/** * litmus/sched_mc_ce.c * * The Cyclic Executive (CE) scheduler used by the mixed criticality scheduling * algorithm. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static struct sched_plugin mc_ce_plugin __cacheline_aligned_in_smp; #define is_active_plugin() (litmus == &mc_ce_plugin) #define get_ce_data(dom_data_ref) (dom_data_ref->domain.data) static atomic_t start_time_set = ATOMIC_INIT(-1); static atomic64_t start_time = ATOMIC64_INIT(0); static struct proc_dir_entry *mc_ce_dir = NULL, *ce_file = NULL; DEFINE_PER_CPU(domain_data_t, mc_ce_doms); DEFINE_PER_CPU(rt_domain_t, mc_ce_rts); DEFINE_PER_CPU(struct ce_dom_data, _mc_ce_dom_data); /* Return the address of the domain_t for this CPU, used by the * mixed-criticality plugin. */ domain_data_t *ce_domain_for(int cpu) { return &per_cpu(mc_ce_doms, cpu); } /* * Get the offset into the cycle taking the start time into account. */ static inline lt_t get_cycle_offset(const lt_t when, const lt_t cycle_time) { long long st = atomic64_read(&start_time); lt_t offset = (when - st) % cycle_time; TRACE("when: %llu cycle_time: %llu start_time: %lld offset %llu\n", when, cycle_time, st, offset); return offset; } /* * The user land job completion call will set the RT_F_SLEEP flag and then * call schedule. This function is used when schedule sleeps a task. * * Do not call prepare_for_next_period on Level-A tasks! */ static void mc_ce_job_completion(struct task_struct *ts) { const domain_data_t *dom_data = &per_cpu(mc_ce_doms, smp_processor_id()); const struct ce_dom_data *ce_data = get_ce_data(dom_data); const int idx = tsk_mc_data(ts)->mc_task.lvl_a_id; const struct ce_dom_pid_entry *pid_entry = &ce_data->pid_entries[idx]; int just_finished; TRACE_TASK(ts, "completed\n"); sched_trace_task_completion(ts, 0); /* post-increment is important here */ just_finished = (tsk_rt(ts)->job_params.job_no)++; /* Job completes in expected window: everything is normal. * Job completes in an earlier window: BUG(), that's wrong. * Job completes in a later window: The job is behind. */ if (just_finished < pid_entry->expected_job) { /* this job is already released because it's running behind */ set_rt_flags(ts, RT_F_RUNNING); TRACE_TASK(ts, "appears behind: the expected job is %d but " "job %d just completed\n", pid_entry->expected_job, just_finished); } else if (pid_entry->expected_job < just_finished) { printk(KERN_CRIT "job %d completed in expected job %d which " "seems too early\n", just_finished, pid_entry->expected_job); BUG(); } } /* * Return the index into the PID entries table of what to schedule next. * Don't call if the table is empty. Assumes the caller has the domain lock. * The offset parameter is the offset into the cycle. * * TODO Currently O(n) in the number of tasks on the CPU. Binary search? */ static int mc_ce_schedule_at(const domain_t *dom, lt_t offset) { const struct ce_dom_data *ce_data = dom->data; const struct ce_dom_pid_entry *pid_entry = NULL; int i; BUG_ON(ce_data->cycle_time < 1); BUG_ON(ce_data->num_pid_entries < 1); for (i = 0; i < ce_data->num_pid_entries; ++i) { pid_entry = &ce_data->pid_entries[i]; if (offset < pid_entry->acc_time) { /* found task to schedule in this window */ break; } } /* can only happen if cycle_time is not right */ BUG_ON(pid_entry->acc_time > ce_data->cycle_time); TRACE("schedule at returned task %d for CPU %d\n", i, ce_data->cpu); return i; } static struct task_struct *mc_ce_schedule(struct task_struct *prev) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, smp_processor_id()); domain_t *dom = &dom_data->domain; struct ce_dom_data *ce_data = get_ce_data(dom_data); struct task_struct *next = NULL; int exists, sleep, should_sched_exists, should_sched_blocked, should_sched_asleep; raw_spin_lock(dom->lock); /* sanity checking */ BUG_ON(ce_data->scheduled && ce_data->scheduled != prev); BUG_ON(ce_data->scheduled && !is_realtime(prev)); BUG_ON(is_realtime(prev) && !ce_data->scheduled); exists = NULL != ce_data->scheduled; sleep = exists && RT_F_SLEEP == get_rt_flags(ce_data->scheduled); TRACE("exists: %d, sleep: %d\n", exists, sleep); if (sleep) mc_ce_job_completion(ce_data->scheduled); /* these checks must go after the call to mc_ce_job_completion in case * a late task needs to be scheduled again right away and its the only * task on a core */ should_sched_exists = NULL != ce_data->should_schedule; should_sched_blocked = should_sched_exists && !is_running(ce_data->should_schedule); should_sched_asleep = should_sched_exists && RT_F_SLEEP == get_rt_flags(ce_data->should_schedule); TRACE("should_sched_exists: %d, should_sched_blocked: %d, " "should_sched_asleep: %d\n", should_sched_exists, should_sched_blocked, should_sched_asleep); if (should_sched_exists && !should_sched_blocked && !should_sched_asleep) { /* * schedule the task that should be executing in the cyclic * schedule if it is not blocked and not sleeping */ next = ce_data->should_schedule; } sched_state_task_picked(); raw_spin_unlock(dom->lock); return next; } static void mc_ce_finish_switch(struct task_struct *prev) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, smp_processor_id()); struct ce_dom_data *ce_data = get_ce_data(dom_data); TRACE("finish switch\n"); if (is_realtime(current) && CRIT_LEVEL_A == tsk_mc_crit(current)) ce_data->scheduled = current; else ce_data->scheduled = NULL; } /* * Called for every local timer interrupt. * Linux calls this with interrupts disabled, AFAIK. */ static void mc_ce_tick(struct task_struct *ts) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, smp_processor_id()); domain_t *dom = &dom_data->domain; struct ce_dom_data *ce_data = get_ce_data(dom_data); struct task_struct *should_schedule; if (is_realtime(ts) && CRIT_LEVEL_A == tsk_mc_crit(ts)) { raw_spin_lock(dom->lock); should_schedule = ce_data->should_schedule; raw_spin_unlock(dom->lock); if (!is_np(ts) && ts != should_schedule) { litmus_reschedule_local(); } else if (is_user_np(ts)) { request_exit_np(ts); } } } /* * Admit task called to see if this task is permitted to enter the system. * Here we look up the task's PID structure and save it in the proper slot on * the CPU this task will run on. */ static long __mc_ce_admit_task(struct task_struct *ts) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, get_partition(ts)); struct ce_dom_data *ce_data = get_ce_data(dom_data); struct mc_data *mcd = tsk_mc_data(ts); struct pid *pid = NULL; long retval = -EINVAL; const int lvl_a_id = mcd->mc_task.lvl_a_id; /* check the task has migrated to the right CPU (like in sched_cedf) */ if (task_cpu(ts) != get_partition(ts)) { printk(KERN_INFO "litmus: %d admitted on CPU %d but want %d ", ts->pid, task_cpu(ts), get_partition(ts)); goto out; } /* only level A tasks can be CE */ if (!mcd || CRIT_LEVEL_A != tsk_mc_crit(ts)) { printk(KERN_INFO "litmus: non-MC or non level A task %d\n", ts->pid); goto out; } /* try and get the task's PID structure */ pid = get_task_pid(ts, PIDTYPE_PID); if (IS_ERR_OR_NULL(pid)) { printk(KERN_INFO "litmus: couldn't get pid struct for %d\n", ts->pid); goto out; } if (lvl_a_id >= ce_data->num_pid_entries) { printk(KERN_INFO "litmus: level A id greater than expected " "number of tasks %d for %d cpu %d\n", ce_data->num_pid_entries, ts->pid, get_partition(ts)); goto out_put_pid; } if (ce_data->pid_entries[lvl_a_id].pid) { printk(KERN_INFO "litmus: have saved pid info id: %d cpu: %d\n", lvl_a_id, get_partition(ts)); goto out_put_pid; } if (get_exec_cost(ts) >= ce_data->pid_entries[lvl_a_id].budget) { printk(KERN_INFO "litmus: execution cost %llu is larger than " "the budget %llu\n", get_exec_cost(ts), ce_data->pid_entries[lvl_a_id].budget); goto out_put_pid; } ce_data->pid_entries[lvl_a_id].pid = pid; retval = 0; /* don't call put_pid if we are successful */ goto out; out_put_pid: put_pid(pid); out: return retval; } static long mc_ce_admit_task(struct task_struct *ts) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, get_partition(ts)); domain_t *dom = &dom_data->domain; unsigned long flags, retval; raw_spin_lock_irqsave(dom->lock, flags); retval = __mc_ce_admit_task(ts); raw_spin_unlock_irqrestore(dom->lock, flags); return retval; } /* * Called to set up a new real-time task (after the admit_task callback). * At this point the task's struct PID is already hooked up on the destination * CPU. The task may already be running. */ static void mc_ce_task_new(struct task_struct *ts, int on_rq, int running) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, task_cpu(ts)); domain_t *dom = &dom_data->domain; struct ce_dom_data *ce_data = get_ce_data(dom_data); struct pid *pid_should_be_running; struct ce_dom_pid_entry *pid_entry; unsigned long flags; int idx, should_be_running; lt_t offset; /* have to call mc_ce_schedule_at because the task only gets a PID * entry after calling admit_task */ raw_spin_lock_irqsave(dom->lock, flags); pid_entry = &ce_data->pid_entries[tsk_mc_data(ts)->mc_task.lvl_a_id]; /* initialize some task state */ set_rt_flags(ts, RT_F_RUNNING); tsk_rt(ts)->job_params.job_no = 0; offset = get_cycle_offset(litmus_clock(), ce_data->cycle_time); idx = mc_ce_schedule_at(dom, offset); pid_should_be_running = ce_data->pid_entries[idx].pid; rcu_read_lock(); should_be_running = (ts == pid_task(pid_should_be_running, PIDTYPE_PID)); rcu_read_unlock(); if (running) { /* admit task checks that the task is not on the wrong CPU */ BUG_ON(task_cpu(ts) != get_partition(ts)); BUG_ON(ce_data->scheduled); ce_data->scheduled = ts; if (should_be_running) ce_data->should_schedule = ts; else preempt_if_preemptable(ce_data->scheduled, ce_data->cpu); } else if (!running && should_be_running) { ce_data->should_schedule = ts; preempt_if_preemptable(ce_data->scheduled, ce_data->cpu); } raw_spin_unlock_irqrestore(dom->lock, flags); } /* * Called to re-introduce a task after blocking. * Can potentailly be called multiple times. */ static void mc_ce_task_wake_up(struct task_struct *ts) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, smp_processor_id()); domain_t *dom = &dom_data->domain; struct ce_dom_data *ce_data = get_ce_data(dom_data); unsigned long flags; TRACE_TASK(ts, "wake up\n"); raw_spin_lock_irqsave(dom->lock, flags); if (ts == ce_data->should_schedule && ts != ce_data->scheduled) preempt_if_preemptable(ts, ce_data->cpu); raw_spin_unlock_irqrestore(dom->lock, flags); } /* * Called to notify the plugin of a blocking real-time tasks. Only called for * real-time tasks and before schedule is called. */ static void mc_ce_task_block(struct task_struct *ts) { /* nothing to do because it will be taken care of in schedule */ TRACE_TASK(ts, "blocked\n"); } /* * Called when a task switches from RT mode back to normal mode. */ void mc_ce_task_exit(struct task_struct *ts) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, get_partition(ts)); domain_t *dom = &dom_data->domain; struct ce_dom_data *ce_data = get_ce_data(dom_data); unsigned long flags; struct pid *pid; const int lvl_a_id = tsk_mc_data(ts)->mc_task.lvl_a_id; TRACE_TASK(ts, "exited\n"); BUG_ON(task_cpu(ts) != get_partition(ts)); BUG_ON(CRIT_LEVEL_A != tsk_mc_crit(ts)); BUG_ON(lvl_a_id >= ce_data->num_pid_entries); raw_spin_lock_irqsave(dom->lock, flags); pid = ce_data->pid_entries[lvl_a_id].pid; BUG_ON(!pid); put_pid(pid); ce_data->pid_entries[lvl_a_id].pid = NULL; if (ce_data->scheduled == ts) ce_data->scheduled = NULL; if (ce_data->should_schedule == ts) ce_data->should_schedule = NULL; raw_spin_unlock_irqrestore(dom->lock, flags); } /*********************************************************** * Timer stuff **********************************************************/ void __mc_ce_timer_callback(struct hrtimer *timer) { /* relative and absolute times for cycles */ lt_t now, offset_rel, cycle_start_abs, next_timer_abs; struct task_struct *should_schedule; struct ce_dom_pid_entry *pid_entry; struct ce_dom_data *ce_data; domain_data_t *dom_data; domain_t *dom; int idx, budget_overrun; ce_data = container_of(timer, struct ce_dom_data, timer); dom_data = &per_cpu(mc_ce_doms, ce_data->cpu); dom = &dom_data->domain; /* Based off of the current time, figure out the offset into the cycle * and the cycle's start time, and determine what should be scheduled. */ now = litmus_clock(); offset_rel = get_cycle_offset(now, ce_data->cycle_time); cycle_start_abs = now - offset_rel; idx = mc_ce_schedule_at(dom, offset_rel); pid_entry = &ce_data->pid_entries[idx]; /* set the timer to fire at the next cycle start */ next_timer_abs = cycle_start_abs + pid_entry->acc_time; hrtimer_set_expires(timer, ns_to_ktime(next_timer_abs)); TRACE("timer: now: %llu offset_rel: %llu cycle_start_abs: %llu " "next_timer_abs: %llu\n", now, offset_rel, cycle_start_abs, next_timer_abs); /* get the task_struct (pid_task can accept a NULL) */ rcu_read_lock(); should_schedule = pid_task(pid_entry->pid, PIDTYPE_PID); rcu_read_unlock(); ce_data->should_schedule = should_schedule; if (should_schedule && 0 == atomic_read(&start_time_set)) { /* * If jobs are not overrunning their budgets, then this * should not happen. */ pid_entry->expected_job++; budget_overrun = pid_entry->expected_job != tsk_rt(should_schedule)->job_params.job_no; if (budget_overrun) TRACE_TASK(should_schedule, "timer expected job number: %d " "but current job: %d\n", pid_entry->expected_job, tsk_rt(should_schedule)->job_params.job_no); } if (ce_data->should_schedule) { tsk_rt(should_schedule)->job_params.deadline = cycle_start_abs + pid_entry->acc_time; tsk_rt(should_schedule)->job_params.release = tsk_rt(should_schedule)->job_params.deadline - pid_entry->budget; tsk_rt(should_schedule)->job_params.exec_time = 0; sched_trace_task_release(should_schedule); set_rt_flags(ce_data->should_schedule, RT_F_RUNNING); } } /* * What to do when a timer fires. The timer should only be armed if the number * of PID entries is positive. */ static enum hrtimer_restart mc_ce_timer_callback(struct hrtimer *timer) { struct ce_dom_data *ce_data; unsigned long flags; domain_data_t *dom_data; domain_t *dom; ce_data = container_of(timer, struct ce_dom_data, timer); dom_data = &per_cpu(mc_ce_doms, ce_data->cpu); dom = &dom_data->domain; TRACE("timer callback on CPU %d (before lock)\n", ce_data->cpu); raw_spin_lock_irqsave(dom->lock, flags); __mc_ce_timer_callback(timer); if (ce_data->scheduled != ce_data->should_schedule) preempt_if_preemptable(ce_data->scheduled, ce_data->cpu); raw_spin_unlock_irqrestore(dom->lock, flags); return HRTIMER_RESTART; } /* * Cancel timers on all CPUs. Returns 1 if any were active. */ static int cancel_all_timers(void) { struct ce_dom_data *ce_data; domain_data_t *dom_data; int cpu, ret = 0, cancel_res; TRACE("cancel all timers\n"); for_each_online_cpu(cpu) { dom_data = &per_cpu(mc_ce_doms, cpu); ce_data = get_ce_data(dom_data); ce_data->should_schedule = NULL; cancel_res = hrtimer_cancel(&ce_data->timer); atomic_set(&ce_data->timer_info.state, HRTIMER_START_ON_INACTIVE); ret = ret || cancel_res; } return ret; } /* * Arm all timers so that they start at the new value of start time. * Any CPU without CE PID entries won't have a timer armed. * All timers should be canceled before calling this. */ static void arm_all_timers(void) { struct ce_dom_data *ce_data; domain_data_t *dom_data; int cpu, idx; const lt_t start = atomic64_read(&start_time); TRACE("arm all timers\n"); for_each_online_cpu(cpu) { dom_data = &per_cpu(mc_ce_doms, cpu); ce_data = get_ce_data(dom_data); if (0 == ce_data->num_pid_entries) continue; for (idx = 0; idx < ce_data->num_pid_entries; idx++) { ce_data->pid_entries[idx].expected_job = -1; } TRACE("arming timer for CPU %d\n", cpu); hrtimer_start_on(cpu, &ce_data->timer_info, &ce_data->timer, ns_to_ktime(start), HRTIMER_MODE_ABS_PINNED); } } /* * There are no real releases in the CE, but the task release syscall will * call this. We can re-set our notion of the CE period start to make * the schedule look pretty. */ void mc_ce_release_at(struct task_struct *ts, lt_t start) { TRACE_TASK(ts, "release at\n"); if (atomic_inc_and_test(&start_time_set)) { /* in this case, we won the race */ cancel_all_timers(); atomic64_set(&start_time, start); arm_all_timers(); } else atomic_dec(&start_time_set); } long mc_ce_activate_plugin(void) { struct ce_dom_data *ce_data; domain_data_t *dom_data; int cpu; for_each_online_cpu(cpu) { dom_data = &per_cpu(mc_ce_doms, cpu); ce_data = get_ce_data(dom_data); ce_data->scheduled = NULL; ce_data->should_schedule = NULL; } atomic_set(&start_time_set, -1); atomic64_set(&start_time, litmus_clock()); /* may not want to arm timers on activation, just after release */ arm_all_timers(); return 0; } static void clear_pid_entries(void) { int cpu, entry; domain_data_t *dom_data; struct ce_dom_data *ce_data; for_each_online_cpu(cpu) { dom_data = &per_cpu(mc_ce_doms, cpu); ce_data = get_ce_data(dom_data); ce_data->num_pid_entries = 0; ce_data->cycle_time = 0; for (entry = 0; entry < CONFIG_PLUGIN_MC_LEVEL_A_MAX_TASKS; ++entry) { if (NULL != ce_data->pid_entries[entry].pid) { put_pid(ce_data->pid_entries[entry].pid); ce_data->pid_entries[entry].pid = NULL; } ce_data->pid_entries[entry].budget = 0; ce_data->pid_entries[entry].acc_time = 0; ce_data->pid_entries[entry].expected_job = -1; } } } long mc_ce_deactivate_plugin(void) { cancel_all_timers(); return 0; } /* Plugin object */ static struct sched_plugin mc_ce_plugin __cacheline_aligned_in_smp = { .plugin_name = "MC-CE", .admit_task = mc_ce_admit_task, .task_new = mc_ce_task_new, .complete_job = complete_job, .release_at = mc_ce_release_at, .task_exit = mc_ce_task_exit, .schedule = mc_ce_schedule, .finish_switch = mc_ce_finish_switch, .tick = mc_ce_tick, .task_wake_up = mc_ce_task_wake_up, .task_block = mc_ce_task_block, .activate_plugin = mc_ce_activate_plugin, .deactivate_plugin = mc_ce_deactivate_plugin, }; static int setup_proc(void); static int __init init_sched_mc_ce(void) { struct ce_dom_data *ce_data; domain_data_t *dom_data; domain_t *dom; rt_domain_t *rt; int cpu, err; for_each_online_cpu(cpu) { dom_data = &per_cpu(mc_ce_doms, cpu); dom = &dom_data->domain; rt = &per_cpu(mc_ce_rts, cpu); pd_domain_init(dom, rt, NULL, NULL, NULL, NULL, NULL); dom->data = &per_cpu(_mc_ce_dom_data, cpu); ce_data = get_ce_data(dom_data); hrtimer_init(&ce_data->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); hrtimer_start_on_info_init(&ce_data->timer_info); ce_data->cpu = cpu; ce_data->timer.function = mc_ce_timer_callback; } clear_pid_entries(); err = setup_proc(); if (!err) err = register_sched_plugin(&mc_ce_plugin); return err; } #define BUF_SIZE PAGE_SIZE static int write_into_proc(char *proc_buf, const int proc_size, char *fmt, ...) { static char buf[BUF_SIZE]; int n; va_list args; /* When writing to procfs, we don't care about the trailing null that * is not included in the count returned by vscnprintf. */ va_start(args, fmt); n = vsnprintf(buf, BUF_SIZE, fmt, args); va_end(args); if (BUF_SIZE <= n || proc_size <= n) { /* too big for formatting buffer or proc (less null byte) */ n = -EINVAL; goto out; } memcpy(proc_buf, buf, n); out: return n; } #undef BUF_SIZE /* * Writes a PID entry to the procfs. * * @page buffer to write into. * @count bytes available in the buffer */ #define PID_SPACE 15 #define TASK_INFO_BUF (PID_SPACE + TASK_COMM_LEN) static int write_pid_entry(char *page, const int count, const int cpu, const int task, struct ce_dom_pid_entry *pid_entry) { static char task_info[TASK_INFO_BUF]; struct task_struct *ts; int n = 0, err, ti_n; char *ti_b; if (pid_entry->pid) { rcu_read_lock(); ts = pid_task(pid_entry->pid, PIDTYPE_PID); rcu_read_unlock(); /* get some information about the task */ if (ts) { ti_b = task_info; ti_n = snprintf(ti_b, PID_SPACE, "%d", ts->pid); if (PID_SPACE <= ti_n) ti_n = PID_SPACE - 1; ti_b += ti_n; *ti_b = ' '; /* nuke the null byte */ ti_b++; get_task_comm(ti_b, ts); } else { strncpy(task_info, "pid_task() failed :(", TASK_INFO_BUF); } } else strncpy(task_info, "no", TASK_INFO_BUF); task_info[TASK_INFO_BUF - 1] = '\0'; /* just to be sure */ err = write_into_proc(page + n, count - n, "# task: %s\n", task_info); if (err < 0) { n = -ENOSPC; goto out; } n += err; err = write_into_proc(page + n, count - n, "%d, %d, %llu\n", cpu, task, pid_entry->budget); if (err < 0) { n = -ENOSPC; goto out; } n += err; out: return n; } #undef PID_SPACE #undef TASK_INFO_BUF /* * Called when the user-land reads from proc. */ static int proc_read_ce_file(char *page, char **start, off_t off, int count, int *eof, void *data) { int n = 0, err, cpu, t; struct ce_dom_data *ce_data; domain_data_t *dom_data; if (off > 0) { printk(KERN_INFO "litmus: MC-CE called read with off > 0\n"); goto out; } for_each_online_cpu(cpu) { dom_data = &per_cpu(mc_ce_doms, cpu); ce_data = get_ce_data(dom_data); for (t = 0; t < ce_data->num_pid_entries; ++t) { err = write_pid_entry(page + n, count - n, cpu, t, &ce_data->pid_entries[t]); if (err < 0) { n = -ENOSPC; goto out; } n += err; } } out: *eof = 1; return n; } /* * Skip a commented line. */ static int skip_comment(const char *buf, const unsigned long max) { unsigned long i = 0; const char *c = buf; if (0 == max || !c || *c != '#') return 0; ++c; ++i; for (; i < max; ++i) { if (*c == '\n') { ++c; ++i; break; } ++c; } return i; } /* a budget of 5 milliseconds is probably reasonable */ #define BUDGET_THRESHOLD 5000000ULL static int setup_pid_entry(const int cpu, const int task, const lt_t budget) { domain_data_t *dom_data = &per_cpu(mc_ce_doms, cpu); struct ce_dom_data *ce_data = get_ce_data(dom_data); struct ce_dom_pid_entry *new_entry; int err = 0; /* check the inputs */ if (cpu < 0 || NR_CPUS <= cpu || task < 0 || CONFIG_PLUGIN_MC_LEVEL_A_MAX_TASKS <= task || budget < 1) { printk(KERN_INFO "litmus: bad cpu, task ID, or budget sent to " "MC-CE proc\n"); err = -EINVAL; goto out; } /* check for small budgets */ if (BUDGET_THRESHOLD > budget) { printk(KERN_CRIT "litmus: you gave a small budget for an " "MC-CE task; that might be an issue.\n"); } /* check that we have space for a new entry */ if (CONFIG_PLUGIN_MC_LEVEL_A_MAX_TASKS <= ce_data->num_pid_entries) { printk(KERN_INFO "litmus: too many MC-CE tasks for cpu " "%d\n", cpu); err = -EINVAL; goto out; } /* add the new entry */ new_entry = &ce_data->pid_entries[ce_data->num_pid_entries]; BUG_ON(NULL != new_entry->pid); new_entry->budget = budget; new_entry->acc_time = ce_data->cycle_time + budget; /* update the domain entry */ ce_data->cycle_time += budget; ce_data->num_pid_entries++; out: return err; } #undef BUDGET_THRESHOLD /* * Called when the user-land writes to proc. * * Error checking is quite minimal. Format is: * , , */ #define PROCFS_MAX_SIZE PAGE_SIZE static int proc_write_ce_file(struct file *file, const char __user *buffer, unsigned long count, void *data) { static char kbuf[PROCFS_MAX_SIZE]; char *c = kbuf, *c_skipped; int cpu, task, cnt = 0, chars_read, converted, err; lt_t budget; if (is_active_plugin()) { printk(KERN_INFO "litmus: can't edit MC-CE proc when plugin " "active\n"); cnt = -EINVAL; goto out; } if (count > PROCFS_MAX_SIZE) { printk(KERN_INFO "litmus: MC-CE procfs got too many bytes " "from user-space.\n"); cnt = -EINVAL; goto out; } if (copy_from_user(kbuf, buffer, count)) { printk(KERN_INFO "litmus: couldn't copy from user %s\n", __FUNCTION__); cnt = -EFAULT; goto out; } clear_pid_entries(); while (cnt < count) { c_skipped = skip_spaces(c); if (c_skipped != c) { chars_read = c_skipped - c; cnt += chars_read; c += chars_read; continue; } if (*c == '#') { chars_read = skip_comment(c, count - cnt); cnt += chars_read; c += chars_read; continue; } converted = sscanf(c, "%d, %d, %llu%n", &cpu, &task, &budget, &chars_read); if (3 != converted) { printk(KERN_INFO "litmus: MC-CE procfs expected three " "arguments, but got %d.\n", converted); cnt = -EINVAL; goto out; } cnt += chars_read; c += chars_read; err = setup_pid_entry(cpu, task, budget); if (err) { cnt = -EINVAL; goto out; } } out: return cnt; } #undef PROCFS_MAX_SIZE #define CE_FILE_PROC_NAME "ce_file" static void tear_down_proc(void) { if (ce_file) remove_proc_entry(CE_FILE_PROC_NAME, mc_ce_dir); if (mc_ce_dir) remove_plugin_proc_dir(&mc_ce_plugin); } static int setup_proc(void) { int err; err = make_plugin_proc_dir(&mc_ce_plugin, &mc_ce_dir); if (err) { printk(KERN_ERR "could not create MC-CE procfs dir.\n"); goto out; } ce_file = create_proc_entry(CE_FILE_PROC_NAME, 0644, mc_ce_dir); if (!ce_file) { printk(KERN_ERR "could not create MC-CE procfs file.\n"); err = -EIO; goto out_remove_proc; } ce_file->read_proc = proc_read_ce_file; ce_file->write_proc = proc_write_ce_file; goto out; out_remove_proc: tear_down_proc(); out: return err; } #undef CE_FILE_PROC_NAME static void clean_sched_mc_ce(void) { tear_down_proc(); } module_init(init_sched_mc_ce); module_exit(clean_sched_mc_ce);