]> err.no Git - linux-2.6/blobdiff - arch/powerpc/kernel/time.c
[POWERPC] Prevent decrementer clockevents from firing early
[linux-2.6] / arch / powerpc / kernel / time.c
index 2c8564d54e4d009a67527463d1aa2a8c7f3f9d6f..64b503c82a31a4d6848a873b0b08aa28d90795ab 100644 (file)
 #include <asm/div64.h>
 #include <asm/smp.h>
 #include <asm/vdso_datapage.h>
-#ifdef CONFIG_PPC64
 #include <asm/firmware.h>
-#endif
 #ifdef CONFIG_PPC_ISERIES
 #include <asm/iseries/it_lp_queue.h>
 #include <asm/iseries/hv_call_xm.h>
 #endif
-#include <asm/smp.h>
 
-/* keep track of when we need to update the rtc */
-time_t last_rtc_update;
+/* powerpc clocksource/clockevent code */
+
+#include <linux/clockchips.h>
+#include <linux/clocksource.h>
+
+static cycle_t rtc_read(void);
+static struct clocksource clocksource_rtc = {
+       .name         = "rtc",
+       .rating       = 400,
+       .flags        = CLOCK_SOURCE_IS_CONTINUOUS,
+       .mask         = CLOCKSOURCE_MASK(64),
+       .shift        = 22,
+       .mult         = 0,      /* To be filled in */
+       .read         = rtc_read,
+};
+
+static cycle_t timebase_read(void);
+static struct clocksource clocksource_timebase = {
+       .name         = "timebase",
+       .rating       = 400,
+       .flags        = CLOCK_SOURCE_IS_CONTINUOUS,
+       .mask         = CLOCKSOURCE_MASK(64),
+       .shift        = 22,
+       .mult         = 0,      /* To be filled in */
+       .read         = timebase_read,
+};
+
+#define DECREMENTER_MAX        0x7fffffff
+
+static int decrementer_set_next_event(unsigned long evt,
+                                     struct clock_event_device *dev);
+static void decrementer_set_mode(enum clock_event_mode mode,
+                                struct clock_event_device *dev);
+
+static struct clock_event_device decrementer_clockevent = {
+       .name           = "decrementer",
+       .rating         = 200,
+       .shift          = 32,
+       .mult           = 0,    /* To be filled in */
+       .irq            = 0,
+       .set_next_event = decrementer_set_next_event,
+       .set_mode       = decrementer_set_mode,
+       .features       = CLOCK_EVT_FEAT_ONESHOT,
+};
+
+static DEFINE_PER_CPU(struct clock_event_device, decrementers);
+void init_decrementer_clockevent(void);
+static DEFINE_PER_CPU(u64, decrementer_next_tb);
+
 #ifdef CONFIG_PPC_ISERIES
-unsigned long iSeries_recal_titan = 0;
-unsigned long iSeries_recal_tb = 0; 
-static unsigned long first_settimeofday = 1;
-#endif
+static unsigned long __initdata iSeries_recal_titan;
+static signed long __initdata iSeries_recal_tb;
 
-/* The decrementer counts down by 128 every 128ns on a 601. */
-#define DECREMENTER_COUNT_601  (1000000000 / HZ)
+/* Forward declaration is only needed for iSereis compiles */
+void __init clocksource_init(void);
+#endif
 
 #define XSEC_PER_SEC (1024*1024)
 
@@ -113,8 +156,9 @@ u64 ticklen_to_xs;  /* 0.64 fraction */
 DEFINE_SPINLOCK(rtc_lock);
 EXPORT_SYMBOL_GPL(rtc_lock);
 
-u64 tb_to_ns_scale;
-unsigned tb_to_ns_shift;
+static u64 tb_to_ns_scale __read_mostly;
+static unsigned tb_to_ns_shift __read_mostly;
+static unsigned long boot_tb __read_mostly;
 
 struct gettimeofday_struct do_gtod;
 
@@ -122,6 +166,7 @@ extern struct timezone sys_tz;
 static long timezone_offset;
 
 unsigned long ppc_proc_freq;
+EXPORT_SYMBOL(ppc_proc_freq);
 unsigned long ppc_tb_freq;
 
 static u64 tb_last_jiffy __cacheline_aligned_in_smp;
@@ -214,7 +259,6 @@ static void account_process_time(struct pt_regs *regs)
        run_posix_cpu_timers(current);
 }
 
-#ifdef CONFIG_PPC_SPLPAR
 /*
  * Stuff for accounting stolen time.
  */
@@ -222,19 +266,28 @@ struct cpu_purr_data {
        int     initialized;                    /* thread is running */
        u64     tb;                     /* last TB value read */
        u64     purr;                   /* last PURR value read */
-       spinlock_t lock;
 };
 
+/*
+ * Each entry in the cpu_purr_data array is manipulated only by its
+ * "owner" cpu -- usually in the timer interrupt but also occasionally
+ * in process context for cpu online.  As long as cpus do not touch
+ * each others' cpu_purr_data, disabling local interrupts is
+ * sufficient to serialize accesses.
+ */
 static DEFINE_PER_CPU(struct cpu_purr_data, cpu_purr_data);
 
 static void snapshot_tb_and_purr(void *data)
 {
+       unsigned long flags;
        struct cpu_purr_data *p = &__get_cpu_var(cpu_purr_data);
 
-       p->tb = mftb();
+       local_irq_save(flags);
+       p->tb = get_tb_or_rtc();
        p->purr = mfspr(SPRN_PURR);
        wmb();
        p->initialized = 1;
+       local_irq_restore(flags);
 }
 
 /*
@@ -242,15 +295,14 @@ static void snapshot_tb_and_purr(void *data)
  */
 void snapshot_timebases(void)
 {
-       int cpu;
-
        if (!cpu_has_feature(CPU_FTR_PURR))
                return;
-       for_each_possible_cpu(cpu)
-               spin_lock_init(&per_cpu(cpu_purr_data, cpu).lock);
        on_each_cpu(snapshot_tb_and_purr, NULL, 0, 1);
 }
 
+/*
+ * Must be called with interrupts disabled.
+ */
 void calculate_steal_time(void)
 {
        u64 tb, purr;
@@ -262,7 +314,6 @@ void calculate_steal_time(void)
        pme = &per_cpu(cpu_purr_data, smp_processor_id());
        if (!pme->initialized)
                return;         /* this can happen in early boot */
-       spin_lock(&pme->lock);
        tb = mftb();
        purr = mfspr(SPRN_PURR);
        stolen = (tb - pme->tb) - (purr - pme->purr);
@@ -270,9 +321,9 @@ void calculate_steal_time(void)
                account_steal_time(current, stolen);
        pme->tb = tb;
        pme->purr = purr;
-       spin_unlock(&pme->lock);
 }
 
+#ifdef CONFIG_PPC_SPLPAR
 /*
  * Must be called before the cpu is added to the online map when
  * a cpu is being brought up at runtime.
@@ -284,12 +335,12 @@ static void snapshot_purr(void)
 
        if (!cpu_has_feature(CPU_FTR_PURR))
                return;
+       local_irq_save(flags);
        pme = &per_cpu(cpu_purr_data, smp_processor_id());
-       spin_lock_irqsave(&pme->lock, flags);
        pme->tb = mftb();
        pme->purr = mfspr(SPRN_PURR);
        pme->initialized = 1;
-       spin_unlock_irqrestore(&pme->lock, flags);
+       local_irq_restore(flags);
 }
 
 #endif /* CONFIG_PPC_SPLPAR */
@@ -310,7 +361,7 @@ static void snapshot_purr(void)
  */
 void snapshot_timebase(void)
 {
-       __get_cpu_var(last_jiffy) = get_tb();
+       __get_cpu_var(last_jiffy) = get_tb_or_rtc();
        snapshot_purr();
 }
 
@@ -342,98 +393,6 @@ void udelay(unsigned long usecs)
 }
 EXPORT_SYMBOL(udelay);
 
-static __inline__ void timer_check_rtc(void)
-{
-        /*
-         * update the rtc when needed, this should be performed on the
-         * right fraction of a second. Half or full second ?
-         * Full second works on mk48t59 clocks, others need testing.
-         * Note that this update is basically only used through 
-         * the adjtimex system calls. Setting the HW clock in
-         * any other way is a /dev/rtc and userland business.
-         * This is still wrong by -0.5/+1.5 jiffies because of the
-         * timer interrupt resolution and possible delay, but here we 
-         * hit a quantization limit which can only be solved by higher
-         * resolution timers and decoupling time management from timer
-         * interrupts. This is also wrong on the clocks
-         * which require being written at the half second boundary.
-         * We should have an rtc call that only sets the minutes and
-         * seconds like on Intel to avoid problems with non UTC clocks.
-         */
-        if (ppc_md.set_rtc_time && ntp_synced() &&
-           xtime.tv_sec - last_rtc_update >= 659 &&
-           abs((xtime.tv_nsec/1000) - (1000000-1000000/HZ)) < 500000/HZ) {
-               struct rtc_time tm;
-               to_tm(xtime.tv_sec + 1 + timezone_offset, &tm);
-               tm.tm_year -= 1900;
-               tm.tm_mon -= 1;
-               if (ppc_md.set_rtc_time(&tm) == 0)
-                       last_rtc_update = xtime.tv_sec + 1;
-               else
-                       /* Try again one minute later */
-                       last_rtc_update += 60;
-        }
-}
-
-/*
- * This version of gettimeofday has microsecond resolution.
- */
-static inline void __do_gettimeofday(struct timeval *tv)
-{
-       unsigned long sec, usec;
-       u64 tb_ticks, xsec;
-       struct gettimeofday_vars *temp_varp;
-       u64 temp_tb_to_xs, temp_stamp_xsec;
-
-       /*
-        * These calculations are faster (gets rid of divides)
-        * if done in units of 1/2^20 rather than microseconds.
-        * The conversion to microseconds at the end is done
-        * without a divide (and in fact, without a multiply)
-        */
-       temp_varp = do_gtod.varp;
-
-       /* Sampling the time base must be done after loading
-        * do_gtod.varp in order to avoid racing with update_gtod.
-        */
-       data_barrier(temp_varp);
-       tb_ticks = get_tb() - temp_varp->tb_orig_stamp;
-       temp_tb_to_xs = temp_varp->tb_to_xs;
-       temp_stamp_xsec = temp_varp->stamp_xsec;
-       xsec = temp_stamp_xsec + mulhdu(tb_ticks, temp_tb_to_xs);
-       sec = xsec / XSEC_PER_SEC;
-       usec = (unsigned long)xsec & (XSEC_PER_SEC - 1);
-       usec = SCALE_XSEC(usec, 1000000);
-
-       tv->tv_sec = sec;
-       tv->tv_usec = usec;
-}
-
-void do_gettimeofday(struct timeval *tv)
-{
-       if (__USE_RTC()) {
-               /* do this the old way */
-               unsigned long flags, seq;
-               unsigned int sec, nsec, usec;
-
-               do {
-                       seq = read_seqbegin_irqsave(&xtime_lock, flags);
-                       sec = xtime.tv_sec;
-                       nsec = xtime.tv_nsec + tb_ticks_since(tb_last_jiffy);
-               } while (read_seqretry_irqrestore(&xtime_lock, seq, flags));
-               usec = nsec / 1000;
-               while (usec >= 1000000) {
-                       usec -= 1000000;
-                       ++sec;
-               }
-               tv->tv_sec = sec;
-               tv->tv_usec = usec;
-               return;
-       }
-       __do_gettimeofday(tv);
-}
-
-EXPORT_SYMBOL(do_gettimeofday);
 
 /*
  * There are two copies of tb_to_xs and stamp_xsec so that no
@@ -479,56 +438,6 @@ static inline void update_gtod(u64 new_tb_stamp, u64 new_stamp_xsec,
        ++(vdso_data->tb_update_count);
 }
 
-/*
- * When the timebase - tb_orig_stamp gets too big, we do a manipulation
- * between tb_orig_stamp and stamp_xsec. The goal here is to keep the
- * difference tb - tb_orig_stamp small enough to always fit inside a
- * 32 bits number. This is a requirement of our fast 32 bits userland
- * implementation in the vdso. If we "miss" a call to this function
- * (interrupt latency, CPU locked in a spinlock, ...) and we end up
- * with a too big difference, then the vdso will fallback to calling
- * the syscall
- */
-static __inline__ void timer_recalc_offset(u64 cur_tb)
-{
-       unsigned long offset;
-       u64 new_stamp_xsec;
-       u64 tlen, t2x;
-       u64 tb, xsec_old, xsec_new;
-       struct gettimeofday_vars *varp;
-
-       if (__USE_RTC())
-               return;
-       tlen = current_tick_length();
-       offset = cur_tb - do_gtod.varp->tb_orig_stamp;
-       if (tlen == last_tick_len && offset < 0x80000000u)
-               return;
-       if (tlen != last_tick_len) {
-               t2x = mulhdu(tlen << TICKLEN_SHIFT, ticklen_to_xs);
-               last_tick_len = tlen;
-       } else
-               t2x = do_gtod.varp->tb_to_xs;
-       new_stamp_xsec = (u64) xtime.tv_nsec * XSEC_PER_SEC;
-       do_div(new_stamp_xsec, 1000000000);
-       new_stamp_xsec += (u64) xtime.tv_sec * XSEC_PER_SEC;
-
-       ++vdso_data->tb_update_count;
-       smp_mb();
-
-       /*
-        * Make sure time doesn't go backwards for userspace gettimeofday.
-        */
-       tb = get_tb();
-       varp = do_gtod.varp;
-       xsec_old = mulhdu(tb - varp->tb_orig_stamp, varp->tb_to_xs)
-               + varp->stamp_xsec;
-       xsec_new = mulhdu(tb - cur_tb, t2x) + new_stamp_xsec;
-       if (xsec_new < xsec_old)
-               new_stamp_xsec += xsec_old - xsec_new;
-
-       update_gtod(cur_tb, new_stamp_xsec, t2x);
-}
-
 #ifdef CONFIG_SMP
 unsigned long profile_pc(struct pt_regs *regs)
 {
@@ -550,10 +459,15 @@ EXPORT_SYMBOL(profile_pc);
  * returned by the service processor for the timebase frequency.  
  */
 
-static void iSeries_tb_recal(void)
+static int __init iSeries_tb_recal(void)
 {
        struct div_result divres;
        unsigned long titan, tb;
+
+       /* Make sure we only run on iSeries */
+       if (!firmware_has_feature(FW_FEATURE_ISERIES))
+               return -ENODEV;
+
        tb = get_tb();
        titan = HvCallXm_loadTod();
        if ( iSeries_recal_titan ) {
@@ -594,8 +508,20 @@ static void iSeries_tb_recal(void)
        }
        iSeries_recal_titan = titan;
        iSeries_recal_tb = tb;
+
+       /* Called here as now we know accurate values for the timebase */
+       clocksource_init();
+       return 0;
 }
-#endif
+late_initcall(iSeries_tb_recal);
+
+/* Called from platform early init */
+void __init iSeries_time_init_early(void)
+{
+       iSeries_recal_tb = get_tb();
+       iSeries_recal_titan = HvCallXm_loadTod();
+}
+#endif /* CONFIG_PPC_ISERIES */
 
 /*
  * For iSeries shared processors, we have to let the hypervisor
@@ -614,20 +540,30 @@ static void iSeries_tb_recal(void)
 void timer_interrupt(struct pt_regs * regs)
 {
        struct pt_regs *old_regs;
-       int next_dec;
        int cpu = smp_processor_id();
-       unsigned long ticks;
-       u64 tb_next_jiffy;
+       struct clock_event_device *evt = &per_cpu(decrementers, cpu);
+       u64 now;
+
+       /* Ensure a positive value is written to the decrementer, or else
+        * some CPUs will continuue to take decrementer exceptions */
+       set_dec(DECREMENTER_MAX);
 
 #ifdef CONFIG_PPC32
        if (atomic_read(&ppc_n_lost_interrupts) != 0)
                do_IRQ(regs);
 #endif
 
+       now = get_tb_or_rtc();
+       if (now < per_cpu(decrementer_next_tb, cpu)) {
+               /* not time for this event yet */
+               now = per_cpu(decrementer_next_tb, cpu) - now;
+               if (now <= DECREMENTER_MAX)
+                       set_dec((unsigned int)now - 1);
+               return;
+       }
        old_regs = set_irq_regs(regs);
        irq_enter();
 
-       profile_tick(CPU_PROFILING);
        calculate_steal_time();
 
 #ifdef CONFIG_PPC_ISERIES
@@ -635,44 +571,20 @@ void timer_interrupt(struct pt_regs * regs)
                get_lppaca()->int_dword.fields.decr_int = 0;
 #endif
 
-       while ((ticks = tb_ticks_since(per_cpu(last_jiffy, cpu)))
-              >= tb_ticks_per_jiffy) {
-               /* Update last_jiffy */
-               per_cpu(last_jiffy, cpu) += tb_ticks_per_jiffy;
-               /* Handle RTCL overflow on 601 */
-               if (__USE_RTC() && per_cpu(last_jiffy, cpu) >= 1000000000)
-                       per_cpu(last_jiffy, cpu) -= 1000000000;
-
-               /*
-                * We cannot disable the decrementer, so in the period
-                * between this cpu's being marked offline in cpu_online_map
-                * and calling stop-self, it is taking timer interrupts.
-                * Avoid calling into the scheduler rebalancing code if this
-                * is the case.
-                */
-               if (!cpu_is_offline(cpu))
-                       account_process_time(regs);
-
-               /*
-                * No need to check whether cpu is offline here; boot_cpuid
-                * should have been fixed up by now.
-                */
-               if (cpu != boot_cpuid)
-                       continue;
+       /*
+        * We cannot disable the decrementer, so in the period
+        * between this cpu's being marked offline in cpu_online_map
+        * and calling stop-self, it is taking timer interrupts.
+        * Avoid calling into the scheduler rebalancing code if this
+        * is the case.
+        */
+       if (!cpu_is_offline(cpu))
+               account_process_time(regs);
 
-               write_seqlock(&xtime_lock);
-               tb_next_jiffy = tb_last_jiffy + tb_ticks_per_jiffy;
-               if (per_cpu(last_jiffy, cpu) >= tb_next_jiffy) {
-                       tb_last_jiffy = tb_next_jiffy;
-                       do_timer(1);
-                       timer_recalc_offset(tb_last_jiffy);
-                       timer_check_rtc();
-               }
-               write_sequnlock(&xtime_lock);
-       }
-       
-       next_dec = tb_ticks_per_jiffy - ticks;
-       set_dec(next_dec);
+       if (evt->event_handler)
+               evt->event_handler(evt);
+       else
+               evt->set_next_event(DECREMENTER_MAX, evt);
 
 #ifdef CONFIG_PPC_ISERIES
        if (firmware_has_feature(FW_FEATURE_ISERIES) && hvlpevent_is_pending())
@@ -735,80 +647,9 @@ unsigned long long sched_clock(void)
 {
        if (__USE_RTC())
                return get_rtc();
-       return mulhdu(get_tb(), tb_to_ns_scale) << tb_to_ns_shift;
-}
-
-int do_settimeofday(struct timespec *tv)
-{
-       time_t wtm_sec, new_sec = tv->tv_sec;
-       long wtm_nsec, new_nsec = tv->tv_nsec;
-       unsigned long flags;
-       u64 new_xsec;
-       unsigned long tb_delta;
-
-       if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
-               return -EINVAL;
-
-       write_seqlock_irqsave(&xtime_lock, flags);
-
-       /*
-        * Updating the RTC is not the job of this code. If the time is
-        * stepped under NTP, the RTC will be updated after STA_UNSYNC
-        * is cleared.  Tools like clock/hwclock either copy the RTC
-        * to the system time, in which case there is no point in writing
-        * to the RTC again, or write to the RTC but then they don't call
-        * settimeofday to perform this operation.
-        */
-#ifdef CONFIG_PPC_ISERIES
-       if (firmware_has_feature(FW_FEATURE_ISERIES) && first_settimeofday) {
-               iSeries_tb_recal();
-               first_settimeofday = 0;
-       }
-#endif
-
-       /* Make userspace gettimeofday spin until we're done. */
-       ++vdso_data->tb_update_count;
-       smp_mb();
-
-       /*
-        * Subtract off the number of nanoseconds since the
-        * beginning of the last tick.
-        */
-       tb_delta = tb_ticks_since(tb_last_jiffy);
-       tb_delta = mulhdu(tb_delta, do_gtod.varp->tb_to_xs); /* in xsec */
-       new_nsec -= SCALE_XSEC(tb_delta, 1000000000);
-
-       wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - new_sec);
-       wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - new_nsec);
-
-       set_normalized_timespec(&xtime, new_sec, new_nsec);
-       set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
-
-       /* In case of a large backwards jump in time with NTP, we want the 
-        * clock to be updated as soon as the PLL is again in lock.
-        */
-       last_rtc_update = new_sec - 658;
-
-       ntp_clear();
-
-       new_xsec = xtime.tv_nsec;
-       if (new_xsec != 0) {
-               new_xsec *= XSEC_PER_SEC;
-               do_div(new_xsec, NSEC_PER_SEC);
-       }
-       new_xsec += (u64)xtime.tv_sec * XSEC_PER_SEC;
-       update_gtod(tb_last_jiffy, new_xsec, do_gtod.varp->tb_to_xs);
-
-       vdso_data->tz_minuteswest = sys_tz.tz_minuteswest;
-       vdso_data->tz_dsttime = sys_tz.tz_dsttime;
-
-       write_sequnlock_irqrestore(&xtime_lock, flags);
-       clock_was_set();
-       return 0;
+       return mulhdu(get_tb() - boot_tb, tb_to_ns_scale) << tb_to_ns_shift;
 }
 
-EXPORT_SYMBOL(do_settimeofday);
-
 static int __init get_freq(char *name, int cells, unsigned long *val)
 {
        struct device_node *cpu;
@@ -851,7 +692,7 @@ void __init generic_calibrate_decr(void)
                                "(not found)\n");
        }
 
-#ifdef CONFIG_BOOKE
+#if defined(CONFIG_BOOKE) || defined(CONFIG_40x)
        /* Set the time base to zero */
        mtspr(SPRN_TBWL, 0);
        mtspr(SPRN_TBWU, 0);
@@ -864,12 +705,35 @@ void __init generic_calibrate_decr(void)
 #endif
 }
 
-unsigned long get_boot_time(void)
+int update_persistent_clock(struct timespec now)
 {
        struct rtc_time tm;
 
-       if (ppc_md.get_boot_time)
-               return ppc_md.get_boot_time();
+       if (!ppc_md.set_rtc_time)
+               return 0;
+
+       to_tm(now.tv_sec + 1 + timezone_offset, &tm);
+       tm.tm_year -= 1900;
+       tm.tm_mon -= 1;
+
+       return ppc_md.set_rtc_time(&tm);
+}
+
+unsigned long read_persistent_clock(void)
+{
+       struct rtc_time tm;
+       static int first = 1;
+
+       /* XXX this is a litle fragile but will work okay in the short term */
+       if (first) {
+               first = 0;
+               if (ppc_md.time_init)
+                       timezone_offset = ppc_md.time_init();
+
+               /* get_boot_time() isn't guaranteed to be safe to call late */
+               if (ppc_md.get_boot_time)
+                       return ppc_md.get_boot_time() -timezone_offset;
+       }
        if (!ppc_md.get_rtc_time)
                return 0;
        ppc_md.get_rtc_time(&tm);
@@ -877,18 +741,128 @@ unsigned long get_boot_time(void)
                      tm.tm_hour, tm.tm_min, tm.tm_sec);
 }
 
+/* clocksource code */
+static cycle_t rtc_read(void)
+{
+       return (cycle_t)get_rtc();
+}
+
+static cycle_t timebase_read(void)
+{
+       return (cycle_t)get_tb();
+}
+
+void update_vsyscall(struct timespec *wall_time, struct clocksource *clock)
+{
+       u64 t2x, stamp_xsec;
+
+       if (clock != &clocksource_timebase)
+               return;
+
+       /* Make userspace gettimeofday spin until we're done. */
+       ++vdso_data->tb_update_count;
+       smp_mb();
+
+       /* XXX this assumes clock->shift == 22 */
+       /* 4611686018 ~= 2^(20+64-22) / 1e9 */
+       t2x = (u64) clock->mult * 4611686018ULL;
+       stamp_xsec = (u64) xtime.tv_nsec * XSEC_PER_SEC;
+       do_div(stamp_xsec, 1000000000);
+       stamp_xsec += (u64) xtime.tv_sec * XSEC_PER_SEC;
+       update_gtod(clock->cycle_last, stamp_xsec, t2x);
+}
+
+void update_vsyscall_tz(void)
+{
+       /* Make userspace gettimeofday spin until we're done. */
+       ++vdso_data->tb_update_count;
+       smp_mb();
+       vdso_data->tz_minuteswest = sys_tz.tz_minuteswest;
+       vdso_data->tz_dsttime = sys_tz.tz_dsttime;
+       smp_mb();
+       ++vdso_data->tb_update_count;
+}
+
+void __init clocksource_init(void)
+{
+       struct clocksource *clock;
+
+       if (__USE_RTC())
+               clock = &clocksource_rtc;
+       else
+               clock = &clocksource_timebase;
+
+       clock->mult = clocksource_hz2mult(tb_ticks_per_sec, clock->shift);
+
+       if (clocksource_register(clock)) {
+               printk(KERN_ERR "clocksource: %s is already registered\n",
+                      clock->name);
+               return;
+       }
+
+       printk(KERN_INFO "clocksource: %s mult[%x] shift[%d] registered\n",
+              clock->name, clock->mult, clock->shift);
+}
+
+static int decrementer_set_next_event(unsigned long evt,
+                                     struct clock_event_device *dev)
+{
+       __get_cpu_var(decrementer_next_tb) = get_tb_or_rtc() + evt;
+       /* The decrementer interrupts on the 0 -> -1 transition */
+       if (evt)
+               --evt;
+       set_dec(evt);
+       return 0;
+}
+
+static void decrementer_set_mode(enum clock_event_mode mode,
+                                struct clock_event_device *dev)
+{
+       if (mode != CLOCK_EVT_MODE_ONESHOT)
+               decrementer_set_next_event(DECREMENTER_MAX, dev);
+}
+
+static void register_decrementer_clockevent(int cpu)
+{
+       struct clock_event_device *dec = &per_cpu(decrementers, cpu);
+
+       *dec = decrementer_clockevent;
+       dec->cpumask = cpumask_of_cpu(cpu);
+
+       printk(KERN_ERR "clockevent: %s mult[%lx] shift[%d] cpu[%d]\n",
+              dec->name, dec->mult, dec->shift, cpu);
+
+       clockevents_register_device(dec);
+}
+
+void init_decrementer_clockevent(void)
+{
+       int cpu = smp_processor_id();
+
+       decrementer_clockevent.mult = div_sc(ppc_tb_freq, NSEC_PER_SEC,
+                                            decrementer_clockevent.shift);
+       decrementer_clockevent.max_delta_ns =
+               clockevent_delta2ns(DECREMENTER_MAX, &decrementer_clockevent);
+       decrementer_clockevent.min_delta_ns = 1000;
+
+       register_decrementer_clockevent(cpu);
+}
+
+void secondary_cpu_time_init(void)
+{
+       /* FIME: Should make unrelatred change to move snapshot_timebase
+        * call here ! */
+       register_decrementer_clockevent(smp_processor_id());
+}
+
 /* This function is only called on the boot processor */
 void __init time_init(void)
 {
        unsigned long flags;
-       unsigned long tm = 0;
        struct div_result res;
        u64 scale, x;
        unsigned shift;
 
-        if (ppc_md.time_init != NULL)
-                timezone_offset = ppc_md.time_init();
-
        if (__USE_RTC()) {
                /* 601 processor: dec counts down by 128 every 128ns */
                ppc_tb_freq = 1000000000;
@@ -960,8 +934,8 @@ void __init time_init(void)
        }
        tb_to_ns_scale = scale;
        tb_to_ns_shift = shift;
-
-       tm = get_boot_time();
+       /* Save the current timebase to pretty up CONFIG_PRINTK_TIME */
+       boot_tb = get_tb_or_rtc();
 
        write_seqlock_irqsave(&xtime_lock, flags);
 
@@ -969,11 +943,8 @@ void __init time_init(void)
         if (timezone_offset) {
                sys_tz.tz_minuteswest = -timezone_offset / 60;
                sys_tz.tz_dsttime = 0;
-               tm -= timezone_offset;
         }
 
-       xtime.tv_sec = tm;
-       xtime.tv_nsec = 0;
        do_gtod.varp = &do_gtod.vars[0];
        do_gtod.var_idx = 0;
        do_gtod.varp->tb_orig_stamp = tb_last_jiffy;
@@ -991,13 +962,13 @@ void __init time_init(void)
 
        time_freq = 0;
 
-       last_rtc_update = xtime.tv_sec;
-       set_normalized_timespec(&wall_to_monotonic,
-                               -xtime.tv_sec, -xtime.tv_nsec);
        write_sequnlock_irqrestore(&xtime_lock, flags);
 
-       /* Not exact, but the timer interrupt takes care of this */
-       set_dec(tb_ticks_per_jiffy);
+       /* Register the clocksource, if we're not running on iSeries */
+       if (!firmware_has_feature(FW_FEATURE_ISERIES))
+               clocksource_init();
+
+       init_decrementer_clockevent();
 }