net/sched/sch_hfsc.c

   1 /*
   2  * Copyright (c) 2003 Patrick McHardy, <kaber@trash.net>
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public License
   6  * as published by the Free Software Foundation; either version 2
   7  * of the License, or (at your option) any later version.
   8  *
   9  * 2003-10-17 - Ported from altq
  10  */
  11 /*
  12  * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
  13  *
  14  * Permission to use, copy, modify, and distribute this software and
  15  * its documentation is hereby granted (including for commercial or
  16  * for-profit use), provided that both the copyright notice and this
  17  * permission notice appear in all copies of the software, derivative
  18  * works, or modified versions, and any portions thereof.
  19  *
  20  * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
  21  * WHICH MAY HAVE SERIOUS CONSEQUENCES.  CARNEGIE MELLON PROVIDES THIS
  22  * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
  23  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  24  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  25  * DISCLAIMED.  IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
  26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  27  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
  28  * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  29  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  30  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
  32  * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  33  * DAMAGE.
  34  *
  35  * Carnegie Mellon encourages (but does not require) users of this
  36  * software to return any improvements or extensions that they make,
  37  * and to grant Carnegie Mellon the rights to redistribute these
  38  * changes without encumbrance.
  39  */
  40 /*
  41  * H-FSC is described in Proceedings of SIGCOMM'97,
  42  * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
  43  * Real-Time and Priority Service"
  44  * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
  45  *
  46  * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
  47  * when a class has an upperlimit, the fit-time is computed from the
  48  * upperlimit service curve.  the link-sharing scheduler does not schedule
  49  * a class whose fit-time exceeds the current time.
  50  */
  51
  52 #include <linux/kernel.h>
  53 #include <linux/module.h>
  54 #include <linux/types.h>
  55 #include <linux/errno.h>
  56 #include <linux/jiffies.h>
  57 #include <linux/compiler.h>
  58 #include <linux/spinlock.h>
  59 #include <linux/skbuff.h>
  60 #include <linux/string.h>
  61 #include <linux/slab.h>
  62 #include <linux/timer.h>
  63 #include <linux/list.h>
  64 #include <linux/rbtree.h>
  65 #include <linux/init.h>
  66 #include <linux/netdevice.h>
  67 #include <linux/rtnetlink.h>
  68 #include <linux/pkt_sched.h>
  69 #include <net/pkt_sched.h>
  70 #include <net/pkt_cls.h>
  71 #include <asm/system.h>
  72 #include <asm/div64.h>
  73
  74 #define HFSC_DEBUG 1
  75
  76 /*
  77  * kernel internal service curve representation:
  78  *   coordinates are given by 64 bit unsigned integers.
  79  *   x-axis: unit is clock count.
  80  *   y-axis: unit is byte.
  81  *
  82  *   The service curve parameters are converted to the internal
  83  *   representation. The slope values are scaled to avoid overflow.
  84  *   the inverse slope values as well as the y-projection of the 1st
  85  *   segment are kept in order to to avoid 64-bit divide operations
  86  *   that are expensive on 32-bit architectures.
  87  */
  88
  89 struct internal_sc
  90 {
  91         u64     sm1;    /* scaled slope of the 1st segment */
  92         u64     ism1;   /* scaled inverse-slope of the 1st segment */
  93         u64     dx;     /* the x-projection of the 1st segment */
  94         u64     dy;     /* the y-projection of the 1st segment */
  95         u64     sm2;    /* scaled slope of the 2nd segment */
  96         u64     ism2;   /* scaled inverse-slope of the 2nd segment */
  97 };
  98
  99 /* runtime service curve */
 100 struct runtime_sc
 101 {
 102         u64     x;      /* current starting position on x-axis */
 103         u64     y;      /* current starting position on y-axis */
 104         u64     sm1;    /* scaled slope of the 1st segment */
 105         u64     ism1;   /* scaled inverse-slope of the 1st segment */
 106         u64     dx;     /* the x-projection of the 1st segment */
 107         u64     dy;     /* the y-projection of the 1st segment */
 108         u64     sm2;    /* scaled slope of the 2nd segment */
 109         u64     ism2;   /* scaled inverse-slope of the 2nd segment */
 110 };
 111
 112 enum hfsc_class_flags
 113 {
 114         HFSC_RSC = 0x1,
 115         HFSC_FSC = 0x2,
 116         HFSC_USC = 0x4
 117 };
 118
 119 struct hfsc_class
 120 {
 121         u32             classid;        /* class id */
 122         unsigned int    refcnt;         /* usage count */
 123
 124         struct gnet_stats_basic bstats;
 125         struct gnet_stats_queue qstats;
 126         struct gnet_stats_rate_est rate_est;
 127         spinlock_t      *stats_lock;
 128         unsigned int    level;          /* class level in hierarchy */
 129         struct tcf_proto *filter_list;  /* filter list */
 130         unsigned int    filter_cnt;     /* filter count */
 131
 132         struct hfsc_sched *sched;       /* scheduler data */
 133         struct hfsc_class *cl_parent;   /* parent class */
 134         struct list_head siblings;      /* sibling classes */
 135         struct list_head children;      /* child classes */
 136         struct Qdisc    *qdisc;         /* leaf qdisc */
 137
 138         struct rb_node el_node;         /* qdisc's eligible tree member */
 139         struct rb_root vt_tree;         /* active children sorted by cl_vt */
 140         struct rb_node vt_node;         /* parent's vt_tree member */
 141         struct rb_root cf_tree;         /* active children sorted by cl_f */
 142         struct rb_node cf_node;         /* parent's cf_heap member */
 143         struct list_head hlist;         /* hash list member */
 144         struct list_head dlist;         /* drop list member */
 145
 146         u64     cl_total;               /* total work in bytes */
 147         u64     cl_cumul;               /* cumulative work in bytes done by
 148                                            real-time criteria */
 149
 150         u64     cl_d;                   /* deadline*/
 151         u64     cl_e;                   /* eligible time */
 152         u64     cl_vt;                  /* virtual time */
 153         u64     cl_f;                   /* time when this class will fit for
 154                                            link-sharing, max(myf, cfmin) */
 155         u64     cl_myf;                 /* my fit-time (calculated from this
 156                                            class's own upperlimit curve) */
 157         u64     cl_myfadj;              /* my fit-time adjustment (to cancel
 158                                            history dependence) */
 159         u64     cl_cfmin;               /* earliest children's fit-time (used
 160                                            with cl_myf to obtain cl_f) */
 161         u64     cl_cvtmin;              /* minimal virtual time among the
 162                                            children fit for link-sharing
 163                                            (monotonic within a period) */
 164         u64     cl_vtadj;               /* intra-period cumulative vt
 165                                            adjustment */
 166         u64     cl_vtoff;               /* inter-period cumulative vt offset */
 167         u64     cl_cvtmax;              /* max child's vt in the last period */
 168         u64     cl_cvtoff;              /* cumulative cvtmax of all periods */
 169         u64     cl_pcvtoff;             /* parent's cvtoff at initalization
 170                                            time */
 171
 172         struct internal_sc cl_rsc;      /* internal real-time service curve */
 173         struct internal_sc cl_fsc;      /* internal fair service curve */
 174         struct internal_sc cl_usc;      /* internal upperlimit service curve */
 175         struct runtime_sc cl_deadline;  /* deadline curve */
 176         struct runtime_sc cl_eligible;  /* eligible curve */
 177         struct runtime_sc cl_virtual;   /* virtual curve */
 178         struct runtime_sc cl_ulimit;    /* upperlimit curve */
 179
 180         unsigned long   cl_flags;       /* which curves are valid */
 181         unsigned long   cl_vtperiod;    /* vt period sequence number */
 182         unsigned long   cl_parentperiod;/* parent's vt period sequence number*/
 183         unsigned long   cl_nactive;     /* number of active children */
 184 };
 185
 186 #define HFSC_HSIZE      16
 187
 188 struct hfsc_sched
 189 {
 190         u16     defcls;                         /* default class id */
 191         struct hfsc_class root;                 /* root class */
 192         struct list_head clhash[HFSC_HSIZE];    /* class hash */
 193         struct rb_root eligible;                /* eligible tree */
 194         struct list_head droplist;              /* active leaf class list (for
 195                                                    dropping) */
 196         struct sk_buff_head requeue;            /* requeued packet */
 197         struct timer_list wd_timer;             /* watchdog timer */
 198 };
 199
 200 /*
 201  * macros
 202  */
 203 #ifdef CONFIG_NET_SCH_CLK_GETTIMEOFDAY
 204 #include <linux/time.h>
 205 #undef PSCHED_GET_TIME
 206 #define PSCHED_GET_TIME(stamp)                                          \
 207 do {                                                                    \
 208         struct timeval tv;                                              \
 209         do_gettimeofday(&tv);                                           \
 210         (stamp) = 1ULL * USEC_PER_SEC * tv.tv_sec + tv.tv_usec;         \
 211 } while (0)
 212 #endif
 213
 214 #if HFSC_DEBUG
 215 #define ASSERT(cond)                                                    \
 216 do {                                                                    \
 217         if (unlikely(!(cond)))                                          \
 218                 printk("assertion %s failed at %s:%i (%s)\n",           \
 219                        #cond, __FILE__, __LINE__, __FUNCTION__);        \
 220 } while (0)
 221 #else
 222 #define ASSERT(cond)
 223 #endif /* HFSC_DEBUG */
 224
 225 #define HT_INFINITY     0xffffffffffffffffULL   /* infinite time value */
 226
 227
 228 /*
 229  * eligible tree holds backlogged classes being sorted by their eligible times.
 230  * there is one eligible tree per hfsc instance.
 231  */
 232
 233 static void
 234 eltree_insert(struct hfsc_class *cl)
 235 {
 236         struct rb_node **p = &cl->sched->eligible.rb_node;
 237         struct rb_node *parent = NULL;
 238         struct hfsc_class *cl1;
 239
 240         while (*p != NULL) {
 241                 parent = *p;
 242                 cl1 = rb_entry(parent, struct hfsc_class, el_node);
 243                 if (cl->cl_e >= cl1->cl_e)
 244                         p = &parent->rb_right;
 245                 else
 246                         p = &parent->rb_left;
 247         }
 248         rb_link_node(&cl->el_node, parent, p);
 249         rb_insert_color(&cl->el_node, &cl->sched->eligible);
 250 }
 251
 252 static inline void
 253 eltree_remove(struct hfsc_class *cl)
 254 {
 255         rb_erase(&cl->el_node, &cl->sched->eligible);
 256 }
 257
 258 static inline void
 259 eltree_update(struct hfsc_class *cl)
 260 {
 261         eltree_remove(cl);
 262         eltree_insert(cl);
 263 }
 264
 265 /* find the class with the minimum deadline among the eligible classes */
 266 static inline struct hfsc_class *
 267 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
 268 {
 269         struct hfsc_class *p, *cl = NULL;
 270         struct rb_node *n;
 271
 272         for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
 273                 p = rb_entry(n, struct hfsc_class, el_node);
 274                 if (p->cl_e > cur_time)
 275                         break;
 276                 if (cl == NULL || p->cl_d < cl->cl_d)
 277                         cl = p;
 278         }
 279         return cl;
 280 }
 281
 282 /* find the class with minimum eligible time among the eligible classes */
 283 static inline struct hfsc_class *
 284 eltree_get_minel(struct hfsc_sched *q)
 285 {
 286         struct rb_node *n;
 287
 288         n = rb_first(&q->eligible);
 289         if (n == NULL)
 290                 return NULL;
 291         return rb_entry(n, struct hfsc_class, el_node);
 292 }
 293
 294 /*
 295  * vttree holds holds backlogged child classes being sorted by their virtual
 296  * time. each intermediate class has one vttree.
 297  */
 298 static void
 299 vttree_insert(struct hfsc_class *cl)
 300 {
 301         struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
 302         struct rb_node *parent = NULL;
 303         struct hfsc_class *cl1;
 304
 305         while (*p != NULL) {
 306                 parent = *p;
 307                 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
 308                 if (cl->cl_vt >= cl1->cl_vt)
 309                         p = &parent->rb_right;
 310                 else
 311                         p = &parent->rb_left;
 312         }
 313         rb_link_node(&cl->vt_node, parent, p);
 314         rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
 315 }
 316
 317 static inline void
 318 vttree_remove(struct hfsc_class *cl)
 319 {
 320         rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
 321 }
 322
 323 static inline void
 324 vttree_update(struct hfsc_class *cl)
 325 {
 326         vttree_remove(cl);
 327         vttree_insert(cl);
 328 }
 329
 330 static inline struct hfsc_class *
 331 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
 332 {
 333         struct hfsc_class *p;
 334         struct rb_node *n;
 335
 336         for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
 337                 p = rb_entry(n, struct hfsc_class, vt_node);
 338                 if (p->cl_f <= cur_time)
 339                         return p;
 340         }
 341         return NULL;
 342 }
 343
 344 /*
 345  * get the leaf class with the minimum vt in the hierarchy
 346  */
 347 static struct hfsc_class *
 348 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
 349 {
 350         /* if root-class's cfmin is bigger than cur_time nothing to do */
 351         if (cl->cl_cfmin > cur_time)
 352                 return NULL;
 353
 354         while (cl->level > 0) {
 355                 cl = vttree_firstfit(cl, cur_time);
 356                 if (cl == NULL)
 357                         return NULL;
 358                 /*
 359                  * update parent's cl_cvtmin.
 360                  */
 361                 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
 362                         cl->cl_parent->cl_cvtmin = cl->cl_vt;
 363         }
 364         return cl;
 365 }
 366
 367 static void
 368 cftree_insert(struct hfsc_class *cl)
 369 {
 370         struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
 371         struct rb_node *parent = NULL;
 372         struct hfsc_class *cl1;
 373
 374         while (*p != NULL) {
 375                 parent = *p;
 376                 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
 377                 if (cl->cl_f >= cl1->cl_f)
 378                         p = &parent->rb_right;
 379                 else
 380                         p = &parent->rb_left;
 381         }
 382         rb_link_node(&cl->cf_node, parent, p);
 383         rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
 384 }
 385
 386 static inline void
 387 cftree_remove(struct hfsc_class *cl)
 388 {
 389         rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
 390 }
 391
 392 static inline void
 393 cftree_update(struct hfsc_class *cl)
 394 {
 395         cftree_remove(cl);
 396         cftree_insert(cl);
 397 }
 398
 399 /*
 400  * service curve support functions
 401  *
 402  *  external service curve parameters
 403  *      m: bps
 404  *      d: us
 405  *  internal service curve parameters
 406  *      sm: (bytes/psched_us) << SM_SHIFT
 407  *      ism: (psched_us/byte) << ISM_SHIFT
 408  *      dx: psched_us
 409  *
 410  * Clock source resolution (CONFIG_NET_SCH_CLK_*)
 411  *  JIFFIES: for 48<=HZ<=1534 resolution is between 0.63us and 1.27us.
 412  *  CPU: resolution is between 0.5us and 1us.
 413  *  GETTIMEOFDAY: resolution is exactly 1us.
 414  *
 415  * sm and ism are scaled in order to keep effective digits.
 416  * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
 417  * digits in decimal using the following table.
 418  *
 419  * Note: We can afford the additional accuracy (altq hfsc keeps at most
 420  * 3 effective digits) thanks to the fact that linux clock is bounded
 421  * much more tightly.
 422  *
 423  *  bits/sec      100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
 424  *  ------------+-------------------------------------------------------
 425  *  bytes/0.5us   6.25e-3    62.5e-3    625e-3     6250e-e    62500e-3
 426  *  bytes/us      12.5e-3    125e-3     1250e-3    12500e-3   125000e-3
 427  *  bytes/1.27us  15.875e-3  158.75e-3  1587.5e-3  15875e-3   158750e-3
 428  *
 429  *  0.5us/byte    160        16         1.6        0.16       0.016
 430  *  us/byte       80         8          0.8        0.08       0.008
 431  *  1.27us/byte   63         6.3        0.63       0.063      0.0063
 432  */
 433 #define SM_SHIFT        20
 434 #define ISM_SHIFT       18
 435
 436 #define SM_MASK         ((1ULL << SM_SHIFT) - 1)
 437 #define ISM_MASK        ((1ULL << ISM_SHIFT) - 1)
 438
 439 static inline u64
 440 seg_x2y(u64 x, u64 sm)
 441 {
 442         u64 y;
 443
 444         /*
 445          * compute
 446          *      y = x * sm >> SM_SHIFT
 447          * but divide it for the upper and lower bits to avoid overflow
 448          */
 449         y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
 450         return y;
 451 }
 452
 453 static inline u64
 454 seg_y2x(u64 y, u64 ism)
 455 {
 456         u64 x;
 457
 458         if (y == 0)
 459                 x = 0;
 460         else if (ism == HT_INFINITY)
 461                 x = HT_INFINITY;
 462         else {
 463                 x = (y >> ISM_SHIFT) * ism
 464                     + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
 465         }
 466         return x;
 467 }
 468
 469 /* Convert m (bps) into sm (bytes/psched us) */
 470 static u64
 471 m2sm(u32 m)
 472 {
 473         u64 sm;
 474
 475         sm = ((u64)m << SM_SHIFT);
 476         sm += PSCHED_JIFFIE2US(HZ) - 1;
 477         do_div(sm, PSCHED_JIFFIE2US(HZ));
 478         return sm;
 479 }
 480
 481 /* convert m (bps) into ism (psched us/byte) */
 482 static u64
 483 m2ism(u32 m)
 484 {
 485         u64 ism;
 486
 487         if (m == 0)
 488                 ism = HT_INFINITY;
 489         else {
 490                 ism = ((u64)PSCHED_JIFFIE2US(HZ) << ISM_SHIFT);
 491                 ism += m - 1;
 492                 do_div(ism, m);
 493         }
 494         return ism;
 495 }
 496
 497 /* convert d (us) into dx (psched us) */
 498 static u64
 499 d2dx(u32 d)
 500 {
 501         u64 dx;
 502
 503         dx = ((u64)d * PSCHED_JIFFIE2US(HZ));
 504         dx += USEC_PER_SEC - 1;
 505         do_div(dx, USEC_PER_SEC);
 506         return dx;
 507 }
 508
 509 /* convert sm (bytes/psched us) into m (bps) */
 510 static u32
 511 sm2m(u64 sm)
 512 {
 513         u64 m;
 514
 515         m = (sm * PSCHED_JIFFIE2US(HZ)) >> SM_SHIFT;
 516         return (u32)m;
 517 }
 518
 519 /* convert dx (psched us) into d (us) */
 520 static u32
 521 dx2d(u64 dx)
 522 {
 523         u64 d;
 524
 525         d = dx * USEC_PER_SEC;
 526         do_div(d, PSCHED_JIFFIE2US(HZ));
 527         return (u32)d;
 528 }
 529
 530 static void
 531 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
 532 {
 533         isc->sm1  = m2sm(sc->m1);
 534         isc->ism1 = m2ism(sc->m1);
 535         isc->dx   = d2dx(sc->d);
 536         isc->dy   = seg_x2y(isc->dx, isc->sm1);
 537         isc->sm2  = m2sm(sc->m2);
 538         isc->ism2 = m2ism(sc->m2);
 539 }
 540
 541 /*
 542  * initialize the runtime service curve with the given internal
 543  * service curve starting at (x, y).
 544  */
 545 static void
 546 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
 547 {
 548         rtsc->x    = x;
 549         rtsc->y    = y;
 550         rtsc->sm1  = isc->sm1;
 551         rtsc->ism1 = isc->ism1;
 552         rtsc->dx   = isc->dx;
 553         rtsc->dy   = isc->dy;
 554         rtsc->sm2  = isc->sm2;
 555         rtsc->ism2 = isc->ism2;
 556 }
 557
 558 /*
 559  * calculate the y-projection of the runtime service curve by the
 560  * given x-projection value
 561  */
 562 static u64
 563 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
 564 {
 565         u64 x;
 566
 567         if (y < rtsc->y)
 568                 x = rtsc->x;
 569         else if (y <= rtsc->y + rtsc->dy) {
 570                 /* x belongs to the 1st segment */
 571                 if (rtsc->dy == 0)
 572                         x = rtsc->x + rtsc->dx;
 573                 else
 574                         x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
 575         } else {
 576                 /* x belongs to the 2nd segment */
 577                 x = rtsc->x + rtsc->dx
 578                     + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
 579         }
 580         return x;
 581 }
 582
 583 static u64
 584 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
 585 {
 586         u64 y;
 587
 588         if (x <= rtsc->x)
 589                 y = rtsc->y;
 590         else if (x <= rtsc->x + rtsc->dx)
 591                 /* y belongs to the 1st segment */
 592                 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
 593         else
 594                 /* y belongs to the 2nd segment */
 595                 y = rtsc->y + rtsc->dy
 596                     + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
 597         return y;
 598 }
 599
 600 /*
 601  * update the runtime service curve by taking the minimum of the current
 602  * runtime service curve and the service curve starting at (x, y).
 603  */
 604 static void
 605 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
 606 {
 607         u64 y1, y2, dx, dy;
 608         u32 dsm;
 609
 610         if (isc->sm1 <= isc->sm2) {
 611                 /* service curve is convex */
 612                 y1 = rtsc_x2y(rtsc, x);
 613                 if (y1 < y)
 614                         /* the current rtsc is smaller */
 615                         return;
 616                 rtsc->x = x;
 617                 rtsc->y = y;
 618                 return;
 619         }
 620
 621         /*
 622          * service curve is concave
 623          * compute the two y values of the current rtsc
 624          *      y1: at x
 625          *      y2: at (x + dx)
 626          */
 627         y1 = rtsc_x2y(rtsc, x);
 628         if (y1 <= y) {
 629                 /* rtsc is below isc, no change to rtsc */
 630                 return;
 631         }
 632
 633         y2 = rtsc_x2y(rtsc, x + isc->dx);
 634         if (y2 >= y + isc->dy) {
 635                 /* rtsc is above isc, replace rtsc by isc */
 636                 rtsc->x = x;
 637                 rtsc->y = y;
 638                 rtsc->dx = isc->dx;
 639                 rtsc->dy = isc->dy;
 640                 return;
 641         }
 642
 643         /*
 644          * the two curves intersect
 645          * compute the offsets (dx, dy) using the reverse
 646          * function of seg_x2y()
 647          *      seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
 648          */
 649         dx = (y1 - y) << SM_SHIFT;
 650         dsm = isc->sm1 - isc->sm2;
 651         do_div(dx, dsm);
 652         /*
 653          * check if (x, y1) belongs to the 1st segment of rtsc.
 654          * if so, add the offset.
 655          */
 656         if (rtsc->x + rtsc->dx > x)
 657                 dx += rtsc->x + rtsc->dx - x;
 658         dy = seg_x2y(dx, isc->sm1);
 659
 660         rtsc->x = x;
 661         rtsc->y = y;
 662         rtsc->dx = dx;
 663         rtsc->dy = dy;
 664         return;
 665 }
 666
 667 static void
 668 init_ed(struct hfsc_class *cl, unsigned int next_len)
 669 {
 670         u64 cur_time;
 671
 672         PSCHED_GET_TIME(cur_time);
 673
 674         /* update the deadline curve */
 675         rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
 676
 677         /*
 678          * update the eligible curve.
 679          * for concave, it is equal to the deadline curve.
 680          * for convex, it is a linear curve with slope m2.
 681          */
 682         cl->cl_eligible = cl->cl_deadline;
 683         if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
 684                 cl->cl_eligible.dx = 0;
 685                 cl->cl_eligible.dy = 0;
 686         }
 687
 688         /* compute e and d */
 689         cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
 690         cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 691
 692         eltree_insert(cl);
 693 }
 694
 695 static void
 696 update_ed(struct hfsc_class *cl, unsigned int next_len)
 697 {
 698         cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
 699         cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 700
 701         eltree_update(cl);
 702 }
 703
 704 static inline void
 705 update_d(struct hfsc_class *cl, unsigned int next_len)
 706 {
 707         cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 708 }
 709
 710 static inline void
 711 update_cfmin(struct hfsc_class *cl)
 712 {
 713         struct rb_node *n = rb_first(&cl->cf_tree);
 714         struct hfsc_class *p;
 715
 716         if (n == NULL) {
 717                 cl->cl_cfmin = 0;
 718                 return;
 719         }
 720         p = rb_entry(n, struct hfsc_class, cf_node);
 721         cl->cl_cfmin = p->cl_f;
 722 }
 723
 724 static void
 725 init_vf(struct hfsc_class *cl, unsigned int len)
 726 {
 727         struct hfsc_class *max_cl;
 728         struct rb_node *n;
 729         u64 vt, f, cur_time;
 730         int go_active;
 731
 732         cur_time = 0;
 733         go_active = 1;
 734         for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
 735                 if (go_active && cl->cl_nactive++ == 0)
 736                         go_active = 1;
 737                 else
 738                         go_active = 0;
 739
 740                 if (go_active) {
 741                         n = rb_last(&cl->cl_parent->vt_tree);
 742                         if (n != NULL) {
 743                                 max_cl = rb_entry(n, struct hfsc_class,vt_node);
 744                                 /*
 745                                  * set vt to the average of the min and max
 746                                  * classes.  if the parent's period didn't
 747                                  * change, don't decrease vt of the class.
 748                                  */
 749                                 vt = max_cl->cl_vt;
 750                                 if (cl->cl_parent->cl_cvtmin != 0)
 751                                         vt = (cl->cl_parent->cl_cvtmin + vt)/2;
 752
 753                                 if (cl->cl_parent->cl_vtperiod !=
 754                                     cl->cl_parentperiod || vt > cl->cl_vt)
 755                                         cl->cl_vt = vt;
 756                         } else {
 757                                 /*
 758                                  * first child for a new parent backlog period.
 759                                  * add parent's cvtmax to cvtoff to make a new
 760                                  * vt (vtoff + vt) larger than the vt in the
 761                                  * last period for all children.
 762                                  */
 763                                 vt = cl->cl_parent->cl_cvtmax;
 764                                 cl->cl_parent->cl_cvtoff += vt;
 765                                 cl->cl_parent->cl_cvtmax = 0;
 766                                 cl->cl_parent->cl_cvtmin = 0;
 767                                 cl->cl_vt = 0;
 768                         }
 769
 770                         cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
 771                                                         cl->cl_pcvtoff;
 772
 773                         /* update the virtual curve */
 774                         vt = cl->cl_vt + cl->cl_vtoff;
 775                         rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
 776                                                       cl->cl_total);
 777                         if (cl->cl_virtual.x == vt) {
 778                                 cl->cl_virtual.x -= cl->cl_vtoff;
 779                                 cl->cl_vtoff = 0;
 780                         }
 781                         cl->cl_vtadj = 0;
 782
 783                         cl->cl_vtperiod++;  /* increment vt period */
 784                         cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
 785                         if (cl->cl_parent->cl_nactive == 0)
 786                                 cl->cl_parentperiod++;
 787                         cl->cl_f = 0;
 788
 789                         vttree_insert(cl);
 790                         cftree_insert(cl);
 791
 792                         if (cl->cl_flags & HFSC_USC) {
 793                                 /* class has upper limit curve */
 794                                 if (cur_time == 0)
 795                                         PSCHED_GET_TIME(cur_time);
 796
 797                                 /* update the ulimit curve */
 798                                 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
 799                                          cl->cl_total);
 800                                 /* compute myf */
 801                                 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
 802                                                       cl->cl_total);
 803                                 cl->cl_myfadj = 0;
 804                         }
 805                 }
 806
 807                 f = max(cl->cl_myf, cl->cl_cfmin);
 808                 if (f != cl->cl_f) {
 809                         cl->cl_f = f;
 810                         cftree_update(cl);
 811                         update_cfmin(cl->cl_parent);
 812                 }
 813         }
 814 }
 815
 816 static void
 817 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
 818 {
 819         u64 f; /* , myf_bound, delta; */
 820         int go_passive = 0;
 821
 822         if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
 823                 go_passive = 1;
 824
 825         for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
 826                 cl->cl_total += len;
 827
 828                 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
 829                         continue;
 830
 831                 if (go_passive && --cl->cl_nactive == 0)
 832                         go_passive = 1;
 833                 else
 834                         go_passive = 0;
 835
 836                 if (go_passive) {
 837                         /* no more active child, going passive */
 838
 839                         /* update cvtmax of the parent class */
 840                         if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
 841                                 cl->cl_parent->cl_cvtmax = cl->cl_vt;
 842
 843                         /* remove this class from the vt tree */
 844                         vttree_remove(cl);
 845
 846                         cftree_remove(cl);
 847                         update_cfmin(cl->cl_parent);
 848
 849                         continue;
 850                 }
 851
 852                 /*
 853                  * update vt and f
 854                  */
 855                 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
 856                             - cl->cl_vtoff + cl->cl_vtadj;
 857
 858                 /*
 859                  * if vt of the class is smaller than cvtmin,
 860                  * the class was skipped in the past due to non-fit.
 861                  * if so, we need to adjust vtadj.
 862                  */
 863                 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
 864                         cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
 865                         cl->cl_vt = cl->cl_parent->cl_cvtmin;
 866                 }
 867
 868                 /* update the vt tree */
 869                 vttree_update(cl);
 870
 871                 if (cl->cl_flags & HFSC_USC) {
 872                         cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
 873                                                               cl->cl_total);
 874 #if 0
 875                         /*
 876                          * This code causes classes to stay way under their
 877                          * limit when multiple classes are used at gigabit
 878                          * speed. needs investigation. -kaber
 879                          */
 880                         /*
 881                          * if myf lags behind by more than one clock tick
 882                          * from the current time, adjust myfadj to prevent
 883                          * a rate-limited class from going greedy.
 884                          * in a steady state under rate-limiting, myf
 885                          * fluctuates within one clock tick.
 886                          */
 887                         myf_bound = cur_time - PSCHED_JIFFIE2US(1);
 888                         if (cl->cl_myf < myf_bound) {
 889                                 delta = cur_time - cl->cl_myf;
 890                                 cl->cl_myfadj += delta;
 891                                 cl->cl_myf += delta;
 892                         }
 893 #endif
 894                 }
 895
 896                 f = max(cl->cl_myf, cl->cl_cfmin);
 897                 if (f != cl->cl_f) {
 898                         cl->cl_f = f;
 899                         cftree_update(cl);
 900                         update_cfmin(cl->cl_parent);
 901                 }
 902         }
 903 }
 904
 905 static void
 906 set_active(struct hfsc_class *cl, unsigned int len)
 907 {
 908         if (cl->cl_flags & HFSC_RSC)
 909                 init_ed(cl, len);
 910         if (cl->cl_flags & HFSC_FSC)
 911                 init_vf(cl, len);
 912
 913         list_add_tail(&cl->dlist, &cl->sched->droplist);
 914 }
 915
 916 static void
 917 set_passive(struct hfsc_class *cl)
 918 {
 919         if (cl->cl_flags & HFSC_RSC)
 920                 eltree_remove(cl);
 921
 922         list_del(&cl->dlist);
 923
 924         /*
 925          * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
 926          * needs to be called explicitly to remove a class from vttree.
 927          */
 928 }
 929
 930 /*
 931  * hack to get length of first packet in queue.
 932  */
 933 static unsigned int
 934 qdisc_peek_len(struct Qdisc *sch)
 935 {
 936         struct sk_buff *skb;
 937         unsigned int len;
 938
 939         skb = sch->dequeue(sch);
 940         if (skb == NULL) {
 941                 if (net_ratelimit())
 942                         printk("qdisc_peek_len: non work-conserving qdisc ?\n");
 943                 return 0;
 944         }
 945         len = skb->len;
 946         if (unlikely(sch->ops->requeue(skb, sch) != NET_XMIT_SUCCESS)) {
 947                 if (net_ratelimit())
 948                         printk("qdisc_peek_len: failed to requeue\n");
 949                 return 0;
 950         }
 951         return len;
 952 }
 953
 954 static void
 955 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
 956 {
 957         unsigned int len = cl->qdisc->q.qlen;
 958
 959         qdisc_reset(cl->qdisc);
 960         if (len > 0) {
 961                 update_vf(cl, 0, 0);
 962                 set_passive(cl);
 963                 sch->q.qlen -= len;
 964         }
 965 }
 966
 967 static void
 968 hfsc_adjust_levels(struct hfsc_class *cl)
 969 {
 970         struct hfsc_class *p;
 971         unsigned int level;
 972
 973         do {
 974                 level = 0;
 975                 list_for_each_entry(p, &cl->children, siblings) {
 976                         if (p->level >= level)
 977                                 level = p->level + 1;
 978                 }
 979                 cl->level = level;
 980         } while ((cl = cl->cl_parent) != NULL);
 981 }
 982
 983 static inline unsigned int
 984 hfsc_hash(u32 h)
 985 {
 986         h ^= h >> 8;
 987         h ^= h >> 4;
 988
 989         return h & (HFSC_HSIZE - 1);
 990 }
 991
 992 static inline struct hfsc_class *
 993 hfsc_find_class(u32 classid, struct Qdisc *sch)
 994 {
 995         struct hfsc_sched *q = qdisc_priv(sch);
 996         struct hfsc_class *cl;
 997
 998         list_for_each_entry(cl, &q->clhash[hfsc_hash(classid)], hlist) {
 999                 if (cl->classid == classid)
1000                         return cl;
1001         }
1002         return NULL;
1003 }
1004
1005 static void
1006 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
1007                 u64 cur_time)
1008 {
1009         sc2isc(rsc, &cl->cl_rsc);
1010         rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
1011         cl->cl_eligible = cl->cl_deadline;
1012         if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
1013                 cl->cl_eligible.dx = 0;
1014                 cl->cl_eligible.dy = 0;
1015         }
1016         cl->cl_flags |= HFSC_RSC;
1017 }
1018
1019 static void
1020 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
1021 {
1022         sc2isc(fsc, &cl->cl_fsc);
1023         rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
1024         cl->cl_flags |= HFSC_FSC;
1025 }
1026
1027 static void
1028 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
1029                 u64 cur_time)
1030 {
1031         sc2isc(usc, &cl->cl_usc);
1032         rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
1033         cl->cl_flags |= HFSC_USC;
1034 }
1035
1036 static int
1037 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
1038                   struct rtattr **tca, unsigned long *arg)
1039 {
1040         struct hfsc_sched *q = qdisc_priv(sch);
1041         struct hfsc_class *cl = (struct hfsc_class *)*arg;
1042         struct hfsc_class *parent = NULL;
1043         struct rtattr *opt = tca[TCA_OPTIONS-1];
1044         struct rtattr *tb[TCA_HFSC_MAX];
1045         struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
1046         u64 cur_time;
1047
1048         if (opt == NULL || rtattr_parse_nested(tb, TCA_HFSC_MAX, opt))
1049                 return -EINVAL;
1050
1051         if (tb[TCA_HFSC_RSC-1]) {
1052                 if (RTA_PAYLOAD(tb[TCA_HFSC_RSC-1]) < sizeof(*rsc))
1053                         return -EINVAL;
1054                 rsc = RTA_DATA(tb[TCA_HFSC_RSC-1]);
1055                 if (rsc->m1 == 0 && rsc->m2 == 0)
1056                         rsc = NULL;
1057         }
1058
1059         if (tb[TCA_HFSC_FSC-1]) {
1060                 if (RTA_PAYLOAD(tb[TCA_HFSC_FSC-1]) < sizeof(*fsc))
1061                         return -EINVAL;
1062                 fsc = RTA_DATA(tb[TCA_HFSC_FSC-1]);
1063                 if (fsc->m1 == 0 && fsc->m2 == 0)
1064                         fsc = NULL;
1065         }
1066
1067         if (tb[TCA_HFSC_USC-1]) {
1068                 if (RTA_PAYLOAD(tb[TCA_HFSC_USC-1]) < sizeof(*usc))
1069                         return -EINVAL;
1070                 usc = RTA_DATA(tb[TCA_HFSC_USC-1]);
1071                 if (usc->m1 == 0 && usc->m2 == 0)
1072                         usc = NULL;
1073         }
1074
1075         if (cl != NULL) {
1076                 if (parentid) {
1077                         if (cl->cl_parent && cl->cl_parent->classid != parentid)
1078                                 return -EINVAL;
1079                         if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1080                                 return -EINVAL;
1081                 }
1082                 PSCHED_GET_TIME(cur_time);
1083
1084                 sch_tree_lock(sch);
1085                 if (rsc != NULL)
1086                         hfsc_change_rsc(cl, rsc, cur_time);
1087                 if (fsc != NULL)
1088                         hfsc_change_fsc(cl, fsc);
1089                 if (usc != NULL)
1090                         hfsc_change_usc(cl, usc, cur_time);
1091
1092                 if (cl->qdisc->q.qlen != 0) {
1093                         if (cl->cl_flags & HFSC_RSC)
1094                                 update_ed(cl, qdisc_peek_len(cl->qdisc));
1095                         if (cl->cl_flags & HFSC_FSC)
1096                                 update_vf(cl, 0, cur_time);
1097                 }
1098                 sch_tree_unlock(sch);
1099
1100 #ifdef CONFIG_NET_ESTIMATOR
1101                 if (tca[TCA_RATE-1])
1102                         gen_replace_estimator(&cl->bstats, &cl->rate_est,
1103                                 cl->stats_lock, tca[TCA_RATE-1]);
1104 #endif
1105                 return 0;
1106         }
1107
1108         if (parentid == TC_H_ROOT)
1109                 return -EEXIST;
1110
1111         parent = &q->root;
1112         if (parentid) {
1113                 parent = hfsc_find_class(parentid, sch);
1114                 if (parent == NULL)
1115                         return -ENOENT;
1116         }
1117
1118         if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1119                 return -EINVAL;
1120         if (hfsc_find_class(classid, sch))
1121                 return -EEXIST;
1122
1123         if (rsc == NULL && fsc == NULL)
1124                 return -EINVAL;
1125
1126         cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1127         if (cl == NULL)
1128                 return -ENOBUFS;
1129
1130         if (rsc != NULL)
1131                 hfsc_change_rsc(cl, rsc, 0);
1132         if (fsc != NULL)
1133                 hfsc_change_fsc(cl, fsc);
1134         if (usc != NULL)
1135                 hfsc_change_usc(cl, usc, 0);
1136
1137         cl->refcnt    = 1;
1138         cl->classid   = classid;
1139         cl->sched     = q;
1140         cl->cl_parent = parent;
1141         cl->qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
1142         if (cl->qdisc == NULL)
1143                 cl->qdisc = &noop_qdisc;
1144         cl->stats_lock = &sch->dev->queue_lock;
1145         INIT_LIST_HEAD(&cl->children);
1146         cl->vt_tree = RB_ROOT;
1147         cl->cf_tree = RB_ROOT;
1148
1149         sch_tree_lock(sch);
1150         list_add_tail(&cl->hlist, &q->clhash[hfsc_hash(classid)]);
1151         list_add_tail(&cl->siblings, &parent->children);
1152         if (parent->level == 0)
1153                 hfsc_purge_queue(sch, parent);
1154         hfsc_adjust_levels(parent);
1155         cl->cl_pcvtoff = parent->cl_cvtoff;
1156         sch_tree_unlock(sch);
1157
1158 #ifdef CONFIG_NET_ESTIMATOR
1159         if (tca[TCA_RATE-1])
1160                 gen_new_estimator(&cl->bstats, &cl->rate_est,
1161                         cl->stats_lock, tca[TCA_RATE-1]);
1162 #endif
1163         *arg = (unsigned long)cl;
1164         return 0;
1165 }
1166
1167 static void
1168 hfsc_destroy_filters(struct tcf_proto **fl)
1169 {
1170         struct tcf_proto *tp;
1171
1172         while ((tp = *fl) != NULL) {
1173                 *fl = tp->next;
1174                 tcf_destroy(tp);
1175         }
1176 }
1177
1178 static void
1179 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1180 {
1181         struct hfsc_sched *q = qdisc_priv(sch);
1182
1183         hfsc_destroy_filters(&cl->filter_list);
1184         qdisc_destroy(cl->qdisc);
1185 #ifdef CONFIG_NET_ESTIMATOR
1186         gen_kill_estimator(&cl->bstats, &cl->rate_est);
1187 #endif
1188         if (cl != &q->root)
1189                 kfree(cl);
1190 }
1191
1192 static int
1193 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1194 {
1195         struct hfsc_sched *q = qdisc_priv(sch);
1196         struct hfsc_class *cl = (struct hfsc_class *)arg;
1197
1198         if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1199                 return -EBUSY;
1200
1201         sch_tree_lock(sch);
1202
1203         list_del(&cl->hlist);
1204         list_del(&cl->siblings);
1205         hfsc_adjust_levels(cl->cl_parent);
1206         hfsc_purge_queue(sch, cl);
1207         if (--cl->refcnt == 0)
1208                 hfsc_destroy_class(sch, cl);
1209
1210         sch_tree_unlock(sch);
1211         return 0;
1212 }
1213
1214 static struct hfsc_class *
1215 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1216 {
1217         struct hfsc_sched *q = qdisc_priv(sch);
1218         struct hfsc_class *cl;
1219         struct tcf_result res;
1220         struct tcf_proto *tcf;
1221         int result;
1222
1223         if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1224             (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1225                 if (cl->level == 0)
1226                         return cl;
1227
1228         *qerr = NET_XMIT_BYPASS;
1229         tcf = q->root.filter_list;
1230         while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1231 #ifdef CONFIG_NET_CLS_ACT
1232                 switch (result) {
1233                 case TC_ACT_QUEUED:
1234                 case TC_ACT_STOLEN:
1235                         *qerr = NET_XMIT_SUCCESS;
1236                 case TC_ACT_SHOT:
1237                         return NULL;
1238                 }
1239 #elif defined(CONFIG_NET_CLS_POLICE)
1240                 if (result == TC_POLICE_SHOT)
1241                         return NULL;
1242 #endif
1243                 if ((cl = (struct hfsc_class *)res.class) == NULL) {
1244                         if ((cl = hfsc_find_class(res.classid, sch)) == NULL)
1245                                 break; /* filter selected invalid classid */
1246                 }
1247
1248                 if (cl->level == 0)
1249                         return cl; /* hit leaf class */
1250
1251                 /* apply inner filter chain */
1252                 tcf = cl->filter_list;
1253         }
1254
1255         /* classification failed, try default class */
1256         cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1257         if (cl == NULL || cl->level > 0)
1258                 return NULL;
1259
1260         return cl;
1261 }
1262
1263 static int
1264 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1265                  struct Qdisc **old)
1266 {
1267         struct hfsc_class *cl = (struct hfsc_class *)arg;
1268
1269         if (cl == NULL)
1270                 return -ENOENT;
1271         if (cl->level > 0)
1272                 return -EINVAL;
1273         if (new == NULL) {
1274                 new = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
1275                 if (new == NULL)
1276                         new = &noop_qdisc;
1277         }
1278
1279         sch_tree_lock(sch);
1280         hfsc_purge_queue(sch, cl);
1281         *old = xchg(&cl->qdisc, new);
1282         sch_tree_unlock(sch);
1283         return 0;
1284 }
1285
1286 static struct Qdisc *
1287 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1288 {
1289         struct hfsc_class *cl = (struct hfsc_class *)arg;
1290
1291         if (cl != NULL && cl->level == 0)
1292                 return cl->qdisc;
1293
1294         return NULL;
1295 }
1296
1297 static unsigned long
1298 hfsc_get_class(struct Qdisc *sch, u32 classid)
1299 {
1300         struct hfsc_class *cl = hfsc_find_class(classid, sch);
1301
1302         if (cl != NULL)
1303                 cl->refcnt++;
1304
1305         return (unsigned long)cl;
1306 }
1307
1308 static void
1309 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1310 {
1311         struct hfsc_class *cl = (struct hfsc_class *)arg;
1312
1313         if (--cl->refcnt == 0)
1314                 hfsc_destroy_class(sch, cl);
1315 }
1316
1317 static unsigned long
1318 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1319 {
1320         struct hfsc_class *p = (struct hfsc_class *)parent;
1321         struct hfsc_class *cl = hfsc_find_class(classid, sch);
1322
1323         if (cl != NULL) {
1324                 if (p != NULL && p->level <= cl->level)
1325                         return 0;
1326                 cl->filter_cnt++;
1327         }
1328
1329         return (unsigned long)cl;
1330 }
1331
1332 static void
1333 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1334 {
1335         struct hfsc_class *cl = (struct hfsc_class *)arg;
1336
1337         cl->filter_cnt--;
1338 }
1339
1340 static struct tcf_proto **
1341 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1342 {
1343         struct hfsc_sched *q = qdisc_priv(sch);
1344         struct hfsc_class *cl = (struct hfsc_class *)arg;
1345
1346         if (cl == NULL)
1347                 cl = &q->root;
1348
1349         return &cl->filter_list;
1350 }
1351
1352 static int
1353 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1354 {
1355         struct tc_service_curve tsc;
1356
1357         tsc.m1 = sm2m(sc->sm1);
1358         tsc.d  = dx2d(sc->dx);
1359         tsc.m2 = sm2m(sc->sm2);
1360         RTA_PUT(skb, attr, sizeof(tsc), &tsc);
1361
1362         return skb->len;
1363
1364  rtattr_failure:
1365         return -1;
1366 }
1367
1368 static inline int
1369 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1370 {
1371         if ((cl->cl_flags & HFSC_RSC) &&
1372             (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1373                 goto rtattr_failure;
1374
1375         if ((cl->cl_flags & HFSC_FSC) &&
1376             (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1377                 goto rtattr_failure;
1378
1379         if ((cl->cl_flags & HFSC_USC) &&
1380             (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1381                 goto rtattr_failure;
1382
1383         return skb->len;
1384
1385  rtattr_failure:
1386         return -1;
1387 }
1388
1389 static int
1390 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1391                 struct tcmsg *tcm)
1392 {
1393         struct hfsc_class *cl = (struct hfsc_class *)arg;
1394         unsigned char *b = skb->tail;
1395         struct rtattr *rta = (struct rtattr *)b;
1396
1397         tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->classid : TC_H_ROOT;
1398         tcm->tcm_handle = cl->classid;
1399         if (cl->level == 0)
1400                 tcm->tcm_info = cl->qdisc->handle;
1401
1402         RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
1403         if (hfsc_dump_curves(skb, cl) < 0)
1404                 goto rtattr_failure;
1405         rta->rta_len = skb->tail - b;
1406         return skb->len;
1407
1408  rtattr_failure:
1409         skb_trim(skb, b - skb->data);
1410         return -1;
1411 }
1412
1413 static int
1414 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1415         struct gnet_dump *d)
1416 {
1417         struct hfsc_class *cl = (struct hfsc_class *)arg;
1418         struct tc_hfsc_stats xstats;
1419
1420         cl->qstats.qlen = cl->qdisc->q.qlen;
1421         xstats.level   = cl->level;
1422         xstats.period  = cl->cl_vtperiod;
1423         xstats.work    = cl->cl_total;
1424         xstats.rtwork  = cl->cl_cumul;
1425
1426         if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1427 #ifdef CONFIG_NET_ESTIMATOR
1428             gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1429 #endif
1430             gnet_stats_copy_queue(d, &cl->qstats) < 0)
1431                 return -1;
1432
1433         return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1434 }
1435
1436
1437
1438 static void
1439 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1440 {
1441         struct hfsc_sched *q = qdisc_priv(sch);
1442         struct hfsc_class *cl;
1443         unsigned int i;
1444
1445         if (arg->stop)
1446                 return;
1447
1448         for (i = 0; i < HFSC_HSIZE; i++) {
1449                 list_for_each_entry(cl, &q->clhash[i], hlist) {
1450                         if (arg->count < arg->skip) {
1451                                 arg->count++;
1452                                 continue;
1453                         }
1454                         if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1455                                 arg->stop = 1;
1456                                 return;
1457                         }
1458                         arg->count++;
1459                 }
1460         }
1461 }
1462
1463 static void
1464 hfsc_watchdog(unsigned long arg)
1465 {
1466         struct Qdisc *sch = (struct Qdisc *)arg;
1467
1468         sch->flags &= ~TCQ_F_THROTTLED;
1469         netif_schedule(sch->dev);
1470 }
1471
1472 static void
1473 hfsc_schedule_watchdog(struct Qdisc *sch, u64 cur_time)
1474 {
1475         struct hfsc_sched *q = qdisc_priv(sch);
1476         struct hfsc_class *cl;
1477         u64 next_time = 0;
1478         long delay;
1479
1480         if ((cl = eltree_get_minel(q)) != NULL)
1481                 next_time = cl->cl_e;
1482         if (q->root.cl_cfmin != 0) {
1483                 if (next_time == 0 || next_time > q->root.cl_cfmin)
1484                         next_time = q->root.cl_cfmin;
1485         }
1486         ASSERT(next_time != 0);
1487         delay = next_time - cur_time;
1488         delay = PSCHED_US2JIFFIE(delay);
1489
1490         sch->flags |= TCQ_F_THROTTLED;
1491         mod_timer(&q->wd_timer, jiffies + delay);
1492 }
1493
1494 static int
1495 hfsc_init_qdisc(struct Qdisc *sch, struct rtattr *opt)
1496 {
1497         struct hfsc_sched *q = qdisc_priv(sch);
1498         struct tc_hfsc_qopt *qopt;
1499         unsigned int i;
1500
1501         if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1502                 return -EINVAL;
1503         qopt = RTA_DATA(opt);
1504
1505         sch->stats_lock = &sch->dev->queue_lock;
1506
1507         q->defcls = qopt->defcls;
1508         for (i = 0; i < HFSC_HSIZE; i++)
1509                 INIT_LIST_HEAD(&q->clhash[i]);
1510         q->eligible = RB_ROOT;
1511         INIT_LIST_HEAD(&q->droplist);
1512         skb_queue_head_init(&q->requeue);
1513
1514         q->root.refcnt  = 1;
1515         q->root.classid = sch->handle;
1516         q->root.sched   = q;
1517         q->root.qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
1518         if (q->root.qdisc == NULL)
1519                 q->root.qdisc = &noop_qdisc;
1520         q->root.stats_lock = &sch->dev->queue_lock;
1521         INIT_LIST_HEAD(&q->root.children);
1522         q->root.vt_tree = RB_ROOT;
1523         q->root.cf_tree = RB_ROOT;
1524
1525         list_add(&q->root.hlist, &q->clhash[hfsc_hash(q->root.classid)]);
1526
1527         init_timer(&q->wd_timer);
1528         q->wd_timer.function = hfsc_watchdog;
1529         q->wd_timer.data = (unsigned long)sch;
1530
1531         return 0;
1532 }
1533
1534 static int
1535 hfsc_change_qdisc(struct Qdisc *sch, struct rtattr *opt)
1536 {
1537         struct hfsc_sched *q = qdisc_priv(sch);
1538         struct tc_hfsc_qopt *qopt;
1539
1540         if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1541                 return -EINVAL;
1542         qopt = RTA_DATA(opt);
1543
1544         sch_tree_lock(sch);
1545         q->defcls = qopt->defcls;
1546         sch_tree_unlock(sch);
1547
1548         return 0;
1549 }
1550
1551 static void
1552 hfsc_reset_class(struct hfsc_class *cl)
1553 {
1554         cl->cl_total        = 0;
1555         cl->cl_cumul        = 0;
1556         cl->cl_d            = 0;
1557         cl->cl_e            = 0;
1558         cl->cl_vt           = 0;
1559         cl->cl_vtadj        = 0;
1560         cl->cl_vtoff        = 0;
1561         cl->cl_cvtmin       = 0;
1562         cl->cl_cvtmax       = 0;
1563         cl->cl_cvtoff       = 0;
1564         cl->cl_pcvtoff      = 0;
1565         cl->cl_vtperiod     = 0;
1566         cl->cl_parentperiod = 0;
1567         cl->cl_f            = 0;
1568         cl->cl_myf          = 0;
1569         cl->cl_myfadj       = 0;
1570         cl->cl_cfmin        = 0;
1571         cl->cl_nactive      = 0;
1572
1573         cl->vt_tree = RB_ROOT;
1574         cl->cf_tree = RB_ROOT;
1575         qdisc_reset(cl->qdisc);
1576
1577         if (cl->cl_flags & HFSC_RSC)
1578                 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1579         if (cl->cl_flags & HFSC_FSC)
1580                 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1581         if (cl->cl_flags & HFSC_USC)
1582                 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1583 }
1584
1585 static void
1586 hfsc_reset_qdisc(struct Qdisc *sch)
1587 {
1588         struct hfsc_sched *q = qdisc_priv(sch);
1589         struct hfsc_class *cl;
1590         unsigned int i;
1591
1592         for (i = 0; i < HFSC_HSIZE; i++) {
1593                 list_for_each_entry(cl, &q->clhash[i], hlist)
1594                         hfsc_reset_class(cl);
1595         }
1596         __skb_queue_purge(&q->requeue);
1597         q->eligible = RB_ROOT;
1598         INIT_LIST_HEAD(&q->droplist);
1599         del_timer(&q->wd_timer);
1600         sch->flags &= ~TCQ_F_THROTTLED;
1601         sch->q.qlen = 0;
1602 }
1603
1604 static void
1605 hfsc_destroy_qdisc(struct Qdisc *sch)
1606 {
1607         struct hfsc_sched *q = qdisc_priv(sch);
1608         struct hfsc_class *cl, *next;
1609         unsigned int i;
1610
1611         for (i = 0; i < HFSC_HSIZE; i++) {
1612                 list_for_each_entry_safe(cl, next, &q->clhash[i], hlist)
1613                         hfsc_destroy_class(sch, cl);
1614         }
1615         __skb_queue_purge(&q->requeue);
1616         del_timer(&q->wd_timer);
1617 }
1618
1619 static int
1620 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1621 {
1622         struct hfsc_sched *q = qdisc_priv(sch);
1623         unsigned char *b = skb->tail;
1624         struct tc_hfsc_qopt qopt;
1625
1626         qopt.defcls = q->defcls;
1627         RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1628         return skb->len;
1629
1630  rtattr_failure:
1631         skb_trim(skb, b - skb->data);
1632         return -1;
1633 }
1634
1635 static int
1636 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1637 {
1638         struct hfsc_class *cl;
1639         unsigned int len;
1640         int err;
1641
1642         cl = hfsc_classify(skb, sch, &err);
1643         if (cl == NULL) {
1644                 if (err == NET_XMIT_BYPASS)
1645                         sch->qstats.drops++;
1646                 kfree_skb(skb);
1647                 return err;
1648         }
1649
1650         len = skb->len;
1651         err = cl->qdisc->enqueue(skb, cl->qdisc);
1652         if (unlikely(err != NET_XMIT_SUCCESS)) {
1653                 cl->qstats.drops++;
1654                 sch->qstats.drops++;
1655                 return err;
1656         }
1657
1658         if (cl->qdisc->q.qlen == 1)
1659                 set_active(cl, len);
1660
1661         cl->bstats.packets++;
1662         cl->bstats.bytes += len;
1663         sch->bstats.packets++;
1664         sch->bstats.bytes += len;
1665         sch->q.qlen++;
1666
1667         return NET_XMIT_SUCCESS;
1668 }
1669
1670 static struct sk_buff *
1671 hfsc_dequeue(struct Qdisc *sch)
1672 {
1673         struct hfsc_sched *q = qdisc_priv(sch);
1674         struct hfsc_class *cl;
1675         struct sk_buff *skb;
1676         u64 cur_time;
1677         unsigned int next_len;
1678         int realtime = 0;
1679
1680         if (sch->q.qlen == 0)
1681                 return NULL;
1682         if ((skb = __skb_dequeue(&q->requeue)))
1683                 goto out;
1684
1685         PSCHED_GET_TIME(cur_time);
1686
1687         /*
1688          * if there are eligible classes, use real-time criteria.
1689          * find the class with the minimum deadline among
1690          * the eligible classes.
1691          */
1692         if ((cl = eltree_get_mindl(q, cur_time)) != NULL) {
1693                 realtime = 1;
1694         } else {
1695                 /*
1696                  * use link-sharing criteria
1697                  * get the class with the minimum vt in the hierarchy
1698                  */
1699                 cl = vttree_get_minvt(&q->root, cur_time);
1700                 if (cl == NULL) {
1701                         sch->qstats.overlimits++;
1702                         hfsc_schedule_watchdog(sch, cur_time);
1703                         return NULL;
1704                 }
1705         }
1706
1707         skb = cl->qdisc->dequeue(cl->qdisc);
1708         if (skb == NULL) {
1709                 if (net_ratelimit())
1710                         printk("HFSC: Non-work-conserving qdisc ?\n");
1711                 return NULL;
1712         }
1713
1714         update_vf(cl, skb->len, cur_time);
1715         if (realtime)
1716                 cl->cl_cumul += skb->len;
1717
1718         if (cl->qdisc->q.qlen != 0) {
1719                 if (cl->cl_flags & HFSC_RSC) {
1720                         /* update ed */
1721                         next_len = qdisc_peek_len(cl->qdisc);
1722                         if (realtime)
1723                                 update_ed(cl, next_len);
1724                         else
1725                                 update_d(cl, next_len);
1726                 }
1727         } else {
1728                 /* the class becomes passive */
1729                 set_passive(cl);
1730         }
1731
1732  out:
1733         sch->flags &= ~TCQ_F_THROTTLED;
1734         sch->q.qlen--;
1735
1736         return skb;
1737 }
1738
1739 static int
1740 hfsc_requeue(struct sk_buff *skb, struct Qdisc *sch)
1741 {
1742         struct hfsc_sched *q = qdisc_priv(sch);
1743
1744         __skb_queue_head(&q->requeue, skb);
1745         sch->q.qlen++;
1746         sch->qstats.requeues++;
1747         return NET_XMIT_SUCCESS;
1748 }
1749
1750 static unsigned int
1751 hfsc_drop(struct Qdisc *sch)
1752 {
1753         struct hfsc_sched *q = qdisc_priv(sch);
1754         struct hfsc_class *cl;
1755         unsigned int len;
1756
1757         list_for_each_entry(cl, &q->droplist, dlist) {
1758                 if (cl->qdisc->ops->drop != NULL &&
1759                     (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1760                         if (cl->qdisc->q.qlen == 0) {
1761                                 update_vf(cl, 0, 0);
1762                                 set_passive(cl);
1763                         } else {
1764                                 list_move_tail(&cl->dlist, &q->droplist);
1765                         }
1766                         cl->qstats.drops++;
1767                         sch->qstats.drops++;
1768                         sch->q.qlen--;
1769                         return len;
1770                 }
1771         }
1772         return 0;
1773 }
1774
1775 static struct Qdisc_class_ops hfsc_class_ops = {
1776         .change         = hfsc_change_class,
1777         .delete         = hfsc_delete_class,
1778         .graft          = hfsc_graft_class,
1779         .leaf           = hfsc_class_leaf,
1780         .get            = hfsc_get_class,
1781         .put            = hfsc_put_class,
1782         .bind_tcf       = hfsc_bind_tcf,
1783         .unbind_tcf     = hfsc_unbind_tcf,
1784         .tcf_chain      = hfsc_tcf_chain,
1785         .dump           = hfsc_dump_class,
1786         .dump_stats     = hfsc_dump_class_stats,
1787         .walk           = hfsc_walk
1788 };
1789
1790 static struct Qdisc_ops hfsc_qdisc_ops = {
1791         .id             = "hfsc",
1792         .init           = hfsc_init_qdisc,
1793         .change         = hfsc_change_qdisc,
1794         .reset          = hfsc_reset_qdisc,
1795         .destroy        = hfsc_destroy_qdisc,
1796         .dump           = hfsc_dump_qdisc,
1797         .enqueue        = hfsc_enqueue,
1798         .dequeue        = hfsc_dequeue,
1799         .requeue        = hfsc_requeue,
1800         .drop           = hfsc_drop,
1801         .cl_ops         = &hfsc_class_ops,
1802         .priv_size      = sizeof(struct hfsc_sched),
1803         .owner          = THIS_MODULE
1804 };
1805
1806 static int __init
1807 hfsc_init(void)
1808 {
1809         return register_qdisc(&hfsc_qdisc_ops);
1810 }
1811
1812 static void __exit
1813 hfsc_cleanup(void)
1814 {
1815         unregister_qdisc(&hfsc_qdisc_ops);
1816 }
1817
1818 MODULE_LICENSE("GPL");
1819 module_init(hfsc_init);
1820 module_exit(hfsc_cleanup);