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/config.h>
  54 #include <linux/module.h>
  55 #include <linux/types.h>
  56 #include <linux/errno.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 tc_stats stats;          /* generic statistics */
 125         unsigned int    level;          /* class level in hierarchy */
 126         struct tcf_proto *filter_list;  /* filter list */
 127         unsigned int    filter_cnt;     /* filter count */
 128
 129         struct hfsc_sched *sched;       /* scheduler data */
 130         struct hfsc_class *cl_parent;   /* parent class */
 131         struct list_head siblings;      /* sibling classes */
 132         struct list_head children;      /* child classes */
 133         struct Qdisc    *qdisc;         /* leaf qdisc */
 134
 135         rb_node_t el_node;              /* qdisc's eligible tree member */
 136         rb_root_t vt_tree;              /* active children sorted by cl_vt */
 137         rb_node_t vt_node;              /* parent's vt_tree member */
 138         rb_root_t cf_tree;              /* active children sorted by cl_f */
 139         rb_node_t cf_node;              /* parent's cf_heap member */
 140         struct list_head hlist;         /* hash list member */
 141         struct list_head dlist;         /* drop list member */
 142
 143         u64     cl_total;               /* total work in bytes */
 144         u64     cl_cumul;               /* cumulative work in bytes done by
 145                                            real-time criteria */
 146
 147         u64     cl_d;                   /* deadline*/
 148         u64     cl_e;                   /* eligible time */
 149         u64     cl_vt;                  /* virtual time */
 150         u64     cl_f;                   /* time when this class will fit for
 151                                            link-sharing, max(myf, cfmin) */
 152         u64     cl_myf;                 /* my fit-time (calculated from this
 153                                            class's own upperlimit curve) */
 154         u64     cl_myfadj;              /* my fit-time adjustment (to cancel
 155                                            history dependence) */
 156         u64     cl_cfmin;               /* earliest children's fit-time (used
 157                                            with cl_myf to obtain cl_f) */
 158         u64     cl_cvtmin;              /* minimal virtual time among the
 159                                            children fit for link-sharing
 160                                            (monotonic within a period) */
 161         u64     cl_vtadj;               /* intra-period cumulative vt
 162                                            adjustment */
 163         u64     cl_vtoff;               /* inter-period cumulative vt offset */
 164         u64     cl_cvtmax;              /* max child's vt in the last period */
 165         u64     cl_cvtoff;              /* cumulative cvtmax of all periods */
 166         u64     cl_pcvtoff;             /* parent's cvtoff at initalization
 167                                            time */
 168
 169         struct internal_sc cl_rsc;      /* internal real-time service curve */
 170         struct internal_sc cl_fsc;      /* internal fair service curve */
 171         struct internal_sc cl_usc;      /* internal upperlimit service curve */
 172         struct runtime_sc cl_deadline;  /* deadline curve */
 173         struct runtime_sc cl_eligible;  /* eligible curve */
 174         struct runtime_sc cl_virtual;   /* virtual curve */
 175         struct runtime_sc cl_ulimit;    /* upperlimit curve */
 176
 177         unsigned long   cl_flags;       /* which curves are valid */
 178         unsigned long   cl_vtperiod;    /* vt period sequence number */
 179         unsigned long   cl_parentperiod;/* parent's vt period sequence number*/
 180         unsigned long   cl_nactive;     /* number of active children */
 181 };
 182
 183 #define HFSC_HSIZE      16
 184
 185 struct hfsc_sched
 186 {
 187         u16     defcls;                         /* default class id */
 188         struct hfsc_class root;                 /* root class */
 189         struct list_head clhash[HFSC_HSIZE];    /* class hash */
 190         rb_root_t eligible;                     /* eligible tree */
 191         struct list_head droplist;              /* active leaf class list (for
 192                                                    dropping) */
 193         struct sk_buff_head requeue;            /* requeued packet */
 194         struct timer_list wd_timer;             /* watchdog timer */
 195 };
 196
 197 /*
 198  * macros
 199  */
 200 #if PSCHED_CLOCK_SOURCE == PSCHED_GETTIMEOFDAY
 201 #include <linux/time.h>
 202 #undef PSCHED_GET_TIME
 203 #define PSCHED_GET_TIME(stamp)                                          \
 204 do {                                                                    \
 205         struct timeval tv;                                              \
 206         do_gettimeofday(&tv);                                           \
 207         (stamp) = 1000000ULL * tv.tv_sec + tv.tv_usec;                  \
 208 } while (0)
 209 #endif
 210
 211 #if HFSC_DEBUG
 212 #define ASSERT(cond)                                                    \
 213 do {                                                                    \
 214         if (unlikely(!(cond)))                                          \
 215                 printk("assertion %s failed at %s:%i (%s)\n",           \
 216                        #cond, __FILE__, __LINE__, __FUNCTION__);        \
 217 } while (0)
 218 #else
 219 #define ASSERT(cond)
 220 #endif /* HFSC_DEBUG */
 221
 222 #define HT_INFINITY     0xffffffffffffffffULL   /* infinite time value */
 223
 224
 225 /*
 226  * eligible tree holds backlogged classes being sorted by their eligible times.
 227  * there is one eligible tree per hfsc instance.
 228  */
 229
 230 static void
 231 eltree_insert(struct hfsc_class *cl)
 232 {
 233         rb_node_t **p = &cl->sched->eligible.rb_node;
 234         rb_node_t *parent = NULL;
 235         struct hfsc_class *cl1;
 236
 237         while (*p != NULL) {
 238                 parent = *p;
 239                 cl1 = rb_entry(parent, struct hfsc_class, el_node);
 240                 if (cl->cl_e >= cl1->cl_e)
 241                         p = &parent->rb_right;
 242                 else
 243                         p = &parent->rb_left;
 244         }
 245         rb_link_node(&cl->el_node, parent, p);
 246         rb_insert_color(&cl->el_node, &cl->sched->eligible);
 247 }
 248
 249 static inline void
 250 eltree_remove(struct hfsc_class *cl)
 251 {
 252         rb_erase(&cl->el_node, &cl->sched->eligible);
 253 }
 254
 255 static inline void
 256 eltree_update(struct hfsc_class *cl)
 257 {
 258         eltree_remove(cl);
 259         eltree_insert(cl);
 260 }
 261
 262 /* find the class with the minimum deadline among the eligible classes */
 263 static inline struct hfsc_class *
 264 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
 265 {
 266         struct hfsc_class *p, *cl = NULL;
 267         rb_node_t *n;
 268
 269         for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
 270                 p = rb_entry(n, struct hfsc_class, el_node);
 271                 if (p->cl_e > cur_time)
 272                         break;
 273                 if (cl == NULL || p->cl_d < cl->cl_d)
 274                         cl = p;
 275         }
 276         return cl;
 277 }
 278
 279 /* find the class with minimum eligible time among the eligible classes */
 280 static inline struct hfsc_class *
 281 eltree_get_minel(struct hfsc_sched *q)
 282 {
 283         rb_node_t *n;
 284
 285         n = rb_first(&q->eligible);
 286         if (n == NULL)
 287                 return NULL;
 288         return rb_entry(n, struct hfsc_class, el_node);
 289 }
 290
 291 /*
 292  * vttree holds holds backlogged child classes being sorted by their virtual
 293  * time. each intermediate class has one vttree.
 294  */
 295 static void
 296 vttree_insert(struct hfsc_class *cl)
 297 {
 298         rb_node_t **p = &cl->cl_parent->vt_tree.rb_node;
 299         rb_node_t *parent = NULL;
 300         struct hfsc_class *cl1;
 301
 302         while (*p != NULL) {
 303                 parent = *p;
 304                 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
 305                 if (cl->cl_vt >= cl1->cl_vt)
 306                         p = &parent->rb_right;
 307                 else
 308                         p = &parent->rb_left;
 309         }
 310         rb_link_node(&cl->vt_node, parent, p);
 311         rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
 312 }
 313
 314 static inline void
 315 vttree_remove(struct hfsc_class *cl)
 316 {
 317         rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
 318 }
 319
 320 static inline void
 321 vttree_update(struct hfsc_class *cl)
 322 {
 323         vttree_remove(cl);
 324         vttree_insert(cl);
 325 }
 326
 327 static inline struct hfsc_class *
 328 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
 329 {
 330         struct hfsc_class *p;
 331         rb_node_t *n;
 332
 333         for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
 334                 p = rb_entry(n, struct hfsc_class, vt_node);
 335                 if (p->cl_f <= cur_time)
 336                         return p;
 337         }
 338         return NULL;
 339 }
 340
 341 /*
 342  * get the leaf class with the minimum vt in the hierarchy
 343  */
 344 static struct hfsc_class *
 345 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
 346 {
 347         /* if root-class's cfmin is bigger than cur_time nothing to do */
 348         if (cl->cl_cfmin > cur_time)
 349                 return NULL;
 350
 351         while (cl->level > 0) {
 352                 cl = vttree_firstfit(cl, cur_time);
 353                 if (cl == NULL)
 354                         return NULL;
 355                 /*
 356                  * update parent's cl_cvtmin.
 357                  */
 358                 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
 359                         cl->cl_parent->cl_cvtmin = cl->cl_vt;
 360         }
 361         return cl;
 362 }
 363
 364 static void
 365 cftree_insert(struct hfsc_class *cl)
 366 {
 367         rb_node_t **p = &cl->cl_parent->cf_tree.rb_node;
 368         rb_node_t *parent = NULL;
 369         struct hfsc_class *cl1;
 370
 371         while (*p != NULL) {
 372                 parent = *p;
 373                 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
 374                 if (cl->cl_f >= cl1->cl_f)
 375                         p = &parent->rb_right;
 376                 else
 377                         p = &parent->rb_left;
 378         }
 379         rb_link_node(&cl->cf_node, parent, p);
 380         rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
 381 }
 382
 383 static inline void
 384 cftree_remove(struct hfsc_class *cl)
 385 {
 386         rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
 387 }
 388
 389 static inline void
 390 cftree_update(struct hfsc_class *cl)
 391 {
 392         cftree_remove(cl);
 393         cftree_insert(cl);
 394 }
 395
 396 /*
 397  * service curve support functions
 398  *
 399  *  external service curve parameters
 400  *      m: bps
 401  *      d: us
 402  *  internal service curve parameters
 403  *      sm: (bytes/psched_us) << SM_SHIFT
 404  *      ism: (psched_us/byte) << ISM_SHIFT
 405  *      dx: psched_us
 406  *
 407  * Time source resolution
 408  *  PSCHED_JIFFIES: for 48<=HZ<=1534 resolution is between 0.63us and 1.27us.
 409  *  PSCHED_CPU: resolution is between 0.5us and 1us.
 410  *  PSCHED_GETTIMEOFDAY: resolution is exactly 1us.
 411  *
 412  * sm and ism are scaled in order to keep effective digits.
 413  * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
 414  * digits in decimal using the following table.
 415  *
 416  * Note: We can afford the additional accuracy (altq hfsc keeps at most
 417  * 3 effective digits) thanks to the fact that linux clock is bounded
 418  * much more tightly.
 419  *
 420  *  bits/sec      100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
 421  *  ------------+-------------------------------------------------------
 422  *  bytes/0.5us   6.25e-3    62.5e-3    625e-3     6250e-e    62500e-3
 423  *  bytes/us      12.5e-3    125e-3     1250e-3    12500e-3   125000e-3
 424  *  bytes/1.27us  15.875e-3  158.75e-3  1587.5e-3  15875e-3   158750e-3
 425  *
 426  *  0.5us/byte    160        16         1.6        0.16       0.016
 427  *  us/byte       80         8          0.8        0.08       0.008
 428  *  1.27us/byte   63         6.3        0.63       0.063      0.0063
 429  */
 430 #define SM_SHIFT        20
 431 #define ISM_SHIFT       18
 432
 433 #define SM_MASK         ((1ULL << SM_SHIFT) - 1)
 434 #define ISM_MASK        ((1ULL << ISM_SHIFT) - 1)
 435
 436 static inline u64
 437 seg_x2y(u64 x, u64 sm)
 438 {
 439         u64 y;
 440
 441         /*
 442          * compute
 443          *      y = x * sm >> SM_SHIFT
 444          * but divide it for the upper and lower bits to avoid overflow
 445          */
 446         y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
 447         return y;
 448 }
 449
 450 static inline u64
 451 seg_y2x(u64 y, u64 ism)
 452 {
 453         u64 x;
 454
 455         if (y == 0)
 456                 x = 0;
 457         else if (ism == HT_INFINITY)
 458                 x = HT_INFINITY;
 459         else {
 460                 x = (y >> ISM_SHIFT) * ism
 461                     + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
 462         }
 463         return x;
 464 }
 465
 466 /* Convert m (bps) into sm (bytes/psched us) */
 467 static u64
 468 m2sm(u32 m)
 469 {
 470         u64 sm;
 471
 472         sm = ((u64)m << SM_SHIFT);
 473         sm += PSCHED_JIFFIE2US(HZ) - 1;
 474         do_div(sm, PSCHED_JIFFIE2US(HZ));
 475         return sm;
 476 }
 477
 478 /* convert m (bps) into ism (psched us/byte) */
 479 static u64
 480 m2ism(u32 m)
 481 {
 482         u64 ism;
 483
 484         if (m == 0)
 485                 ism = HT_INFINITY;
 486         else {
 487                 ism = ((u64)PSCHED_JIFFIE2US(HZ) << ISM_SHIFT);
 488                 ism += m - 1;
 489                 do_div(ism, m);
 490         }
 491         return ism;
 492 }
 493
 494 /* convert d (us) into dx (psched us) */
 495 static u64
 496 d2dx(u32 d)
 497 {
 498         u64 dx;
 499
 500         dx = ((u64)d * PSCHED_JIFFIE2US(HZ));
 501         dx += 1000000 - 1;
 502         do_div(dx, 1000000);
 503         return dx;
 504 }
 505
 506 /* convert sm (bytes/psched us) into m (bps) */
 507 static u32
 508 sm2m(u64 sm)
 509 {
 510         u64 m;
 511
 512         m = (sm * PSCHED_JIFFIE2US(HZ)) >> SM_SHIFT;
 513         return (u32)m;
 514 }
 515
 516 /* convert dx (psched us) into d (us) */
 517 static u32
 518 dx2d(u64 dx)
 519 {
 520         u64 d;
 521
 522         d = dx * 1000000;
 523         do_div(d, PSCHED_JIFFIE2US(HZ));
 524         return (u32)d;
 525 }
 526
 527 static void
 528 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
 529 {
 530         isc->sm1  = m2sm(sc->m1);
 531         isc->ism1 = m2ism(sc->m1);
 532         isc->dx   = d2dx(sc->d);
 533         isc->dy   = seg_x2y(isc->dx, isc->sm1);
 534         isc->sm2  = m2sm(sc->m2);
 535         isc->ism2 = m2ism(sc->m2);
 536 }
 537
 538 /*
 539  * initialize the runtime service curve with the given internal
 540  * service curve starting at (x, y).
 541  */
 542 static void
 543 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
 544 {
 545         rtsc->x    = x;
 546         rtsc->y    = y;
 547         rtsc->sm1  = isc->sm1;
 548         rtsc->ism1 = isc->ism1;
 549         rtsc->dx   = isc->dx;
 550         rtsc->dy   = isc->dy;
 551         rtsc->sm2  = isc->sm2;
 552         rtsc->ism2 = isc->ism2;
 553 }
 554
 555 /*
 556  * calculate the y-projection of the runtime service curve by the
 557  * given x-projection value
 558  */
 559 static u64
 560 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
 561 {
 562         u64 x;
 563
 564         if (y < rtsc->y)
 565                 x = rtsc->x;
 566         else if (y <= rtsc->y + rtsc->dy) {
 567                 /* x belongs to the 1st segment */
 568                 if (rtsc->dy == 0)
 569                         x = rtsc->x + rtsc->dx;
 570                 else
 571                         x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
 572         } else {
 573                 /* x belongs to the 2nd segment */
 574                 x = rtsc->x + rtsc->dx
 575                     + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
 576         }
 577         return x;
 578 }
 579
 580 static u64
 581 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
 582 {
 583         u64 y;
 584
 585         if (x <= rtsc->x)
 586                 y = rtsc->y;
 587         else if (x <= rtsc->x + rtsc->dx)
 588                 /* y belongs to the 1st segment */
 589                 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
 590         else
 591                 /* y belongs to the 2nd segment */
 592                 y = rtsc->y + rtsc->dy
 593                     + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
 594         return y;
 595 }
 596
 597 /*
 598  * update the runtime service curve by taking the minimum of the current
 599  * runtime service curve and the service curve starting at (x, y).
 600  */
 601 static void
 602 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
 603 {
 604         u64 y1, y2, dx, dy;
 605         u32 dsm;
 606
 607         if (isc->sm1 <= isc->sm2) {
 608                 /* service curve is convex */
 609                 y1 = rtsc_x2y(rtsc, x);
 610                 if (y1 < y)
 611                         /* the current rtsc is smaller */
 612                         return;
 613                 rtsc->x = x;
 614                 rtsc->y = y;
 615                 return;
 616         }
 617
 618         /*
 619          * service curve is concave
 620          * compute the two y values of the current rtsc
 621          *      y1: at x
 622          *      y2: at (x + dx)
 623          */
 624         y1 = rtsc_x2y(rtsc, x);
 625         if (y1 <= y) {
 626                 /* rtsc is below isc, no change to rtsc */
 627                 return;
 628         }
 629
 630         y2 = rtsc_x2y(rtsc, x + isc->dx);
 631         if (y2 >= y + isc->dy) {
 632                 /* rtsc is above isc, replace rtsc by isc */
 633                 rtsc->x = x;
 634                 rtsc->y = y;
 635                 rtsc->dx = isc->dx;
 636                 rtsc->dy = isc->dy;
 637                 return;
 638         }
 639
 640         /*
 641          * the two curves intersect
 642          * compute the offsets (dx, dy) using the reverse
 643          * function of seg_x2y()
 644          *      seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
 645          */
 646         dx = (y1 - y) << SM_SHIFT;
 647         dsm = isc->sm1 - isc->sm2;
 648         do_div(dx, dsm);
 649         /*
 650          * check if (x, y1) belongs to the 1st segment of rtsc.
 651          * if so, add the offset.
 652          */
 653         if (rtsc->x + rtsc->dx > x)
 654                 dx += rtsc->x + rtsc->dx - x;
 655         dy = seg_x2y(dx, isc->sm1);
 656
 657         rtsc->x = x;
 658         rtsc->y = y;
 659         rtsc->dx = dx;
 660         rtsc->dy = dy;
 661         return;
 662 }
 663
 664 static void
 665 init_ed(struct hfsc_class *cl, unsigned int next_len)
 666 {
 667         u64 cur_time;
 668
 669         PSCHED_GET_TIME(cur_time);
 670
 671         /* update the deadline curve */
 672         rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
 673
 674         /*
 675          * update the eligible curve.
 676          * for concave, it is equal to the deadline curve.
 677          * for convex, it is a linear curve with slope m2.
 678          */
 679         cl->cl_eligible = cl->cl_deadline;
 680         if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
 681                 cl->cl_eligible.dx = 0;
 682                 cl->cl_eligible.dy = 0;
 683         }
 684
 685         /* compute e and d */
 686         cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
 687         cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 688
 689         eltree_insert(cl);
 690 }
 691
 692 static void
 693 update_ed(struct hfsc_class *cl, unsigned int next_len)
 694 {
 695         cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
 696         cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 697
 698         eltree_update(cl);
 699 }
 700
 701 static inline void
 702 update_d(struct hfsc_class *cl, unsigned int next_len)
 703 {
 704         cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
 705 }
 706
 707 static inline void
 708 update_cfmin(struct hfsc_class *cl)
 709 {
 710         rb_node_t *n = rb_first(&cl->cf_tree);
 711         struct hfsc_class *p;
 712
 713         if (n == NULL) {
 714                 cl->cl_cfmin = 0;
 715                 return;
 716         }
 717         p = rb_entry(n, struct hfsc_class, cf_node);
 718         cl->cl_cfmin = p->cl_f;
 719 }
 720
 721 static void
 722 init_vf(struct hfsc_class *cl, unsigned int len)
 723 {
 724         struct hfsc_class *max_cl;
 725         rb_node_t *n;
 726         u64 vt, f, cur_time;
 727         int go_active;
 728
 729         cur_time = 0;
 730         go_active = 1;
 731         for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
 732                 if (go_active && cl->cl_nactive++ == 0)
 733                         go_active = 1;
 734                 else
 735                         go_active = 0;
 736
 737                 if (go_active) {
 738                         n = rb_last(&cl->cl_parent->vt_tree);
 739                         if (n != NULL) {
 740                                 max_cl = rb_entry(n, struct hfsc_class,vt_node);
 741                                 /*
 742                                  * set vt to the average of the min and max
 743                                  * classes.  if the parent's period didn't
 744                                  * change, don't decrease vt of the class.
 745                                  */
 746                                 vt = max_cl->cl_vt;
 747                                 if (cl->cl_parent->cl_cvtmin != 0)
 748                                         vt = (cl->cl_parent->cl_cvtmin + vt)/2;
 749
 750                                 if (cl->cl_parent->cl_vtperiod !=
 751                                     cl->cl_parentperiod || vt > cl->cl_vt)
 752                                         cl->cl_vt = vt;
 753                         } else {
 754                                 /*
 755                                  * first child for a new parent backlog period.
 756                                  * add parent's cvtmax to cvtoff to make a new
 757                                  * vt (vtoff + vt) larger than the vt in the
 758                                  * last period for all children.
 759                                  */
 760                                 vt = cl->cl_parent->cl_cvtmax;
 761                                 cl->cl_parent->cl_cvtoff += vt;
 762                                 cl->cl_parent->cl_cvtmax = 0;
 763                                 cl->cl_parent->cl_cvtmin = 0;
 764                                 cl->cl_vt = 0;
 765                         }
 766
 767                         cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
 768                                                         cl->cl_pcvtoff;
 769
 770                         /* update the virtual curve */
 771                         vt = cl->cl_vt + cl->cl_vtoff;
 772                         rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
 773                                                       cl->cl_total);
 774                         if (cl->cl_virtual.x == vt) {
 775                                 cl->cl_virtual.x -= cl->cl_vtoff;
 776                                 cl->cl_vtoff = 0;
 777                         }
 778                         cl->cl_vtadj = 0;
 779
 780                         cl->cl_vtperiod++;  /* increment vt period */
 781                         cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
 782                         if (cl->cl_parent->cl_nactive == 0)
 783                                 cl->cl_parentperiod++;
 784                         cl->cl_f = 0;
 785
 786                         vttree_insert(cl);
 787                         cftree_insert(cl);
 788
 789                         if (cl->cl_flags & HFSC_USC) {
 790                                 /* class has upper limit curve */
 791                                 if (cur_time == 0)
 792                                         PSCHED_GET_TIME(cur_time);
 793
 794                                 /* update the ulimit curve */
 795                                 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
 796                                          cl->cl_total);
 797                                 /* compute myf */
 798                                 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
 799                                                       cl->cl_total);
 800                                 cl->cl_myfadj = 0;
 801                         }
 802                 }
 803
 804                 f = max(cl->cl_myf, cl->cl_cfmin);
 805                 if (f != cl->cl_f) {
 806                         cl->cl_f = f;
 807                         cftree_update(cl);
 808                         update_cfmin(cl->cl_parent);
 809                 }
 810         }
 811 }
 812
 813 static void
 814 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
 815 {
 816         u64 f; /* , myf_bound, delta; */
 817         int go_passive = 0;
 818
 819         if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
 820                 go_passive = 1;
 821
 822         for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
 823                 cl->cl_total += len;
 824
 825                 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
 826                         continue;
 827
 828                 if (go_passive && --cl->cl_nactive == 0)
 829                         go_passive = 1;
 830                 else
 831                         go_passive = 0;
 832
 833                 if (go_passive) {
 834                         /* no more active child, going passive */
 835
 836                         /* update cvtmax of the parent class */
 837                         if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
 838                                 cl->cl_parent->cl_cvtmax = cl->cl_vt;
 839
 840                         /* remove this class from the vt tree */
 841                         vttree_remove(cl);
 842
 843                         cftree_remove(cl);
 844                         update_cfmin(cl->cl_parent);
 845
 846                         continue;
 847                 }
 848
 849                 /*
 850                  * update vt and f
 851                  */
 852                 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
 853                             - cl->cl_vtoff + cl->cl_vtadj;
 854
 855                 /*
 856                  * if vt of the class is smaller than cvtmin,
 857                  * the class was skipped in the past due to non-fit.
 858                  * if so, we need to adjust vtadj.
 859                  */
 860                 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
 861                         cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
 862                         cl->cl_vt = cl->cl_parent->cl_cvtmin;
 863                 }
 864
 865                 /* update the vt tree */
 866                 vttree_update(cl);
 867
 868                 if (cl->cl_flags & HFSC_USC) {
 869                         cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
 870                                                               cl->cl_total);
 871 #if 0
 872                         /*
 873                          * This code causes classes to stay way under their
 874                          * limit when multiple classes are used at gigabit
 875                          * speed. needs investigation. -kaber
 876                          */
 877                         /*
 878                          * if myf lags behind by more than one clock tick
 879                          * from the current time, adjust myfadj to prevent
 880                          * a rate-limited class from going greedy.
 881                          * in a steady state under rate-limiting, myf
 882                          * fluctuates within one clock tick.
 883                          */
 884                         myf_bound = cur_time - PSCHED_JIFFIE2US(1);
 885                         if (cl->cl_myf < myf_bound) {
 886                                 delta = cur_time - cl->cl_myf;
 887                                 cl->cl_myfadj += delta;
 888                                 cl->cl_myf += delta;
 889                         }
 890 #endif
 891                 }
 892
 893                 f = max(cl->cl_myf, cl->cl_cfmin);
 894                 if (f != cl->cl_f) {
 895                         cl->cl_f = f;
 896                         cftree_update(cl);
 897                         update_cfmin(cl->cl_parent);
 898                 }
 899         }
 900 }
 901
 902 static void
 903 set_active(struct hfsc_class *cl, unsigned int len)
 904 {
 905         if (cl->cl_flags & HFSC_RSC)
 906                 init_ed(cl, len);
 907         if (cl->cl_flags & HFSC_FSC)
 908                 init_vf(cl, len);
 909
 910         list_add_tail(&cl->dlist, &cl->sched->droplist);
 911 }
 912
 913 static void
 914 set_passive(struct hfsc_class *cl)
 915 {
 916         if (cl->cl_flags & HFSC_RSC)
 917                 eltree_remove(cl);
 918
 919         list_del(&cl->dlist);
 920
 921         /*
 922          * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
 923          * needs to be called explicitly to remove a class from vttree.
 924          */
 925 }
 926
 927 /*
 928  * hack to get length of first packet in queue.
 929  */
 930 static unsigned int
 931 qdisc_peek_len(struct Qdisc *sch)
 932 {
 933         struct sk_buff *skb;
 934         unsigned int len;
 935
 936         skb = sch->dequeue(sch);
 937         if (skb == NULL) {
 938                 if (net_ratelimit())
 939                         printk("qdisc_peek_len: non work-conserving qdisc ?\n");
 940                 return 0;
 941         }
 942         len = skb->len;
 943         if (unlikely(sch->ops->requeue(skb, sch) != NET_XMIT_SUCCESS)) {
 944                 if (net_ratelimit())
 945                         printk("qdisc_peek_len: failed to requeue\n");
 946                 return 0;
 947         }
 948         return len;
 949 }
 950
 951 static void
 952 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
 953 {
 954         unsigned int len = cl->qdisc->q.qlen;
 955
 956         qdisc_reset(cl->qdisc);
 957         if (len > 0) {
 958                 update_vf(cl, 0, 0);
 959                 set_passive(cl);
 960                 sch->q.qlen -= len;
 961         }
 962 }
 963
 964 static void
 965 hfsc_adjust_levels(struct hfsc_class *cl)
 966 {
 967         struct hfsc_class *p;
 968         unsigned int level;
 969
 970         do {
 971                 level = 0;
 972                 list_for_each_entry(p, &cl->children, siblings) {
 973                         if (p->level > level)
 974                                 level = p->level;
 975                 }
 976                 cl->level = level + 1;
 977         } while ((cl = cl->cl_parent) != NULL);
 978 }
 979
 980 static inline unsigned int
 981 hfsc_hash(u32 h)
 982 {
 983         h ^= h >> 8;
 984         h ^= h >> 4;
 985
 986         return h & (HFSC_HSIZE - 1);
 987 }
 988
 989 static inline struct hfsc_class *
 990 hfsc_find_class(u32 classid, struct Qdisc *sch)
 991 {
 992         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
 993         struct hfsc_class *cl;
 994
 995         list_for_each_entry(cl, &q->clhash[hfsc_hash(classid)], hlist) {
 996                 if (cl->classid == classid)
 997                         return cl;
 998         }
 999         return NULL;
1000 }
1001
1002 static void
1003 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
1004                 u64 cur_time)
1005 {
1006         sc2isc(rsc, &cl->cl_rsc);
1007         rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
1008         cl->cl_eligible = cl->cl_deadline;
1009         if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
1010                 cl->cl_eligible.dx = 0;
1011                 cl->cl_eligible.dy = 0;
1012         }
1013         cl->cl_flags |= HFSC_RSC;
1014 }
1015
1016 static void
1017 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
1018 {
1019         sc2isc(fsc, &cl->cl_fsc);
1020         rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
1021         cl->cl_flags |= HFSC_FSC;
1022 }
1023
1024 static void
1025 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
1026                 u64 cur_time)
1027 {
1028         sc2isc(usc, &cl->cl_usc);
1029         rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
1030         cl->cl_flags |= HFSC_USC;
1031 }
1032
1033 static int
1034 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
1035                   struct rtattr **tca, unsigned long *arg)
1036 {
1037         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1038         struct hfsc_class *cl = (struct hfsc_class *)*arg;
1039         struct hfsc_class *parent = NULL;
1040         struct rtattr *opt = tca[TCA_OPTIONS-1];
1041         struct rtattr *tb[TCA_HFSC_MAX];
1042         struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
1043         u64 cur_time;
1044
1045         if (opt == NULL ||
1046             rtattr_parse(tb, TCA_HFSC_MAX, RTA_DATA(opt), RTA_PAYLOAD(opt)))
1047                 return -EINVAL;
1048
1049         if (tb[TCA_HFSC_RSC-1]) {
1050                 if (RTA_PAYLOAD(tb[TCA_HFSC_RSC-1]) < sizeof(*rsc))
1051                         return -EINVAL;
1052                 rsc = RTA_DATA(tb[TCA_HFSC_RSC-1]);
1053                 if (rsc->m1 == 0 && rsc->m2 == 0)
1054                         rsc = NULL;
1055         }
1056
1057         if (tb[TCA_HFSC_FSC-1]) {
1058                 if (RTA_PAYLOAD(tb[TCA_HFSC_FSC-1]) < sizeof(*fsc))
1059                         return -EINVAL;
1060                 fsc = RTA_DATA(tb[TCA_HFSC_FSC-1]);
1061                 if (fsc->m1 == 0 && fsc->m2 == 0)
1062                         fsc = NULL;
1063         }
1064
1065         if (tb[TCA_HFSC_USC-1]) {
1066                 if (RTA_PAYLOAD(tb[TCA_HFSC_USC-1]) < sizeof(*usc))
1067                         return -EINVAL;
1068                 usc = RTA_DATA(tb[TCA_HFSC_USC-1]);
1069                 if (usc->m1 == 0 && usc->m2 == 0)
1070                         usc = NULL;
1071         }
1072
1073         if (cl != NULL) {
1074                 if (parentid) {
1075                         if (cl->cl_parent && cl->cl_parent->classid != parentid)
1076                                 return -EINVAL;
1077                         if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1078                                 return -EINVAL;
1079                 }
1080                 PSCHED_GET_TIME(cur_time);
1081
1082                 sch_tree_lock(sch);
1083                 if (rsc != NULL)
1084                         hfsc_change_rsc(cl, rsc, cur_time);
1085                 if (fsc != NULL)
1086                         hfsc_change_fsc(cl, fsc);
1087                 if (usc != NULL)
1088                         hfsc_change_usc(cl, usc, cur_time);
1089
1090                 if (cl->qdisc->q.qlen != 0) {
1091                         if (cl->cl_flags & HFSC_RSC)
1092                                 update_ed(cl, qdisc_peek_len(cl->qdisc));
1093                         if (cl->cl_flags & HFSC_FSC)
1094                                 update_vf(cl, 0, cur_time);
1095                 }
1096                 sch_tree_unlock(sch);
1097
1098 #ifdef CONFIG_NET_ESTIMATOR
1099                 if (tca[TCA_RATE-1]) {
1100                         qdisc_kill_estimator(&cl->stats);
1101                         qdisc_new_estimator(&cl->stats, tca[TCA_RATE-1]);
1102                 }
1103 #endif
1104                 return 0;
1105         }
1106
1107         if (parentid == TC_H_ROOT)
1108                 return -EEXIST;
1109
1110         parent = &q->root;
1111         if (parentid) {
1112                 parent = hfsc_find_class(parentid, sch);
1113                 if (parent == NULL)
1114                         return -ENOENT;
1115         }
1116
1117         if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1118                 return -EINVAL;
1119         if (hfsc_find_class(classid, sch))
1120                 return -EEXIST;
1121
1122         if (rsc == NULL && fsc == NULL)
1123                 return -EINVAL;
1124
1125         cl = kmalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1126         if (cl == NULL)
1127                 return -ENOBUFS;
1128         memset(cl, 0, sizeof(struct hfsc_class));
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                 qdisc_new_estimator(&cl->stats, tca[TCA_RATE-1]);
1161 #endif
1162         *arg = (unsigned long)cl;
1163         return 0;
1164 }
1165
1166 static void
1167 hfsc_destroy_filters(struct tcf_proto **fl)
1168 {
1169         struct tcf_proto *tp;
1170
1171         while ((tp = *fl) != NULL) {
1172                 *fl = tp->next;
1173                 tcf_destroy(tp);
1174         }
1175 }
1176
1177 static void
1178 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1179 {
1180         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1181
1182         hfsc_destroy_filters(&cl->filter_list);
1183         qdisc_destroy(cl->qdisc);
1184 #ifdef CONFIG_NET_ESTIMATOR
1185         qdisc_kill_estimator(&cl->stats);
1186 #endif
1187         if (cl != &q->root)
1188                 kfree(cl);
1189 }
1190
1191 static int
1192 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1193 {
1194         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1195         struct hfsc_class *cl = (struct hfsc_class *)arg;
1196
1197         if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1198                 return -EBUSY;
1199
1200         sch_tree_lock(sch);
1201
1202         list_del(&cl->hlist);
1203         list_del(&cl->siblings);
1204         hfsc_adjust_levels(cl->cl_parent);
1205         hfsc_purge_queue(sch, cl);
1206         if (--cl->refcnt == 0)
1207                 hfsc_destroy_class(sch, cl);
1208
1209         sch_tree_unlock(sch);
1210         return 0;
1211 }
1212
1213 static struct hfsc_class *
1214 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch)
1215 {
1216         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1217         struct hfsc_class *cl;
1218         struct tcf_result res;
1219         struct tcf_proto *tcf;
1220         int result;
1221
1222         if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1223             (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1224                 if (cl->level == 0)
1225                         return cl;
1226
1227         tcf = q->root.filter_list;
1228         while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1229 #ifdef CONFIG_NET_CLS_POLICE
1230                 if (result == TC_POLICE_SHOT)
1231                         return NULL;
1232 #endif
1233                 if ((cl = (struct hfsc_class *)res.class) == NULL) {
1234                         if ((cl = hfsc_find_class(res.classid, sch)) == NULL)
1235                                 break; /* filter selected invalid classid */
1236                 }
1237
1238                 if (cl->level == 0)
1239                         return cl; /* hit leaf class */
1240
1241                 /* apply inner filter chain */
1242                 tcf = cl->filter_list;
1243         }
1244
1245         /* classification failed, try default class */
1246         cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1247         if (cl == NULL || cl->level > 0)
1248                 return NULL;
1249
1250         return cl;
1251 }
1252
1253 static int
1254 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1255                  struct Qdisc **old)
1256 {
1257         struct hfsc_class *cl = (struct hfsc_class *)arg;
1258
1259         if (cl == NULL)
1260                 return -ENOENT;
1261         if (cl->level > 0)
1262                 return -EINVAL;
1263         if (new == NULL) {
1264                 new = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
1265                 if (new == NULL)
1266                         new = &noop_qdisc;
1267         }
1268
1269         sch_tree_lock(sch);
1270         hfsc_purge_queue(sch, cl);
1271         *old = xchg(&cl->qdisc, new);
1272         sch_tree_unlock(sch);
1273         return 0;
1274 }
1275
1276 static struct Qdisc *
1277 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1278 {
1279         struct hfsc_class *cl = (struct hfsc_class *)arg;
1280
1281         if (cl != NULL && cl->level == 0)
1282                 return cl->qdisc;
1283
1284         return NULL;
1285 }
1286
1287 static unsigned long
1288 hfsc_get_class(struct Qdisc *sch, u32 classid)
1289 {
1290         struct hfsc_class *cl = hfsc_find_class(classid, sch);
1291
1292         if (cl != NULL)
1293                 cl->refcnt++;
1294
1295         return (unsigned long)cl;
1296 }
1297
1298 static void
1299 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1300 {
1301         struct hfsc_class *cl = (struct hfsc_class *)arg;
1302
1303         if (--cl->refcnt == 0)
1304                 hfsc_destroy_class(sch, cl);
1305 }
1306
1307 static unsigned long
1308 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1309 {
1310         struct hfsc_class *p = (struct hfsc_class *)parent;
1311         struct hfsc_class *cl = hfsc_find_class(classid, sch);
1312
1313         if (cl != NULL) {
1314                 if (p != NULL && p->level <= cl->level)
1315                         return 0;
1316                 cl->filter_cnt++;
1317         }
1318
1319         return (unsigned long)cl;
1320 }
1321
1322 static void
1323 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1324 {
1325         struct hfsc_class *cl = (struct hfsc_class *)arg;
1326
1327         cl->filter_cnt--;
1328 }
1329
1330 static struct tcf_proto **
1331 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1332 {
1333         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1334         struct hfsc_class *cl = (struct hfsc_class *)arg;
1335
1336         if (cl == NULL)
1337                 cl = &q->root;
1338
1339         return &cl->filter_list;
1340 }
1341
1342 static int
1343 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1344 {
1345         struct tc_service_curve tsc;
1346
1347         tsc.m1 = sm2m(sc->sm1);
1348         tsc.d  = dx2d(sc->dx);
1349         tsc.m2 = sm2m(sc->sm2);
1350         RTA_PUT(skb, attr, sizeof(tsc), &tsc);
1351
1352         return skb->len;
1353
1354  rtattr_failure:
1355         return -1;
1356 }
1357
1358 static inline int
1359 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1360 {
1361         if ((cl->cl_flags & HFSC_RSC) &&
1362             (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1363                 goto rtattr_failure;
1364
1365         if ((cl->cl_flags & HFSC_FSC) &&
1366             (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1367                 goto rtattr_failure;
1368
1369         if ((cl->cl_flags & HFSC_USC) &&
1370             (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1371                 goto rtattr_failure;
1372
1373         return skb->len;
1374
1375  rtattr_failure:
1376         return -1;
1377 }
1378
1379 static inline int
1380 hfsc_dump_stats(struct sk_buff *skb, struct hfsc_class *cl)
1381 {
1382         cl->stats.qlen = cl->qdisc->q.qlen;
1383         if (qdisc_copy_stats(skb, &cl->stats) < 0)
1384                 goto rtattr_failure;
1385
1386         return skb->len;
1387
1388  rtattr_failure:
1389         return -1;
1390 }
1391
1392 static inline int
1393 hfsc_dump_xstats(struct sk_buff *skb, struct hfsc_class *cl)
1394 {
1395         struct tc_hfsc_stats xstats;
1396
1397         xstats.level  = cl->level;
1398         xstats.period = cl->cl_vtperiod;
1399         xstats.work   = cl->cl_total;
1400         xstats.rtwork = cl->cl_cumul;
1401         RTA_PUT(skb, TCA_XSTATS, sizeof(xstats), &xstats);
1402
1403         return skb->len;
1404
1405  rtattr_failure:
1406         return -1;
1407 }
1408
1409 static int
1410 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1411                 struct tcmsg *tcm)
1412 {
1413         struct hfsc_class *cl = (struct hfsc_class *)arg;
1414         unsigned char *b = skb->tail;
1415         struct rtattr *rta = (struct rtattr *)b;
1416
1417         tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->classid : TC_H_ROOT;
1418         tcm->tcm_handle = cl->classid;
1419         if (cl->level == 0)
1420                 tcm->tcm_info = cl->qdisc->handle;
1421
1422         RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
1423         if (hfsc_dump_curves(skb, cl) < 0)
1424                 goto rtattr_failure;
1425         rta->rta_len = skb->tail - b;
1426
1427         if ((hfsc_dump_stats(skb, cl) < 0) ||
1428             (hfsc_dump_xstats(skb, cl) < 0))
1429                 goto rtattr_failure;
1430
1431         return skb->len;
1432
1433  rtattr_failure:
1434         skb_trim(skb, b - skb->data);
1435         return -1;
1436 }
1437
1438 static void
1439 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1440 {
1441         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
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 = (struct hfsc_sched *)sch->data;
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 = (struct hfsc_sched *)sch->data;
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         MOD_INC_USE_COUNT;
1532         return 0;
1533 }
1534
1535 static int
1536 hfsc_change_qdisc(struct Qdisc *sch, struct rtattr *opt)
1537 {
1538         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1539         struct tc_hfsc_qopt *qopt;
1540
1541         if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1542                 return -EINVAL;;
1543         qopt = RTA_DATA(opt);
1544
1545         sch_tree_lock(sch);
1546         q->defcls = qopt->defcls;
1547         sch_tree_unlock(sch);
1548
1549         return 0;
1550 }
1551
1552 static void
1553 hfsc_reset_class(struct hfsc_class *cl)
1554 {
1555         cl->cl_total        = 0;
1556         cl->cl_cumul        = 0;
1557         cl->cl_d            = 0;
1558         cl->cl_e            = 0;
1559         cl->cl_vt           = 0;
1560         cl->cl_vtadj        = 0;
1561         cl->cl_vtoff        = 0;
1562         cl->cl_cvtmin       = 0;
1563         cl->cl_cvtmax       = 0;
1564         cl->cl_cvtoff       = 0;
1565         cl->cl_pcvtoff      = 0;
1566         cl->cl_vtperiod     = 0;
1567         cl->cl_parentperiod = 0;
1568         cl->cl_f            = 0;
1569         cl->cl_myf          = 0;
1570         cl->cl_myfadj       = 0;
1571         cl->cl_cfmin        = 0;
1572         cl->cl_nactive      = 0;
1573
1574         cl->vt_tree = RB_ROOT;
1575         cl->cf_tree = RB_ROOT;
1576         qdisc_reset(cl->qdisc);
1577
1578         if (cl->cl_flags & HFSC_RSC)
1579                 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1580         if (cl->cl_flags & HFSC_FSC)
1581                 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1582         if (cl->cl_flags & HFSC_USC)
1583                 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1584 }
1585
1586 static void
1587 hfsc_reset_qdisc(struct Qdisc *sch)
1588 {
1589         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1590         struct hfsc_class *cl;
1591         unsigned int i;
1592
1593         for (i = 0; i < HFSC_HSIZE; i++) {
1594                 list_for_each_entry(cl, &q->clhash[i], hlist)
1595                         hfsc_reset_class(cl);
1596         }
1597         __skb_queue_purge(&q->requeue);
1598         q->eligible = RB_ROOT;
1599         INIT_LIST_HEAD(&q->droplist);
1600         del_timer(&q->wd_timer);
1601         sch->flags &= ~TCQ_F_THROTTLED;
1602         sch->q.qlen = 0;
1603 }
1604
1605 static void
1606 hfsc_destroy_qdisc(struct Qdisc *sch)
1607 {
1608         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1609         struct hfsc_class *cl, *next;
1610         unsigned int i;
1611
1612         for (i = 0; i < HFSC_HSIZE; i++) {
1613                 list_for_each_entry_safe(cl, next, &q->clhash[i], hlist)
1614                         hfsc_destroy_class(sch, cl);
1615         }
1616         __skb_queue_purge(&q->requeue);
1617         del_timer(&q->wd_timer);
1618         MOD_DEC_USE_COUNT;
1619 }
1620
1621 static int
1622 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1623 {
1624         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1625         unsigned char *b = skb->tail;
1626         struct tc_hfsc_qopt qopt;
1627
1628         qopt.defcls = q->defcls;
1629         RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1630
1631         return skb->len;
1632
1633  rtattr_failure:
1634         skb_trim(skb, b - skb->data);
1635         return -1;
1636 }
1637
1638 static int
1639 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1640 {
1641         struct hfsc_class *cl = hfsc_classify(skb, sch);
1642         unsigned int len = skb->len;
1643         int err;
1644
1645         if (cl == NULL) {
1646                 kfree_skb(skb);
1647                 sch->stats.drops++;
1648                 return NET_XMIT_DROP;
1649         }
1650
1651         err = cl->qdisc->enqueue(skb, cl->qdisc);
1652         if (unlikely(err != NET_XMIT_SUCCESS)) {
1653                 cl->stats.drops++;
1654                 sch->stats.drops++;
1655                 return err;
1656         }
1657
1658         if (cl->qdisc->q.qlen == 1)
1659                 set_active(cl, len);
1660
1661         cl->stats.packets++;
1662         cl->stats.bytes += len;
1663         sch->stats.packets++;
1664         sch->stats.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 = (struct hfsc_sched *)sch->data;
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->stats.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 = (struct hfsc_sched *)sch->data;
1743
1744         __skb_queue_head(&q->requeue, skb);
1745         sch->q.qlen++;
1746         return NET_XMIT_SUCCESS;
1747 }
1748
1749 static unsigned int
1750 hfsc_drop(struct Qdisc *sch)
1751 {
1752         struct hfsc_sched *q = (struct hfsc_sched *)sch->data;
1753         struct hfsc_class *cl;
1754         unsigned int len;
1755
1756         list_for_each_entry(cl, &q->droplist, dlist) {
1757                 if (cl->qdisc->ops->drop != NULL &&
1758                     (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1759                         if (cl->qdisc->q.qlen == 0) {
1760                                 update_vf(cl, 0, 0);
1761                                 set_passive(cl);
1762                         } else {
1763                                 list_move_tail(&cl->dlist, &q->droplist);
1764                         }
1765                         cl->stats.drops++;
1766                         sch->stats.drops++;
1767                         sch->q.qlen--;
1768                         return len;
1769                 }
1770         }
1771         return 0;
1772 }
1773
1774 static struct Qdisc_class_ops hfsc_class_ops = {
1775         .change         = hfsc_change_class,
1776         .delete         = hfsc_delete_class,
1777         .graft          = hfsc_graft_class,
1778         .leaf           = hfsc_class_leaf,
1779         .get            = hfsc_get_class,
1780         .put            = hfsc_put_class,
1781         .bind_tcf       = hfsc_bind_tcf,
1782         .unbind_tcf     = hfsc_unbind_tcf,
1783         .tcf_chain      = hfsc_tcf_chain,
1784         .dump           = hfsc_dump_class,
1785         .walk           = hfsc_walk
1786 };
1787
1788 struct Qdisc_ops hfsc_qdisc_ops = {
1789         .id             = "hfsc",
1790         .init           = hfsc_init_qdisc,
1791         .change         = hfsc_change_qdisc,
1792         .reset          = hfsc_reset_qdisc,
1793         .destroy        = hfsc_destroy_qdisc,
1794         .dump           = hfsc_dump_qdisc,
1795         .enqueue        = hfsc_enqueue,
1796         .dequeue        = hfsc_dequeue,
1797         .requeue        = hfsc_requeue,
1798         .drop           = hfsc_drop,
1799         .cl_ops         = &hfsc_class_ops,
1800         .priv_size      = sizeof(struct hfsc_sched)
1801 };
1802
1803 static int __init
1804 hfsc_init(void)
1805 {
1806         return register_qdisc(&hfsc_qdisc_ops);
1807 }
1808
1809 static void __exit
1810 hfsc_cleanup(void)
1811 {
1812         unregister_qdisc(&hfsc_qdisc_ops);
1813 }
1814
1815 MODULE_LICENSE("GPL");
1816 module_init(hfsc_init);
1817 module_exit(hfsc_cleanup);