Skip to contentSkip to conversation

p4-clockmod: Calculate frequency based on TSC value for P4 models 0 and 1

PATCH In review Frans Pop 2 participants 4 messages Feb 24, 2009 – Feb 25, 2009 x86

lore
Share & tools
lore

Frans Pop

author
#0 lore raw

PATCH 1/1 p4-clockmod: Calculate frequency based on TSC value for P4 models 0 and 1

From: Mattias-Christian Ott <[email protected]>

Commit ed9cbcd40004904dbe61ccc16d6106a7de38c998 reverted an incorrect change that resulted in random frequencies being displayed in some cases for early P4 models because the MSR_FBC_REGISTER_ID ratio is undefined.

Result of the revert is that for some users the CPU frequency is displayed multiplied by a factor 8 in /proc/cpuinfo when p4-clockmod is loaded.

Calculating the frequency from the TSC value should work in all cases.

---
This patch has been submitted before, but seems to have gotten lost:
http://bugzilla.kernel.org/show_bug.cgi?id=10968

I'm resubmitting it as it has been verified to solve a post-2.6.24 regression reported by Joerg Platte:
http://www.gossamer-threads.com/lists/linux/kernel/1034291?page=last

I've been unable to find the actual submission mail, so I've created a new patch description and added the appropriate signed-off lines.

arch/x86/kernel/cpu/cpufreq/p4-clockmod.c +6 -0
index b585e04..a4dcc65 100644
@@ -30,6 +30,7 @@
 
 #include <asm/processor.h>
 #include <asm/msr.h>
+#include <asm/timer.h>
 #include <asm/timex.h>
 
 #include "speedstep-lib.h"
@@ -217,6 +218,11 @@ static int cpufreq_p4_cpu_init(struct cpufreq_policy *policy)
 		dprintk("has errata -- disabling low frequencies\n");
 	}
 
+	if (speedstep_detect_processor() == SPEEDSTEP_PROCESSOR_P4D && c->x86_model < 2) {
+		/* switch to maximum frequency and measure result */
+		cpufreq_p4_setdc(policy->cpu, DC_DISABLE);
+		recalibrate_cpu_khz();
+	}
 	/* get max frequency */
 	stock_freq = cpufreq_p4_get_frequency(c);
 	if (!stock_freq)
arch/x86/kernel/cpu/cpufreq/speedstep-lib.c +20 -14
index cdac7d6..919644c 100644
@@ -16,6 +16,7 @@
 #include <linux/slab.h>
 
 #include <asm/msr.h>
+#include <asm/tsc.h>
 #include "speedstep-lib.h"
 
 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, "speedstep-lib", msg)
@@ -167,6 +168,15 @@ static unsigned int pentium4_get_frequency(void)
 	struct cpuinfo_x86 *c = &boot_cpu_data;
 	u32 msr_lo, msr_hi, mult;
 	unsigned int fsb = 0;
+	u8 fsb_code;
+
+	/* Pentium 4 Model 0 and 1 do not have the Core Clock Frequency
+	 * to System Bus Frequency Ratio Field in the Processor Frequency
+	 * Configuration Register of the MSR. Therefore the current
+	 * frequency cannot be calculated and has to be measured.
+	 */
+	if (c->x86_model < 2)
+		return cpu_khz;
 
 	rdmsr(0x2c, msr_lo, msr_hi);
 
@@ -177,21 +187,17 @@ static unsigned int pentium4_get_frequency(void)
 	 * revision #12 in Table B-1: MSRs in the Pentium 4 and
 	 * Intel Xeon Processors, on page B-4 and B-5.
 	 */
-	if (c->x86_model < 2)
+	fsb_code = (msr_lo >> 16) & 0x7;
+	switch (fsb_code) {
+	case 0:
 		fsb = 100 * 1000;
-	else {
-		u8 fsb_code = (msr_lo >> 16) & 0x7;
-		switch (fsb_code) {
-		case 0:
-			fsb = 100 * 1000;
-			break;
-		case 1:
-			fsb = 13333 * 10;
-			break;
-		case 2:
-			fsb = 200 * 1000;
-			break;
-		}
+		break;
+	case 1:
+		fsb = 13333 * 10;
+		break;
+	case 2:
+		fsb = 200 * 1000;
+		break;
 	}
 
 	if (!fsb)
arch/x86/kernel/tsc.c +0 -1
index 599e581..19ccb3b 100644
@@ -524,7 +524,6 @@ unsigned long native_calibrate_tsc(void)
 }
 
 #ifdef CONFIG_X86_32
-/* Only called from the Powernow K7 cpu freq driver */
 int recalibrate_cpu_khz(void)
 {
 #ifndef CONFIG_SMP
#1 lore raw
Quoted 10 lines · Frans Pop

From: Mattias-Christian Ott <[email protected]>

Commit ed9cbcd40004904dbe61ccc16d6106a7de38c998 reverted an incorrect change that resulted in random frequencies being displayed in some cases for early P4 models because the MSR_FBC_REGISTER_ID ratio is undefined.

Result of the revert is that for some users the CPU frequency is displayed multiplied by a factor 8 in /proc/cpuinfo when p4-clockmod is loaded.

Calculating the frequency from the TSC value should work in all cases.

I already picked this up in cpufreq-next.
Additionally, it needs the chunk below.

(checkpatch also complains about the overlong line in the original
patch which I fixed up before merging).

Dave

arch/x86/include/asm/timer.h +1 -2
index 2bb6a83..4f5c247 100644
@@ -11,8 +11,8 @@ unsigned long native_calibrate_tsc(void);
 
 #ifdef CONFIG_X86_32
 extern int timer_ack;
+#endif
 extern int recalibrate_cpu_khz(void);
-#endif /* CONFIG_X86_32 */
 
-- 

http://www.codemonkey.org.uk

author
#2 lore raw
Quoted 8 lines · Dave Jones
Quoted 7 lines · Frans Pop

From: Mattias-Christian Ott <[email protected]>

Commit ed9cbcd40004904dbe61ccc16d6106a7de38c998 reverted an incorrect change that resulted in random frequencies being displayed in some cases for early P4 models because the MSR_FBC_REGISTER_ID ratio is undefined.

Quoted 8 lines
Quoted 6 lines

Result of the revert is that for some users the CPU frequency is displayed multiplied by a factor 8 in /proc/cpuinfo when p4-clockmod is loaded.

Calculating the frequency from the TSC value should work in all cases.

I already picked this up in cpufreq-next.

OK, thanks. The status of the patch was somewhat unclear as, from what I could see in bugzilla, it had been submitted well before the merge window for .29 was opened. Good to know it's in the pipeline.

#3 lore raw
In reply to Dave Jones
Quoted 2 lines · Dave Jones

I already picked this up in cpufreq-next.
Additionally, it needs the chunk below.

nggh, it also needs the ifdef X86_32 guards around recalibrate_cpu_khz removed or it still fails allmodconfig on 64bit.

Dave

Signature · 1 line
http://www.codemonkey.org.uk