include/linux/ckrm_sched.h

   1 /* include/linux/ckrm_sched.h - Supports CKRM scheduling
   2  *
   3  * Copyright (C) Haoqiang Zheng,  IBM Corp. 2004
   4  * Copyright (C) Hubertus Franke,  IBM Corp. 2004
   5  *
   6  * Latest version, more details at http://ckrm.sf.net
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  */
  14
  15 #ifndef _CKRM_SCHED_H
  16 #define _CKRM_SCHED_H
  17
  18 #include <linux/sched.h>
  19 #include <linux/ckrm_rc.h>
  20 #include <linux/ckrm_classqueue.h>
  21
  22 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
  23
  24 struct prio_array {
  25         unsigned int nr_active;
  26         unsigned long bitmap[BITMAP_SIZE];
  27         struct list_head queue[MAX_PRIO];
  28 };
  29
  30 #ifdef CONFIG_CKRM_CPU_SCHEDULE
  31 #define rq_active(p,rq)   (get_task_lrq(p)->active)
  32 #define rq_expired(p,rq)  (get_task_lrq(p)->expired)
  33 int __init init_ckrm_sched_res(void);
  34 #else
  35 #define rq_active(p,rq)   (rq->active)
  36 #define rq_expired(p,rq)  (rq->expired)
  37 static inline void init_ckrm_sched_res(void) {}
  38 static inline int ckrm_cpu_monitor_init(void) {return 0;}
  39 #endif //CONFIG_CKRM_CPU_SCHEDULE
  40
  41 #ifdef CONFIG_CKRM_CPU_SCHEDULE
  42 struct ckrm_runqueue {
  43         cq_node_t classqueue_linkobj;   /*links in classqueue */
  44         struct ckrm_cpu_class *cpu_class;       // class it belongs to
  45         struct classqueue_struct *classqueue;   // classqueue it belongs tow
  46         unsigned long long uncounted_ns;
  47
  48         prio_array_t *active, *expired, arrays[2];
  49         /*
  50            set to 0 on init, become null or array switch
  51            set to jiffies whenever an non-interactive job expires
  52            reset to jiffies if expires
  53          */
  54         unsigned long expired_timestamp;
  55
  56         /*
  57          * highest priority of tasks in active
  58          * initialized to be MAX_PRIO
  59          * updated on enqueue, dequeue
  60          */
  61         int top_priority;
  62         CVT_t local_cvt;
  63
  64         unsigned long lrq_load;
  65         int local_weight;
  66
  67
  68         /*
  69          * unused CPU time accumulated while thoe class
  70          * is inactive goes to savings
  71          *
  72          * initialized to be 0
  73          * a class can't accumulate more than SAVING_THRESHOLD of savings
  74          * savings are kept in normalized form (like cvt)
  75          * so when task share change the savings should be scaled accordingly
  76          */
  77         unsigned long long savings;
  78
  79         unsigned long magic;    //for debugging
  80 };
  81
  82 typedef struct ckrm_runqueue ckrm_lrq_t;
  83
  84 /**
  85  * ckrm_cpu_class_stat - cpu usage statistics maintained for each class
  86  *
  87  */
  88 struct ckrm_cpu_class_stat {
  89         spinlock_t stat_lock;
  90
  91         unsigned long long total_ns;    /*how much nano-secs it has consumed */
  92
  93         struct ckrm_cpu_demand_stat local_stats[NR_CPUS];
  94
  95         /*
  96          *
  97          */
  98         unsigned long max_demand; /* the maximun a class can consume */
  99         int egrt,megrt; /*effective guarantee*/
 100         int ehl,mehl; /*effective hard limit, my effective hard limit*/
 101
 102         /*
 103          * eshare: for both default class and its children
 104          * meshare: just for the default class
 105          */
 106         int eshare;
 107         int meshare;
 108 };
 109
 110 #define CKRM_CPU_CLASS_MAGIC 0x7af2abe3
 111
 112 #define USAGE_SAMPLE_FREQ HZ  //sample every 1 seconds
 113 #define NS_PER_SAMPLE (USAGE_SAMPLE_FREQ*(NSEC_PER_SEC/HZ))
 114 #define USAGE_WINDOW_SIZE 60  //keep the last 60 sample
 115
 116 struct ckrm_usage {
 117         unsigned long samples[USAGE_WINDOW_SIZE]; //record usages
 118         unsigned long sample_pointer; //pointer for the sliding window
 119         unsigned long long last_ns; //ns for last sample
 120         long long last_sample_jiffies; //in number of jiffies
 121 };
 122
 123 /*
 124  * manages the class status
 125  * there should be only one instance of this object for each class in the whole system
 126  */
 127 struct ckrm_cpu_class {
 128         struct ckrm_core_class *core;
 129         struct ckrm_core_class *parent;
 130         struct ckrm_shares shares;
 131         spinlock_t cnt_lock;    // always grab parent's lock first and then child's
 132         struct ckrm_cpu_class_stat stat;
 133         struct list_head links; // for linking up in cpu classes
 134         ckrm_lrq_t local_queues[NR_CPUS];       // runqueues
 135         struct ckrm_usage usage;
 136         unsigned long magic;    //for debugging
 137 };
 138
 139 #define cpu_class_weight(cls) (cls->stat.meshare)
 140 #define local_class_weight(lrq) (lrq->local_weight)
 141
 142 static inline int valid_cpu_class(struct ckrm_cpu_class * cls)
 143 {
 144         return (cls && cls->magic == CKRM_CPU_CLASS_MAGIC);
 145 }
 146
 147 struct classqueue_struct *get_cpu_classqueue(int cpu);
 148 struct ckrm_cpu_class * get_default_cpu_class(void);
 149
 150
 151 static inline void ckrm_usage_init(struct ckrm_usage* usage)
 152 {
 153         int i;
 154
 155         for (i=0; i < USAGE_WINDOW_SIZE; i++)
 156                 usage->samples[i] = 0;
 157         usage->sample_pointer = 0;
 158         usage->last_ns = 0;
 159         usage->last_sample_jiffies = 0;
 160 }
 161
 162 /*
 163  * this function can be called at any frequency
 164  * it's self-contained
 165  */
 166 static inline void ckrm_sample_usage(struct ckrm_cpu_class* clsptr)
 167 {
 168         struct ckrm_usage* usage = &clsptr->usage;
 169         unsigned long long cur_sample;
 170         int duration = jiffies - usage->last_sample_jiffies;
 171
 172         //jiffies wasn't start from 0
 173         //so it need to be properly handled
 174         if (unlikely(!usage->last_sample_jiffies))
 175                 usage->last_sample_jiffies = jiffies;
 176
 177         //called too frequenctly
 178         if (duration < USAGE_SAMPLE_FREQ)
 179                 return;
 180
 181         usage->last_sample_jiffies = jiffies;
 182
 183         cur_sample = clsptr->stat.total_ns - usage->last_ns;
 184         usage->last_ns = clsptr->stat.total_ns;
 185
 186         //scale it based on the sample duration
 187         cur_sample *= ((USAGE_SAMPLE_FREQ<< 15)/duration);
 188         cur_sample >>= 15;
 189         usage->samples[usage->sample_pointer] = cur_sample;
 190         //      printk("sample = %llu jiffies=%lu \n",cur_sample, jiffies);
 191
 192         usage->sample_pointer ++;
 193         if (usage->sample_pointer >= USAGE_WINDOW_SIZE)
 194                 usage->sample_pointer = 0;
 195 }
 196
 197 //duration is specified in number of jiffies
 198 //return the usage in percentage
 199 static inline int get_ckrm_usage(struct ckrm_cpu_class* clsptr, int duration)
 200 {
 201         int nr_samples = duration/USAGE_SAMPLE_FREQ?:1;
 202         struct ckrm_usage* usage = &clsptr->usage;
 203         unsigned long long total = 0;
 204         int i, idx;
 205
 206         if (nr_samples > USAGE_WINDOW_SIZE)
 207                 nr_samples = USAGE_WINDOW_SIZE;
 208
 209         idx = usage->sample_pointer;
 210         for (i = 0; i< nr_samples; i++) {
 211                 if (! idx)
 212                         idx = USAGE_WINDOW_SIZE;
 213                 idx --;
 214                 total += usage->samples[idx];
 215         }
 216         total *= 100;
 217         do_div(total,nr_samples);
 218         do_div(total,NS_PER_SAMPLE);
 219         do_div(total,cpus_weight(cpu_online_map));
 220         return total;
 221 }
 222
 223
 224 #define lrq_nr_running(lrq) \
 225              (lrq->active->nr_active + lrq->expired->nr_active)
 226
 227 static inline ckrm_lrq_t *
 228 get_ckrm_lrq(struct ckrm_cpu_class*cls, int cpu)
 229 {
 230         return &(cls->local_queues[cpu]);
 231 }
 232
 233 static inline ckrm_lrq_t *get_task_lrq(struct task_struct *p)
 234 {
 235         return &(p->cpu_class->local_queues[task_cpu(p)]);
 236 }
 237
 238 #define task_list_entry(list)  list_entry(list,struct task_struct,run_list)
 239 #define class_list_entry(list) list_entry(list,struct ckrm_runqueue,classqueue_linkobj)
 240
 241 /* some additional interfaces exported from sched.c */
 242 struct runqueue;
 243 extern rwlock_t class_list_lock;
 244 extern struct list_head active_cpu_classes;
 245 unsigned int task_timeslice(task_t *p);
 246 void _ckrm_cpu_change_class(task_t *task, struct ckrm_cpu_class *newcls);
 247
 248 void init_cpu_classes(void);
 249 void init_cpu_class(struct ckrm_cpu_class *cls,ckrm_shares_t* shares);
 250 void ckrm_cpu_change_class(void *task, void *old, void *new);
 251
 252
 253 #define CPU_DEMAND_ENQUEUE 0
 254 #define CPU_DEMAND_DEQUEUE 1
 255 #define CPU_DEMAND_DESCHEDULE 2
 256 #define CPU_DEMAND_INIT 3
 257
 258 /*functions exported by ckrm_cpu_monitor.c*/
 259 void ckrm_cpu_monitor(void);
 260 int ckrm_cpu_monitor_init(void);
 261 void ckrm_cpu_stat_init(struct ckrm_cpu_class_stat *stat);
 262 void cpu_demand_event(struct ckrm_cpu_demand_stat* local_stat, int event, unsigned long long len);
 263 void adjust_local_weight(void);
 264
 265 #define get_task_lrq_stat(p) (&(p)->cpu_class->stat.local_stats[task_cpu(p)])
 266 #define get_cls_local_stat(cls,cpu) (&(cls)->stat.local_stats[cpu])
 267 #define get_rq_local_stat(lrq,cpu) (get_cls_local_stat((lrq)->cpu_class,cpu))
 268
 269 /********************************************************************
 270  * Parameters that determine how quickly CVT's progress and how
 271  * priority can impact a LRQ's runqueue position. See also
 272  * get_effective_prio(). These parameters need to adjusted
 273  * in accordance to the following example and understanding.
 274  *
 275  * CLASS_QUANTIZER:
 276  *
 277  * A class with 5% share, can execute 50M nsecs / per sec ~ 2^28.
 278  * It's share will be set to 512 = 2^9. The globl CLASSQUEUE_SIZE is set to 2^7.
 279  * With CLASS_QUANTIZER=16, the local_cvt of this class will increase
 280  * by 2^28/2^9 = 2^19 = 512K.
 281  * Setting CLASS_QUANTIZER to 16, 2^(19-16) = 8 slots / per second.
 282  * A class with 5% shares, will cover 80 slots / per second.
 283  *
 284  * PRIORITY_QUANTIZER:
 285  *
 286  * How much can top priorities of class impact slot bonus.
 287  * There are 40 nice priorities. "2" will allow upto 10 slots improvement
 288  * in the RQ thus for 50% class it can perform ~1sec starvation.
 289  *
 290  *******************************************************************/
 291
 292 #define CLASS_QUANTIZER 16      //shift from ns to increase class bonus
 293 #define PRIORITY_QUANTIZER 2    //controls how much a high prio task can borrow
 294
 295 #define CKRM_SHARE_ACCURACY 10
 296 #define NSEC_PER_MS 1000000
 297 #define NSEC_PER_JIFFIES (NSEC_PER_SEC/HZ)
 298
 299
 300 #define MAX_SAVINGS_ABSOLUTE (10LLU*NSEC_PER_SEC)  // 10 seconds
 301
 302 #define CVT_UPDATE_TICK     ((HZ/2)?:1)
 303
 304 // ABSOLUTE_CKRM_TUNING determines whether classes can make up
 305 // lost time in absolute time or in relative values
 306
 307 #define ABSOLUTE_CKRM_TUNING         // preferred due to more predictable behavior
 308
 309 #ifdef ABSOLUTE_CKRM_TUNING
 310
 311 #define MAX_SAVINGS        MAX_SAVINGS_ABSOLUTE
 312 //an absolute bonus of 200ms for classes when reactivated
 313 #define INTERACTIVE_BONUS(lrq) ((200*NSEC_PER_MS)/local_class_weight(lrq))
 314 #define SAVINGS_LEAK_SPEED (CVT_UPDATE_TICK/10*NSEC_PER_JIFFIES)
 315
 316 #define scale_cvt(val,lrq)   ((val)*local_class_weight(lrq))
 317 #define unscale_cvt(val,lrq) (do_div(val,local_class_weight(lrq)))
 318
 319 #else
 320
 321 #define MAX_SAVINGS (MAX_SAVINGS_ABSOLUTE >> CKRM_SHARE_ACCURACY)
 322 /*
 323  * to improve system responsiveness
 324  * an inactive class is put a little bit ahead of the current class when it wakes up
 325  * the amount is set in normalized termis to simplify the calculation
 326  * for class with 100% share, it can be 2s ahead
 327  * while for class with 10% share, it can be 200ms ahead
 328  */
 329 #define INTERACTIVE_BONUS(lrq) (2*NSEC_PER_MS)
 330
 331 /*
 332  * normalized savings can't be more than MAX_NORMALIZED_SAVINGS
 333  * based on the current configuration
 334  * this means that a class with share 100% will accumulate 10s at most
 335  * while a class with 1% of the share can only accumulate 100ms
 336  */
 337
 338 //a class with share 100% can get 100ms every 500ms
 339 //while a class with share 10% can only get 10ms every 500ms
 340 #define SAVINGS_LEAK_SPEED ((CVT_UPDATE_TICK/5*NSEC_PER_JIFFIES) >> CKRM_SHARE_ACCURACY)
 341
 342 #define scale_cvt(val,lrq)   (val)
 343 #define unscale_cvt(val,lrq) (val)
 344
 345 #endif
 346
 347
 348 /**
 349  * get_effective_prio: return the effective priority of a class local queue
 350  *
 351  * class priority = progress * a + urgency * b
 352  * progress = queue cvt
 353  * urgency = queue top priority
 354  * a and b are scaling factors
 355  * currently, prio increases by 1 if either: top_priority increase by one
 356  *                                   or, local_cvt increases by 4ms
 357  */
 358 static inline int get_effective_prio(ckrm_lrq_t * lrq)
 359 {
 360         int prio;
 361
 362         prio = lrq->local_cvt >> CLASS_QUANTIZER;  // cumulative usage
 363         prio += lrq->top_priority >> PRIORITY_QUANTIZER; // queue urgency
 364
 365         return prio;
 366 }
 367
 368 CVT_t get_local_cur_cvt(int cpu);
 369
 370 /**
 371  * update_class_priority:
 372  *
 373  * called whenever cvt or top_priority changes
 374  *
 375  * internal: (calling structure)
 376  * update_class_priority
 377  *   -- set_top_priority
 378  *      -- class_enqueue_task
 379  *      -- class_dequeue_task
 380  *      -- rq_get_next_task (queue switch)
 381  *   -- update_local_cvt
 382  *      -- schedule
 383  */
 384 static inline void update_class_priority(ckrm_lrq_t *local_rq)
 385 {
 386         int effective_prio = get_effective_prio(local_rq);
 387         classqueue_update_prio(local_rq->classqueue,
 388                                &local_rq->classqueue_linkobj,
 389                                effective_prio);
 390 }
 391
 392 /*
 393  *  set the new top priority and reposition the queue
 394  *  called when: task enqueue/dequeue and queue switch
 395  */
 396 static inline void set_top_priority(ckrm_lrq_t *lrq,
 397                                     int new_priority)
 398 {
 399         lrq->top_priority = new_priority;
 400         update_class_priority(lrq);
 401 }
 402
 403 /*
 404  * task_load: how much load this task counts
 405  */
 406 static inline unsigned long task_load(struct task_struct* p)
 407 {
 408         return (task_timeslice(p) * p->demand_stat.cpu_demand);
 409 }
 410
 411 /*
 412  * runqueue load is the local_weight of all the classes on this cpu
 413  * must be called with class_list_lock held
 414  */
 415 static inline unsigned long ckrm_cpu_load(int cpu)
 416 {
 417         struct ckrm_cpu_class *clsptr;
 418         ckrm_lrq_t* lrq;
 419         struct ckrm_cpu_demand_stat* l_stat;
 420         int total_load = 0;
 421         int load;
 422
 423         list_for_each_entry(clsptr,&active_cpu_classes,links) {
 424                 lrq =  get_ckrm_lrq(clsptr,cpu);
 425                 l_stat = get_cls_local_stat(clsptr,cpu);
 426                 load = lrq->local_weight;
 427                 if (l_stat->cpu_demand < load)
 428                         load = l_stat->cpu_demand;
 429                 total_load += load;
 430         }
 431         return total_load;
 432 }
 433
 434 static inline void class_enqueue_task(struct task_struct *p,
 435                                       prio_array_t * array)
 436 {
 437         ckrm_lrq_t *lrq;
 438         int effective_prio;
 439
 440         lrq = get_task_lrq(p);
 441
 442         cpu_demand_event(&p->demand_stat,CPU_DEMAND_ENQUEUE,0);
 443         lrq->lrq_load += task_load(p);
 444
 445         if ((p->prio < lrq->top_priority) && (array == lrq->active))
 446                 set_top_priority(lrq, p->prio);
 447
 448         if (! cls_in_classqueue(&lrq->classqueue_linkobj)) {
 449                 cpu_demand_event(get_task_lrq_stat(p),CPU_DEMAND_ENQUEUE,0);
 450                 effective_prio = get_effective_prio(lrq);
 451                 classqueue_enqueue(lrq->classqueue, &lrq->classqueue_linkobj, effective_prio);
 452         }
 453
 454 }
 455
 456 static inline void class_dequeue_task(struct task_struct *p,
 457                                       prio_array_t * array)
 458 {
 459         ckrm_lrq_t *lrq = get_task_lrq(p);
 460         unsigned long load = task_load(p);
 461
 462         BUG_ON(lrq->lrq_load < load);
 463         lrq->lrq_load -= load;
 464
 465         cpu_demand_event(&p->demand_stat,CPU_DEMAND_DEQUEUE,0);
 466
 467         if ((array == lrq->active) && (p->prio == lrq->top_priority)
 468             && list_empty(&(array->queue[p->prio])))
 469                 set_top_priority(lrq,
 470                                  find_next_bit(array->bitmap, MAX_PRIO,
 471                                                p->prio));
 472 }
 473
 474 /*
 475  *  called after a task is switched out. Update the local cvt accounting
 476  *  we need to stick with long instead of long long due to nonexistent 64-bit division
 477  */
 478 static inline void update_local_cvt(struct task_struct *p, unsigned long nsec)
 479 {
 480         ckrm_lrq_t * lrq = get_task_lrq(p);
 481
 482         unsigned long cvt_inc = nsec / local_class_weight(lrq);
 483
 484         lrq->local_cvt += cvt_inc;
 485         lrq->uncounted_ns += nsec;
 486
 487         update_class_priority(lrq);
 488 }
 489
 490 static inline int class_preempts_curr(struct task_struct * p, struct task_struct* curr)
 491 {
 492         struct cq_node_struct* node1 = &(get_task_lrq(p)->classqueue_linkobj);
 493         struct cq_node_struct* node2 = &(get_task_lrq(curr)->classqueue_linkobj);
 494
 495         return (class_compare_prio(node1,node2) < 0);
 496 }
 497
 498 /*
 499  * return a random value with range [0, (val-1)]
 500  */
 501 static inline int get_ckrm_rand(unsigned long val)
 502 {
 503         int rand;
 504         static int last_rand[NR_CPUS];
 505         int cpu = smp_processor_id();
 506
 507         rand = last_rand[cpu];
 508         rand ++;
 509         if (rand >= val)
 510                 rand = 0;
 511
 512         last_rand[cpu] = rand;
 513         return rand;
 514 }
 515
 516 void update_class_cputime(int this_cpu);
 517
 518 /**********************************************/
 519 /*          PID_LOAD_BALANCING                */
 520 /**********************************************/
 521 struct ckrm_load_struct {
 522         unsigned long load_p;   /*propotional*/
 523         unsigned long load_i;   /*integral   */
 524         long load_d;   /*derivative */
 525 };
 526
 527 typedef struct ckrm_load_struct ckrm_load_t;
 528
 529 static inline void ckrm_load_init(ckrm_load_t* ckrm_load) {
 530         ckrm_load->load_p = 0;
 531         ckrm_load->load_i = 0;
 532         ckrm_load->load_d = 0;
 533 }
 534
 535 void ckrm_load_sample(ckrm_load_t* ckrm_load,int cpu);
 536 long pid_get_pressure(ckrm_load_t* ckrm_load, int local_group);
 537 #define rq_ckrm_load(rq) (&((rq)->ckrm_load))
 538
 539 static inline void ckrm_sched_tick(unsigned long j,int this_cpu,struct ckrm_load_struct* ckrm_load)
 540 {
 541         read_lock(&class_list_lock);
 542
 543 #ifdef CONFIG_SMP
 544         ckrm_load_sample(ckrm_load,this_cpu);
 545 #endif
 546
 547         if (! (j % CVT_UPDATE_TICK)) {
 548                 //              printk("ckrm_sched j=%lu\n",j);
 549                 classqueue_update_base(get_cpu_classqueue(this_cpu));
 550                 update_class_cputime(this_cpu);
 551         }
 552
 553         read_unlock(&class_list_lock);
 554 }
 555
 556 #endif //CONFIG_CKRM_CPU_SCHEDULE
 557
 558 #endif