[PATCH] per-cpu timer changes

[PATCH] per-cpu timer changes

[Don Fry]

Here are some patches to help xen run on a system where the CPUs are not
synchronized so that the tsc counters drift.  Without this patch I get
lots of 'time went backwards' messages in the linux logs.  I still get
an occasional message, but not the number I did previously.

It has been tested on an x86 box with 4 hyper threaded cpus (8 logical),
but has not been tested with x86-64 or any other processor.  The patches
are against today's xeno-unstable.bk bits.

I am still trying to understand why I get any 'time went backwards'
messages, but thought I would post this now to get feedback as it is.


--- xeno-unstable.bk/xen/include/public/xen.h.orig	2005-05-23 17:24:21.000000000 -0700
+++ xeno-unstable.bk/xen/include/public/xen.h	2005-05-23 16:37:22.000000000 -0700
@@ -338,6 +338,21 @@ typedef struct
 } PACKED vcpu_info_t;                   /* 8 + arch */
 
 /*
+ * Xen/kernel shared data
+ *  per cpu timing information.
+ */
+typedef struct time_info_st
+{
+    u32			time_version1;
+    u32			time_version2;
+    tsc_timestamp_t	tsc_timestamp;	/* TSC at last update */
+    u64			system_time;	/* time, in nanoseconds, since boot */
+    u64			cpu_freq;	/* CPU frequency (Hz) */
+    u32			wc_sec;		/* Secs  00:00:00 UTC,  Jan 1, 1970. */
+    u32			wc_usec;	/* Usecs 00:00:00 UTC,  Jan 1, 1970. */
+} PACKED time_info_t;
+
+/*
  * Xen/kernel shared data -- pointer provided in start_info.
  * NB. We expect that this struct is smaller than a page.
  */
@@ -412,6 +427,7 @@ typedef struct shared_info_st
 
     arch_shared_info_t arch;
 
+    time_info_t vcpu_time[MAX_VIRT_CPUS];
 } PACKED shared_info_t;
 
 /*

--- xeno-unstable.bk/xen/common/schedule.c.orig	2005-05-23 17:25:39.000000000 -0700
+++ xeno-unstable.bk/xen/common/schedule.c	2005-05-23 16:42:24.000000000 -0700
@@ -57,6 +57,7 @@ string_param("sched", opt_sched);
 /* Various timer handlers. */
 static void s_timer_fn(unsigned long unused);
 static void t_timer_fn(unsigned long unused);
+static void tsc_timer_fn(unsigned long unused);
 static void dom_timer_fn(unsigned long data);
 
 /* This is global for now so that private implementations can reach it */
@@ -80,6 +81,7 @@ static struct scheduler ops;
 
 /* Per-CPU periodic timer sends an event to the currently-executing domain. */
 static struct ac_timer t_timer[NR_CPUS]; 
+static struct ac_timer tsc_timer[NR_CPUS];
 
 void free_domain_struct(struct domain *d)
 {
@@ -514,6 +516,7 @@ int idle_cpu(int cpu)
  * Timers: the scheduler utilises a number of timers
  * - s_timer: per CPU timer for preemption and scheduling decisions
  * - t_timer: per CPU periodic timer to send timer interrupt to current dom
+ * - tsc_timer: per CPU periodic timer to update time bases
  * - dom_timer: per domain timer to specifiy timeout values
  ****************************************************************************/
 
@@ -548,6 +551,18 @@ static void t_timer_fn(unsigned long unu
     add_ac_timer(&t_timer[cpu]);
 }
 
+/* Periodic tick timer: update time bases for per-cpu timing. */
+static void tsc_timer_fn(unsigned long unused)
+{
+    unsigned int        cpu = current->processor;
+
+    extern void percpu_ticks(void);
+    percpu_ticks();
+
+    tsc_timer[cpu].expires = NOW() + MILLISECS(500);
+    add_ac_timer(&tsc_timer[cpu]);
+}
+
 /* Domain timer function, sends a virtual timer interrupt to domain */
 static void dom_timer_fn(unsigned long data)
 {
@@ -578,6 +593,11 @@ void __init scheduler_init(void)
         t_timer[i].cpu      = i;
         t_timer[i].data     = 3;
         t_timer[i].function = &t_timer_fn;
+
+        init_ac_timer(&tsc_timer[i]);
+        tsc_timer[i].cpu      = i;
+        tsc_timer[i].data     = 4;
+        tsc_timer[i].function = &tsc_timer_fn;
     }
 
     schedule_data[0].idle = &idle0_exec_domain;
@@ -609,6 +629,9 @@ void schedulers_start(void) 
 
     t_timer_fn(0);
     smp_call_function((void *)t_timer_fn, NULL, 1, 1);
+
+    tsc_timer_fn(0);
+    smp_call_function((void *)tsc_timer_fn, NULL, 1, 1);
 }
 
 
--- xeno-unstable.bk/xen/arch/x86/smpboot.c.orig	2005-05-23 15:33:50.000000000 -0700
+++ xeno-unstable.bk/xen/arch/x86/smpboot.c	2005-05-23 16:41:56.000000000 -0700
@@ -400,6 +400,7 @@ void __init start_secondary(void)
 
     extern void percpu_traps_init(void);
     extern void cpu_init(void);
+    extern void setup_percpu_time(void);
 
     set_current(idle_task[cpu]);
     set_processor_id(cpu);
@@ -421,6 +422,8 @@ void __init start_secondary(void)
     construct_percpu_idt(cpu);
 #endif
 
+    setup_percpu_time();
+
     local_flush_tlb();
 
     startup_cpu_idle_loop();

--- xeno-unstable.bk/xen/arch/x86/time.c.orig	2005-05-23 17:25:12.000000000 -0700
+++ xeno-unstable.bk/xen/arch/x86/time.c	2005-05-23 16:42:35.000000000 -0700
@@ -50,6 +50,29 @@ static u64             full_tsc_irq;    
 static s_time_t        stime_irq;       /* System time at last 'time update' */
 static unsigned long   wc_sec, wc_usec; /* UTC time at last 'time update'.   */
 static rwlock_t        time_lock = RW_LOCK_UNLOCKED;
+static time_info_t     percpu_time_info[NR_CPUS];
+
+void percpu_ticks(void)
+{
+    int cpu = smp_processor_id();
+    time_info_t *t = &percpu_time_info[cpu];
+    u64 tsc, delta;
+    u64 quarter = t->cpu_freq >> 2;
+
+    rdtscll(tsc);
+    delta = tsc - t->tsc_timestamp;
+    while (delta >= quarter) {
+	t->wc_usec += 1000000UL / 4;
+	t->system_time += 1000000000ULL / 4;
+	t->tsc_timestamp += quarter;
+	delta -= quarter;
+    }
+
+    while (t->wc_usec > 1000000UL) {
+	t->wc_sec += 1;
+	t->wc_usec -= 10000000UL;
+    }
+}
 
 void timer_interrupt(int irq, void *dev_id, struct cpu_user_regs *regs)
 {
@@ -278,20 +301,29 @@ static inline void __update_dom_time(str
 {
     struct domain *d  = ed->domain;
     shared_info_t *si = d->shared_info;
+    time_info_t *dom = &si->vcpu_time[ed->processor];
+    time_info_t *xen = &percpu_time_info[smp_processor_id()];
 
     spin_lock(&d->time_lock);
 
     si->time_version1++;
+    dom->time_version1++;
     wmb();
 
     si->cpu_freq       = cpu_freq;
+    dom->cpu_freq      = xen->cpu_freq;
     si->tsc_timestamp  = full_tsc_irq;
+    dom->tsc_timestamp = xen->tsc_timestamp;
     si->system_time    = stime_irq;
+    dom->system_time   = xen->system_time;
     si->wc_sec         = wc_sec;
+    dom->wc_sec        = xen->wc_sec;
     si->wc_usec        = wc_usec;
+    dom->wc_usec       = xen->wc_usec;
 
     wmb();
     si->time_version2++;
+    dom->time_version2++;
 
     spin_unlock(&d->time_lock);
 }
@@ -299,8 +331,11 @@ static inline void __update_dom_time(str
 void update_dom_time(struct exec_domain *ed)
 {
     unsigned long flags;
+    int cpu = smp_processor_id();
 
-    if ( ed->domain->shared_info->tsc_timestamp != full_tsc_irq )
+    if ( ed->domain->shared_info->tsc_timestamp != full_tsc_irq 
+    ||   ed->domain->shared_info->vcpu_time[ed->processor].tsc_timestamp !=
+	percpu_time_info[cpu].tsc_timestamp)
     {
         read_lock_irqsave(&time_lock, flags);
         __update_dom_time(ed);
@@ -313,6 +348,7 @@ void do_settime(unsigned long secs, unsi
 {
     s64 delta;
     long _usecs = (long)usecs;
+    int i;
 
     write_lock_irq(&time_lock);
 
@@ -327,6 +363,10 @@ void do_settime(unsigned long secs, unsi
 
     wc_sec  = secs;
     wc_usec = _usecs;
+    for (i=0; i<NR_CPUS; i++) {
+	percpu_time_info[i].wc_sec = wc_sec;
+	percpu_time_info[i].wc_usec = wc_usec;
+    }
 
     /* Others will pick up the change at the next tick. */
     __update_dom_time(current);
@@ -336,16 +376,39 @@ void do_settime(unsigned long secs, unsi
 }
 
 
+spinlock_t tsc_lock = SPIN_LOCK_UNLOCKED;
+
+/*
+ * Time setup for this processor.
+ */
+void __init setup_percpu_time(void)
+{
+    unsigned long flags;
+    unsigned long ticks_per_frac;
+    int cpu = smp_processor_id();
+
+    /* only have 1 cpu calibrate at a time */
+    spin_lock_irqsave(&tsc_lock, flags);
+    ticks_per_frac = calibrate_tsc();
+    spin_unlock_irqrestore(&tsc_lock, flags);
+
+    if (!ticks_per_frac)
+	panic("Error calibrating TSC\n");
+    percpu_time_info[cpu].cpu_freq = (u64)ticks_per_frac * (u64)CALIBRATE_FRAC;
+    rdtscll(percpu_time_info[cpu].tsc_timestamp);
+    percpu_time_info[cpu].system_time = stime_irq;
+}
+
 /* Late init function (after all CPUs are booted). */
 int __init init_xen_time()
 {
     u64      scale;
     unsigned int cpu_ghz;
+    int i;
 
     cpu_ghz = (unsigned int)(cpu_freq / 1000000000ULL);
     for ( rdtsc_bitshift = 0; cpu_ghz != 0; rdtsc_bitshift++, cpu_ghz >>= 1 )
         continue;
-
     scale  = 1000000000LL << (32 + rdtsc_bitshift);
     scale /= cpu_freq;
     st_scale_f = scale & 0xffffffff;
@@ -358,6 +421,12 @@ int __init init_xen_time()
 
     /* Wallclock time starts as the initial RTC time. */
     wc_sec  = get_cmos_time();
+    for (i=0; i<NR_CPUS; i++) {
+	percpu_time_info[i].wc_sec = wc_sec;
+	percpu_time_info[i].wc_usec = 0;
+	percpu_time_info[i].system_time = stime_irq;
+	percpu_time_info[i].cpu_freq = cpu_freq;	// default speed
+    }
 
     printk("Time init:\n");
     printk(".... cpu_freq:    %08X:%08X\n", (u32)(cpu_freq>>32),(u32)cpu_freq);

--- xeno-unstable.bk/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/time.c.orig	2005-05-23 17:28:47.000000000 -0700
+++ xeno-unstable.bk/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/time.c	2005-05-23 17:06:18.000000000 -0700
@@ -105,9 +105,13 @@ struct timer_opts *cur_timer = &timer_ts
 
 /* These are peridically updated in shared_info, and then copied here. */
 u32 shadow_tsc_stamp;
+DEFINE_PER_CPU(u32, shadow_tsc_stamp);
 u64 shadow_system_time;
+DEFINE_PER_CPU(u64, shadow_system_time);
 static u32 shadow_time_version;
+DEFINE_PER_CPU(u32, shadow_time_version);
 static struct timeval shadow_tv;
+static DEFINE_PER_CPU(struct timeval, shadow_tv);
 
 /*
  * We use this to ensure that gettimeofday() is monotonically increasing. We
@@ -171,23 +175,29 @@ __setup("independent_wallclock", __indep
 static void __get_time_values_from_xen(void)
 {
 	shared_info_t *s = HYPERVISOR_shared_info;
+	int cpu = smp_processor_id();
 
 	do {
 		shadow_time_version = s->time_version2;
+		per_cpu(shadow_time_version, cpu) = s->vcpu_time[cpu].time_version2;
 		rmb();
 		shadow_tv.tv_sec    = s->wc_sec;
 		shadow_tv.tv_usec   = s->wc_usec;
 		shadow_tsc_stamp    = (u32)s->tsc_timestamp;
 		shadow_system_time  = s->system_time;
+		per_cpu(shadow_tv.tv_sec, cpu)	= s->vcpu_time[cpu].wc_sec;
+		per_cpu(shadow_tv.tv_usec, cpu)	= s->vcpu_time[cpu].wc_usec;
+		per_cpu(shadow_tsc_stamp, cpu)	= (u32)s->vcpu_time[cpu].tsc_timestamp;
+		per_cpu(shadow_system_time, cpu) = s->vcpu_time[cpu].system_time;
 		rmb();
 	}
-	while (shadow_time_version != s->time_version1);
+	while (shadow_time_version != s->time_version1 ||  per_cpu(shadow_time_version, cpu) != s->vcpu_time[cpu].time_version1);
 
 	cur_timer->mark_offset();
 }
 
 #define TIME_VALUES_UP_TO_DATE \
- ({ rmb(); (shadow_time_version == HYPERVISOR_shared_info->time_version2); })
+ ({ rmb(); ((per_cpu(shadow_time_version, cpu) == HYPERVISOR_shared_info->vcpu_time[cpu].time_version2) && (shadow_time_version == HYPERVISOR_shared_info->time_version2)); })
 
 /*
  * This version of gettimeofday has microsecond resolution
@@ -200,6 +210,7 @@ void do_gettimeofday(struct timeval *tv)
 	unsigned long max_ntp_tick;
 	unsigned long flags;
 	s64 nsec;
+	int cpu = smp_processor_id();
 
 	do {
 		unsigned long lost;
@@ -227,7 +238,7 @@ void do_gettimeofday(struct timeval *tv)
 		sec = xtime.tv_sec;
 		usec += (xtime.tv_nsec / NSEC_PER_USEC);
 
-		nsec = shadow_system_time - processed_system_time;
+		nsec = per_cpu(shadow_system_time, cpu) - per_cpu(processed_system_time, cpu);
 		__normalize_time(&sec, &nsec);
 		usec += (long)nsec / NSEC_PER_USEC;
 
@@ -273,6 +284,7 @@ int do_settimeofday(struct timespec *tv)
 	long wtm_nsec;
 	s64 nsec;
 	struct timespec xentime;
+	int cpu = smp_processor_id();
 
 	if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
 		return -EINVAL;
@@ -306,7 +318,7 @@ int do_settimeofday(struct timespec *tv)
 	 */
 	nsec -= (jiffies - wall_jiffies) * TICK_NSEC;
 
-	nsec -= (shadow_system_time - processed_system_time);
+	nsec -= (per_cpu(shadow_system_time, cpu) - per_cpu(processed_system_time, cpu));
 
 	__normalize_time(&sec, &nsec);
 	wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
@@ -392,27 +404,25 @@ static inline void do_timer_interrupt(in
 					struct pt_regs *regs)
 {
 	time_t wtm_sec, sec;
-	s64 delta, delta_cpu, nsec;
+	s64 delta_cpu, nsec;
 	long sec_diff, wtm_nsec;
 	int cpu = smp_processor_id();
 
 	do {
 		__get_time_values_from_xen();
 
-		delta = delta_cpu = (s64)shadow_system_time +
-			((s64)cur_timer->get_offset() * (s64)NSEC_PER_USEC);
-		delta     -= processed_system_time;
-		delta_cpu -= per_cpu(processed_system_time, cpu);
+		delta_cpu = (s64)per_cpu(shadow_system_time, cpu) +
+			((s64)cur_timer->get_offset() * (s64)NSEC_PER_USEC)
+			- per_cpu(processed_system_time, cpu);
 	}
 	while (!TIME_VALUES_UP_TO_DATE);
 
-	if (unlikely(delta < 0) || unlikely(delta_cpu < 0)) {
+	if (unlikely(delta_cpu < 0)) {
 		printk("Timer ISR/%d: Time went backwards: "
-		       "delta=%lld cpu_delta=%lld shadow=%lld "
-		       "off=%lld processed=%lld cpu_processed=%lld\n",
-		       cpu, delta, delta_cpu, shadow_system_time,
+		       "cpu_delta=%lld cpu_shadow=%lld "
+		       "off=%lld cpu_processed=%lld\n",
+		       cpu, delta_cpu, per_cpu(shadow_system_time, cpu),
 		       ((s64)cur_timer->get_offset() * (s64)NSEC_PER_USEC), 
-		       processed_system_time,
 		       per_cpu(processed_system_time, cpu));
 		for (cpu = 0; cpu < num_online_cpus(); cpu++)
 			printk(" %d: %lld\n", cpu,
@@ -420,19 +430,15 @@ static inline void do_timer_interrupt(in
 		return;
 	}
 
-	/* System-wide jiffy work. */
-	while (delta >= NS_PER_TICK) {
-		delta -= NS_PER_TICK;
-		processed_system_time += NS_PER_TICK;
-		do_timer(regs);
-	}
-
 	/* Local CPU jiffy work. */
 	while (delta_cpu >= NS_PER_TICK) {
 		delta_cpu -= NS_PER_TICK;
 		per_cpu(processed_system_time, cpu) += NS_PER_TICK;
 		update_process_times(user_mode(regs));
 		profile_tick(CPU_PROFILING, regs);
+		/* System-wide jiffy work. */
+		if (cpu == 0)
+			do_timer(regs);
 	}
 
 	if (cpu != 0)
@@ -447,19 +453,19 @@ static inline void do_timer_interrupt(in
 	    ((time_status & STA_UNSYNC) != 0) &&
 	    (xtime.tv_sec > (last_update_from_xen + 60))) {
 		/* Adjust shadow for jiffies that haven't updated xtime yet. */
-		shadow_tv.tv_usec -= 
+		per_cpu(shadow_tv.tv_usec, cpu) -= 
 			(jiffies - wall_jiffies) * (USEC_PER_SEC / HZ);
-		HANDLE_USEC_UNDERFLOW(shadow_tv);
+		HANDLE_USEC_UNDERFLOW(per_cpu(shadow_tv, cpu));
 
 		/*
 		 * Reset our running time counts if they are invalidated by
 		 * a warp backwards of more than 500ms.
 		 */
-		sec_diff = xtime.tv_sec - shadow_tv.tv_sec;
+		sec_diff = xtime.tv_sec - per_cpu(shadow_tv.tv_sec, cpu);
 		if (unlikely(abs(sec_diff) > 1) ||
 		    unlikely(((sec_diff * USEC_PER_SEC) +
 			      (xtime.tv_nsec / NSEC_PER_USEC) -
-			      shadow_tv.tv_usec) > 500000)) {
+			      per_cpu(shadow_tv.tv_usec, cpu)) > 500000)) {
 #ifdef CONFIG_XEN_PRIVILEGED_GUEST
 			last_rtc_update = last_update_to_xen = 0;
 #endif
@@ -467,8 +473,8 @@ static inline void do_timer_interrupt(in
 		}
 
 		/* Update our unsynchronised xtime appropriately. */
-		sec = shadow_tv.tv_sec;
-		nsec = shadow_tv.tv_usec * NSEC_PER_USEC;
+		sec = per_cpu(shadow_tv.tv_sec, cpu);
+		nsec = per_cpu(shadow_tv.tv_usec, cpu) * NSEC_PER_USEC;
 
 		__normalize_time(&sec, &nsec);
 		wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
@@ -498,7 +504,7 @@ static inline void do_timer_interrupt(in
 		op.cmd = DOM0_SETTIME;
 		op.u.settime.secs        = tv.tv_sec;
 		op.u.settime.usecs       = tv.tv_usec;
-		op.u.settime.system_time = shadow_system_time;
+		op.u.settime.system_time = per_cpu(shadow_system_time, cpu);
 		HYPERVISOR_dom0_op(&op);
 
 		last_update_to_xen = xtime.tv_sec;
@@ -670,7 +676,7 @@ void __init time_init(void)
 	set_normalized_timespec(&wall_to_monotonic,
 		-xtime.tv_sec, -xtime.tv_nsec);
 	processed_system_time = shadow_system_time;
-	per_cpu(processed_system_time, 0) = processed_system_time;
+	per_cpu(processed_system_time, 0) = per_cpu(shadow_system_time, 0);
 
 	if (timer_tsc_init.init(NULL) != 0)
 		BUG();
@@ -759,7 +765,7 @@ void time_resume(void)
 
 	/* Reset our own concept of passage of system time. */
 	processed_system_time = shadow_system_time;
-	per_cpu(processed_system_time, 0) = processed_system_time;
+	per_cpu(processed_system_time, 0) = per_cpu(shadow_system_time, 0);
 
 	/* Accept a warp in UTC (wall-clock) time. */
 	last_seen_tv.tv_sec = 0;
@@ -776,7 +782,7 @@ void local_setup_timer(void)
 
 	do {
 		seq = read_seqbegin(&xtime_lock);
-		per_cpu(processed_system_time, cpu) = shadow_system_time;
+		per_cpu(processed_system_time, cpu) = per_cpu(shadow_system_time, cpu);
 	} while (read_seqretry(&xtime_lock, seq));
 
 	per_cpu(timer_irq, cpu) = bind_virq_to_irq(VIRQ_TIMER);

--- xeno-unstable.bk/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/timers/timer_tsc.c.orig	2005-05-23 17:29:10.000000000 -0700
+++ xeno-unstable.bk/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/timers/timer_tsc.c	2005-05-23 17:04:11.000000000 -0700
@@ -10,6 +10,7 @@
 #include <linux/cpufreq.h>
 #include <linux/string.h>
 #include <linux/jiffies.h>
+#include <linux/percpu.h>
 
 #include <asm/timer.h>
 #include <asm/io.h>
@@ -35,8 +36,8 @@ extern spinlock_t i8253_lock;
 
 static int use_tsc;
 
-static unsigned long long monotonic_base;
-static u32 monotonic_offset;
+static DEFINE_PER_CPU(unsigned long long, monotonic_base);
+static DEFINE_PER_CPU(u32, monotonic_offset);
 static seqlock_t monotonic_lock = SEQLOCK_UNLOCKED;
 
 /* convert from cycles(64bits) => nanoseconds (64bits)
@@ -74,8 +75,8 @@ static inline unsigned long long cycles_
  */
 static unsigned long fast_gettimeoffset_quotient;
 
-extern u32 shadow_tsc_stamp;
-extern u64 shadow_system_time;
+extern DEFINE_PER_CPU(u32, shadow_tsc_stamp);
+extern DEFINE_PER_CPU(u64, shadow_system_time);
 
 static unsigned long get_offset_tsc(void)
 {
@@ -86,7 +87,7 @@ static unsigned long get_offset_tsc(void
 	rdtsc(eax,edx);
 
 	/* .. relative to previous jiffy (32 bits is enough) */
-	eax -= shadow_tsc_stamp;
+	eax -= per_cpu(shadow_tsc_stamp, smp_processor_id());
 
 	/*
          * Time offset = (tsc_low delta) * fast_gettimeoffset_quotient
@@ -110,12 +111,13 @@ static unsigned long long monotonic_cloc
 {
 	unsigned long long last_offset, this_offset, base;
 	unsigned seq;
+	int cpu = smp_processor_id();
 	
 	/* atomically read monotonic base & last_offset */
 	do {
 		seq = read_seqbegin(&monotonic_lock);
-		last_offset = monotonic_offset;
-		base = monotonic_base;
+		last_offset = per_cpu(monotonic_offset, cpu);
+		base = per_cpu(monotonic_base, cpu);
 	} while (read_seqretry(&monotonic_lock, seq));
 
 	/* Read the Time Stamp Counter */
@@ -152,11 +154,12 @@ unsigned long long sched_clock(void)
 
 static void mark_offset_tsc(void)
 {
+	int cpu = smp_processor_id();
 
 	/* update the monotonic base value */
 	write_seqlock(&monotonic_lock);
-	monotonic_base = shadow_system_time;
-	monotonic_offset = shadow_tsc_stamp;
+	per_cpu(monotonic_base, cpu) = per_cpu(shadow_system_time, cpu);
+	per_cpu(monotonic_offset, cpu) = per_cpu(shadow_tsc_stamp, cpu);
 	write_sequnlock(&monotonic_lock);
 }
 
-- 
Don Fry
brazilnut@us.ibm.com

RE: [PATCH] per-cpu timer changes

[Ian Pratt]

> Here are some patches to help xen run on a system where the 
> CPUs are not synchronized so that the tsc counters drift.  
> Without this patch I get lots of 'time went backwards' 
> messages in the linux logs.  I still get an occasional 
> message, but not the number I did previously.

Don,

This is looking good. To help other people review the patch, it might be
a good idea to post some of the design discussion we had off list as I
think the approach will be new to most people. (Perhaps put some of the
text in a comment in the hypervisor interface).

As regards the time going backwards messages, if you're seeing small -ve
deltas, I'm not surprised -- you need to round to some precision as we
won't be nanosecond accurate. Experience suggests we'll be good for a
few 10's of ns with any kind of decent crystal. We could round to e.g.
512ns or 1024ns to make sure.

Best,
Ian

Re: [PATCH] per-cpu timer changes

[Don Fry]

On Tue, May 24, 2005 at 02:20:36AM +0100, Ian Pratt wrote:
> 
> Don,
> 
> This is looking good. To help other people review the patch, it might be
> a good idea to post some of the design discussion we had off list as I
> think the approach will be new to most people. (Perhaps put some of the
> text in a comment in the hypervisor interface).
> 
> As regards the time going backwards messages, if you're seeing small -ve
> deltas, I'm not surprised -- you need to round to some precision as we
> won't be nanosecond accurate. Experience suggests we'll be good for a
> few 10's of ns with any kind of decent crystal. We could round to e.g.
> 512ns or 1024ns to make sure.
> 
> Best,
> Ian
> 

I am including the email that we exchanged off-list.  I started to edit
it, but decided that something I thought wasn't important, others would
find vital, so I include all the email.

The time going backwards was only occasionally, and it was a BIG jump
backwards.  I tracked it down yesterday to a problem with doing 32-bit
arithmetic in Linux on the tsc values.  For some reason, every 5-20
minutes xen seems to pause for about 5 seconds.  This causes the tsc to
wrap if only 32-bits are used, and the 'time went backwards' message is
printed.  I changed to use 64-bit tsc deltas and have been running since
yesterday afternoon without any 'time went backwards' messages.  I want
to do some more cleanup (remove my debugging code) and will post all my
changes to the list this afternoon.


----- Forwarded by Don Fry/Beaverton/IBM on 05/26/2005 09:29 AM -----

Bruce Jones/Beaverton/IBM wrote on 04/21/2005 09:07:26 AM:

> John, can you provide some additional technical guidance here?
>
> Ian, Keir: John is the implementor of our Linux changes for Summit
> and understands these issues better than anyone.
>
> I've added Don to the cc: list but he's on vacation this week and
> not reading email.
>
>  -- brucej
>
> Ian Pratt <Ian.Pratt@cl.cam.ac.uk> wrote on 04/20/2005 05:42:47 PM:
>
> > > "Ian Pratt" <m+Ian.Pratt@cl.cam.ac.uk> wrote on 04/20/2005 04:47:44 PM:
> > > > Please could Don write a paragraph explaining why cyclone timer support
> > > > is useful. Do summit systems have different frequency CPUs in the same
> > > > box?
> > > Bruce writes:
> > > I can write that paragraph myself.  IBM's high end xSeries systems are
> > > NUMA systems, each node is a separate machine with it's own front side
> > > bus, I/O buses, etc...  The chipset provides a cache-coherent interconnect
> > > to allow them to be cabled together into one big system.
> >
> > OK, so even the FSB clocks come from different crystals.
>
> Yes, and the hardware intentionally skews their frequencies, for reasons
> only the chipset designers understand. :)
>
> > > We had a boatload of problems with Linux when we first shipped it, with
> > > time moving around forward and backward for applications.  The processors
> > > in the various nodes  run at different frequencies and the on-processor
> > > timers do not run in sync.  We needed to modify Linux to use a system-wide
> > > timer.  Our chipset (code-named Cyclone) provides one, for newer systems
> > > Intel has defined the HPET that we can use.  We need to make similar
> > > changes to Xen.
> >
> > This needs some agreement on the design.
> >
> > My gut feeling is that it should still be possible for guests to use
> > the TSC to calculate the time offset relative to the published
> > Xen system time record (which is updated every couple of
> > seconds). The TSC calibration should be good enough to mean that
> > the relative drift over the period between records is tiny (and
> > errors can't accumulate beyond the period).
>
> My gut feeling is that your gut feeling is wrong.  We can't ever
> use the TSC on these systems - even a tiny amount of relative drift
> causes problems.
>
> But I'm no expert.  John, this is your cue.  Please join in.
>
> > The 'TSC when time record created' and 'TSC frequency' will have
> > to be per VCPU and updated to reflect the real CPU that the VCPU
> > is running on.
>
> As long as these are virtual and not read using the readTSC instruction,
> we may be OK.
>
> >
> > Ian
> >
> >
> >

----- Forwarded by Don Fry/Beaverton/IBM on 05/26/2005 09:29 AM -----

"Ian Pratt" <m+Ian.Pratt@cl.cam.ac.uk> wrote on 04/21/2005 09:24:54 AM: 

> > Yes, and the hardware intentionally skews their frequencies,
> > for reasons only the chipset designers understand. :)
>
> It's to be sneaky as regards FCC EMC emissions regulations.
>
> Some systems even modulate the PCI bus frequency.
>
> > > My gut feeling is that it should still be possible for
> > guests to use
> > > the TSC to calculate the time offset relative to the published Xen
> > > system time record (which is updated every couple of
> > seconds). The TSC
> > > calibration should be good enough to mean that the relative
> > drift over
> > > the period between records is tiny (and errors can't
> > accumulate beyond
> > > the period).
> >
> > My gut feeling is that your gut feeling is wrong.  We can't
> > ever use the TSC on these systems - even a tiny amount of
> > relative drift causes problems.
>
> It depends on the crystal stability, the accuracy with which the
> calibration is done, and the frequency of publishing new absoloute time
> records.
>
> The latter can be made quite frequent if need be.
>
> I'd much prefer avoiding having to expose linux to the HPET/cyclone by
> hiding it Xen, and having the guest use TSC extrapolation from the the
> time record published by Xen.
> We'd just need to update the current interface to have per-CPU records
> (and TSC frequency calibration).
>
> > But I'm no expert.  John, this is your cue.  Please join in.
> >
> > > The 'TSC when time record created' and 'TSC frequency' will
> > have to be
> > > per VCPU and updated to reflect the real CPU that the VCPU
> > is running
> > > on.
> >
> > As long as these are virtual and not read using the readTSC
> > instruction, we may be OK.
>
> Using readTSC should be fine, since we're only using it to extrapolate
> from the last Xen supplied time record, and we've calibrated the
> frequency of the particular CPU we're running on. We only have to worry
> about rapid clock drift due to sudden temperature changes etc, but even
> then we can just update the calibration frequency periodically. Using
> this approach we get to keep gettimeofday very fast, and avoid
> complicating the hypervisor API -- it's exactly what we need for
> migrating a domain between physical servers with different frequency
> CPUs.
>
> Ian
>

----- Forwarded by Don Fry/Beaverton/IBM on 05/26/2005 09:29 AM -----

"Ian Pratt" <m+Ian.Pratt@cl.cam.ac.uk> wrote on 04/21/2005 01:12:51 PM: 

> > First, forgive my lack of knowledge about Xen. Since I don't
> > know the details of what you're proposing, let me make a
> > straw-man and let you correct my assumptions.
> >
> > Lets say you're proposing that time be calculated with the
> > following formula:
> >
> > timefoday = xen_time_base +  rdtsc() - xen_last_tsc[CPUNUM]
> >
> > Given a guest domain with two cpus, the issue is managing
> > xen_last_tsc[] and xen_time_base. For the equation to work,
> > xen_last_tsc[0] must hold the TSC value from CPU0 at exactly
> > the time stored in xen_time_base. Additionally the same is
> > true with xen_las_tsc[1].
>
> I'm proposing:
>
> timeofday  = round_to_precision( last_xen_time_base[VCPU] +
>              ( rdtsc() - last_xen_tsc[VCPU] ) * xen_tsc_calibrate[VCPU]
> )
>
> We update last_xen_time_base and last_xen_tsc on each CPU every second
> or so.
> xen_tsc_calibrate is calculated for each CPU at start of day. For
> completeness, we could recalculate the calibration every 30s or so to 
> cope with crystal temperature drift if we wanted ultimate precision.
>
> > The difficult question is how do you ensure that the two
> > values in xen_last_tsc[] are linked with the time in
> > xen_time_base? This requires reading the TSC on two cpus at
> > the exact same time. Additionally, this sync point must
> > happen frequently enough so that the continuing drift between
> > cpus isn't noticed.
>
> Nope, we would set the time_base on each CPU independently, but relative
> to the same timer.
> This could be the cyclone, HPET, or even the PIT if its possible to read
> the same PIT from any node (though I'm guessing you probably have a PIT
> per node and can't read the remote one).
>
> > Then you'll have to weigh that solution against just using an
> > alternate global timesource like HPET/Cyclone.
>
> I'd prefer to avoid this as it would mean that there'd be a different 
> hypervisor API for guests on cylcone/hpet systems vs. normal synchronous
> CPU systems.
> Using the TSC will probably give a lower cost gettimeofday, we can also
> trap it and emulate if we want to lie to guests about the progress of
> time.
>
> Best,
> Ian
>
>
>
>
>
>

----- Forwarded by Don Fry/Beaverton/IBM on 05/26/2005 09:29 AM -----

John Stultz/Beaverton/IBM wrote on 04/21/2005 01:49:54 PM:

> I'm just resending this with proper addresses as something got futzed up
in the CC list on that last mail.
>
> "Ian Pratt" <m+Ian.Pratt@cl.cam.ac.uk> wrote on 04/21/2005 01:12:51 PM:
>
> > > First, forgive my lack of knowledge about Xen. Since I don't
> > > know the details of what you're proposing, let me make a
> > > straw-man and let you correct my assumptions.
> > >
> > > Lets say you're proposing that time be calculated with the
> > > following formula:
> > >
> > > timefoday = xen_time_base +  rdtsc() - xen_last_tsc[CPUNUM]
> > >
> > > Given a guest domain with two cpus, the issue is managing
> > > xen_last_tsc[] and xen_time_base. For the equation to work,
> > > xen_last_tsc[0] must hold the TSC value from CPU0 at exactly
> > > the time stored in xen_time_base. Additionally the same is
> > > true with xen_las_tsc[1].
> >
> > I'm proposing:
> >
> > timeofday  = round_to_precision( last_xen_time_base[VCPU] +
> >              ( rdtsc() - last_xen_tsc[VCPU] ) * xen_tsc_calibrate[VCPU]
> > )
> >
> > We update last_xen_time_base and last_xen_tsc on each CPU every second
> > or so.
>
> Or possibly more frequently, as on a 4Ghz cpu the 32bit TSC will wrap 
each second. Alternatively you could use the full 64bits.
>
> > xen_tsc_calibrate is calculated for each CPU at start of day. For
> > completeness, we could recalculate the calibration every 30s or so to
> > cope with crystal temperature drift if we wanted ultimate precision.
> >
> > > The difficult question is how do you ensure that the two
> > > values in xen_last_tsc[] are linked with the time in
> > > xen_time_base? This requires reading the TSC on two cpus at
> > > the exact same time. Additionally, this sync point must
> > > happen frequently enough so that the continuing drift between
> > > cpus isn't noticed.
> >
> > Nope, we would set the time_base on each CPU independently, but
relative
> > to the same timer.

> Hmmm. That sounds like it could work. Just be sure that preempt won't 
bite you in the timeofday calculation. The bit about still using the
cyclone/HPET to sync the different xen_time_base[] values is the real key.
>
> > This could be the cyclone, HPET, or even the PIT if its possible to read
> > the same PIT from any node (though I'm guessing you probably have a PIT
> > per node and can't read the remote one).

> The ioport space is unified by the BIOS so there is one global PIT shared
by all nodes. Although as you'll need a continuous timesource that doesn't
loop between xen_time_base updates, the PIT would not work.
>
> thanks
> -john
----- Forwarded by Don Fry/Beaverton/IBM on 05/26/2005 09:29 AM -----

"Ian Pratt" <m+Ian.Pratt@cl.cam.ac.uk> wrote on 04/28/2005 07:08:05 PM:

>
> > First I apologize for not being involved in this email
> > exchange last week.
> > I am also just learning about Xen so my questions may be
> > obvious to others.
> >
> > What is the last_xen_time_base referred to in Ian's email? Is
> > this the stime_irq or wc_sec,wc_usec or something else?
>
> I was referring to the wc_ wall clock and system time values.
> We'll need to make these per VPU, or perhaps slightly more cleanly,
> store an offset in ns for each VCPU.
>
> > When would the last_xen_tsc[VCPU] values be captured by Xen?
> > Currently, the tsc for cpu 0 is obtained during
> > timer_interrupt as full_tsc_irq.
>
> These just need to be captured periodically on each real CPU -- every
> couple of seconds would do it, though more frequently woulnd't hurt.
>
> > When updating the domain's shared_info structure mapping the
> > physical CPU to the domain's view of the CPU would need to be
> > done. For example if domain2 was running on CPU3 and CPU2 and
> > the domain's view was cpu0 and cpu1, the saved tsc value for
> > CPU3 would be copied to last_xen_tsc[0] and CPU2 to
> > last_xen_tsc[1] before sending the interrupt to the domain.
>
> Yep, this shouldn't be hard -- there's already some code to spot when
> they need to be updated.
>
> > From the last algorithm from Ian, I don't see anything that
> > refers to the Cyclone/HPET value directly. Is that because
> > Xen is the only thing that reads the Cyclone/HPET counter and
> > the domain just uses the TSC?
>
> Yep, we don't want to expose the cyclone/hpet to guests. There's no
> need, and it would have implications for migrating VMs between different
> systems.
>
> Strictly speaking, Xen wouldn't even need support for the hpet/cyclone
> as it could just use the shared PIT, though I have no objection to
> adding such support.
>
> Are you happy with this design? It's a little more work, but I believe
> better in the long run. We need to get the hypervisor interface change
> incorporated ASAP.
>
> Cheers,
> Ian

----- Forwarded by Don Fry/Beaverton/IBM on 05/26/2005 09:29 AM -----

"Ian Pratt" <m+Ian.Pratt@cl.cam.ac.uk> wrote on 04/30/2005 12:04:57 AM:

> > It sounds like the per-cpu changes should be sufficient.
> >
> > Having a time base and ns deltas for each CPU sounds
> > interesting, but wouldn't you have do a subtraction to
> > generate the delta in Xen, and then add it back in, in the
> > domain? Just saving the per-cpu value would save the extra
> > add and subtract.
>
> Sure, but the add/subtract won't cost much, and it saves some space in
> the shared info page, which might be an issue if we have a lot of VCPUs.
>
> Not a big deal either way.
>
> > The bottom line is that it can all be done with the TSC,
> > without needing to use the Cyclone or HPET hardware, which
> > isn't available on all systems like the TSC.
>
> Great, we're in agreement. I think the first stage is just to do the per
> [V]CPU calibration and time vals. Could you work something up?
>
> Thanks,
> Ian


-- 
Don Fry
brazilnut@us.ibm.com

Re: [PATCH] per-cpu timer changes

[Keir Fraser]

On 26 May 2005, at 18:34, Don Fry wrote:

> The time going backwards was only occasionally, and it was a BIG jump
> backwards.  I tracked it down yesterday to a problem with doing 32-bit
> arithmetic in Linux on the tsc values.  For some reason, every 5-20
> minutes xen seems to pause for about 5 seconds.  This causes the tsc to
> wrap if only 32-bits are used, and the 'time went backwards' message is
> printed.  I changed to use 64-bit tsc deltas and have been running 
> since
> yesterday afternoon without any 'time went backwards' messages.  I want
> to do some more cleanup (remove my debugging code) and will post all my
> changes to the list this afternoon.

It would be good to know where those 5 seconds go. Doing 64-bit 
arithmetic on x86 will be slower and possibly fudge over the underlying 
problem.

  -- Keir

Re: [PATCH] per-cpu timer changes

[Nivedita Singhvi]

Don Fry wrote:

> The time going backwards was only occasionally, and it was a BIG jump
> backwards.  I tracked it down yesterday to a problem with doing 32-bit
> arithmetic in Linux on the tsc values.  For some reason, every 5-20
> minutes xen seems to pause for about 5 seconds.  This causes the tsc to
> wrap if only 32-bits are used, and the 'time went backwards' message is

What happens when there is activity on the system? Do you
still see the pause? 5 seconds is an inordinately long time,
and will impact networking if it happens during normal
workloads too..

thanks,
Nivedita

Re: [PATCH] per-cpu timer changes

[Andrew Thompson]

[-- Attachment #1: Type: text/plain, Size: 536 bytes --]

Keir Fraser wrote:
> On 26 May 2005, at 18:34, Don Fry wrote:
>> The time going backwards was only occasionally, and it was a BIG jump
>> backwards.  I tracked it down yesterday to a problem with doing 32-bit
>> arithmetic in Linux on the tsc values.  For some reason, every 5-20
>> minutes xen seems to pause for about 5 seconds.  
 >> <snip>

Have you noticed this directly in any dom's, or just in xen?

> It would be good to know where those 5 seconds go. 

See: UFOs, abductions, etc... ;)

-- 
Andrew Thompson
http://aktzero.com/

[-- Attachment #2: andrewkt.vcf --]
[-- Type: text/x-vcard, Size: 153 bytes --]

begin:vcard
fn:Andrew Thompson
n:Thompson;Andrew
email;internet:andrewkt@aktzero.com
x-mozilla-html:FALSE
url:http://aktzero.com/
version:2.1
end:vcard


[-- Attachment #3: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Re: [PATCH] per-cpu timer changes

[Don Fry]

I am adding some debugging to figure this out.  Right now, I have
nothing running on the system except dom0.  Xwindows is not running.
The system has 8 cpus, (4 HT) and I see this on dom0 when it is compiled
with and without SMP support.  The only thing running on dom0 is
tail -f /var/log/messages and that is very quiet.

For example, between 11:34:01 and 11:54:01 I had a 5 second pause occur
3 times.  Between 11:59:00 and 12:14:01 it occurred 4 times.

Since May 25 16:53:31 and now it has happened about 180 times.

On Thu, May 26, 2005 at 11:03:52AM -0700, Nivedita Singhvi wrote:
> Don Fry wrote:
> 
> >The time going backwards was only occasionally, and it was a BIG jump
> >backwards.  I tracked it down yesterday to a problem with doing 32-bit
> >arithmetic in Linux on the tsc values.  For some reason, every 5-20
> >minutes xen seems to pause for about 5 seconds.  This causes the tsc to
> >wrap if only 32-bits are used, and the 'time went backwards' message is
> 
> What happens when there is activity on the system? Do you
> still see the pause? 5 seconds is an inordinately long time,
> and will impact networking if it happens during normal
> workloads too..
> 
> thanks,
> Nivedita
-- 
Don Fry
brazilnut@us.ibm.com

[PATCH] per-cpu timer changes

[Don Fry]

These patches implement per-cpu relative time.  They have been tested on
32-bit x86 platforms but not even compiled on x86_64 or ia64.

They solve the 'time went backwards' problems on systems with
unsynchronized cpus.  Using last week's xeno-unstable I had no problems
with these patches.  Using today's bits, the system will sometimes panic
when booting dom0 between 'PCI: IRQ init' and 'Grant table initialized'.

I would like to get wider testing of these changes.

Signed-off-by:  Don Fry <brazilnut@us.ibm.com>


--- xeno-unstable.bk/xen/include/public/xen.h.orig	2005-06-07 09:50:08.000000000 -0700
+++ xeno-unstable.bk/xen/include/public/xen.h	2005-06-07 09:50:08.000000000 -0700
@@ -329,6 +329,21 @@ typedef struct
 } PACKED vcpu_info_t;                   /* 8 + arch */
 
 /*
+ * Xen/kernel shared data
+ *  per cpu timing information.
+ */
+typedef struct time_info_st
+{
+    u32			time_version1;
+    u32			time_version2;
+    tsc_timestamp_t	tsc_timestamp;	/* TSC at last update */
+    u64			system_time;	/* time, in nanoseconds, since boot */
+    u64			cpu_freq;	/* CPU frequency (Hz) */
+    u32			wc_sec;		/* Secs  00:00:00 UTC,  Jan 1, 1970. */
+    u32			wc_usec;	/* Usecs 00:00:00 UTC,  Jan 1, 1970. */
+} PACKED time_info_t;
+
+/*
  * Xen/kernel shared data -- pointer provided in start_info.
  * NB. We expect that this struct is smaller than a page.
  */
@@ -403,6 +418,7 @@ typedef struct shared_info_st
 
     arch_shared_info_t arch;
 
+    time_info_t vcpu_time[MAX_VIRT_CPUS];
 } PACKED shared_info_t;
 
 /*
--- xeno-unstable.bk/xen/common/schedule.c.orig	2005-06-07 10:40:50.000000000 -0700
+++ xeno-unstable.bk/xen/common/schedule.c	2005-06-07 11:23:57.000000000 -0700
@@ -53,6 +53,7 @@ string_param("sched", opt_sched);
 /* Various timer handlers. */
 static void s_timer_fn(void *unused);
 static void t_timer_fn(void *unused);
+static void tsc_timer_fn(void *unused);
 static void dom_timer_fn(void *data);
 
 /* This is global for now so that private implementations can reach it */
@@ -76,6 +77,7 @@ static struct scheduler ops;
 
 /* Per-CPU periodic timer sends an event to the currently-executing domain. */
 static struct ac_timer t_timer[NR_CPUS]; 
+static struct ac_timer tsc_timer[NR_CPUS]; 
 
 void free_domain_struct(struct domain *d)
 {
@@ -487,6 +489,7 @@ int idle_cpu(int cpu)
  * Timers: the scheduler utilises a number of timers
  * - s_timer: per CPU timer for preemption and scheduling decisions
  * - t_timer: per CPU periodic timer to send timer interrupt to current dom
+ * - tsc_timer: per CPU periodic timer to update time bases
  * - dom_timer: per domain timer to specifiy timeout values
  ****************************************************************************/
 
@@ -516,6 +519,17 @@ static void t_timer_fn(void *unused)
     set_ac_timer(&t_timer[cpu], NOW() + MILLISECS(10));
 }
 
+/* Periodic tick timer: update time bases for per-cpu timing. */
+static void tsc_timer_fn(void *unused)
+{
+    unsigned int        cpu = current->processor;
+
+    extern void percpu_ticks(void);
+    percpu_ticks();
+
+    set_ac_timer(&tsc_timer[cpu], NOW() + MILLISECS(250));
+}
+
 /* Domain timer function, sends a virtual timer interrupt to domain */
 static void dom_timer_fn(void *data)
 {
@@ -537,6 +551,7 @@ void __init scheduler_init(void)
         spin_lock_init(&schedule_data[i].schedule_lock);
         init_ac_timer(&schedule_data[i].s_timer, s_timer_fn, NULL, i);
         init_ac_timer(&t_timer[i], t_timer_fn, NULL, i);
+        init_ac_timer(&tsc_timer[i], tsc_timer_fn, NULL, i);
     }
 
     schedule_data[0].curr = idle_task[0];
@@ -566,6 +581,9 @@ void schedulers_start(void) 
 {   
     t_timer_fn(0);
     smp_call_function((void *)t_timer_fn, NULL, 1, 1);
+
+    tsc_timer_fn(0);
+    smp_call_function((void *)tsc_timer_fn, NULL, 1, 1);
 }
 
 void dump_runq(unsigned char key)


--- xeno-unstable.bk/xen/arch/x86/smpboot.c.orig	2005-06-07 10:41:14.000000000 -0700
+++ xeno-unstable.bk/xen/arch/x86/smpboot.c	2005-06-07 11:23:51.000000000 -0700
@@ -431,6 +431,7 @@ void __init start_secondary(void *unused
 
 	extern void percpu_traps_init(void);
 	extern void cpu_init(void);
+	extern void setup_percpu_time(void);
 
 	set_current(idle_task[cpu]);
 	set_processor_id(cpu);
@@ -452,6 +453,7 @@ void __init start_secondary(void *unused
 
 	setup_secondary_APIC_clock();
 	enable_APIC_timer();
+	setup_percpu_time();
 
 	/*
 	 * low-memory mappings have been cleared, flush them from


--- xeno-unstable.bk/xen/arch/x86/time.c.orig	2005-06-07 10:40:57.000000000 -0700
+++ xeno-unstable.bk/xen/arch/x86/time.c	2005-06-07 11:40:22.000000000 -0700
@@ -49,6 +49,29 @@ static u64             full_tsc_irq;    
 static s_time_t        stime_irq;       /* System time at last 'time update' */
 static unsigned long   wc_sec, wc_usec; /* UTC time at last 'time update'.   */
 static rwlock_t        time_lock = RW_LOCK_UNLOCKED;
+static time_info_t     percpu_time_info[NR_CPUS];
+
+void percpu_ticks(void)
+{
+    int cpu = smp_processor_id();
+    time_info_t *t = &percpu_time_info[cpu];
+    u64 tsc, delta;
+    u64 quarter = t->cpu_freq >> 2;
+
+    rdtscll(tsc);
+    delta = tsc - t->tsc_timestamp;
+    while (delta >= quarter) {
+	t->wc_usec += 1000000UL / 4;
+	t->system_time += 1000000000ULL / 4;
+	t->tsc_timestamp += quarter;
+	delta -= quarter;
+    }
+
+    while (t->wc_usec > 1000000UL) {
+	t->wc_sec += 1;
+	t->wc_usec -= 10000000UL;
+    }
+}
 
 void timer_interrupt(int irq, void *dev_id, struct cpu_user_regs *regs)
 {
@@ -277,20 +300,29 @@ static inline void __update_dom_time(str
 {
     struct domain *d  = v->domain;
     shared_info_t *si = d->shared_info;
+    time_info_t *dom = &si->vcpu_time[v->processor];
+    time_info_t *xen = &percpu_time_info[smp_processor_id()];
 
     spin_lock(&d->time_lock);
 
     si->time_version1++;
+    dom->time_version1++;
     wmb();
 
     si->cpu_freq       = cpu_freq;
+    dom->cpu_freq      = xen->cpu_freq;
     si->tsc_timestamp  = full_tsc_irq;
+    dom->tsc_timestamp = xen->tsc_timestamp;
     si->system_time    = stime_irq;
+    dom->system_time   = xen->system_time;
     si->wc_sec         = wc_sec;
+    dom->wc_sec        = xen->wc_sec;
     si->wc_usec        = wc_usec;
+    dom->wc_usec       = xen->wc_usec;
 
     wmb();
     si->time_version2++;
+    dom->time_version2++;
 
     spin_unlock(&d->time_lock);
 }
@@ -298,8 +330,11 @@ static inline void __update_dom_time(str
 void update_dom_time(struct vcpu *v)
 {
     unsigned long flags;
+    int cpu = smp_processor_id();
 
-    if ( v->domain->shared_info->tsc_timestamp != full_tsc_irq )
+    if ( v->domain->shared_info->tsc_timestamp != full_tsc_irq 
+    ||   v->domain->shared_info->vcpu_time[v->processor].tsc_timestamp !=
+	percpu_time_info[cpu].tsc_timestamp)
     {
         read_lock_irqsave(&time_lock, flags);
         __update_dom_time(v);
@@ -312,6 +347,7 @@ void do_settime(unsigned long secs, unsi
 {
     s64 delta;
     long _usecs = (long)usecs;
+    int i;
 
     write_lock_irq(&time_lock);
 
@@ -326,6 +362,10 @@ void do_settime(unsigned long secs, unsi
 
     wc_sec  = secs;
     wc_usec = _usecs;
+    for (i=0; i<NR_CPUS; i++) {
+	percpu_time_info[i].wc_sec = wc_sec;
+	percpu_time_info[i].wc_usec = wc_usec;
+    }
 
     /* Others will pick up the change at the next tick. */
     __update_dom_time(current);
@@ -335,16 +375,39 @@ void do_settime(unsigned long secs, unsi
 }
 
 
+spinlock_t tsc_lock = SPIN_LOCK_UNLOCKED;
+
+/*
+ * Time setup for this processor.
+ */
+void __init setup_percpu_time(void)
+{
+    unsigned long flags;
+    unsigned long ticks_per_frac;
+    int cpu = smp_processor_id();
+
+    /* only have 1 cpu calibrate at a time */
+    spin_lock_irqsave(&tsc_lock, flags);
+    ticks_per_frac = calibrate_tsc();
+    spin_unlock_irqrestore(&tsc_lock, flags);
+
+    if (!ticks_per_frac)
+	panic("Error calibrating TSC\n");
+    percpu_time_info[cpu].cpu_freq = (u64)ticks_per_frac * (u64)CALIBRATE_FRAC;
+    rdtscll(percpu_time_info[cpu].tsc_timestamp);
+    percpu_time_info[cpu].system_time = stime_irq;
+}
+
 /* Late init function (after all CPUs are booted). */
 int __init init_xen_time()
 {
     u64      scale;
     unsigned int cpu_ghz;
+    int i;
 
     cpu_ghz = (unsigned int)(cpu_freq / 1000000000ULL);
     for ( rdtsc_bitshift = 0; cpu_ghz != 0; rdtsc_bitshift++, cpu_ghz >>= 1 )
         continue;
-
     scale  = 1000000000LL << (32 + rdtsc_bitshift);
     scale /= cpu_freq;
     st_scale_f = scale & 0xffffffff;
@@ -357,6 +420,12 @@ int __init init_xen_time()
 
     /* Wallclock time starts as the initial RTC time. */
     wc_sec  = get_cmos_time();
+    for (i=0; i<NR_CPUS; i++) {
+	percpu_time_info[i].wc_sec = wc_sec;
+	percpu_time_info[i].wc_usec = 0;
+	percpu_time_info[i].system_time = stime_irq;
+	percpu_time_info[i].cpu_freq = cpu_freq;	// default speed
+    }
 
     printk("Time init:\n");
     printk(".... cpu_freq:    %08X:%08X\n", (u32)(cpu_freq>>32),(u32)cpu_freq);


--- xeno-unstable.bk/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/time.c.orig	2005-06-07 10:21:21.000000000 -0700
+++ xeno-unstable.bk/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/time.c	2005-06-07 13:06:01.000000000 -0700
@@ -105,9 +105,13 @@ struct timer_opts *cur_timer = &timer_ts
 
 /* These are peridically updated in shared_info, and then copied here. */
 u32 shadow_tsc_stamp;
+DEFINE_PER_CPU(u64, shadow_tsc_stamp);
 u64 shadow_system_time;
+DEFINE_PER_CPU(u64, shadow_system_time);
 static u32 shadow_time_version;
+DEFINE_PER_CPU(u32, shadow_time_version);
 static struct timeval shadow_tv;
+static DEFINE_PER_CPU(struct timeval, shadow_tv);
 
 /*
  * We use this to ensure that gettimeofday() is monotonically increasing. We
@@ -171,23 +175,29 @@ __setup("independent_wallclock", __indep
 static void __get_time_values_from_xen(void)
 {
 	shared_info_t *s = HYPERVISOR_shared_info;
+	int cpu = smp_processor_id();
 
 	do {
 		shadow_time_version = s->time_version2;
+		per_cpu(shadow_time_version, cpu) = s->vcpu_time[cpu].time_version2;
 		rmb();
 		shadow_tv.tv_sec    = s->wc_sec;
 		shadow_tv.tv_usec   = s->wc_usec;
 		shadow_tsc_stamp    = (u32)s->tsc_timestamp;
 		shadow_system_time  = s->system_time;
+		per_cpu(shadow_tv.tv_sec, cpu)	= s->vcpu_time[cpu].wc_sec;
+		per_cpu(shadow_tv.tv_usec, cpu)	= s->vcpu_time[cpu].wc_usec;
+		per_cpu(shadow_tsc_stamp, cpu)	= s->vcpu_time[cpu].tsc_timestamp;
+		per_cpu(shadow_system_time, cpu) = s->vcpu_time[cpu].system_time;
 		rmb();
 	}
-	while (shadow_time_version != s->time_version1);
+	while (shadow_time_version != s->time_version1 ||  per_cpu(shadow_time_version, cpu) != s->vcpu_time[cpu].time_version1);
 
 	cur_timer->mark_offset();
 }
 
 #define TIME_VALUES_UP_TO_DATE \
- ({ rmb(); (shadow_time_version == HYPERVISOR_shared_info->time_version2); })
+ ({ rmb(); ((per_cpu(shadow_time_version, cpu) == HYPERVISOR_shared_info->vcpu_time[cpu].time_version2) && (shadow_time_version == HYPERVISOR_shared_info->time_version2)); })
 
 /*
  * This version of gettimeofday has microsecond resolution
@@ -200,6 +210,7 @@ void do_gettimeofday(struct timeval *tv)
 	unsigned long max_ntp_tick;
 	unsigned long flags;
 	s64 nsec;
+	int cpu = smp_processor_id();
 
 	do {
 		unsigned long lost;
@@ -227,7 +238,7 @@ void do_gettimeofday(struct timeval *tv)
 		sec = xtime.tv_sec;
 		usec += (xtime.tv_nsec / NSEC_PER_USEC);
 
-		nsec = shadow_system_time - processed_system_time;
+		nsec = per_cpu(shadow_system_time, cpu) - per_cpu(processed_system_time, cpu);
 		__normalize_time(&sec, &nsec);
 		usec += (long)nsec / NSEC_PER_USEC;
 
@@ -273,6 +284,7 @@ int do_settimeofday(struct timespec *tv)
 	long wtm_nsec;
 	s64 nsec;
 	struct timespec xentime;
+	int cpu = smp_processor_id();
 
 	if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
 		return -EINVAL;
@@ -306,7 +318,7 @@ int do_settimeofday(struct timespec *tv)
 	 */
 	nsec -= (jiffies - wall_jiffies) * TICK_NSEC;
 
-	nsec -= (shadow_system_time - processed_system_time);
+	nsec -= (per_cpu(shadow_system_time, cpu) - per_cpu(processed_system_time, cpu));
 
 	__normalize_time(&sec, &nsec);
 	wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
@@ -384,6 +396,7 @@ unsigned long profile_pc(struct pt_regs 
 EXPORT_SYMBOL(profile_pc);
 #endif
 
+extern unsigned long long get_full_tsc_offset(void);
 /*
  * timer_interrupt() needs to keep up the real-time clock,
  * as well as call the "do_timer()" routine every clocktick
@@ -392,27 +405,25 @@ static inline void do_timer_interrupt(in
 					struct pt_regs *regs)
 {
 	time_t wtm_sec, sec;
-	s64 delta, delta_cpu, nsec;
+	s64 delta_cpu, nsec;
 	long sec_diff, wtm_nsec;
 	int cpu = smp_processor_id();
 
 	do {
 		__get_time_values_from_xen();
 
-		delta = delta_cpu = (s64)shadow_system_time +
-			((s64)cur_timer->get_offset() * (s64)NSEC_PER_USEC);
-		delta     -= processed_system_time;
-		delta_cpu -= per_cpu(processed_system_time, cpu);
+		delta_cpu = (s64)per_cpu(shadow_system_time, cpu) +
+			((s64)get_full_tsc_offset())
+			- per_cpu(processed_system_time, cpu);
 	}
 	while (!TIME_VALUES_UP_TO_DATE);
 
-	if (unlikely(delta < 0) || unlikely(delta_cpu < 0)) {
+	if (unlikely(delta_cpu < 0)) {
 		printk("Timer ISR/%d: Time went backwards: "
-		       "delta=%lld cpu_delta=%lld shadow=%lld "
-		       "off=%lld processed=%lld cpu_processed=%lld\n",
-		       cpu, delta, delta_cpu, shadow_system_time,
-		       ((s64)cur_timer->get_offset() * (s64)NSEC_PER_USEC), 
-		       processed_system_time,
+		       "cpu_delta=%lld cpu_shadow=%lld "
+		       "off=%lld cpu_processed=%lld\n",
+		       cpu, delta_cpu, per_cpu(shadow_system_time, cpu),
+		       (s64)get_full_tsc_offset(), 
 		       per_cpu(processed_system_time, cpu));
 		for (cpu = 0; cpu < num_online_cpus(); cpu++)
 			printk(" %d: %lld\n", cpu,
@@ -420,19 +431,15 @@ static inline void do_timer_interrupt(in
 		return;
 	}
 
-	/* System-wide jiffy work. */
-	while (delta >= NS_PER_TICK) {
-		delta -= NS_PER_TICK;
-		processed_system_time += NS_PER_TICK;
-		do_timer(regs);
-	}
-
 	/* Local CPU jiffy work. */
 	while (delta_cpu >= NS_PER_TICK) {
 		delta_cpu -= NS_PER_TICK;
 		per_cpu(processed_system_time, cpu) += NS_PER_TICK;
 		update_process_times(user_mode(regs));
 		profile_tick(CPU_PROFILING, regs);
+		/* System-wide jiffy work. */
+		if (cpu == 0)
+			do_timer(regs);
 	}
 
 	if (cpu != 0)
@@ -447,19 +454,19 @@ static inline void do_timer_interrupt(in
 	    ((time_status & STA_UNSYNC) != 0) &&
 	    (xtime.tv_sec > (last_update_from_xen + 60))) {
 		/* Adjust shadow for jiffies that haven't updated xtime yet. */
-		shadow_tv.tv_usec -= 
+		per_cpu(shadow_tv.tv_usec, cpu) -= 
 			(jiffies - wall_jiffies) * (USEC_PER_SEC / HZ);
-		HANDLE_USEC_UNDERFLOW(shadow_tv);
+		HANDLE_USEC_UNDERFLOW(per_cpu(shadow_tv, cpu));
 
 		/*
 		 * Reset our running time counts if they are invalidated by
 		 * a warp backwards of more than 500ms.
 		 */
-		sec_diff = xtime.tv_sec - shadow_tv.tv_sec;
+		sec_diff = xtime.tv_sec - per_cpu(shadow_tv.tv_sec, cpu);
 		if (unlikely(abs(sec_diff) > 1) ||
 		    unlikely(((sec_diff * USEC_PER_SEC) +
 			      (xtime.tv_nsec / NSEC_PER_USEC) -
-			      shadow_tv.tv_usec) > 500000)) {
+			      per_cpu(shadow_tv.tv_usec, cpu)) > 500000)) {
 #ifdef CONFIG_XEN_PRIVILEGED_GUEST
 			last_rtc_update = last_update_to_xen = 0;
 #endif
@@ -467,8 +474,8 @@ static inline void do_timer_interrupt(in
 		}
 
 		/* Update our unsynchronised xtime appropriately. */
-		sec = shadow_tv.tv_sec;
-		nsec = shadow_tv.tv_usec * NSEC_PER_USEC;
+		sec = per_cpu(shadow_tv.tv_sec, cpu);
+		nsec = per_cpu(shadow_tv.tv_usec, cpu) * NSEC_PER_USEC;
 
 		__normalize_time(&sec, &nsec);
 		wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
@@ -498,7 +505,7 @@ static inline void do_timer_interrupt(in
 		op.cmd = DOM0_SETTIME;
 		op.u.settime.secs        = tv.tv_sec;
 		op.u.settime.usecs       = tv.tv_usec;
-		op.u.settime.system_time = shadow_system_time;
+		op.u.settime.system_time = per_cpu(shadow_system_time, cpu);
 		HYPERVISOR_dom0_op(&op);
 
 		last_update_to_xen = xtime.tv_sec;
@@ -670,7 +677,7 @@ void __init time_init(void)
 	set_normalized_timespec(&wall_to_monotonic,
 		-xtime.tv_sec, -xtime.tv_nsec);
 	processed_system_time = shadow_system_time;
-	per_cpu(processed_system_time, 0) = processed_system_time;
+	per_cpu(processed_system_time, 0) = per_cpu(shadow_system_time, 0);
 
 	if (timer_tsc_init.init(NULL) != 0)
 		BUG();
@@ -753,7 +760,7 @@ void time_resume(void)
 
 	/* Reset our own concept of passage of system time. */
 	processed_system_time = shadow_system_time;
-	per_cpu(processed_system_time, 0) = processed_system_time;
+	per_cpu(processed_system_time, 0) = per_cpu(shadow_system_time, 0);
 
 	/* Accept a warp in UTC (wall-clock) time. */
 	last_seen_tv.tv_sec = 0;
@@ -770,7 +777,7 @@ void local_setup_timer(void)
 
 	do {
 		seq = read_seqbegin(&xtime_lock);
-		per_cpu(processed_system_time, cpu) = shadow_system_time;
+		per_cpu(processed_system_time, cpu) = per_cpu(shadow_system_time, cpu);
 	} while (read_seqretry(&xtime_lock, seq));
 
 	per_cpu(timer_irq, cpu) = bind_virq_to_irq(VIRQ_TIMER);


--- xeno-unstable.bk/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/timers/timer_tsc.c.orig	2005-06-07 10:21:29.000000000 -0700
+++ xeno-unstable.bk/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/timers/timer_tsc.c	2005-06-07 10:21:29.000000000 -0700
@@ -10,6 +10,7 @@
 #include <linux/cpufreq.h>
 #include <linux/string.h>
 #include <linux/jiffies.h>
+#include <linux/percpu.h>
 
 #include <asm/timer.h>
 #include <asm/io.h>
@@ -35,8 +36,8 @@ extern spinlock_t i8253_lock;
 
 static int use_tsc;
 
-static unsigned long long monotonic_base;
-static u32 monotonic_offset;
+static DEFINE_PER_CPU(unsigned long long, monotonic_base);
+static DEFINE_PER_CPU(u32, monotonic_offset);
 static seqlock_t monotonic_lock = SEQLOCK_UNLOCKED;
 
 /* convert from cycles(64bits) => nanoseconds (64bits)
@@ -74,8 +75,20 @@ static inline unsigned long long cycles_
  */
 static unsigned long fast_gettimeoffset_quotient;
 
-extern u32 shadow_tsc_stamp;
-extern u64 shadow_system_time;
+extern DEFINE_PER_CPU(u64, shadow_tsc_stamp);
+extern DEFINE_PER_CPU(u64, shadow_system_time);
+
+unsigned long long get_full_tsc_offset(void)
+{
+	unsigned long long tsc;
+
+	/* Read the Time Stamp Counter */
+	rdtscll(tsc);
+
+	tsc -= per_cpu(shadow_tsc_stamp, smp_processor_id());
+
+	return cycles_2_ns(tsc);
+}
 
 static unsigned long get_offset_tsc(void)
 {
@@ -86,7 +99,7 @@ static unsigned long get_offset_tsc(void
 	rdtsc(eax,edx);
 
 	/* .. relative to previous jiffy (32 bits is enough) */
-	eax -= shadow_tsc_stamp;
+	eax -= per_cpu(shadow_tsc_stamp, smp_processor_id());
 
 	/*
          * Time offset = (tsc_low delta) * fast_gettimeoffset_quotient
@@ -110,12 +123,13 @@ static unsigned long long monotonic_cloc
 {
 	unsigned long long last_offset, this_offset, base;
 	unsigned seq;
+	int cpu = smp_processor_id();
 	
 	/* atomically read monotonic base & last_offset */
 	do {
 		seq = read_seqbegin(&monotonic_lock);
-		last_offset = monotonic_offset;
-		base = monotonic_base;
+		last_offset = per_cpu(monotonic_offset, cpu);
+		base = per_cpu(monotonic_base, cpu);
 	} while (read_seqretry(&monotonic_lock, seq));
 
 	/* Read the Time Stamp Counter */
@@ -152,11 +166,12 @@ unsigned long long sched_clock(void)
 
 static void mark_offset_tsc(void)
 {
+	int cpu = smp_processor_id();
 
 	/* update the monotonic base value */
 	write_seqlock(&monotonic_lock);
-	monotonic_base = shadow_system_time;
-	monotonic_offset = shadow_tsc_stamp;
+	per_cpu(monotonic_base, cpu) = per_cpu(shadow_system_time, cpu);
+	per_cpu(monotonic_offset, cpu) = per_cpu(shadow_tsc_stamp, cpu);
 	write_sequnlock(&monotonic_lock);
 }
 
-- 
Don Fry
brazilnut@us.ibm.com