_data *q, esfq_index x)
 377 +{
 378 +       esfq_index p, n;
 379 +
 380 +       n = q->dep[x].next;
 381 +       p = q->dep[x].prev;
 382 +       q->dep[p].next = n;
 383 +       q->dep[n].prev = p;
 384 +
 385 +       if (n == p && q->max_depth == q->qs[x].qlen + 1)
 386 +               q->max_depth--;
 387 +
 388 +       esfq_link(q, x);
 389 +}
 390 +
 391 +static inline void esfq_inc(struct esfq_sched_data *q, esfq_index x)
 392 +{
 393 +       esfq_index p, n;
 394 +       int d;
 395 +
 396 +       n = q->dep[x].next;
 397 +       p = q->dep[x].prev;
 398 +       q->dep[p].next = n;
 399 +       q->dep[n].prev = p;
 400 +       d = q->qs[x].qlen;
 401 +       if (q->max_depth < d)
 402 +               q->max_depth = d;
 403 +
 404 +       esfq_link(q, x);
 405 +}
 406 +
 407 +static unsigned int esfq_drop(struct Qdisc *sch)
 408 +{
 409 +       struct esfq_sched_data *q = qdisc_priv(sch);
 410 +       esfq_index d = q->max_depth;
 411 +       struct sk_buff *skb;
 412 +       unsigned int len;
 413 +
 414 +       /* Queue is full! Find the longest slot and
 415 +          drop a packet from it */
 416 +
 417 +       if (d > 1) {
 418 +               esfq_index x = q->dep[d+q->depth].next;
 419 +               skb = q->qs[x].prev;
 420 +               len = skb->len;
 421 +               __skb_unlink(skb, &q->qs[x]);
 422 +               kfree_skb(skb);
 423 +               esfq_dec(q, x);
 424 +               sch->q.qlen--;
 425 +               sch->qstats.drops++;
 426 +               sch->qstats.backlog -= len;
 427 +               return len;
 428 +       }
 429 +
 430 +       if (d == 1) {
 431 +               /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
 432 +               d = q->next[q->tail];
 433 +               q->next[q->tail] = q->next[d];
 434 +               q->allot[q->next[d]] += q->quantum;
 435 +               skb = q->qs[d].prev;
 436 +               len = skb->len;
 437 +               __skb_unlink(skb, &q->qs[d]);
 438 +               kfree_skb(skb);
 439 +               esfq_dec(q, d);
 440 +               sch->q.qlen--;
 441 +               q->ht[q->hash[d]] = q->depth;
 442 +               sch->qstats.drops++;
 443 +               sch->qstats.backlog -= len;
 444 +               return len;
 445 +       }
 446 +
 447 +       return 0;
 448 +}
 449 +
 450 +static int
 451 +esfq_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 452 +{
 453 +       struct esfq_sched_data *q = qdisc_priv(sch);
 454 +       unsigned hash = esfq_hash(q, skb);
 455 +       unsigned depth = q->depth;
 456 +       esfq_index x;
 457 +
 458 +       x = q->ht[hash];
 459 +       if (x == depth) {
 460 +               q->ht[hash] = x = q->dep[depth].next;
 461 +               q->hash[x] = hash;
 462 +       }
 463 +       sch->qstats.backlog += skb->len;
 464 +       __skb_queue_tail(&q->qs[x], skb);
 465 +       esfq_inc(q, x);
 466 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 467 +               if (q->tail == depth) { /* It is the first flow */
 468 +                       q->tail = x;
 469 +                       q->next[x] = x;
 470 +                       q->allot[x] = q->quantum;
 471 +               } else {
 472 +                       q->next[x] = q->next[q->tail];
 473 +                       q->next[q->tail] = x;
 474 +                       q->tail = x;
 475 +               }
 476 +       }
 477 +       if (++sch->q.qlen < q->limit-1) {
 478 +               sch->bstats.bytes += skb->len;
 479 +               sch->bstats.packets++;
 480 +               return 0;
 481 +       }
 482 +
 483 +       esfq_drop(sch);
 484 +       return NET_XMIT_CN;
 485 +}
 486 +
 487 +static int
 488 +esfq_requeue(struct sk_buff *skb, struct Qdisc* sch)
 489 +{
 490 +       struct esfq_sched_data *q = qdisc_priv(sch);
 491 +       unsigned hash = esfq_hash(q, skb);
 492 +       unsigned depth = q->depth;
 493 +       esfq_index x;
 494 +
 495 +       x = q->ht[hash];
 496 +       if (x == depth) {
 497 +               q->ht[hash] = x = q->dep[depth].next;
 498 +               q->hash[x] = hash;
 499 +       }
 500 +       sch->qstats.backlog += skb->len;
 501 +       __skb_queue_head(&q->qs[x], skb);
 502 +       esfq_inc(q, x);
 503 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 504 +               if (q->tail == depth) { /* It is the first flow */
 505 +                       q->tail = x;
 506 +                       q->next[x] = x;
 507 +                       q->allot[x] = q->quantum;
 508 +               } else {
 509 +                       q->next[x] = q->next[q->tail];
 510 +                       q->next[q->tail] = x;
 511 +                       q->tail = x;
 512 +               }
 513 +       }
 514 +       if (++sch->q.qlen < q->limit - 1) {
 515 +               sch->qstats.requeues++;
 516 +               return 0;
 517 +       }
 518 +
 519 +       sch->qstats.drops++;
 520 +       esfq_drop(sch);
 521 +       return NET_XMIT_CN;
 522 +}
 523 +
 524 +
 525 +
 526 +
 527 +static struct sk_buff *
 528 +esfq_dequeue(struct Qdisc* sch)
 529 +{
 530 +       struct esfq_sched_data *q = qdisc_priv(sch);
 531 +       struct sk_buff *skb;
 532 +       unsigned depth = q->depth;
 533 +       esfq_index a, old_a;
 534 +
 535 +       /* No active slots */
 536 +       if (q->tail == depth)
 537 +               return NULL;
 538 +
 539 +       a = old_a = q->next[q->tail];
 540 +
 541 +       /* Grab packet */
 542 +       skb = __skb_dequeue(&q->qs[a]);
 543 +       esfq_dec(q, a);
 544 +       sch->q.qlen--;
 545 +       sch->qstats.backlog -= skb->len;
 546 +
 547 +       /* Is the slot empty? */
 548 +       if (q->qs[a].qlen == 0) {
 549 +               q->ht[q->hash[a]] = depth;
 550 +               a = q->next[a];
 551 +               if (a == old_a) {
 552 +                       q->tail = depth;
 553 +                       return skb;
 554 +               }
 555 +               q->next[q->tail] = a;
 556 +               q->allot[a] += q->quantum;
 557 +       } else if ((q->allot[a] -= skb->len) <= 0) {
 558 +               q->tail = a;
 559 +               a = q->next[a];
 560 +               q->allot[a] += q->quantum;
 561 +       }
 562 +
 563 +       return skb;
 564 +}
 565 +
 566 +static void
 567 +esfq_reset(struct Qdisc* sch)
 568 +{
 569 +       struct sk_buff *skb;
 570 +
 571 +       while ((skb = esfq_dequeue(sch)) != NULL)
 572 +               kfree_skb(skb);
 573 +}
 574 +
 575 +static void esfq_perturbation(unsigned long arg)
 576 +{
 577 +       struct Qdisc *sch = (struct Qdisc*)arg;
 578 +       struct esfq_sched_data *q = qdisc_priv(sch);
 579 +
 580 +       q->perturbation = net_random()&0x1F;
 581 +
 582 +       if (q->perturb_period) {
 583 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 584 +               add_timer(&q->perturb_timer);
 585 +       }
 586 +}
 587 +
 588 +static int esfq_change(struct Qdisc *sch, struct rtattr *opt)
 589 +{
 590 +       struct esfq_sched_data *q = qdisc_priv(sch);
 591 +       struct tc_esfq_qopt *ctl = RTA_DATA(opt);
 592 +       int old_perturb = q->perturb_period;
 593 +
 594 +       if (opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 595 +               return -EINVAL;
 596 +
 597 +       sch_tree_lock(sch);
 598 +       q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 599 +       q->perturb_period = ctl->perturb_period*HZ;
 600 +//     q->hash_divisor = ctl->divisor;
 601 +//     q->tail = q->limit = q->depth = ctl->flows;
 602 +
 603 +       if (ctl->limit)
 604 +               q->limit = min_t(u32, ctl->limit, q->depth);
 605 +
 606 +       if (ctl->hash_kind) {
 607 +               q->hash_kind = ctl->hash_kind;
 608 +               if (q->hash_kind != TCA_SFQ_HASH_CLASSIC)
 609 +                       q->perturb_period = 0;
 610 +       }
 611 +
 612 +       // is sch_tree_lock enough to do this ?
 613 +       while (sch->q.qlen >= q->limit-1)
 614 +               esfq_drop(sch);
 615 +
 616 +       if (old_perturb)
 617 +               del_timer(&q->perturb_timer);
 618 +       if (q->perturb_period) {
 619 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 620 +               add_timer(&q->perturb_timer);
 621 +       } else {
 622 +               q->perturbation = 0;
 623 +       }
 624 +       sch_tree_unlock(sch);
 625 +       return 0;
 626 +}
 627 +
 628 +static int esfq_init(struct Qdisc *sch, struct rtattr *opt)
 629 +{
 630 +       struct esfq_sched_data *q = qdisc_priv(sch);
 631 +       struct tc_esfq_qopt *ctl;
 632 +       esfq_index p = ~0U/2;
 633 +       int i;
 634 +
 635 +       if (opt && opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
 636 +               return -EINVAL;
 637 +
 638 +       init_timer(&q->perturb_timer);
 639 +       q->perturb_timer.data = (unsigned long)sch;
 640 +       q->perturb_timer.function = esfq_perturbation;
 641 +       q->perturbation = 0;
 642 +       q->hash_kind = TCA_SFQ_HASH_CLASSIC;
 643 +       q->max_depth = 0;
 644 +       q->dyn_min = ~0U; /* maximum value for this type */
 645 +       q->dyn_max = 0;  /* dyn_min/dyn_max will be set properly upon first packet */
 646 +       if (opt == NULL) {
 647 +               q->quantum = psched_mtu(sch->dev);
 648 +               q->perturb_period = 0;
 649 +               q->hash_divisor = 1024;
 650 +               q->tail = q->limit = q->depth = 128;
 651 +
 652 +       } else {
 653 +               ctl = RTA_DATA(opt);
 654 +               q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
 655 +               q->perturb_period = ctl->perturb_period*HZ;
 656 +               q->hash_divisor = ctl->divisor ? : 1024;
 657 +               q->tail = q->limit = q->depth = ctl->flows ? : 128;
 658 +
 659 +               if ( q->depth > p - 1 )
 660 +                       return -EINVAL;
 661 +
 662 +               if (ctl->limit)
 663 +                       q->limit = min_t(u32, ctl->limit, q->depth);
 664 +
 665 +               if (ctl->hash_kind) {
 666 +                       q->hash_kind = ctl->hash_kind;
 667 +               }
 668 +
 669 +               if (q->perturb_period) {
 670 +                       q->perturb_timer.expires = jiffies + q->perturb_period;
 671 +                       add_timer(&q->perturb_timer);
 672 +               }
 673 +       }
 674 +
 675 +       q->ht = kmalloc(q->hash_divisor*sizeof(esfq_index), GFP_KERNEL);
 676 +       if (!q->ht)
 677 +               goto err_case;
 678 +
 679 +       q->dep = kmalloc((1+q->depth*2)*sizeof(struct esfq_head), GFP_KERNEL);
 680 +       if (!q->dep)
 681 +               goto err_case;
 682 +       q->next = kmalloc(q->depth*sizeof(esfq_index), GFP_KERNEL);
 683 +       if (!q->next)
 684 +               goto err_case;
 685 +
 686 +       q->allot = kmalloc(q->depth*sizeof(short), GFP_KERNEL);
 687 +       if (!q->allot)
 688 +               goto err_case;
 689 +       q->hash = kmalloc(q->depth*sizeof(unsigned short), GFP_KERNEL);
 690 +       if (!q->hash)
 691 +               goto err_case;
 692 +       q->qs = kmalloc(q->depth*sizeof(struct sk_buff_head), GFP_KERNEL);
 693 +       if (!q->qs)
 694 +               goto err_case;
 695 +
 696 +       for (i=0; i< q->hash_divisor; i++)
 697 +               q->ht[i] = q->depth;
 698 +       for (i=0; i<q->depth; i++) {
 699 +               skb_queue_head_init(&q->qs[i]);
 700 +               q->dep[i+q->depth].next = i+q->depth;
 701 +               q->dep[i+q->depth].prev = i+q->depth;
 702 +       }
 703 +
 704 +       for (i=0; i<q->depth; i++)
 705 +               esfq_link(q, i);
 706 +       return 0;
 707 +err_case:
 708 +       del_timer(&q->perturb_timer);
 709 +       if (q->ht)
 710 +               kfree(q->ht);
 711 +       if (q->dep)
 712 +               kfree(q->dep);
 713 +       if (q->next)
 714 +               kfree(q->next);
 715 +       if (q->allot)
 716 +               kfree(q->allot);
 717 +       if (q->hash)
 718 +               kfree(q->hash);
 719 +       if (q->qs)
 720 +               kfree(q->qs);
 721 +       return -ENOBUFS;
 722 +}
 723 +
 724 +static void esfq_destroy(struct Qdisc *sch)
 725 +{
 726 +       struct esfq_sched_data *q = qdisc_priv(sch);
 727 +       del_timer(&q->perturb_timer);
 728 +       if(q->ht)
 729 +               kfree(q->ht);
 730 +       if(q->dep)
 731 +               kfree(q->dep);
 732 +       if(q->next)
 733 +               kfree(q->next);
 734 +       if(q->allot)
 735 +               kfree(q->allot);
 736 +       if(q->hash)
 737 +               kfree(q->hash);
 738 +       if(q->qs)
 739 +               kfree(q->qs);
 740 +}
 741 +
 742 +static int esfq_dump(struct Qdisc *sch, struct sk_buff *skb)
 743 +{
 744 +       struct esfq_sched_data *q = qdisc_priv(sch);
 745 +       unsigned char    *b = skb->tail;
 746 +       struct tc_esfq_qopt opt;
 747 +
 748 +       opt.quantum = q->quantum;
 749 +       opt.perturb_period = q->perturb_period/HZ;
 750 +
 751 +       opt.limit = q->limit;
 752 +       opt.divisor = q->hash_divisor;
 753 +       opt.flows = q->depth;
 754 +       opt.hash_kind = q->hash_kind;
 755 +
 756 +       RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 757 +
 758 +       return skb->len;
 759 +
 760 +rtattr_failure:
 761 +       skb_trim(skb, b - skb->data);
 762 +       return -1;
 763 +}
 764 +
 765 +static struct Qdisc_ops esfq_qdisc_ops =
 766 +{
 767 +       .next           =       NULL,
 768 +       .cl_ops         =       NULL,
 769 +       .id             =       "esfq",
 770 +       .priv_size      =       sizeof(struct esfq_sched_data),
 771 +       .enqueue        =       esfq_enqueue,
 772 +       .dequeue        =       esfq_dequeue,
 773 +       .requeue        =       esfq_requeue,
 774 +       .drop           =       esfq_drop,
 775 +       .init           =       esfq_init,
 776 +       .reset          =       esfq_reset,
 777 +       .destroy        =       esfq_destroy,
 778 +       .change         =       NULL, /* esfq_change - needs more work */
 779 +       .dump           =       esfq_dump,
 780 +       .owner          =       THIS_MODULE,
 781 +};
 782 +
 783 +static int __init esfq_module_init(void)
 784 +{
 785 +       return register_qdisc(&esfq_qdisc_ops);
 786 +}
 787 +static void __exit esfq_module_exit(void)
 788 +{
 789 +       unregister_qdisc(&esfq_qdisc_ops);
 790 +}
 791 +module_init(esfq_module_init)
 792 +module_exit(esfq_module_exit)
 793 +MODULE_LICENSE("GPL");