Subversion Repositories HelenOS

Rev

Rev 1229 | Rev 1264 | Go to most recent revision | Blame | Compare with Previous | Last modification | View Log | Download | RSS feed

  1. /*
  2.  * Copyright (C) 2001-2004 Jakub Jermar
  3.  * All rights reserved.
  4.  *
  5.  * Redistribution and use in source and binary forms, with or without
  6.  * modification, are permitted provided that the following conditions
  7.  * are met:
  8.  *
  9.  * - Redistributions of source code must retain the above copyright
  10.  *   notice, this list of conditions and the following disclaimer.
  11.  * - Redistributions in binary form must reproduce the above copyright
  12.  *   notice, this list of conditions and the following disclaimer in the
  13.  *   documentation and/or other materials provided with the distribution.
  14.  * - The name of the author may not be used to endorse or promote products
  15.  *   derived from this software without specific prior written permission.
  16.  *
  17.  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  18.  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  19.  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  20.  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  21.  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  22.  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  23.  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  24.  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  25.  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  26.  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  27.  */
  28.  
  29. /**
  30.  * @file    scheduler.c
  31.  * @brief   Scheduler and load balancing.
  32.  *
  33.  * This file contains the scheduler and kcpulb kernel thread wich
  34.  * performs load-balancing of per-CPU run queues.
  35.  */
  36.  
  37. #include <proc/scheduler.h>
  38. #include <proc/thread.h>
  39. #include <proc/task.h>
  40. #include <mm/frame.h>
  41. #include <mm/page.h>
  42. #include <mm/as.h>
  43. #include <arch/asm.h>
  44. #include <arch/faddr.h>
  45. #include <atomic.h>
  46. #include <synch/spinlock.h>
  47. #include <config.h>
  48. #include <context.h>
  49. #include <func.h>
  50. #include <arch.h>
  51. #include <adt/list.h>
  52. #include <panic.h>
  53. #include <typedefs.h>
  54. #include <cpu.h>
  55. #include <print.h>
  56. #include <debug.h>
  57.  
  58. static void before_task_runs(void);
  59. static void before_thread_runs(void);
  60. static void after_thread_ran(void);
  61. static void scheduler_separated_stack(void);
  62.  
  63. atomic_t nrdy;  /**< Number of ready threads in the system. */
  64.  
  65. /** Carry out actions before new task runs. */
  66. void before_task_runs(void)
  67. {
  68.     before_task_runs_arch();
  69. }
  70.  
  71. /** Take actions before new thread runs.
  72.  *
  73.  * Perform actions that need to be
  74.  * taken before the newly selected
  75.  * tread is passed control.
  76.  *
  77.  * THREAD->lock is locked on entry
  78.  *
  79.  */
  80. void before_thread_runs(void)
  81. {
  82.     before_thread_runs_arch();
  83. #ifdef CONFIG_FPU_LAZY
  84.     if(THREAD==CPU->fpu_owner)
  85.         fpu_enable();
  86.     else
  87.         fpu_disable();
  88. #else
  89.     fpu_enable();
  90.     if (THREAD->fpu_context_exists)
  91.         fpu_context_restore(THREAD->saved_fpu_context);
  92.     else {
  93.         fpu_init();
  94.         THREAD->fpu_context_exists=1;
  95.     }
  96. #endif
  97. }
  98.  
  99. /** Take actions after THREAD had run.
  100.  *
  101.  * Perform actions that need to be
  102.  * taken after the running thread
  103.  * had been preempted by the scheduler.
  104.  *
  105.  * THREAD->lock is locked on entry
  106.  *
  107.  */
  108. void after_thread_ran(void)
  109. {
  110.     after_thread_ran_arch();
  111. }
  112.  
  113. #ifdef CONFIG_FPU_LAZY
  114. void scheduler_fpu_lazy_request(void)
  115. {
  116. restart:
  117.     fpu_enable();
  118.     spinlock_lock(&CPU->lock);
  119.  
  120.     /* Save old context */
  121.     if (CPU->fpu_owner != NULL) {  
  122.         spinlock_lock(&CPU->fpu_owner->lock);
  123.         fpu_context_save(CPU->fpu_owner->saved_fpu_context);
  124.         /* don't prevent migration */
  125.         CPU->fpu_owner->fpu_context_engaged=0;
  126.         spinlock_unlock(&CPU->fpu_owner->lock);
  127.         CPU->fpu_owner = NULL;
  128.     }
  129.  
  130.     spinlock_lock(&THREAD->lock);
  131.     if (THREAD->fpu_context_exists) {
  132.         fpu_context_restore(THREAD->saved_fpu_context);
  133.     } else {
  134.         /* Allocate FPU context */
  135.         if (!THREAD->saved_fpu_context) {
  136.             /* Might sleep */
  137.             spinlock_unlock(&THREAD->lock);
  138.             spinlock_unlock(&CPU->lock);
  139.             THREAD->saved_fpu_context = slab_alloc(fpu_context_slab,
  140.                                    0);
  141.             /* We may have switched CPUs during slab_alloc */
  142.             goto restart;
  143.         }
  144.         fpu_init();
  145.         THREAD->fpu_context_exists=1;
  146.     }
  147.     CPU->fpu_owner=THREAD;
  148.     THREAD->fpu_context_engaged = 1;
  149.     spinlock_unlock(&THREAD->lock);
  150.  
  151.     spinlock_unlock(&CPU->lock);
  152. }
  153. #endif
  154.  
  155. /** Initialize scheduler
  156.  *
  157.  * Initialize kernel scheduler.
  158.  *
  159.  */
  160. void scheduler_init(void)
  161. {
  162. }
  163.  
  164. /** Get thread to be scheduled
  165.  *
  166.  * Get the optimal thread to be scheduled
  167.  * according to thread accounting and scheduler
  168.  * policy.
  169.  *
  170.  * @return Thread to be scheduled.
  171.  *
  172.  */
  173. static thread_t *find_best_thread(void)
  174. {
  175.     thread_t *t;
  176.     runq_t *r;
  177.     int i;
  178.  
  179.     ASSERT(CPU != NULL);
  180.  
  181. loop:
  182.     interrupts_enable();
  183.    
  184.     if (atomic_get(&CPU->nrdy) == 0) {
  185.         /*
  186.          * For there was nothing to run, the CPU goes to sleep
  187.          * until a hardware interrupt or an IPI comes.
  188.          * This improves energy saving and hyperthreading.
  189.          */
  190.  
  191.         /*
  192.          * An interrupt might occur right now and wake up a thread.
  193.          * In such case, the CPU will continue to go to sleep
  194.          * even though there is a runnable thread.
  195.          */
  196.  
  197.          cpu_sleep();
  198.          goto loop;
  199.     }
  200.  
  201.     interrupts_disable();
  202.    
  203.     for (i = 0; i<RQ_COUNT; i++) {
  204.         r = &CPU->rq[i];
  205.         spinlock_lock(&r->lock);
  206.         if (r->n == 0) {
  207.             /*
  208.              * If this queue is empty, try a lower-priority queue.
  209.              */
  210.             spinlock_unlock(&r->lock);
  211.             continue;
  212.         }
  213.  
  214.         atomic_dec(&CPU->nrdy);
  215.         atomic_dec(&nrdy);
  216.         r->n--;
  217.  
  218.         /*
  219.          * Take the first thread from the queue.
  220.          */
  221.         t = list_get_instance(r->rq_head.next, thread_t, rq_link);
  222.         list_remove(&t->rq_link);
  223.  
  224.         spinlock_unlock(&r->lock);
  225.  
  226.         spinlock_lock(&t->lock);
  227.         t->cpu = CPU;
  228.  
  229.         t->ticks = us2ticks((i+1)*10000);
  230.         t->priority = i;    /* correct rq index */
  231.  
  232.         /*
  233.          * Clear the X_STOLEN flag so that t can be migrated when load balancing needs emerge.
  234.          */
  235.         t->flags &= ~X_STOLEN;
  236.         spinlock_unlock(&t->lock);
  237.  
  238.         return t;
  239.     }
  240.     goto loop;
  241.  
  242. }
  243.  
  244. /** Prevent rq starvation
  245.  *
  246.  * Prevent low priority threads from starving in rq's.
  247.  *
  248.  * When the function decides to relink rq's, it reconnects
  249.  * respective pointers so that in result threads with 'pri'
  250.  * greater or equal @start are moved to a higher-priority queue.
  251.  *
  252.  * @param start Threshold priority.
  253.  *
  254.  */
  255. static void relink_rq(int start)
  256. {
  257.     link_t head;
  258.     runq_t *r;
  259.     int i, n;
  260.  
  261.     list_initialize(&head);
  262.     spinlock_lock(&CPU->lock);
  263.     if (CPU->needs_relink > NEEDS_RELINK_MAX) {
  264.         for (i = start; i<RQ_COUNT-1; i++) {
  265.             /* remember and empty rq[i + 1] */
  266.             r = &CPU->rq[i + 1];
  267.             spinlock_lock(&r->lock);
  268.             list_concat(&head, &r->rq_head);
  269.             n = r->n;
  270.             r->n = 0;
  271.             spinlock_unlock(&r->lock);
  272.        
  273.             /* append rq[i + 1] to rq[i] */
  274.             r = &CPU->rq[i];
  275.             spinlock_lock(&r->lock);
  276.             list_concat(&r->rq_head, &head);
  277.             r->n += n;
  278.             spinlock_unlock(&r->lock);
  279.         }
  280.         CPU->needs_relink = 0;
  281.     }
  282.     spinlock_unlock(&CPU->lock);
  283.  
  284. }
  285.  
  286. /** The scheduler
  287.  *
  288.  * The thread scheduling procedure.
  289.  * Passes control directly to
  290.  * scheduler_separated_stack().
  291.  *
  292.  */
  293. void scheduler(void)
  294. {
  295.     volatile ipl_t ipl;
  296.  
  297.     ASSERT(CPU != NULL);
  298.  
  299.     ipl = interrupts_disable();
  300.  
  301.     if (atomic_get(&haltstate))
  302.         halt();
  303.    
  304.     if (THREAD) {
  305.         spinlock_lock(&THREAD->lock);
  306. #ifndef CONFIG_FPU_LAZY
  307.         fpu_context_save(THREAD->saved_fpu_context);
  308. #endif
  309.         if (!context_save(&THREAD->saved_context)) {
  310.             /*
  311.              * This is the place where threads leave scheduler();
  312.              */
  313.             spinlock_unlock(&THREAD->lock);
  314.             interrupts_restore(THREAD->saved_context.ipl);
  315.            
  316.             return;
  317.         }
  318.  
  319.         /*
  320.          * Interrupt priority level of preempted thread is recorded here
  321.          * to facilitate scheduler() invocations from interrupts_disable()'d
  322.          * code (e.g. waitq_sleep_timeout()).
  323.          */
  324.         THREAD->saved_context.ipl = ipl;
  325.     }
  326.  
  327.     /*
  328.      * Through the 'THE' structure, we keep track of THREAD, TASK, CPU, VM
  329.      * and preemption counter. At this point THE could be coming either
  330.      * from THREAD's or CPU's stack.
  331.      */
  332.     the_copy(THE, (the_t *) CPU->stack);
  333.  
  334.     /*
  335.      * We may not keep the old stack.
  336.      * Reason: If we kept the old stack and got blocked, for instance, in
  337.      * find_best_thread(), the old thread could get rescheduled by another
  338.      * CPU and overwrite the part of its own stack that was also used by
  339.      * the scheduler on this CPU.
  340.      *
  341.      * Moreover, we have to bypass the compiler-generated POP sequence
  342.      * which is fooled by SP being set to the very top of the stack.
  343.      * Therefore the scheduler() function continues in
  344.      * scheduler_separated_stack().
  345.      */
  346.     context_save(&CPU->saved_context);
  347.     context_set(&CPU->saved_context, FADDR(scheduler_separated_stack), (__address) CPU->stack, CPU_STACK_SIZE);
  348.     context_restore(&CPU->saved_context);
  349.     /* not reached */
  350. }
  351.  
  352. /** Scheduler stack switch wrapper
  353.  *
  354.  * Second part of the scheduler() function
  355.  * using new stack. Handling the actual context
  356.  * switch to a new thread.
  357.  *
  358.  * Assume THREAD->lock is held.
  359.  */
  360. void scheduler_separated_stack(void)
  361. {
  362.     int priority;
  363.    
  364.     ASSERT(CPU != NULL);
  365.    
  366.     if (THREAD) {
  367.         /* must be run after the switch to scheduler stack */
  368.         after_thread_ran();
  369.  
  370.         switch (THREAD->state) {
  371.             case Running:
  372.             spinlock_unlock(&THREAD->lock);
  373.             thread_ready(THREAD);
  374.             break;
  375.  
  376.             case Exiting:
  377.             thread_destroy(THREAD);
  378.             break;
  379.            
  380.             case Sleeping:
  381.             /*
  382.              * Prefer the thread after it's woken up.
  383.              */
  384.             THREAD->priority = -1;
  385.  
  386.             /*
  387.              * We need to release wq->lock which we locked in waitq_sleep().
  388.              * Address of wq->lock is kept in THREAD->sleep_queue.
  389.              */
  390.             spinlock_unlock(&THREAD->sleep_queue->lock);
  391.  
  392.             /*
  393.              * Check for possible requests for out-of-context invocation.
  394.              */
  395.             if (THREAD->call_me) {
  396.                 THREAD->call_me(THREAD->call_me_with);
  397.                 THREAD->call_me = NULL;
  398.                 THREAD->call_me_with = NULL;
  399.             }
  400.  
  401.             spinlock_unlock(&THREAD->lock);
  402.  
  403.             break;
  404.  
  405.             default:
  406.             /*
  407.              * Entering state is unexpected.
  408.              */
  409.             panic("tid%d: unexpected state %s\n", THREAD->tid, thread_states[THREAD->state]);
  410.             break;
  411.         }
  412.  
  413.         THREAD = NULL;
  414.     }
  415.  
  416.     THREAD = find_best_thread();
  417.    
  418.     spinlock_lock(&THREAD->lock);
  419.     priority = THREAD->priority;
  420.     spinlock_unlock(&THREAD->lock);
  421.  
  422.     relink_rq(priority);       
  423.  
  424.     spinlock_lock(&THREAD->lock);  
  425.  
  426.     /*
  427.      * If both the old and the new task are the same, lots of work is avoided.
  428.      */
  429.     if (TASK != THREAD->task) {
  430.         as_t *as1 = NULL;
  431.         as_t *as2;
  432.  
  433.         if (TASK) {
  434.             spinlock_lock(&TASK->lock);
  435.             as1 = TASK->as;
  436.             spinlock_unlock(&TASK->lock);
  437.         }
  438.  
  439.         spinlock_lock(&THREAD->task->lock);
  440.         as2 = THREAD->task->as;
  441.         spinlock_unlock(&THREAD->task->lock);
  442.        
  443.         /*
  444.          * Note that it is possible for two tasks to share one address space.
  445.          */
  446.         if (as1 != as2) {
  447.             /*
  448.              * Both tasks and address spaces are different.
  449.              * Replace the old one with the new one.
  450.              */
  451.             as_switch(as1, as2);
  452.         }
  453.         TASK = THREAD->task;
  454.         before_task_runs();
  455.     }
  456.  
  457.     THREAD->state = Running;
  458.  
  459. #ifdef SCHEDULER_VERBOSE
  460.     printf("cpu%d: tid %d (priority=%d,ticks=%lld,nrdy=%ld)\n", CPU->id, THREAD->tid, THREAD->priority, THREAD->ticks, atomic_get(&CPU->nrdy));
  461. #endif 
  462.  
  463.     /*
  464.      * Some architectures provide late kernel PA2KA(identity)
  465.      * mapping in a page fault handler. However, the page fault
  466.      * handler uses the kernel stack of the running thread and
  467.      * therefore cannot be used to map it. The kernel stack, if
  468.      * necessary, is to be mapped in before_thread_runs(). This
  469.      * function must be executed before the switch to the new stack.
  470.      */
  471.     before_thread_runs();
  472.  
  473.     /*
  474.      * Copy the knowledge of CPU, TASK, THREAD and preemption counter to thread's stack.
  475.      */
  476.     the_copy(THE, (the_t *) THREAD->kstack);
  477.    
  478.     context_restore(&THREAD->saved_context);
  479.     /* not reached */
  480. }
  481.  
  482. #ifdef CONFIG_SMP
  483. /** Load balancing thread
  484.  *
  485.  * SMP load balancing thread, supervising thread supplies
  486.  * for the CPU it's wired to.
  487.  *
  488.  * @param arg Generic thread argument (unused).
  489.  *
  490.  */
  491. void kcpulb(void *arg)
  492. {
  493.     thread_t *t;
  494.     int count, average, i, j, k = 0;
  495.     ipl_t ipl;
  496.  
  497. loop:
  498.     /*
  499.      * Work in 1s intervals.
  500.      */
  501.     thread_sleep(1);
  502.  
  503. not_satisfied:
  504.     /*
  505.      * Calculate the number of threads that will be migrated/stolen from
  506.      * other CPU's. Note that situation can have changed between two
  507.      * passes. Each time get the most up to date counts.
  508.      */
  509.     average = atomic_get(&nrdy) / config.cpu_active + 1;
  510.     count = average - atomic_get(&CPU->nrdy);
  511.  
  512.     if (count <= 0)
  513.         goto satisfied;
  514.  
  515.     /*
  516.      * Searching least priority queues on all CPU's first and most priority queues on all CPU's last.
  517.      */
  518.     for (j=RQ_COUNT-1; j >= 0; j--) {
  519.         for (i=0; i < config.cpu_active; i++) {
  520.             link_t *l;
  521.             runq_t *r;
  522.             cpu_t *cpu;
  523.  
  524.             cpu = &cpus[(i + k) % config.cpu_active];
  525.  
  526.             /*
  527.              * Not interested in ourselves.
  528.              * Doesn't require interrupt disabling for kcpulb is X_WIRED.
  529.              */
  530.             if (CPU == cpu)
  531.                 continue;
  532.             if (atomic_get(&cpu->nrdy) <= average)
  533.                 continue;
  534.  
  535.             ipl = interrupts_disable();
  536.             r = &cpu->rq[j];
  537.             spinlock_lock(&r->lock);
  538.             if (r->n == 0) {
  539.                 spinlock_unlock(&r->lock);
  540.                 interrupts_restore(ipl);
  541.                 continue;
  542.             }
  543.        
  544.             t = NULL;
  545.             l = r->rq_head.prev;    /* search rq from the back */
  546.             while (l != &r->rq_head) {
  547.                 t = list_get_instance(l, thread_t, rq_link);
  548.                 /*
  549.                  * We don't want to steal CPU-wired threads neither threads already stolen.
  550.                  * The latter prevents threads from migrating between CPU's without ever being run.
  551.                  * We don't want to steal threads whose FPU context is still in CPU.
  552.                  */
  553.                 spinlock_lock(&t->lock);
  554.                 if ( (!(t->flags & (X_WIRED | X_STOLEN))) && (!(t->fpu_context_engaged)) ) {
  555.                     /*
  556.                      * Remove t from r.
  557.                      */
  558.                     spinlock_unlock(&t->lock);
  559.                    
  560.                     atomic_dec(&cpu->nrdy);
  561.                     atomic_dec(&nrdy);
  562.  
  563.                     r->n--;
  564.                     list_remove(&t->rq_link);
  565.  
  566.                     break;
  567.                 }
  568.                 spinlock_unlock(&t->lock);
  569.                 l = l->prev;
  570.                 t = NULL;
  571.             }
  572.             spinlock_unlock(&r->lock);
  573.  
  574.             if (t) {
  575.                 /*
  576.                  * Ready t on local CPU
  577.                  */
  578.                 spinlock_lock(&t->lock);
  579. #ifdef KCPULB_VERBOSE
  580.                 printf("kcpulb%d: TID %d -> cpu%d, nrdy=%ld, avg=%nd\n", CPU->id, t->tid, CPU->id, atomic_get(&CPU->nrdy), atomic_get(&nrdy) / config.cpu_active);
  581. #endif
  582.                 t->flags |= X_STOLEN;
  583.                 t->state = Entering;
  584.                 spinlock_unlock(&t->lock);
  585.    
  586.                 thread_ready(t);
  587.  
  588.                 interrupts_restore(ipl);
  589.    
  590.                 if (--count == 0)
  591.                     goto satisfied;
  592.                    
  593.                 /*
  594.                  * We are not satisfied yet, focus on another CPU next time.
  595.                  */
  596.                 k++;
  597.                
  598.                 continue;
  599.             }
  600.             interrupts_restore(ipl);
  601.         }
  602.     }
  603.  
  604.     if (atomic_get(&CPU->nrdy)) {
  605.         /*
  606.          * Be a little bit light-weight and let migrated threads run.
  607.          */
  608.         scheduler();
  609.     } else {
  610.         /*
  611.          * We failed to migrate a single thread.
  612.          * Give up this turn.
  613.          */
  614.         goto loop;
  615.     }
  616.        
  617.     goto not_satisfied;
  618.  
  619. satisfied:
  620.     goto loop;
  621. }
  622.  
  623. #endif /* CONFIG_SMP */
  624.  
  625.  
  626. /** Print information about threads & scheduler queues */
  627. void sched_print_list(void)
  628. {
  629.     ipl_t ipl;
  630.     int cpu,i;
  631.     runq_t *r;
  632.     thread_t *t;
  633.     link_t *cur;
  634.  
  635.     /* We are going to mess with scheduler structures,
  636.      * let's not be interrupted */
  637.     ipl = interrupts_disable();
  638.     for (cpu=0;cpu < config.cpu_count; cpu++) {
  639.  
  640.         if (!cpus[cpu].active)
  641.             continue;
  642.  
  643.         spinlock_lock(&cpus[cpu].lock);
  644.         printf("cpu%d: address=%p, nrdy=%ld, needs_relink=%ld\n",
  645.                cpus[cpu].id, &cpus[cpu], atomic_get(&cpus[cpu].nrdy), cpus[cpu].needs_relink);
  646.        
  647.         for (i=0; i<RQ_COUNT; i++) {
  648.             r = &cpus[cpu].rq[i];
  649.             spinlock_lock(&r->lock);
  650.             if (!r->n) {
  651.                 spinlock_unlock(&r->lock);
  652.                 continue;
  653.             }
  654.             printf("\trq[%d]: ", i);
  655.             for (cur=r->rq_head.next; cur!=&r->rq_head; cur=cur->next) {
  656.                 t = list_get_instance(cur, thread_t, rq_link);
  657.                 printf("%d(%s) ", t->tid,
  658.                        thread_states[t->state]);
  659.             }
  660.             printf("\n");
  661.             spinlock_unlock(&r->lock);
  662.         }
  663.         spinlock_unlock(&cpus[cpu].lock);
  664.     }
  665.    
  666.     interrupts_restore(ipl);
  667. }
  668.