arch/ia64/xen/time.c

   1 /******************************************************************************
   2  * arch/ia64/xen/time.c
   3  *
   4  * Copyright (c) 2008 Isaku Yamahata <yamahata at valinux co jp>
   5  *                    VA Linux Systems Japan K.K.
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License as published by
   9  * the Free Software Foundation; either version 2 of the License, or
  10  * (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  20  *
  21  */
  22
  23 #include <linux/delay.h>
  24 #include <linux/kernel_stat.h>
  25 #include <linux/posix-timers.h>
  26 #include <linux/irq.h>
  27 #include <linux/clocksource.h>
  28
  29 #include <asm/timex.h>
  30
  31 #include <asm/xen/hypervisor.h>
  32
  33 #include <xen/interface/vcpu.h>
  34
  35 #include "../kernel/fsyscall_gtod_data.h"
  36
  37 static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate);
  38 static DEFINE_PER_CPU(unsigned long, xen_stolen_time);
  39 static DEFINE_PER_CPU(unsigned long, xen_blocked_time);
  40
  41 /* taken from i386/kernel/time-xen.c */
  42 static void xen_init_missing_ticks_accounting(int cpu)
  43 {
  44         struct vcpu_register_runstate_memory_area area;
  45         struct vcpu_runstate_info *runstate = &per_cpu(xen_runstate, cpu);
  46         int rc;
  47
  48         memset(runstate, 0, sizeof(*runstate));
  49
  50         area.addr.v = runstate;
  51         rc = HYPERVISOR_vcpu_op(VCPUOP_register_runstate_memory_area, cpu,
  52                                 &area);
  53         WARN_ON(rc && rc != -ENOSYS);
  54
  55         per_cpu(xen_blocked_time, cpu) = runstate->time[RUNSTATE_blocked];
  56         per_cpu(xen_stolen_time, cpu) = runstate->time[RUNSTATE_runnable]
  57                                             + runstate->time[RUNSTATE_offline];
  58 }
  59
  60 /*
  61  * Runstate accounting
  62  */
  63 /* stolen from arch/x86/xen/time.c */
  64 static void get_runstate_snapshot(struct vcpu_runstate_info *res)
  65 {
  66         u64 state_time;
  67         struct vcpu_runstate_info *state;
  68
  69         BUG_ON(preemptible());
  70
  71         state = &__get_cpu_var(xen_runstate);
  72
  73         /*
  74          * The runstate info is always updated by the hypervisor on
  75          * the current CPU, so there's no need to use anything
  76          * stronger than a compiler barrier when fetching it.
  77          */
  78         do {
  79                 state_time = state->state_entry_time;
  80                 rmb();
  81                 *res = *state;
  82                 rmb();
  83         } while (state->state_entry_time != state_time);
  84 }
  85
  86 #define NS_PER_TICK (1000000000LL/HZ)
  87
  88 static unsigned long
  89 consider_steal_time(unsigned long new_itm)
  90 {
  91         unsigned long stolen, blocked;
  92         unsigned long delta_itm = 0, stolentick = 0;
  93         int cpu = smp_processor_id();
  94         struct vcpu_runstate_info runstate;
  95         struct task_struct *p = current;
  96
  97         get_runstate_snapshot(&runstate);
  98
  99         /*
 100          * Check for vcpu migration effect
 101          * In this case, itc value is reversed.
 102          * This causes huge stolen value.
 103          * This function just checks and reject this effect.
 104          */
 105         if (!time_after_eq(runstate.time[RUNSTATE_blocked],
 106                            per_cpu(xen_blocked_time, cpu)))
 107                 blocked = 0;
 108
 109         if (!time_after_eq(runstate.time[RUNSTATE_runnable] +
 110                            runstate.time[RUNSTATE_offline],
 111                            per_cpu(xen_stolen_time, cpu)))
 112                 stolen = 0;
 113
 114         if (!time_after(delta_itm + new_itm, ia64_get_itc()))
 115                 stolentick = ia64_get_itc() - new_itm;
 116
 117         do_div(stolentick, NS_PER_TICK);
 118         stolentick++;
 119
 120         do_div(stolen, NS_PER_TICK);
 121
 122         if (stolen > stolentick)
 123                 stolen = stolentick;
 124
 125         stolentick -= stolen;
 126         do_div(blocked, NS_PER_TICK);
 127
 128         if (blocked > stolentick)
 129                 blocked = stolentick;
 130
 131         if (stolen > 0 || blocked > 0) {
 132                 account_steal_ticks(stolen);
 133                 account_idle_ticks(blocked);
 134                 run_local_timers();
 135
 136                 rcu_check_callbacks(cpu, user_mode(get_irq_regs()));
 137
 138                 scheduler_tick();
 139                 run_posix_cpu_timers(p);
 140                 delta_itm += local_cpu_data->itm_delta * (stolen + blocked);
 141
 142                 if (cpu == time_keeper_id) {
 143                         write_seqlock(&xtime_lock);
 144                         do_timer(stolen + blocked);
 145                         local_cpu_data->itm_next = delta_itm + new_itm;
 146                         write_sequnlock(&xtime_lock);
 147                 } else {
 148                         local_cpu_data->itm_next = delta_itm + new_itm;
 149                 }
 150                 per_cpu(xen_stolen_time, cpu) += NS_PER_TICK * stolen;
 151                 per_cpu(xen_blocked_time, cpu) += NS_PER_TICK * blocked;
 152         }
 153         return delta_itm;
 154 }
 155
 156 static int xen_do_steal_accounting(unsigned long *new_itm)
 157 {
 158         unsigned long delta_itm;
 159         delta_itm = consider_steal_time(*new_itm);
 160         *new_itm += delta_itm;
 161         if (time_after(*new_itm, ia64_get_itc()) && delta_itm)
 162                 return 1;
 163
 164         return 0;
 165 }
 166
 167 static void xen_itc_jitter_data_reset(void)
 168 {
 169         u64 lcycle, ret;
 170
 171         do {
 172                 lcycle = itc_jitter_data.itc_lastcycle;
 173                 ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, 0);
 174         } while (unlikely(ret != lcycle));
 175 }
 176
 177 /* based on xen_sched_clock() in arch/x86/xen/time.c. */
 178 /*
 179  * This relies on HAVE_UNSTABLE_SCHED_CLOCK. If it can't be defined,
 180  * something similar logic should be implemented here.
 181  */
 182 /*
 183  * Xen sched_clock implementation.  Returns the number of unstolen
 184  * nanoseconds, which is nanoseconds the VCPU spent in RUNNING+BLOCKED
 185  * states.
 186  */
 187 static unsigned long long xen_sched_clock(void)
 188 {
 189         struct vcpu_runstate_info runstate;
 190
 191         unsigned long long now;
 192         unsigned long long offset;
 193         unsigned long long ret;
 194
 195         /*
 196          * Ideally sched_clock should be called on a per-cpu basis
 197          * anyway, so preempt should already be disabled, but that's
 198          * not current practice at the moment.
 199          */
 200         preempt_disable();
 201
 202         /*
 203          * both ia64_native_sched_clock() and xen's runstate are
 204          * based on mAR.ITC. So difference of them makes sense.
 205          */
 206         now = ia64_native_sched_clock();
 207
 208         get_runstate_snapshot(&runstate);
 209
 210         WARN_ON(runstate.state != RUNSTATE_running);
 211
 212         offset = 0;
 213         if (now > runstate.state_entry_time)
 214                 offset = now - runstate.state_entry_time;
 215         ret = runstate.time[RUNSTATE_blocked] +
 216                 runstate.time[RUNSTATE_running] +
 217                 offset;
 218
 219         preempt_enable();
 220
 221         return ret;
 222 }
 223
 224 struct pv_time_ops xen_time_ops __initdata = {
 225         .init_missing_ticks_accounting  = xen_init_missing_ticks_accounting,
 226         .do_steal_accounting            = xen_do_steal_accounting,
 227         .clocksource_resume             = xen_itc_jitter_data_reset,
 228         .sched_clock                    = xen_sched_clock,
 229 };
 230
 231 /* Called after suspend, to resume time.  */
 232 static void xen_local_tick_resume(void)
 233 {
 234         /* Just trigger a tick.  */
 235         ia64_cpu_local_tick();
 236         touch_softlockup_watchdog();
 237 }
 238
 239 void
 240 xen_timer_resume(void)
 241 {
 242         unsigned int cpu;
 243
 244         xen_local_tick_resume();
 245
 246         for_each_online_cpu(cpu)
 247                 xen_init_missing_ticks_accounting(cpu);
 248 }
 249
 250 static void ia64_cpu_local_tick_fn(void *unused)
 251 {
 252         xen_local_tick_resume();
 253         xen_init_missing_ticks_accounting(smp_processor_id());
 254 }
 255
 256 void
 257 xen_timer_resume_on_aps(void)
 258 {
 259         smp_call_function(&ia64_cpu_local_tick_fn, NULL, 1);
 260 }