From: Bhavana Nagendra <bnagendr@redhat.com>
Subject: Re: [RHEL5.1 PATCH] : Fix TSC reporting for processors with constant  TSC
Date: Thu, 07 Jun 2007 15:02:51 -0400
Bugzilla: 236821
Message-Id: <4668565B.3050600@redhat.com>
Changelog: [x86_64] Fix TSC reporting for processors with constant TSC

BZ 236821

TSC has moved to the NB in AMD's GH processors and so the code that thinks
TSC
is the same as CPU frequency needs to be modified appropriately. This
bugzilla
will fix the reporting of cpu_mhz in /proc/cpuinfo on CPUs with a constant
TSC rate and a kernel with disabled cpufreq, ie Power Now is disabled.

The backport to RHEL5.1 has been tested extensively on GH architectures.

http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commi
tdiff_plain;h=6b37f5a20c0e5c334c010a587058354215433e92

RHEL5 code base does not include perfcrt_nmi changes from upstream and
tsc.c, tsc_sync.c files.  The backport is the minimum changes necessary
without bringing in additional subsystem changes to RHEL5.1.   Please keep
this in mind as you review the patch.

Please review and ACK.

There is an additional change that AMD is working on getting finalized with
Andi Kleen.
http://lkml.org/lkml/2007/6/4/362

The CONSTANT_TSC bit needs to be set earlier in early_identify_cpu() rather
than setup.c so that time_init() has the value set.   Andi thinks this
change will go into 2.6.22 as well. I will submit this snippet of code soon
as it's accepted.

--- linux-2.6.18.x86_64/arch/x86_64/kernel/setup.c.tscorig	2007-06-07 14:58:17.000000000 -0400
+++ linux-2.6.18.x86_64/arch/x86_64/kernel/setup.c	2007-06-07 08:41:57.000000000 -0400
@@ -1084,6 +1084,7 @@ void __cpuinit early_identify_cpu(struct
 			c->x86_model += ((tfms >> 16) & 0xF) << 4;
 		if (c->x86_capability[0] & (1<<19)) 
 			c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
+		c->extended_cpuid_level = cpuid_eax(0x80000000);
 	} else {
 		/* Have CPUID level 0 only - unheard of */
 		c->x86 = 4;
@@ -1092,6 +1093,12 @@ void __cpuinit early_identify_cpu(struct
 #ifdef CONFIG_SMP
 	c->phys_proc_id = (cpuid_ebx(1) >> 24) & 0xff;
 #endif
+
+	/* power flags are 8000_0007 edx. Bit 8 is constant TSC */
+	if ((c->x86_vendor == X86_VENDOR_AMD) &&
+	    (c->extended_cpuid_level >= 0x80000007) &&
+	    (cpuid_edx(0x80000007) & (1<<8)))
+		set_bit(X86_FEATURE_CONSTANT_TSC, &c->x86_capability);
 }
 
 /*
--- linux-2.6.18.x86_64/arch/x86_64/kernel/apic.c.tscorig	2007-05-24 15:54:38.000000000 -0400
+++ linux-2.6.18.x86_64/arch/x86_64/kernel/apic.c	2007-05-30 14:25:46.000000000 -0400
@@ -802,7 +802,7 @@ static int __init calibrate_APIC_clock(v
 		} while ((tsc - tsc_start) < TICK_COUNT &&
 				(apic - apic_start) < TICK_COUNT);
 
-		result = (apic_start - apic) * 1000L * cpu_khz /
+		result = (apic_start - apic) * 1000L * tsc_khz /
 					(tsc - tsc_start);
 	}
 	printk("result %d\n", result);
--- linux-2.6.18.x86_64/arch/x86_64/kernel/time.c.tscorig	2007-05-24 15:54:50.000000000 -0400
+++ linux-2.6.18.x86_64/arch/x86_64/kernel/time.c	2007-06-07 15:00:03.000000000 -0400
@@ -39,6 +39,7 @@
 #include <asm/proto.h>
 #include <asm/hpet.h>
 #include <asm/sections.h>
+#include <asm/nmi.h>
 #include <linux/cpufreq.h>
 #include <linux/hpet.h>
 #ifdef CONFIG_X86_LOCAL_APIC
@@ -69,6 +70,8 @@ static int notsc __initdata = 0;
 
 unsigned int cpu_khz;					/* TSC clocks / usec, not used here */
 EXPORT_SYMBOL(cpu_khz);
+unsigned int tsc_khz;
+EXPORT_SYMBOL(tsc_khz);
 static unsigned long hpet_period;			/* fsecs / HPET clock */
 unsigned long hpet_tick;				/* HPET clocks / HZ */
 unsigned long hpet_tick_real;				/* HPET clocks / interrupt */
@@ -576,6 +579,37 @@ static unsigned long get_cmos_time(void)
 	return mktime(year, mon, day, hour, min, sec);
 }
 
+/* calibrate_cpu is used on systems with fixed rate TSCs to determine
+ * processor frequency */
+#define TICK_COUNT 100000000
+static unsigned int __init tsc_calibrate_cpu_khz(void)
+{
+       int tsc_start, tsc_now;
+       int no_ctr_free;
+       unsigned long evntsel3 = 0, pmc3 = 0, pmc_now = 0;
+       unsigned long flags;
+
+       rdmsrl(MSR_K7_EVNTSEL3, evntsel3);
+       wrmsrl(MSR_K7_EVNTSEL3, 0);
+       rdmsrl(MSR_K7_PERFCTR3, pmc3);
+       local_irq_save(flags);
+       /* start meauring cycles, incrementing from 0 */
+       wrmsrl(MSR_K7_PERFCTR3, 0);
+       wrmsrl(MSR_K7_EVNTSEL3, 1 << 22 | 3 << 16 | 0x76);
+       rdtscl(tsc_start);
+       do {
+               rdmsrl(MSR_K7_PERFCTR3, pmc_now);
+               tsc_now = get_cycles_sync();
+       } while ((tsc_now - tsc_start) < TICK_COUNT);
+
+       local_irq_restore(flags);
+       wrmsrl(MSR_K7_EVNTSEL3, 0);
+       wrmsrl(MSR_K7_PERFCTR3, pmc3);
+       wrmsrl(MSR_K7_EVNTSEL3, evntsel3);
+
+       return pmc_now * tsc_khz / (tsc_now - tsc_start);
+}
+
 #ifdef CONFIG_CPU_FREQ
 
 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
@@ -624,7 +658,7 @@ static void cpufreq_delayed_get(void)
 static unsigned int  ref_freq = 0;
 static unsigned long loops_per_jiffy_ref = 0;
 
-static unsigned long cpu_khz_ref = 0;
+static unsigned long tsc_khz_ref = 0;
 
 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
 				 void *data)
@@ -646,7 +680,7 @@ static int time_cpufreq_notifier(struct 
 	if (!ref_freq) {
 		ref_freq = freq->old;
 		loops_per_jiffy_ref = *lpj;
-		cpu_khz_ref = cpu_khz;
+		tsc_khz_ref = tsc_khz;
 	}
         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
@@ -654,12 +688,12 @@ static int time_cpufreq_notifier(struct 
                 *lpj =
 		cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
 
-		cpu_khz = cpufreq_scale(cpu_khz_ref, ref_freq, freq->new);
+		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
 		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 			vxtime.tsc_quot = (USEC_PER_MSEC << US_SCALE) / cpu_khz;
 	}
 	
-	set_cyc2ns_scale(cpu_khz_ref);
+	set_cyc2ns_scale(tsc_khz_ref);
 
 	return 0;
 }
@@ -939,28 +973,34 @@ void __init time_init(void)
 	if (hpet_use_timer) {
 		/* set tick_nsec to use the proper rate for HPET */
 	  	tick_nsec = TICK_NSEC_HPET;
-		cpu_khz = hpet_calibrate_tsc();
+		tsc_khz = hpet_calibrate_tsc();
 		timename = "HPET";
 #ifdef CONFIG_X86_PM_TIMER
 	} else if (pmtmr_ioport && !vxtime.hpet_address) {
 		vxtime_hz = PM_TIMER_FREQUENCY;
 		timename = "PM";
 		pit_init();
-		cpu_khz = pit_calibrate_tsc();
+		tsc_khz = pit_calibrate_tsc();
 #endif
 	} else {
 		pit_init();
-		cpu_khz = pit_calibrate_tsc();
+		tsc_khz = pit_calibrate_tsc();
 		timename = "PIT";
 	}
 
+	cpu_khz = tsc_khz;
+	if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
+		boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
+		boot_cpu_data.x86 == 16)
+		cpu_khz = tsc_calibrate_cpu_khz();
+
 	vxtime.mode = VXTIME_TSC;
 	vxtime.quot = (USEC_PER_SEC << US_SCALE) / vxtime_hz;
 	vxtime.tsc_quot = (USEC_PER_MSEC << US_SCALE) / cpu_khz;
 	vxtime.last_tsc = get_cycles_sync();
 	setup_irq(0, &irq0);
 
-	set_cyc2ns_scale(cpu_khz);
+	set_cyc2ns_scale(tsc_khz);
 
 	hotcpu_notifier(time_cpu_notifier, 0);
 	time_cpu_notifier(NULL, CPU_ONLINE, (void *)(long)smp_processor_id());
--- linux-2.6.18.x86_64/include/asm-x86_64/proto.h.tscorig	2007-05-24 15:56:07.000000000 -0400
+++ linux-2.6.18.x86_64/include/asm-x86_64/proto.h	2007-05-30 14:25:46.000000000 -0400
@@ -105,6 +105,7 @@ extern int exception_trace;
 extern int using_apic_timer;
 extern int disable_apic;
 extern unsigned cpu_khz;
+extern unsigned tsc_khz;
 extern int ioapic_force;
 extern int skip_ioapic_setup;
 extern int acpi_ht;