target/linux/generic/patches-3.3/620-sched_esfq.patch

   1 --- a/include/linux/pkt_sched.h
   2 +++ b/include/linux/pkt_sched.h
   3 @@ -193,6 +193,33 @@ struct tc_sfq_xstats {
   4         __s32           allot;
   5  };
   6
   7 +/* ESFQ section */
   8 +
   9 +enum
  10 +{
  11 +        /* traditional */
  12 +       TCA_SFQ_HASH_CLASSIC,
  13 +       TCA_SFQ_HASH_DST,
  14 +       TCA_SFQ_HASH_SRC,
  15 +       TCA_SFQ_HASH_FWMARK,
  16 +       /* conntrack */
  17 +       TCA_SFQ_HASH_CTORIGDST,
  18 +       TCA_SFQ_HASH_CTORIGSRC,
  19 +       TCA_SFQ_HASH_CTREPLDST,
  20 +       TCA_SFQ_HASH_CTREPLSRC,
  21 +       TCA_SFQ_HASH_CTNATCHG,
  22 +};
  23 +
  24 +struct tc_esfq_qopt
  25 +{
  26 +       unsigned        quantum;        /* Bytes per round allocated to flow */
  27 +       int             perturb_period; /* Period of hash perturbation */
  28 +       __u32           limit;          /* Maximal packets in queue */
  29 +       unsigned        divisor;        /* Hash divisor  */
  30 +       unsigned        flows;          /* Maximal number of flows  */
  31 +       unsigned        hash_kind;      /* Hash function to use for flow identification */
  32 +};
  33 +
  34  /* RED section */
  35
  36  enum {
  37 --- a/net/sched/Kconfig
  38 +++ b/net/sched/Kconfig
  39 @@ -148,6 +148,37 @@ config NET_SCH_SFQ
  40           To compile this code as a module, choose M here: the
  41           module will be called sch_sfq.
  42
  43 +config NET_SCH_ESFQ
  44 +       tristate "Enhanced Stochastic Fairness Queueing (ESFQ)"
  45 +       ---help---
  46 +         Say Y here if you want to use the Enhanced Stochastic Fairness
  47 +         Queueing (ESFQ) packet scheduling algorithm for some of your network
  48 +         devices or as a leaf discipline for a classful qdisc such as HTB or
  49 +         CBQ (see the top of <file:net/sched/sch_esfq.c> for details and
  50 +         references to the SFQ algorithm).
  51 +
  52 +         This is an enchanced SFQ version which allows you to control some
  53 +         hardcoded values in the SFQ scheduler.
  54 +
  55 +         ESFQ also adds control of the hash function used to identify packet
  56 +         flows. The original SFQ discipline hashes by connection; ESFQ add
  57 +         several other hashing methods, such as by src IP or by dst IP, which
  58 +         can be more fair to users in some networking situations.
  59 +
  60 +         To compile this code as a module, choose M here: the
  61 +         module will be called sch_esfq.
  62 +
  63 +config NET_SCH_ESFQ_NFCT
  64 +       bool "Connection Tracking Hash Types"
  65 +       depends on NET_SCH_ESFQ && NF_CONNTRACK
  66 +       ---help---
  67 +         Say Y here to enable support for hashing based on netfilter connection
  68 +         tracking information. This is useful for a router that is also using
  69 +         NAT to connect privately-addressed hosts to the Internet. If you want
  70 +         to provide fair distribution of upstream bandwidth, ESFQ must use
  71 +         connection tracking information, since all outgoing packets will share
  72 +         the same source address.
  73 +
  74  config NET_SCH_TEQL
  75         tristate "True Link Equalizer (TEQL)"
  76         ---help---
  77 --- a/net/sched/Makefile
  78 +++ b/net/sched/Makefile
  79 @@ -26,6 +26,7 @@ obj-$(CONFIG_NET_SCH_INGRESS) += sch_ing
  80  obj-$(CONFIG_NET_SCH_DSMARK)   += sch_dsmark.o
  81  obj-$(CONFIG_NET_SCH_SFB)      += sch_sfb.o
  82  obj-$(CONFIG_NET_SCH_SFQ)      += sch_sfq.o
  83 +obj-$(CONFIG_NET_SCH_ESFQ)     += sch_esfq.o
  84  obj-$(CONFIG_NET_SCH_TBF)      += sch_tbf.o
  85  obj-$(CONFIG_NET_SCH_TEQL)     += sch_teql.o
  86  obj-$(CONFIG_NET_SCH_PRIO)     += sch_prio.o
  87 --- /dev/null
  88 +++ b/net/sched/sch_esfq.c
  89 @@ -0,0 +1,702 @@
  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 usage of jhash.
 111 + *             Added conntrack support.
 112 + *             Added ctnatchg hash (thanks to Ben Pfountz).
 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 <net/netlink.h>
 137 +#include <linux/ipv6.h>
 138 +#include <net/route.h>
 139 +#include <linux/skbuff.h>
 140 +#include <net/sock.h>
 141 +#include <net/pkt_sched.h>
 142 +#include <linux/jhash.h>
 143 +#ifdef CONFIG_NET_SCH_ESFQ_NFCT
 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 +       ctorigdst:      original destination IP address
 157 +       ctorigsrc:      original source IP address
 158 +       ctrepldst:      reply destination IP address
 159 +       ctreplsrc:      reply source IP
 160 +
 161 +*/
 162 +
 163 +#define ESFQ_HEAD 0
 164 +#define ESFQ_TAIL 1
 165 +
 166 +/* This type should contain at least SFQ_DEPTH*2 values */
 167 +typedef unsigned int esfq_index;
 168 +
 169 +struct esfq_head
 170 +{
 171 +       esfq_index      next;
 172 +       esfq_index      prev;
 173 +};
 174 +
 175 +struct esfq_sched_data
 176 +{
 177 +/* Parameters */
 178 +       int             perturb_period;
 179 +       unsigned        quantum;        /* Allotment per round: MUST BE >= MTU */
 180 +       int             limit;
 181 +       unsigned        depth;
 182 +       unsigned        hash_divisor;
 183 +       unsigned        hash_kind;
 184 +/* Variables */
 185 +       struct timer_list perturb_timer;
 186 +       int             perturbation;
 187 +       esfq_index      tail;           /* Index of current slot in round */
 188 +       esfq_index      max_depth;      /* Maximal depth */
 189 +
 190 +       esfq_index      *ht;                    /* Hash table */
 191 +       esfq_index      *next;                  /* Active slots link */
 192 +       short           *allot;                 /* Current allotment per slot */
 193 +       unsigned short  *hash;                  /* Hash value indexed by slots */
 194 +       struct sk_buff_head     *qs;            /* Slot queue */
 195 +       struct esfq_head        *dep;           /* Linked list of slots, indexed by depth */
 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 +static __inline__ unsigned esfq_jhash_1word(struct esfq_sched_data *q,u32 a)
 212 +{
 213 +       return jhash_1word(a, q->perturbation) & (q->hash_divisor-1);
 214 +}
 215 +
 216 +static __inline__ unsigned esfq_jhash_2words(struct esfq_sched_data *q, u32 a, u32 b)
 217 +{
 218 +       return jhash_2words(a, b, q->perturbation) & (q->hash_divisor-1);
 219 +}
 220 +
 221 +static __inline__ unsigned esfq_jhash_3words(struct esfq_sched_data *q, u32 a, u32 b, u32 c)
 222 +{
 223 +       return jhash_3words(a, b, c, q->perturbation) & (q->hash_divisor-1);
 224 +}
 225 +
 226 +static unsigned esfq_hash(struct esfq_sched_data *q, struct sk_buff *skb)
 227 +{
 228 +       struct esfq_packet_info info;
 229 +#ifdef CONFIG_NET_SCH_ESFQ_NFCT
 230 +       enum ip_conntrack_info ctinfo;
 231 +       struct nf_conn *ct = nf_ct_get(skb, &ctinfo);
 232 +#endif
 233 +
 234 +       switch (skb->protocol) {
 235 +       case __constant_htons(ETH_P_IP):
 236 +       {
 237 +               struct iphdr *iph = ip_hdr(skb);
 238 +               info.dst = iph->daddr;
 239 +               info.src = iph->saddr;
 240 +               if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) &&
 241 +                   (iph->protocol == IPPROTO_TCP ||
 242 +                    iph->protocol == IPPROTO_UDP ||
 243 +                    iph->protocol == IPPROTO_SCTP ||
 244 +                    iph->protocol == IPPROTO_DCCP ||
 245 +                    iph->protocol == IPPROTO_ESP))
 246 +                       info.proto = *(((u32*)iph) + iph->ihl);
 247 +               else
 248 +                       info.proto = iph->protocol;
 249 +               break;
 250 +       }
 251 +       case __constant_htons(ETH_P_IPV6):
 252 +       {
 253 +               struct ipv6hdr *iph = ipv6_hdr(skb);
 254 +               /* Hash ipv6 addresses into a u32. This isn't ideal,
 255 +                * but the code is simple. */
 256 +               info.dst = jhash2(iph->daddr.s6_addr32, 4, q->perturbation);
 257 +               info.src = jhash2(iph->saddr.s6_addr32, 4, q->perturbation);
 258 +               if (iph->nexthdr == IPPROTO_TCP ||
 259 +                   iph->nexthdr == IPPROTO_UDP ||
 260 +                   iph->nexthdr == IPPROTO_SCTP ||
 261 +                   iph->nexthdr == IPPROTO_DCCP ||
 262 +                   iph->nexthdr == IPPROTO_ESP)
 263 +                       info.proto = *(u32*)&iph[1];
 264 +               else
 265 +                       info.proto = iph->nexthdr;
 266 +               break;
 267 +       }
 268 +       default:
 269 +               info.dst   = (u32)(unsigned long)skb_dst(skb);
 270 +               info.src   = (u32)(unsigned long)skb->sk;
 271 +               info.proto = skb->protocol;
 272 +       }
 273 +
 274 +       info.mark = skb->mark;
 275 +
 276 +#ifdef CONFIG_NET_SCH_ESFQ_NFCT
 277 +       /* defaults if there is no conntrack info */
 278 +       info.ctorigsrc = info.src;
 279 +       info.ctorigdst = info.dst;
 280 +       info.ctreplsrc = info.dst;
 281 +       info.ctrepldst = info.src;
 282 +       /* collect conntrack info */
 283 +       if (ct && ct != &nf_conntrack_untracked) {
 284 +               if (skb->protocol == __constant_htons(ETH_P_IP)) {
 285 +                       info.ctorigsrc = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip;
 286 +                       info.ctorigdst = ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip;
 287 +                       info.ctreplsrc = ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip;
 288 +                       info.ctrepldst = ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip;
 289 +               }
 290 +               else if (skb->protocol == __constant_htons(ETH_P_IPV6)) {
 291 +                       /* Again, hash ipv6 addresses into a single u32. */
 292 +                       info.ctorigsrc = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip6, 4, q->perturbation);
 293 +                       info.ctorigdst = jhash2(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.ip6, 4, q->perturbation);
 294 +                       info.ctreplsrc = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.src.u3.ip6, 4, q->perturbation);
 295 +                       info.ctrepldst = jhash2(ct->tuplehash[IP_CT_DIR_REPLY].tuple.dst.u3.ip6, 4, q->perturbation);
 296 +               }
 297 +
 298 +       }
 299 +#endif
 300 +
 301 +       switch(q->hash_kind) {
 302 +       case TCA_SFQ_HASH_CLASSIC:
 303 +               return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 304 +       case TCA_SFQ_HASH_DST:
 305 +               return esfq_jhash_1word(q, info.dst);
 306 +       case TCA_SFQ_HASH_SRC:
 307 +               return esfq_jhash_1word(q, info.src);
 308 +       case TCA_SFQ_HASH_FWMARK:
 309 +               return esfq_jhash_1word(q, info.mark);
 310 +#ifdef CONFIG_NET_SCH_ESFQ_NFCT
 311 +       case TCA_SFQ_HASH_CTORIGDST:
 312 +               return esfq_jhash_1word(q, info.ctorigdst);
 313 +       case TCA_SFQ_HASH_CTORIGSRC:
 314 +               return esfq_jhash_1word(q, info.ctorigsrc);
 315 +       case TCA_SFQ_HASH_CTREPLDST:
 316 +               return esfq_jhash_1word(q, info.ctrepldst);
 317 +       case TCA_SFQ_HASH_CTREPLSRC:
 318 +               return esfq_jhash_1word(q, info.ctreplsrc);
 319 +       case TCA_SFQ_HASH_CTNATCHG:
 320 +       {
 321 +               if (info.ctorigdst == info.ctreplsrc)
 322 +                       return esfq_jhash_1word(q, info.ctorigsrc);
 323 +               return esfq_jhash_1word(q, info.ctreplsrc);
 324 +       }
 325 +#endif
 326 +       default:
 327 +               if (net_ratelimit())
 328 +                       printk(KERN_WARNING "ESFQ: Unknown hash method. Falling back to classic.\n");
 329 +       }
 330 +       return esfq_jhash_3words(q, info.dst, info.src, info.proto);
 331 +}
 332 +
 333 +static inline void esfq_link(struct esfq_sched_data *q, esfq_index x)
 334 +{
 335 +       esfq_index p, n;
 336 +       int d = q->qs[x].qlen + q->depth;
 337 +
 338 +       p = d;
 339 +       n = q->dep[d].next;
 340 +       q->dep[x].next = n;
 341 +       q->dep[x].prev = p;
 342 +       q->dep[p].next = q->dep[n].prev = x;
 343 +}
 344 +
 345 +static inline void esfq_dec(struct esfq_sched_data *q, esfq_index x)
 346 +{
 347 +       esfq_index p, n;
 348 +
 349 +       n = q->dep[x].next;
 350 +       p = q->dep[x].prev;
 351 +       q->dep[p].next = n;
 352 +       q->dep[n].prev = p;
 353 +
 354 +       if (n == p && q->max_depth == q->qs[x].qlen + 1)
 355 +               q->max_depth--;
 356 +
 357 +       esfq_link(q, x);
 358 +}
 359 +
 360 +static inline void esfq_inc(struct esfq_sched_data *q, esfq_index x)
 361 +{
 362 +       esfq_index p, n;
 363 +       int d;
 364 +
 365 +       n = q->dep[x].next;
 366 +       p = q->dep[x].prev;
 367 +       q->dep[p].next = n;
 368 +       q->dep[n].prev = p;
 369 +       d = q->qs[x].qlen;
 370 +       if (q->max_depth < d)
 371 +               q->max_depth = d;
 372 +
 373 +       esfq_link(q, x);
 374 +}
 375 +
 376 +static unsigned int esfq_drop(struct Qdisc *sch)
 377 +{
 378 +       struct esfq_sched_data *q = qdisc_priv(sch);
 379 +       esfq_index d = q->max_depth;
 380 +       struct sk_buff *skb;
 381 +       unsigned int len;
 382 +
 383 +       /* Queue is full! Find the longest slot and
 384 +          drop a packet from it */
 385 +
 386 +       if (d > 1) {
 387 +               esfq_index x = q->dep[d+q->depth].next;
 388 +               skb = q->qs[x].prev;
 389 +               len = skb->len;
 390 +               __skb_unlink(skb, &q->qs[x]);
 391 +               kfree_skb(skb);
 392 +               esfq_dec(q, x);
 393 +               sch->q.qlen--;
 394 +               sch->qstats.drops++;
 395 +               sch->qstats.backlog -= len;
 396 +               return len;
 397 +       }
 398 +
 399 +       if (d == 1) {
 400 +               /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
 401 +               d = q->next[q->tail];
 402 +               q->next[q->tail] = q->next[d];
 403 +               q->allot[q->next[d]] += q->quantum;
 404 +               skb = q->qs[d].prev;
 405 +               len = skb->len;
 406 +               __skb_unlink(skb, &q->qs[d]);
 407 +               kfree_skb(skb);
 408 +               esfq_dec(q, d);
 409 +               sch->q.qlen--;
 410 +               q->ht[q->hash[d]] = q->depth;
 411 +               sch->qstats.drops++;
 412 +               sch->qstats.backlog -= len;
 413 +               return len;
 414 +       }
 415 +
 416 +       return 0;
 417 +}
 418 +
 419 +static void esfq_q_enqueue(struct sk_buff *skb, struct esfq_sched_data *q, unsigned int end)
 420 +{
 421 +       unsigned hash = esfq_hash(q, skb);
 422 +       unsigned depth = q->depth;
 423 +       esfq_index x;
 424 +
 425 +       x = q->ht[hash];
 426 +       if (x == depth) {
 427 +               q->ht[hash] = x = q->dep[depth].next;
 428 +               q->hash[x] = hash;
 429 +       }
 430 +
 431 +       if (end == ESFQ_TAIL)
 432 +               __skb_queue_tail(&q->qs[x], skb);
 433 +       else
 434 +               __skb_queue_head(&q->qs[x], skb);
 435 +
 436 +       esfq_inc(q, x);
 437 +       if (q->qs[x].qlen == 1) {               /* The flow is new */
 438 +               if (q->tail == depth) { /* It is the first flow */
 439 +                       q->tail = x;
 440 +                       q->next[x] = x;
 441 +                       q->allot[x] = q->quantum;
 442 +               } else {
 443 +                       q->next[x] = q->next[q->tail];
 444 +                       q->next[q->tail] = x;
 445 +                       q->tail = x;
 446 +               }
 447 +       }
 448 +}
 449 +
 450 +static int esfq_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 451 +{
 452 +       struct esfq_sched_data *q = qdisc_priv(sch);
 453 +       esfq_q_enqueue(skb, q, ESFQ_TAIL);
 454 +       sch->qstats.backlog += skb->len;
 455 +       if (++sch->q.qlen < q->limit-1) {
 456 +               sch->bstats.bytes += skb->len;
 457 +               sch->bstats.packets++;
 458 +               return 0;
 459 +       }
 460 +
 461 +       sch->qstats.drops++;
 462 +       esfq_drop(sch);
 463 +       return NET_XMIT_CN;
 464 +}
 465 +
 466 +static struct sk_buff *esfq_peek(struct Qdisc* sch)
 467 +{
 468 +       struct esfq_sched_data *q = qdisc_priv(sch);
 469 +       esfq_index a;
 470 +
 471 +       /* No active slots */
 472 +       if (q->tail == q->depth)
 473 +               return NULL;
 474 +
 475 +       a = q->next[q->tail];
 476 +       return skb_peek(&q->qs[a]);
 477 +}
 478 +
 479 +static struct sk_buff *esfq_q_dequeue(struct esfq_sched_data *q)
 480 +{
 481 +       struct sk_buff *skb;
 482 +       unsigned depth = q->depth;
 483 +       esfq_index a, old_a;
 484 +
 485 +       /* No active slots */
 486 +       if (q->tail == depth)
 487 +               return NULL;
 488 +
 489 +       a = old_a = q->next[q->tail];
 490 +
 491 +       /* Grab packet */
 492 +       skb = __skb_dequeue(&q->qs[a]);
 493 +       esfq_dec(q, a);
 494 +
 495 +       /* Is the slot empty? */
 496 +       if (q->qs[a].qlen == 0) {
 497 +               q->ht[q->hash[a]] = depth;
 498 +               a = q->next[a];
 499 +               if (a == old_a) {
 500 +                       q->tail = depth;
 501 +                       return skb;
 502 +               }
 503 +               q->next[q->tail] = a;
 504 +               q->allot[a] += q->quantum;
 505 +       } else if ((q->allot[a] -= skb->len) <= 0) {
 506 +               q->tail = a;
 507 +               a = q->next[a];
 508 +               q->allot[a] += q->quantum;
 509 +       }
 510 +
 511 +       return skb;
 512 +}
 513 +
 514 +static struct sk_buff *esfq_dequeue(struct Qdisc* sch)
 515 +{
 516 +       struct esfq_sched_data *q = qdisc_priv(sch);
 517 +       struct sk_buff *skb;
 518 +
 519 +       skb = esfq_q_dequeue(q);
 520 +       if (skb == NULL)
 521 +               return NULL;
 522 +       sch->q.qlen--;
 523 +       sch->qstats.backlog -= skb->len;
 524 +       return skb;
 525 +}
 526 +
 527 +static void esfq_q_destroy(struct esfq_sched_data *q)
 528 +{
 529 +       del_timer(&q->perturb_timer);
 530 +       if(q->ht)
 531 +               kfree(q->ht);
 532 +       if(q->dep)
 533 +               kfree(q->dep);
 534 +       if(q->next)
 535 +               kfree(q->next);
 536 +       if(q->allot)
 537 +               kfree(q->allot);
 538 +       if(q->hash)
 539 +               kfree(q->hash);
 540 +       if(q->qs)
 541 +               kfree(q->qs);
 542 +}
 543 +
 544 +static void esfq_destroy(struct Qdisc *sch)
 545 +{
 546 +       struct esfq_sched_data *q = qdisc_priv(sch);
 547 +       esfq_q_destroy(q);
 548 +}
 549 +
 550 +
 551 +static void esfq_reset(struct Qdisc* sch)
 552 +{
 553 +       struct sk_buff *skb;
 554 +
 555 +       while ((skb = esfq_dequeue(sch)) != NULL)
 556 +               kfree_skb(skb);
 557 +}
 558 +
 559 +static void esfq_perturbation(unsigned long arg)
 560 +{
 561 +       struct Qdisc *sch = (struct Qdisc*)arg;
 562 +       struct esfq_sched_data *q = qdisc_priv(sch);
 563 +
 564 +       q->perturbation = net_random()&0x1F;
 565 +
 566 +       if (q->perturb_period) {
 567 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 568 +               add_timer(&q->perturb_timer);
 569 +       }
 570 +}
 571 +
 572 +static unsigned int esfq_check_hash(unsigned int kind)
 573 +{
 574 +       switch (kind) {
 575 +       case TCA_SFQ_HASH_CTORIGDST:
 576 +       case TCA_SFQ_HASH_CTORIGSRC:
 577 +       case TCA_SFQ_HASH_CTREPLDST:
 578 +       case TCA_SFQ_HASH_CTREPLSRC:
 579 +       case TCA_SFQ_HASH_CTNATCHG:
 580 +#ifndef CONFIG_NET_SCH_ESFQ_NFCT
 581 +       {
 582 +               if (net_ratelimit())
 583 +                       printk(KERN_WARNING "ESFQ: Conntrack hash types disabled in kernel config. Falling back to classic.\n");
 584 +               return TCA_SFQ_HASH_CLASSIC;
 585 +       }
 586 +#endif
 587 +       case TCA_SFQ_HASH_CLASSIC:
 588 +       case TCA_SFQ_HASH_DST:
 589 +       case TCA_SFQ_HASH_SRC:
 590 +       case TCA_SFQ_HASH_FWMARK:
 591 +               return kind;
 592 +       default:
 593 +       {
 594 +               if (net_ratelimit())
 595 +                       printk(KERN_WARNING "ESFQ: Unknown hash type. Falling back to classic.\n");
 596 +               return TCA_SFQ_HASH_CLASSIC;
 597 +       }
 598 +       }
 599 +}
 600 +
 601 +static int esfq_q_init(struct esfq_sched_data *q, struct nlattr *opt)
 602 +{
 603 +       struct tc_esfq_qopt *ctl = nla_data(opt);
 604 +       esfq_index p = ~0U/2;
 605 +       int i;
 606 +
 607 +       if (opt && opt->nla_len < nla_attr_size(sizeof(*ctl)))
 608 +               return -EINVAL;
 609 +
 610 +       q->perturbation = 0;
 611 +       q->hash_kind = TCA_SFQ_HASH_CLASSIC;
 612 +       q->max_depth = 0;
 613 +       if (opt == NULL) {
 614 +               q->perturb_period = 0;
 615 +               q->hash_divisor = 1024;
 616 +               q->tail = q->limit = q->depth = 128;
 617 +
 618 +       } else {
 619 +               struct tc_esfq_qopt *ctl = nla_data(opt);
 620 +               if (ctl->quantum)
 621 +                       q->quantum = ctl->quantum;
 622 +               q->perturb_period = ctl->perturb_period*HZ;
 623 +               q->hash_divisor = ctl->divisor ? : 1024;
 624 +               q->tail = q->limit = q->depth = ctl->flows ? : 128;
 625 +
 626 +               if ( q->depth > p - 1 )
 627 +                       return -EINVAL;
 628 +
 629 +               if (ctl->limit)
 630 +                       q->limit = min_t(u32, ctl->limit, q->depth);
 631 +
 632 +               if (ctl->hash_kind) {
 633 +                       q->hash_kind = esfq_check_hash(ctl->hash_kind);
 634 +               }
 635 +       }
 636 +
 637 +       q->ht = kmalloc(q->hash_divisor*sizeof(esfq_index), GFP_KERNEL);
 638 +       if (!q->ht)
 639 +               goto err_case;
 640 +       q->dep = kmalloc((1+q->depth*2)*sizeof(struct esfq_head), GFP_KERNEL);
 641 +       if (!q->dep)
 642 +               goto err_case;
 643 +       q->next = kmalloc(q->depth*sizeof(esfq_index), GFP_KERNEL);
 644 +       if (!q->next)
 645 +               goto err_case;
 646 +       q->allot = kmalloc(q->depth*sizeof(short), GFP_KERNEL);
 647 +       if (!q->allot)
 648 +               goto err_case;
 649 +       q->hash = kmalloc(q->depth*sizeof(unsigned short), GFP_KERNEL);
 650 +       if (!q->hash)
 651 +               goto err_case;
 652 +       q->qs = kmalloc(q->depth*sizeof(struct sk_buff_head), GFP_KERNEL);
 653 +       if (!q->qs)
 654 +               goto err_case;
 655 +
 656 +       for (i=0; i< q->hash_divisor; i++)
 657 +               q->ht[i] = q->depth;
 658 +       for (i=0; i<q->depth; i++) {
 659 +               skb_queue_head_init(&q->qs[i]);
 660 +               q->dep[i+q->depth].next = i+q->depth;
 661 +               q->dep[i+q->depth].prev = i+q->depth;
 662 +       }
 663 +
 664 +       for (i=0; i<q->depth; i++)
 665 +               esfq_link(q, i);
 666 +       return 0;
 667 +err_case:
 668 +       esfq_q_destroy(q);
 669 +       return -ENOBUFS;
 670 +}
 671 +
 672 +static int esfq_init(struct Qdisc *sch, struct nlattr *opt)
 673 +{
 674 +       struct esfq_sched_data *q = qdisc_priv(sch);
 675 +       int err;
 676 +
 677 +       q->quantum = psched_mtu(qdisc_dev(sch)); /* default */
 678 +       if ((err = esfq_q_init(q, opt)))
 679 +               return err;
 680 +
 681 +       init_timer(&q->perturb_timer);
 682 +       q->perturb_timer.data = (unsigned long)sch;
 683 +       q->perturb_timer.function = esfq_perturbation;
 684 +       if (q->perturb_period) {
 685 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 686 +               add_timer(&q->perturb_timer);
 687 +       }
 688 +
 689 +       return 0;
 690 +}
 691 +
 692 +static int esfq_change(struct Qdisc *sch, struct nlattr *opt)
 693 +{
 694 +       struct esfq_sched_data *q = qdisc_priv(sch);
 695 +       struct esfq_sched_data new;
 696 +       struct sk_buff *skb;
 697 +       int err;
 698 +
 699 +       /* set up new queue */
 700 +       memset(&new, 0, sizeof(struct esfq_sched_data));
 701 +       new.quantum = psched_mtu(qdisc_dev(sch)); /* default */
 702 +       if ((err = esfq_q_init(&new, opt)))
 703 +               return err;
 704 +
 705 +       /* copy all packets from the old queue to the new queue */
 706 +       sch_tree_lock(sch);
 707 +       while ((skb = esfq_q_dequeue(q)) != NULL)
 708 +               esfq_q_enqueue(skb, &new, ESFQ_TAIL);
 709 +
 710 +       /* clean up the old queue */
 711 +       esfq_q_destroy(q);
 712 +
 713 +       /* copy elements of the new queue into the old queue */
 714 +       q->perturb_period = new.perturb_period;
 715 +       q->quantum        = new.quantum;
 716 +       q->limit          = new.limit;
 717 +       q->depth          = new.depth;
 718 +       q->hash_divisor   = new.hash_divisor;
 719 +       q->hash_kind      = new.hash_kind;
 720 +       q->tail           = new.tail;
 721 +       q->max_depth      = new.max_depth;
 722 +       q->ht    = new.ht;
 723 +       q->dep   = new.dep;
 724 +       q->next  = new.next;
 725 +       q->allot = new.allot;
 726 +       q->hash  = new.hash;
 727 +       q->qs    = new.qs;
 728 +
 729 +       /* finish up */
 730 +       if (q->perturb_period) {
 731 +               q->perturb_timer.expires = jiffies + q->perturb_period;
 732 +               add_timer(&q->perturb_timer);
 733 +       } else {
 734 +               q->perturbation = 0;
 735 +       }
 736 +       sch_tree_unlock(sch);
 737 +       return 0;
 738 +}
 739 +
 740 +static int esfq_dump(struct Qdisc *sch, struct sk_buff *skb)
 741 +{
 742 +       struct esfq_sched_data *q = qdisc_priv(sch);
 743 +       unsigned char *b = skb_tail_pointer(skb);
 744 +       struct tc_esfq_qopt opt;
 745 +
 746 +       opt.quantum = q->quantum;
 747 +       opt.perturb_period = q->perturb_period/HZ;
 748 +
 749 +       opt.limit = q->limit;
 750 +       opt.divisor = q->hash_divisor;
 751 +       opt.flows = q->depth;
 752 +       opt.hash_kind = q->hash_kind;
 753 +
 754 +       NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 755 +
 756 +       return skb->len;
 757 +
 758 +nla_put_failure:
 759 +       nlmsg_trim(skb, b);
 760 +       return -1;
 761 +}
 762 +
 763 +static struct Qdisc_ops esfq_qdisc_ops =
 764 +{
 765 +       .next           =       NULL,
 766 +       .cl_ops         =       NULL,
 767 +       .id             =       "esfq",
 768 +       .priv_size      =       sizeof(struct esfq_sched_data),
 769 +       .enqueue        =       esfq_enqueue,
 770 +       .dequeue        =       esfq_dequeue,
 771 +       .peek           =       esfq_peek,
 772 +       .drop           =       esfq_drop,
 773 +       .init           =       esfq_init,
 774 +       .reset          =       esfq_reset,
 775 +       .destroy        =       esfq_destroy,
 776 +       .change         =       esfq_change,
 777 +       .dump           =       esfq_dump,
 778 +       .owner          =       THIS_MODULE,
 779 +};
 780 +
 781 +static int __init esfq_module_init(void)
 782 +{
 783 +       return register_qdisc(&esfq_qdisc_ops);
 784 +}
 785 +static void __exit esfq_module_exit(void)
 786 +{
 787 +       unregister_qdisc(&esfq_qdisc_ops);
 788 +}
 789 +module_init(esfq_module_init)
 790 +module_exit(esfq_module_exit)
 791 +MODULE_LICENSE("GPL");